We just returned from an action-packed week of discussion with AI leaders across industry, investing, and research at AI House Davos.

We couldn’t help but notice that the AI narrative has shifted a bit in Davos since we first co-founded AI House three years ago. In our first year, AI was a shiny and mysterious “new” object. Last year, it was about trying to separate the hype from the reality. This year, we’ve begun calling it the “year of AI ROI”. We heard a lot of variations of, “What can AI do for my business?” A real push for tangible, bottom line results.

That’s just one of our themes for the year. Below are 9 more. Ultimately, Merantix Capital is an investor. Our portfolio = our predictions. Nevertheless, here are a few educated guesses of what we think will shape AI in 2026.

For this post, I asked for some help from my colleague Thomas Wollmann, the CTO of Merantix Momentum, our AI solutions business.

1. Sovereignty and the European AI stack get real

The debate around European AI sovereignty is finally getting some momentum. One big topic in Davos was European Commission president Ursula von der Leyen officially endorsing the EU Inc initiative, an initiative we signed at its 2024 launch, which would institute the same capital regime across Europe and make it much easier for startups to register here, raise capital, and incentivize talent.

This should be the bare minimum, but it’s a real step in the right direction.

Meanwhile, by the end of 2026, we expect multiple European foundation models to be essentially good enough for most enterprise use cases, and for there to be a greater emphasis on European sovereign cloud providers. In that regard, we expect the narrative focus to shift away from whether Europe can (or should) build its own models, and more towards how the ecosystem actually works together, with a greater emphasis on infrastructure, deployment, public-private partnerships, and distribution.

2. Beyond (Transformer) LLMs

Three years after ChatGPT took the world by storm, we think 2026 will finally be the year where the conversation catches up with the research world and pushes past LLMs, particularly transformer LLMs.

LLMs are powerful interfaces, but their cost and weak grounding in the physical world are pushing researchers and companies to chip away and replace parts of the stack. LLMs will not go away, but state space models and flow models will play a bigger role, moving beyond transformer architecture.

And when it comes to the layperson’s understanding of AI, more people will see AI deeply implemented into their workplaces and personal lives in other modalities, shifting how they understand and talk about what AI really is

3. The Year of AI ROI

As we saw in the marketing installations along the Davos promenade, businesses are pushing to convey how AI can help the bottom line in real terms.

Enterprises are done asking if they should use AI. Now they want to know what it returns. We expect this year that more companies will be able to quantify this, measuring AI’s impact on the P&L. Earnings reports won’t just mention AI, they will quantify productivity as AI implementation becomes a board-level priority. Companies that cannot demonstrate AI ROI will stall.

4. Cybersecurity has a wake up call moment.

If 2025 was the year of “vibe coding,” 2026 is shaping up to be the year of security breaches. As AI-generated code and autonomous agents go mainstream, attackers are gaining leverage. We expect to see more scenarios where one skilled hacker (armed with AI) takes on an entire security team. Will 2026 produce “the big one?” A major AI-enabled cyberattack that forces enterprises and governments to completely rethink their assumptions around software trust, security, and resilience? We think so.

Every threat is an opportunity, which is why we back entrepreneurs with deep domain expertise in cybersecurity working to address these challenges head on.

5. AI’s job impact becomes impossible to ignore

For years, people have talked about the labor impact of AI in theory. So far, it’s been largely marginal or role-specific. Last year you started to see more storylines about how junior or entry-level white-collar work has been hollowed out, especially in knowledge work like software, consulting, or marketing.

This year, we think things will become significantly more noticeable. New junior roles in knowledge work will disappear as AI juices productivity gains in functions like customer support, sales, and content. For high-growth startups, this will mean scaling revenue without proportional hiring. For traditional businesses, it will mean layoffs or headcount freezing.

As a result, the importance of upskilling and reskilling and the future of education will become a more clear topic of discussion, as both new and existing workers in the job force look for alternative roles.

6. World models

If “beyond LLMs” is one theme, the natural follow up question is: What next?

Much of the talk at AI House, including a session with Yann LeCun, centered on embodied AI and world models. World models learn how objects move and how actions lead to consequences, making them foundational for general-purpose robotics. At the same time, world models can also power digital twins for factories, supply chains, cities, and financial systems, enabling decision making through simulation.

The next big leap will not come from words, but from machines that can imagine, simulate, and reason about the world before acting in it.

7. Death of search and and an e-commerce revolution

Consumers are increasingly asking, deciding, and buying through AI.

This may spell the slow death of the open web. But it also unlocks entirely new business models: LLM-native search, agentic commerce, and AI-first marketplaces that are not optimized for clicks.

2026 will be an inflection point. As consumer habits tip toward AI-driven shopping, the economics of the internet and who captures the value will get rewritten.

8. Constraints will drive the next wave of innovation

Constraints breed creativity. The era of “just add more GPUs” will have to end at some point.

We expect that constraints, whether they are geopolitical or financial, will turn compute into an even bigger battlefield in AI with more than just one dominant chip vendor winning via brute-force scaling.

Meanwhile a wave of innovation is expected to hit the hardware side this year. New, specialized chips will challenge the incumbents. On the software side, advances in inference optimization (compression) and efficient model design will push what’s possible with limited compute.

Doing more with less will be the name of the game, which will reshape the power dynamics and dependencies within the AI stack

.

9. AI for Science breaks through

AI is a powerful scientific assistant. That’s really just the beginning. Rather than just helping analyze or generate, AI systems are starting to decide themselves what experiments to run next, creating hypotheses, simulating outcomes, and iterating autonomously. The upside here is huge: AI systems that produce new knowledge and self-improve to accelerate progress.

This will have major implications in materials science, physics, mathematics, and beyond. We anticipate at least one major scientific breakthrough this year where AI is not a footnote or an assistant, but the primary driver. This will get significant press attention, accolades, and help people understand AI’s potential in the scientific realm.

10. And the Oscar goes to AI…

AI-generated content is having an interesting moment. On the one hand, our feeds are full of “AI slop” meant to farm engagement (like videos of Friedrich Merz doing a TikTok dance). Still, some AI content is indistinguishable from human output.

Oftentimes the dam breaks in the commercial realm. Coca-Cola released holiday ads created with the help of AI, which earned the soda maker some public criticism, but was more cost efficient.

We think 2026 is the year society will start to accept that (some) AI content can be high-quality. AI interviews will appear in major newspapers. AI-assisted research will distinguish itself in conferences. AI art will break through. An AI-generated film will receive awards.

This has happened with the advent of nearly every new technology. Radio scoffed at television. Print laughed at the internet. Once the quality bar is met, the acceptance of AI generation in the entertainment realm will unlock new formats and personalization.

Looking Ahead

Once again we couldn’t be more excited to see what this year brings in AI. What are we missing? And if you’re building something that helps solve the world’s most pressing issues with AI, let us know.

AI has become both cybersecurity’s biggest threat and its best defense. Attackers now use it to automate phishing, generate malicious code, and create deepfakes that can fool even advanced systems. Defenders are adopting the same technology to predict breaches, trace anomalies, and verify authenticity, but offensive methods still have a decisive advantage.

This tension defines the next wave of opportunity. Every advance in generative models expands the attack surface but also increases the need for smarter defenses. At the same time, it opens the door for new ventures building predictive, adaptive, and trusted systems.

In this article, I’ll explore where those opportunities lie and how founders can turn AI’s volatility into an advantage. I’ll also share insights from the recent “Deepfakes, Scams and Cybercrime” panel at the Merantix Campus in Berlin, where engineers, researchers, and founders discussed how these threats are evolving in real time, along with what that means for the companies racing to stop them.

AI’s Opportunity for Startups

AI is changing cybersecurity from a reactive discipline into a predictive one. It enables systems that model likely attack paths, anticipate exploits, and block them before they happen. It compresses the skills gap by embedding expert capability into software and brings enterprise-grade protection within reach of smaller firms. It is also unlocking new frontiers, from content authentication and fraud detection to the protection of autonomous systems. Despite the start of this defensive transformation, the scale of the challenge is enormous.

A World Economic Forum report found that 42 percent of organizations were successfully targeted by social engineering attacks in 2024. Nearly half of AI-generated code contains vulnerabilities, exposing products before they even launch. Model-serving APIs and large language model endpoints have become key targets, yet security tools for them remain rudimentary. Deepfake detectors perform well in controlled environments but fail in the wild. Together, these signals point to the same truth: attack surfaces are expanding faster than legacy vendors can adapt.

Even the titans of tech are feeling the squeeze. During the discussion at the Deepfakes and Cybercrime event, a former YouTube security lead in the audience described the pace of this dynamic: any detection system could be bypassed within six weeks. Their team measured success not by perfection but by halving damage with every iteration. That cadence should be a design goal for founders: ship, measure, retrain, repeat.

For founders, this widening gap represents both a technical and commercial opportunity. Security architecture built for static networks cannot protect self-learning systems. The next generation of cybersecurity companies will build for constant motion, predicting, learning, and neutralizing threats as they emerge. In that shift lies the foundation for an entirely new class of defensible: AI-native security ventures.

Where Founders Can Build

Each layer of the tech stack, from prediction and prevention to authentication and autonomy, is being rebuilt for a world where threats evolve faster than any human can track. The founders moving first in this shift are already defining how the next generation of security will work.

Predictive defense sits at the center of this change. Most existing tools still react after a breach, generating alerts once the damage is done. AI enables a different approach: systems that anticipate exploits, simulate attacks, and disrupt them in advance. BforeAI, based in France, is an example. Its predictive threat intelligence platform analyses billions of data points to identify malicious domains before they go live, serving clients that include Fortune 500 firms and NATO.

Revel8, from our own portfolio, uses AI to quantify exposure and simulate potential financial impact. The goal here is time compression. If attackers iterate in weeks, platforms need detection, simulation, and policy updates measured in days. Julius Muth, Co-Founder of Revel8, pointed to recent call-center breaches at large European insurers as examples of how scams now exploit workflows as much as systems, which makes predictive controls more valuable than reactive alerts. But not all companies enjoy the luxury of a robust security team.

Smaller companies are disproportionately exposed to cyber risks: one in three was attacked last year, with average breach costs of $255,000. Nonetheless, fewer than a third employ dedicated security staff. Larger companies typically employ a Chief Information Security Officer (CISO), along with dedicated security teams, but this decades-old security gap between SMEs and enterprises could soon come to an end. AI now makes it possible to embed those functions into SaaS platforms.

Cynomi has harnessed this model with an AI-CISO platform that automates assessments, generates policies, and provides compliance dashboards. In 2024, it signed deals with more than 100 service providers, including Deutsche Telekom, distributing its platform to thousands of SMBs, showing that these services serve a real need and a huge addressable market. These platforms should embed opinionated process controls, guardrails that automatically enforce secure behaviour inside workflows, such as requiring multi-party approval for payments or verifying caller identity through a trusted channel.

Trust infrastructure represents another frontier. Cheap deepfakes and even “deepfake-as-a-service” make detection alone a losing strategy, something that Hasna Najmi, Senior AI Solutions Architect at Merantix Momentum, pointed out during our panel. Najmi said that these services are so efficient that turnkey providers can now onboard users via chat and generate realistic videos in minutes. Detection also remains probabilistic. Muth explained that CISOs must still define thresholds and decision rules for false positives and negatives, while Dr. Nicolas Müller, Senior Engineer at the Fraunhofer Institute, noted that audio deepfake detectors fail to generalize to new model families. Müller added that if generative AI consolidates around a few providers, watermarking and training on real outputs could become practical, improving provenance and verification at scale.

Bigger threats, bigger opportunities

AI is both the nemesis and knight in shining armour for cybersecurity. But it’s also more complicated than that: AI adoption itself is also creating risks deeper in the infrastructure layer. Coding assistants now write large portions of production code, often with hidden vulnerabilities that pass untested into live systems. APIs are multiplying, model-serving endpoints are exposed, and the infrastructure that supports AI models has become a new target. The OWASP 2025 LLM Top 10 and NIST’s Generative AI Profile both highlight the same weak points: injection flaws, insecure outputs, and supply-chain risks linked to AI-generated code.

The consequences are already visible. Ray, an open-source framework designed to scale AI and Python applications, experienced a hack in which its clusters were hijacked for crypto-mining in 2024, and NVIDIA’s Triton inference server required an emergency patch in 2025 after attackers found remote-execution flaws. These problems continue to exist at scale, but a few European startups are beginning to respond. HarfangLab has grown quickly with its AI-powered endpoint detection platform, now surpassing €40 million in annual revenue, while Lupovis, spun out of the University of Strathclyde, uses deception-based AI honeypots to lure and trap attackers.

For founders, the opportunity runs deeper than patches or point solutions. The market still lacks foundational tools that secure AI code and infrastructure from the inside out: firewalls, continuous validation systems, and automated runtime protection built for a world where models write and deploy themselves.

Autonomy and counter-drone security are becoming the new front line of cybersecurity. As the boundary between digital and physical systems dissolves, attacks on drones, satellites, and autonomous vehicles can now trigger real-world consequences. The war in Ukraine exposed how quickly off-the-shelf drones, powered by commercial AI models, can be repurposed into weapons or surveillance tools. The UK NCSC has warned that AI will shorten exploit timelines for critical infrastructure, while NATO and EASA have elevated counter-drone initiatives as strategic priorities.

This shift turns autonomy itself into an attack surface. AI systems that once operated in isolation—navigation, perception, targeting—are increasingly networked and therefore vulnerable to interference, spoofing, and data poisoning. Keeping these systems secure demands continuous model updates, telemetry, and field learning. Europe is beginning to respond. Helsing, Quantum Systems, Dedrone and Origin Robotics represent the first generation of companies building or defending AI-powered autonomy. Their rise signals both the urgency of the challenge and the gap still left to fill. For founders, the opportunity lies in securing this frontier: embedding cybersecurity into the fabric of autonomous systems before adversaries do.

Why This Matters for Europe

Europe’s growing focus on digital sovereignty is reshaping the cybersecurity market. The AI Act sets new standards for trustworthy and transparent AI, while the NIS2 Directive raises security requirements for critical sectors. Together, they are tightening compliance expectations across the continent and creating strong demand for trusted infrastructure. Nearly half of European organizations plan to adopt sovereign cloud solutions by 2025, yet the market remains dominated by a few non-European providers. This imbalance leaves governments and enterprises exposed to both cyber threats and external jurisdiction through laws such as the U.S. Cloud Act.

Momentum is now shifting toward self-reliance. Initiatives like Gaia-X and proposals for a European Tech Sovereignty Fund reflect how sovereignty has become both an economic priority and an investment theme. Enterprises and critical-infrastructure operators are seeking EU-governed systems that guarantee control over data and operations. For founders, this environment provides a natural tailwind. Every new rule, funding program, and procurement policy strengthens the case for European-built, AI-native security platforms.

The same dynamic is playing out among smaller companies. Regulation is pushing for local control, while AI is making it possible to deliver advanced security capabilities at scale. Automated platforms can now offer CISO-level oversight at a fraction of the cost, giving SMEs access to protection once reserved for large enterprises. As these tools spread through supply chains and business networks, they will define how Europe achieves digital autonomy and long-term resilience.

Build the future of cybersecurity with us

The cybersecurity arms race is already underway, and the winners will not only protect Europe but also define new global categories. The challenge is immense, but so is the upside. Founders who build sovereign, AI-native platforms that predict and neutralize threats, democratize security leadership, and anchor trust infrastructure will create the next generation of cybersecurity leaders.

In our venture studio, we partner with founders (like revel8) who are ready to take on this challenge. If you are thinking about building in this space and want to shape the future of AI-first security, get in touch with us.

The earliest applications of AI have made us used to thinking of it in the context of software: intangible, abstract, algorithmic. Or how I just asked ChatGPT to make the picture above for this article.

But AI’s most transformative applications will most likely emerge in the physical world, where it can reshape how we design and scale new materials.

That’s why materials discovery is one of our biggest areas of interest at Merantix Capital. Our portfolio already includes Cambrium, a company born out of our venture studio that designs novel proteins.

If you are in Berlin, we are hosting an AI Breakfast on Thursday 21st August at the Merantix AI Campus to explore these ideas in more depth, with speakers from Cambrium, CuspAI, Dunia, and our own GP Adrian Locher.

If you’re already building in this space (or are keen to), we’d love to see you there!

Why we like the space

Materials have long been the hidden foundation of progress, underpinning advances in semiconductors, solar cells, batteries, and carbon capture. Now, AI is accelerating every stage of their development. It proposes novel compounds, predicts their properties, and links design directly to manufacturing constraints. This convergence opens the possibility of building entirely new materials platforms that are faster, more cost-effective, and more adaptable than ever before.

The next generation of materials giants will fuse deep technology with industrial relevance and, most importantly, help address some of the most urgent challenges of our time, like clean energy and sustainable production.

For years, computational chemistry served as a tool to simulate a few candidate molecules before testing them in the lab. The process still relied on human intuition to select what to simulate but, today, large-scale AI models don’t just evaluate known materials. They propose novel ones at unprecedented scale.

DeepMind’s GNoME project, for instance, trained a graph neural network to predict the stability of inorganic crystal structures. Researchers at Berkeley Lab for their part built A-Lab, an autonomous discovery platform combining machine learning, robotic synthesis, and automated measurement. In a 17-day trial, it completed 355 experiments and discovered 41 new inorganic compounds, all guided by active learning algorithms that improved in real time.

Designing for Scale

Big as these developments in R&D, they’re only going to help drive forward science if they can be applied at scale. That’s why a new cohort of startups are building their discovery platforms with production in mind from the beginning.

Cambrium is taking this scale-oriented approach into the world of protein materials. The company uses AI to design novel proteins that do not exist in nature and produces them through microbial fermentation, allowing for a more flexible and sustainable production process. From the start, its design process has been built around manufacturability, with a focus on properties like thermal stability, fermentation yield, and compatibility with existing bioproduction infrastructure.

This approach led to a fivefold increase in development speed and a twentyfold reduction in cost compared to traditional protein engineering workflows. One of Cambrium’s first products, NovaColl, is a sustainable collagen alternative that achieves a 95 percent lower carbon footprint and was developed in under two years.

In climate-focused materials, CuspAI is using generative AI to design bespoke porous compounds for applications like carbon capture and gas separation. The platform functions like a “search engine” for materials, generating and evaluating candidates that match specific performance criteria, including CO₂ selectivity and release dynamics. This approach has significantly shortened development timelines, reducing discovery cycles from years to months and achieving success rates of around 90 percent in generating viable materials. While still early, CuspAI is prioritising industrial deployment by designing materials that not only perform well in theory but also meet practical requirements such as regeneration efficiency and compatibility with existing capture systems.

Materials Nexus is another example addressing the supply chain risks associated with rare-earth elements by using AI to design alternative permanent magnets. In collaboration with the University of Sheffield and the Henry Royce Institute, the company screened more than 100 million candidate materials to develop MagNex, a rare-earth-free magnet suitable for use in electric vehicles and wind turbines. The material was taken from concept to synthesis in just three months and can be manufactured at around 20 percent of the cost of typical rare-earth magnets, with a 70 percent reduction in carbon emissions per kilogram produced.

XtalPi may have started in pharmaceutical discovery, but its AI-powered R&D platform is now proving adaptable to materials science, including solid-state research. The company offers tools tailored to structure prediction, experimental solid-state chemistry, and automated crystallography, collectively marketed under the “XtalGazer” and digital chemistry suite. Since 2024, XtalPi’s public filings confirm that its platform is being deployed for electrochemical and new energy material development, such as battery electrolytes and advanced ceramics, often in collaboration with partners like GCL Group and Fangda Carbon. These developments demonstrate how an AI-native discovery stack, once proven in one sector, can be repurposed to serve adjacent industries with minimal changes.

Together, these companies illustrate a broader shift: materials discovery is no longer just about predicting what might work in a lab. It’s about building entire pipelines that move from concept to manufacturable product with speed, precision, and commercial focus. By starting with real-world constraints such as cost, scale, and performance, this new wave of ventures is showing that AI-native materials platforms can make a meaningful impact far beyond the simulation stage. The next question is where new founders can still add value.

Where Founders can build

Founders entering the materials space face distinct challenges and opportunities. Initial traction tends to emerge in industries where performance matters more than cost. Markets such as cosmetics, fashion, and specialty coatings offer fast sales cycles and customers willing to pay for novelty, differentiation, or sustainability. These sectors provide early revenue, which can fund further technical development and create defensible IP.

Over time, the opportunity expands into heavier industries. Sectors like construction, packaging, agriculture and energy require greater investment in scale and compliance but offer outsized long-term value. In these markets, success depends not just on novel discovery but on the ability to integrate with supply chains and meet regulatory standards.

Strategically, founders can choose different entry points across the stack:

• Some will focus on software platforms, offering prediction tools to chemical companies or research institutions.

• Others will partner with OEMs, co-developing materials while relying on partners for scale-up.

• A third category will go full stack, owning discovery, validation, and production in-house.

Each approach carries trade-offs in capital requirements and speed, but all benefit from tighter control over learning cycles.

What is still missing in the ecosystem presents new possibilities. Lab automation remains capital-intensive and geographically limited. There is a shortage of multimodal models capable of integrating structural, textual, and experimental data. And very few infrastructure tools exist to bridge gram-scale discovery with ton-scale manufacturing.

In this landscape, speed of iteration becomes the key source of defensibility. Teams that can cycle faster, learn from data, and apply those insights across markets will build the next category-defining materials companies.

I’m always looking to hear from founders in materials discovery — if this article speaks to you or you have an idea you want to develop, get in touch or come along to our AI Breakfast this Thursday in Berlin.

What we’re starting to see now is that the real leverage is moving to the interface layer (as in: how people are actually interacting with AI). This is where users make decisions and generate the feedback that improves the system. It is where value is created and, for startups and investors, where it can be captured.

In this piece, I’ll be looking at why this shift is happening and how AI-native interfaces are becoming the new strategic layer across industries like productivity, legal tech, logistics, and fund operations.

The shift toward interface-level innovation is already happening.

OpenAI recently acquired former Apple designer Jony Ive’s hardware studio io for $6.5 billion. The real signal wasn’t the deal size, but the focus: ambient AI designed to blend into everyday life without screens. We’re witnessing interaction move from something active and controlled to something embedded and passive.

Another signal for the changing interface of AI was the ElevenLabs hackathon, where a team built GibberLink, a protocol for AI agents to communicate using compressed audio. It’s early, but it points to a world where agents coordinate directly, without waiting for human input. How will agent-agent communication be structured, especially on platforms in which humans typically interacted — for example, over the phone? What happens when your personal booking agent calls your doctor's office's booking agent to arrange an appointment? How they interact is a big question.

The shift is changing how people build. Tools like CustomGPT and Google AutoML are making natural language the default interface for software creation. The rise of “vibe coding”, describing an outcome in plain terms and letting the system handle the structure, is lowering the barrier to entry. Gartner estimates that by 2025, 70 percent of enterprise apps will be built using no-code or low-code tools.

Each of these examples offers a view into how interaction with AI is evolving. Interfaces are no longer just access points. They define how we unlock the value of backend AI systems, where a good user interface allows the true potential of AI functionality to be unlocked. The first steps have been made and this shift is already influencing how tools are being designed, adopted, and embedded into daily work.

Innovation and the decision layer

As AI moves from experimentation into real-world use, interfaces have become the main point of leverage. We can see this play out across three layers: general-purpose assistants, embedded enterprise copilots, and domain-specific platforms.

Chat-based assistants like ChatGPT and Claude’s strength lies not just in language generation, but in how they turn complex inputs into simple, guided exchanges. But that was just the beginning. Chatbots reduce user autonomy and increase cognitive load. Why should we ask AI to do something when ambient computing can unlock AI pre-empting actions based on what you are already doing.

OpenAI’s move with Jony Ive shows they are thinking about what comes next: how people interact with AI will (and needs to) shift to AI-native devices and processes.

Enterprise copilots take this further by embedding themselves inside the tools where people already work. Microsoft 365 Copilot and Notion AI are not trying to replace existing applications. Instead, they surface suggestions and automate tasks within the same interfaces teams already use to write, calculate, meet and plan.In other words, everyone is realizing that the interface is the control point. While one might think that reality favors incumbents, this also creates a huge startup opportunity. The challenge is that giants like Microsoft are more easily able to embed AI into daily workflows, but startups that are able to “own the interface” in underserved workflows, such as procurement, logistics, construction, R&D, etc., will be positioned to become new category leaders in every domain.

Indeed, in more specialized environments, vertical platforms are emerging that target specific roles and decision points. Libra, a legal tech and Merantix Capital portfolio company, supports contract analysis, drafting, and research in one system. This year, the firm reached over €1 million in ARR thanks to an interface that understands legal language, firm-specific workflows and decision logic.

Canoe is doing something similar in fund operations. It automates data ingestion from custodians and portfolio systems, flags anomalies, and creates compliance-ready reports. The system integrates directly into operational workflows, which makes it easier to trust and easier to adopt.

Interfaces are becoming decision environments. When that happens, they start to accumulate operational data, model edge cases and adapt in ways that horizontal platforms cannot easily replicate.

Vertical Interfaces: Where AI Meets Real Work

Some of the most important changes are happening in domains that have historically been underserved by software, from fund administration to supply chain planning and everything in between.

These workflows tend to be complex, manual, and spread across systems that were never designed to work together. That has created an opening for AI-native platforms that lead with interface design. I’ll give a few examples from outside of our portfolio.

Procurement is one of the clearest cases. Mercanis has built a system that combines sourcing, supplier management, and contract workflows into one environment. The platform automates routine tasks like RFx processes and risk checks and continuously analyzes spend data to suggest savings opportunities. According to EU Startups, the firm raised €17.3 million in June, and reported that some customers experienced a 40 percent process cost savings and a 2.5 times improvement in efficiency.

In supply chain operations, Pelico helps manufacturers and logistics teams detect bottlenecks, manage disruptions and coordinate responses across departments. The system integrates live data from enterprise resource planning (ERP) and factory systems, as well as proposing real-time adjustments. For companies like Airbus and Safran, this has led to a 40 percent reduction in missing parts and a 15 percent lift in on-time delivery. By giving teams a shared interface to make decisions and act quickly, the platform improves both visibility and execution.

Raft is doing something similar in freight forwarding. The platform replaces manual processes with an integrated workflow that handles shipment tracking, invoice reconciliation and customs documentation. A customer-facing portal provides visibility into bookings, emissions data and quotes. The firm recently raised $30 million in Series B funding and has gone on to become a daily workspace for logistics teams.

What these platforms have in common is not a particular technical breakthrough, but a focus on owning the interface where decisions happen.

By embedding themselves in the day-to-day work of operational teams, they gather data that improves the product, reinforce habits that are hard to unlearn, and position themselves as the primary system of record. In environments that are structured but nuanced, where decisions happen continuously, interface ownership will continue to be a durable strategic advantage.

Our thesis: Interface ownership will create new category leaders

Across every example, one idea keeps showing up: the most meaningful value accrues where decisions are made. Interfaces that support that moment, where information turns into action, end up shaping behavior.

That’s why AI-native UX consistently outperforms tools that simply layer models onto old workflows. Retrofitting intelligence into legacy software often misses the timing, the structure, or the real context of the task. In contrast, deeply integrated vertical copilots are proving more resilient by becoming part of the decision process. We believe that the next generation of category leaders will be built at this layer.

If you’re building interfaces that change how work happens, we want to hear from you.

It’s no secret that “Agentic” has been the AI buzzword of 2025. The problem is, most of the conversation still feels abstract. It's hard to know what’s real, what’s working, and what’s just noise.

In this piece, I want to cut through that. I’ll walk through examples of agentic AI that we find most interesting at Merantix Capital — from financial services to science to healthcare — and share a few thoughts on what makes the best systems stand out.

What is Agentic AI

For the uninitiated, let's briefly define Agentic AI: it’s essentially an AI system that can independently assess a problem, plan a series of actions, execute that plan across tools, and learn from the outcome.

But not everything branded as "agentic AI" today actually earns the title.

Cutting Through the Noise

The term “agentic” has been stretched so far it covers everything from basic automation tools to genuinely autonomous systems. Buyers that think they are getting an AI agent can sometimes end up with a narrow tool with no feedback loop, no memory, and no ability to improve. (Microsoft Copilot, for instance, is a useful service but not yet capable of managing complex workflows without human oversight).

At Merantix Capital, we look for agentic systems that are solving painful, high-friction problems inside the industries that run the world. The teams that matter combine technical firepower with deep domain expertise. They build agents that are not just clever, but dependable, battle-tested and capable of delivering compounding results.

Payments infrastructure, scientific research, healthcare delivery, cybersecurity — these are just a few of the frontiers where agentic AI will prove itself. I want to provide a few examples of innovative solutions in these spaces.

Payments

Payments is one of the first sectors where agentic AI is stepping beyond mere recommendation into real-world action. Traditionally, even the most capable AI systems were bottlenecked: they could suggest transactions but not actually execute them. The problem was not intelligence, it was infrastructure. Existing payment systems were built for human users, not autonomous agents.

Skyfire, Nevermined, and Payman are three startups that represent different but complementary solutions to this bottleneck.

Skyfire built a payment network designed specifically for AI agents to spend money safely. Rather than tying agents directly to a user's bank account, Skyfire issues each agent a wallet ID funded via traditional methods like ACH or wire, but all transactions conducted by agents happen in digital assets like USDC. This architecture allows businesses to tightly control spending through real-time dashboards while enabling machine-speed execution. In beta deployments, enterprises like car parts firm Denso used Skyfire to automate procurement payments, allowing agents to complete supply chain transactions without human intervention.

Nevermined, meanwhile, is building the "PayPal for AI agents," focusing on AI-to-AI commerce. It allows agents to pay one another for services autonomously, using crypto rails optimized for dynamic pricing, metered usage, and trustless transactions. Their approach emphasizes decentralized financial infrastructure, helping agents bypass the rigid, account-based systems that slow down human-driven commerce.

Payman tackles the other side of the problem: letting AI agents pay humans directly. Payman’s platform links agents to traditional banking systems and supports payments not only in USDC but also through fiat rails like ACH and wires. This makes it one of the first to bridge agents into the human economy, enabling scenarios where an AI agent could automatically hire a freelancer, monitor task completion, and release payment without manual coordination. Unlike Skyfire and Nevermined, which still transact primarily in digital assets, Payman enables actual bank transfers alongside crypto.

Most early agentic payment systems rely on digital assets like USDC or stablecoins because traditional banking infrastructure is not yet designed for autonomous, machine-initiated transactions. While Payman has made inroads into traditional rails, the broader financial ecosystem is still catching up.

This emerging layer of infrastructure, including wallets, safeguards and programmable transaction limits hints at what broader autonomous commerce could look like. As the rails mature, there is enormous potential for founders to build orchestration layers, audit and compliance tools, fraud prevention systems, and seamless agent-to-agent marketplaces on top.

Science and Healthcare

Scientific research and healthcare both suffer from massive knowledge bottlenecks: too many papers to read, too many variables to test, too many patients to monitor. Traditional workflows are slow, heavily manual, and ripe for transformation by agentic systems.

AI Scientist-v2 is an agentic system that performs the entire scientific method autonomously. It generates hypotheses, designs and simulates experiments, evaluates results, and iterates toward better answers. Then, it writes up those findings as full academic papers.

Introduced in ‘The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search’, by Yutaro Yamada et al in April 2025, this model contains a remarkable leap forward both AI and science: a fully AI-generated paper with no human ghostwriting that passed peer review. That’s a first in scientific publishing.

Here’s how it works:

1. The system explores multiple experimental pathways in parallel, simulating experiments based on probabilistic models and refining hypotheses over time.

2. Each hypothesis is scored for novelty and plausibility.

3. Promising results are bundled into a structured research paper using a modular writing agent trained on scientific publication norms.

4. Then, the paper is submitted for blind peer review. This time it got in.

This is autonomous research. AI that contributes new knowledge, not just insights. And it raises critical questions not just about what AI can do, but what roles remain essential for humans in the scientific process.

Another important example is Causaly, a biomedical research company that developed an agentic AI platform autonomously planning and executing multi-step research workflows by drawing from a 500 million–fact biomedical knowledge graph and external scientific literature.

Causaly’s agentic system reportedly prioritizes relevant evidence, updates findings dynamically as new data emerges and traces every insight back to its source to maintain scientific transparency.

In healthcare, Hippocratic AI, co-founded by Munjal Shah alongside experts from Johns Hopkins, Stanford, and Google, is tackling the shortage of frontline medical staff.

Hippocratic AI’s platform trains agentic healthcare assistants to handle non-diagnostic tasks like post-discharge follow-ups, chronic disease coaching, and medication reminders, while maintaining a rigorous safety-first approach validated through clinical trials involving over 5,000 nurses and 500 physicians. By late 2024, Hippocratic AI reported that its agents had achieved safety parity with trained nurses and pharmacists, formed partnerships with major health systems like Memorial Hermann and Universal Health Services.

These examples show that agentic systems aren’t just accelerating research or augmenting healthcare, they are starting to expand the boundaries of what is possible by performing tasks that were once considered uniquely human.

That said, there are still major opportunities ahead. Today, most scientific and healthcare agentic systems operate mainly in digital environments, working with existing data rather than interacting directly with the physical world.

This is partly because it’s easier to coordinate digital tools than to manage real-world variables like lab equipment or patient care. However, there is a huge need for experts to build agentic systems that can also automate wet lab experiments, manage clinical trials, and coordinate ongoing patient care — areas where sensitivity to real-world context, safety, and adaptability are essential.

Cybersecurity

Cybersecurity is a race where traditional defenses are falling behind. New vulnerabilities emerge daily, software evolves constantly, and attackers adapt faster than human teams can respond. Quarterly penetration tests, static firewalls, and reactive patching are no longer enough. The next wave of defense needs systems that think, move, and adapt on their own.

Terra Security deploys fleets of autonomous agents that continuously probe enterprise environments like web apps, APIs, and cloud assets without waiting for human instructions. Each agent specializes in a class of exploits such as SQL injection, authentication bypass, or privilege escalation, and plans new attacks dynamically as it learns about the target’s structure.

This shifts cybersecurity from point-in-time testing to continuous, self-directed red teaming. Instead of waiting months for a vulnerability report, companies receive real-time findings prioritized by exploitability and business risk.

Terra uses a hybrid model where human analysts verify AI-generated findings to ensure accuracy and relevance. This combination of machine scale with human oversight addresses one of the biggest weaknesses in traditional penetration testing: inconsistent accuracy and lack of operational context.

Despite these advances, cybersecurity remains wide open for agentic innovation. Most systems today focus on offense. Defensive agentic systems, capable of predicting threats, patching vulnerabilities, and neutralizing attacks without human help, are still rare. There is enormous opportunity for security-native founders to build the next generation of autonomous, self-healing defenses. (In our own portfolio, check out Company Shield).

The best is yet to come

The best agentic projects rise above the noise because they deliver genuine autonomy where others still depend on human intervention. The breakthroughs in payments, research, healthcare, and cybersecurity show that when agentic systems are built for real-world complexity, they do not just streamline workflows, they open up entirely new ways of operating (check out our recent Substack with LogicStarAI’s founder for another example in the engineering space).

Founders who combine technical strength with deep industry knowledge, and who focus relentlessly on solving high-friction problems, will define the next era. The future belongs to those who can turn agentic AI from a buzzword into real-world performance.

As always - if you are ideating or building in this space, we should talk!

For an in-depth pulse check on the Agentic AI hype, be sure to join our AI Breakfast on this topic, being held this coming Thursday, June 12th on the Merantix AI Campus: https://lu.ma/7oazc49z

Good health is a privilege, not a guarantee.

That’s why healthcare is unique as an industry that we will all interact with throughout our lives, whether we like it or not.

Beyond the sisyphean task of day-to-day patient care, healthcare is a creaking sector of ancient IT systems, poor tech integration, mountains of siloed data and manual processes crying out for automation.

It’s also one of the most exciting for innovative AI applications, as technology is finally maturing to a point where it can make a meaningful impact for patients, hospitals and beyond.

In this post, I want to highlight some use cases I think are overdue for their time in the AI sunshine. Thanks also to my colleague Bertram Weiss, VP Health at Merantix Momentum, for providing his input.

Hospital Logistics and Bed Management

Inefficiencies in bed allocation, medication supply chains, and staff scheduling all lead to wasted resources and delayed patient care.

Bed availability is one of many operational pain points for hospitals. But where administrators see a chaotic mess of complex data and poorly integrated systems, AI can streamline and provide actionable insights, as a 2019 Harvard Business Review found.

• A John Hopkins Hospital implemented an AI-powered control center that improved bed assignment efficiency by 30%, reduced the need to keep surgery patients in recovery rooms longer than necessary by 80%, and cut the wait time for beds for incoming emergency room patients by 20%.

AI is also being trialled in transforming predictive inventory control, particularly for medications and critical medical supplies. AI-powered time series forecasting models, including ARIMA (AutoRegressive Integrated Moving Average) and LSTMs (Long Short-Term Memory networks), anticipate fluctuations in medication demand, helping hospitals prevent stockouts and avoid costly overstocking.

A study published in Acta Logistica evaluated the performance of ARIMA and LSTM models in forecasting the demand for five commonly supplied medicines in Rwanda's public health system. The LSTM model demonstrated superior accuracy, achieving a 95.2% performance rate in demand prediction, indicating its potential to significantly reduce stockouts and overstocking.

If you’ve ever tried to get a doctor to give you a firm answer on whether they can attend a future event, you will know how astoundingly chaotic hospital staffing methods are. Days off have to be booked months in advance and understaffing means those precious days off for overworked medical staff can simply disappear into thin air.

Fortunately, AI-driven reinforcement learning algorithms are improving hospital staffing models, dynamically adjusting workforce deployment based on expected patient inflows, peak hours, and emergency events. Bioengineering Basel’s 2024 report demonstrated that AI applications could predict patient admissions and discharges, allowing for dynamic staffing adjustments. This proactive approach helps in maintaining optimal nurse-to-patient ratios, thereby improving patient care and reducing staff burnout. The report also suggested that this data could help improve patient experience by giving accurate information about appointment times, wait periods and treatment schedules.

AI-driven scheduling is also cutting costs and improving efficiency in the operating room, according to a 2024 report from the Fort Lauderdale Eye Institute.

• A machine learning algorithm designed to optimize operating room booking times led to a 21% reduction in nursing overtime, resulting in significant cost savings.

• Over a three-year period, the reduction in overtime hours translated to $469,000 in savings, highlighting AI’s potential to alleviate staff burnout and improve operational efficiency.

These improvements aren’t just reducing costs—they are fundamentally reshaping hospital management by ensuring that every resource, from beds to staff to equipment, is utilized at peak efficiency, directly improving patient outcomes and hospital profitability.

AI-Powered Medical Scribes

Doctors spend a third of their working hours filling in paperwork, a Hett Insights report found. It doesn’t stop when they clock off either, with some spending nearly 5 hours of their personal time per week wrestling with documentation. This is a staggeringly poor use of time that contributes towards burnout and something that can easily be taken over by AI scribes.

AI scribes leverage automatic speech recognition (ASR), natural language processing (NLP), and large language models (LLMs) to extract key information from conversations in real time. AI-generated notes follow standardized formats such as SOAP (Subjective, Objective, Assessment, Plan) and BIRP (Behavior, Intervention, Response, Plan), ensuring consistency and accuracy, according to an NEJM Catalyst study.

A pilot study on AI scribes in dermatology found that they significantly reduced documentation time, enabling doctors to focus more on patient interactions instead of administrative tasks. The American Medical Association (AMA) reported that AI scribes could save doctors one hour on average, drastically improving efficiency and reducing burnout among healthcare professionals.

AI scribes aren’t just a helpful notetaker for doctors making their rounds, they’re also a valuable source of health record and insurance data that can be integrated into future multimodal AI.

Multimodal AI will allow for data from text, such as doctor’s notes and EHRs, to be examined along with images and time series data like X-rays, MRIs and monitoring data from wearables and hospital equipment. AI will assist doctors by quickly synthesizing complex patient histories and identifying correlations that would take a human much longer to process.

Scribes have huge potential to cut back on paperwork and increase the time doctors share with their patients. Any opportunity to automate bureaucracy faced by clinical staff should be taken up by innovators with the dual goal of helping to increase patient rate and generate a valuable data pool from the associated claims data that can help with personalized patient interventions, disease forecasting and working towards a 360 degree multimodal approach to patient care.

AI for Epidemic Detection and Pandemic Preparedness

The global spread of COVID-19 laid bare the grim reality of pandemic preparedness. Thankfully, the worst of the coronavirus is in the rearview mirror, but experts warn that we are far from prepared for the next global outbreak.

The bad news is that we don’t know when the next one will strike. The good news is that AI might be our best chance to minimise the impact when the time comes.

A February 2025 study from the University of Oxford highlights the transformative role of AI in predicting, detecting, and containing pandemics before they escalate. The study found that AI-powered epidemiological modeling, genomic tracking, and real-time surveillance are essential tools in strengthening global pandemic preparedness:

• Improving current models of disease spread, aiming to make modelling more robust, accurate and realistic.

• Pinpointing areas of high-transmission potential, helping ensure limited healthcare resources can be allocated in the most efficient possible way.

• Improving genetic data in disease surveillance, ultimately accelerating vaccine development and the identification of new variants. Real-time surveillance using genomics such as PCR means wastewater can now be monitored as one of the earliest sources for detecting pathogens, something that was successfully implemented during the pandemic.

• Helping determine the properties of new pathogens, predict their traits and identify whether cross species jumps are likely.

• Predicting which new variants of already-circulating pathogens – such as SARS-CoV-2 and influenza viruses – might arise, and which treatments and vaccines are best in reducing their impact.

• Possible AI-aided integration of population-level data with data from individual-level sources – including wearable technologies such as heart rate and step counts – to better detect and monitor outbreaks.

• AI can create a new interface between the highly technical science and healthcare professionals with limited training, improving capacity in settings that need these tools the most. This can bridge between staff that went into the sector for the ‘human interaction’ and help speed up the relatively slow adoption of technology in healthcare.

AI is also proving essential in vaccine development, where it accelerates genomic analysis and drug repurposing, cutting traditional vaccine research timelines by up to 50%. Pharmaceutical firms such as Moderna and Pfizer are actively integrating AI into their R&D processes, allowing for faster identification of high-risk mutations and more precise vaccine targeting.

Epidemiologists are quick to remind us that the next pandemic won’t politely wait around the corner until we’re ready: it’s only a matter of time before the next infectious disease takes hold. The time for AI adoption must be between pandemics, not during one.

In my book, that’s a pretty urgent call to action. For ambitious builders and healthcare experts looking to create tools that will help manage the next pandemic, yesterday was the best time to start, today is the next best opportunity.

AI-Powered Wearables and Health Monitoring

AI-enabled wearable devices are revolutionizing real-time health monitoring, providing continuous, non-invasive insights into patient well-being. These wearables offer early detection of health conditions, personalized interventions, and enhanced patient care. Recent research highlights their growing role in cardiovascular health, human activity recognition, and intensive care monitoring.

A 2023 study showed that wearable ECG devices demonstrate high efficiency in processing single-lead ECGs. The research shows that the devices can deliver near real-time cardiac assessments in 33 to 35.7 seconds, ensuring faster diagnosis and early intervention. These ECGs improve arrhythmia detection and reduce reliance on manual interpretation, helping healthcare professionals manage cardiac patients more effectively.

A 2025 study published in npj Digital Medicine found that a deep learning model trained on a dataset of over 150,000 individuals found associations with 1,600 diseases, outperforming traditional ECGs:

“When compared to standard ECG intervals, the autoencoder-based latent space models reveal a substantially greater number of associations, suggesting that the rich representations afforded by the architectural complexity of the deep learning model provide more information about disease status than routinely ascertained ECG intervals.”

The paper also notes that consumer-based ECG devices could enable large-scale, remote disease detection.

This powerful predictive is backed up by another study from the American Heart Foundation which reported that devices in the ‘Multisensor Monitoring in Congestive Heart Failure (MUSIC) study’ developed and validated a personalized algorithm to predict acute decompensation in 543 patients with heart failure. The wearable device predicted impending hospitalization with 76% to 88% sensitivity at a median of 6.5 days before hospital admission. That’s a huge window for preventative action.

These wearables are also a sign of another trend: building systems outside of traditional healthcare infrastructure.

Traditional infrastructure, such as hospitals and surgeries, often use outdated information systems such as EHR/EMR and in some cases wouldn’t be able to integrate live data from wearable devices. That’s why some innovators are trying to reinvent healthcare end-to-end: to allow the AI to work to its fullest potential and rethink the experience entirely. It’s also another reason why it is easier to set up this independent infrastructure in prevention.

Just look at Spotify co-founder Daniel Ek’s new AI-powered full-body scan and preventative healthcare platform, Neko Health. The firm’s clinics use AI and non-invasive sensor technology to generate millions of data points across cardiovascular, metabolic, and dermatological health indicators. Patients receive comprehensive AI-assisted insights into potential early-stage health risks such as cancer and cardiovascular disease.

They say it’s the next step towards preventative health, although I think there are some fairly large structural and incentive challenges in the health industry before doctors start prescribing jogging and earning from healthy patients rather than sick ones. That said, a pilot test is underway in Kinzig Tal doing just that.

Either way, wearables have enormous potential to improve the doctor’s ability to diagnose and treat patient’s illnesses, in some cases allowing for preventative action. On top of this, new AI tools could connect siloed databases and insurance code to improve tracking in wearables and immediately send alarming data to doctors. We will also see new kinds of sensors that are able to monitor different variables. This unprecedented availability of patient data should be mined to bring real-time nowcasting into healthcare and make trend-spotting, diagnosis and treatment as dynamic as they can be.

An industry more than ripe for AI

Healthcare isn’t just another industry—it’s a system we all rely on that remains crippled by inefficiencies and missed opportunities.

AI presents a rare opportunity to rebuild critical processes from the ground up, not just optimize broken ones. For founders and innovators, this isn’t about incremental improvements; it’s about reshaping how care is delivered, predicted, and personalized around the world.

The first stages of preventative health are finally within reach, with AI wearables catching cardiac events before they happen and predictive models identifying disease risks early. Hospitals are no longer bound by guesswork, as machine learning streamlines logistics, staffing, and resource allocation. Meanwhile, AI-driven automation is freeing doctors from paperwork, returning them to patient care where they belong.

The challenge now is execution. Legacy infrastructure, regulatory barriers, low IT budgets and entrenched incentives won’t make AI adoption seamless - but they do create massive openings for those ready to build smarter, faster, and more patient-centric systems.

Healthcare won’t fix itself. The innovators who recognize this moment as an inflection point will define the next decade of medicine. I’m looking forward to seeing— and helping drive - this transformation.

Yet, recent developments suggest we are inching closer to a transformative moment for robotics. With breakthroughs in reinforcement learning, simulation, and multimodal AI, embodied AI - the fusion of AI with physical systems - is set to redefine automation. Following the panel discussion on “Embodied AI” at our AI House Davos 2025, I wanted to share some of the key advancements shaping the future of robotics. Are we on the verge of a robotics revolution, or are there still fundamental barriers to overcome?

Breakthroughs and Innovations in Embodied AI

The Role of Simulation in Robotics

One of the biggest obstacles in robotics has been the high cost and complexity of training AI-driven physical systems in the real world. Enter simulation. Today, advanced neural simulators allow AI models to learn and refine their behaviors before ever setting foot (or wheel) in the real world.

• Synthetic environments provide scalable training: AI systems can practice in high-fidelity virtual worlds, learning how to navigate, manipulate objects, and interact with dynamic environments.

• AI models can generate their own training data: With self-play techniques, simulators act as adversaries, constantly pushing AI to learn from novel and extreme scenarios that are difficult to collect from real-world data.

• Simulation bridges the gap between virtual and physical: The latest advances enable AI-trained in simulation to transfer to real-world settings with minimal adaptation, accelerating deployment timelines.

The increasing sophistication of simulation is a game-changer for industries like autonomous driving and industrial robotics.

AI-Powered Autonomy: From Trucks to Industrial Robots

Self-Driving Trucks Are Nearing a Commercial Reality

Autonomous vehicles have been in development for over a decade, but one of the biggest hurdles has been achieving full-scale deployment. While passenger car autonomy remains a work in progress, self-driving trucks are poised for a breakthrough.

Case Study: Waabi

• Waabi is set to launch driverless Class 8 trucks by the end of 2025, a major milestone for AI-driven logistics.

• The trucking industry has a severe labor shortage: With nearly 900,000 truck driver vacancies in the U.S. alone, automation offers a necessary and scalable solution.

• Simulation-driven safety testing: Waabi’s AI continuously trains in an advanced virtual environment, ensuring it can handle real-world unpredictability with unprecedented reliability.

Autonomous trucking represents one of the first commercially viable applications of embodied AI at scale, promising efficiency gains and enhanced road safety.

Industrial Robotics: Expanding Beyond Manufacturing

Robots in industrial environments are evolving beyond traditional fixed-arm automation. The latest generation of mobile robots is bringing autonomy to settings where human labor is scarce, such as inspections in oil and gas facilities, chemical plants, and warehouses.

Case Study: ANYbotics:

• ANYbotics’ quadruped robot, ANYmal, is already deployed globally to inspect and monitor industrial environments.

• Deep reinforcement learning enables adaptability: AI-powered robots can now handle stairs, dynamic obstacles, and real-world terrain with over 99% reliability.

• Embedded AI allows real-time decision-making: Unlike cloud-based AI models, these robots process sensor data locally, making instant adjustments based on their surroundings.

With industrial settings facing labor shortages and increasing safety regulations, robots like ANYmal offer a practical and scalable solution for physical-world automation.

The Challenges of Embodied AI

Why Robotics Is Harder Than Software-Only AI

Bringing AI into the physical world introduces new complexities beyond what we’ve seen with language models and software automation.

• Physical Constraints: Unlike AI that operates in the digital realm, robots must deal with physics: Weight, balance, power consumption, and durability all matter.

• Perception and Multimodal Inputs: Unlike software that processes text or images in isolation, embodied AI must integrate multiple sensor inputs: Cameras, LiDAR, radar, microphones—to accurately interpret the world.

• Safety and Regulation: Unlike software bugs, physical-world AI failures can lead to accidents. Proving safety before deployment is crucial, requiring robust testing and regulatory approval.

Humanoid Robots: Still a Long Road Ahead?

The hype around humanoid robots has grown rapidly, with major players investing heavily in this space. But are we truly close to humanoids performing useful work at scale?

Current humanoid robots excel in controlled environments, but they are not yet ready for complex, unstructured human interactions. Additionaly, Hardware limitations persist: Human-like dexterity, grip strength, and manipulation remain unsolved challenges. Practical applications will likely start in industrial settings before expanding to broader use cases.

So, while humanoid robots are making strides, we are still several years away from seeing them functionally integrated into everyday life.

The Future of Embodied AI: What Comes Next?

Safer, More Adaptive AI Systems

The future of embodied AI lies in creating systems that can reason, adapt, and interact more naturally with humans and environments.

• Contextual awareness will improve: Future AI-powered robots will better understand human movement, gestures, and speech, making them easier to deploy in collaborative settings.

• Hardware and AI will co-evolve: Smarter AI will reduce the need for expensive, high-precision hardware by compensating for minor mechanical inaccuracies.

• Regulatory advancements will shape adoption: Governments and industry stakeholders must work together to create safety standards that enable broader AI deployment while mitigating risk.

Final Thoughts

We are witnessing the dawn of embodied AI. While there are still many challenges ahead, industries like trucking, industrial automation, and logistics are already seeing tangible benefits from AI-powered robots. The dream of general-purpose humanoid robots may still be a decade away, but in many ways, embodied AI is already here—changing the way we work, move, and interact with technology.

As we continue to invest in this space at Merantix Capital, I’d love to hear your thoughts. Where do you see embodied AI making the biggest impact in the next few years?

We just got back from another year of AI House at Davos, a project our Merantix Group co-initiated with other great partners to drive forward progress in artificial intelligence.AI House is four fast-paced days of debate and collaboration on the most important technology in the world today — and I wanted to share some of the themes that were on the minds of the many investors, academics, entrepreneurs, and policy-makers that we had the fortune of meeting.

For this newsletter, I asked for some help from my colleague Thomas Wollmann, the CTO of Merantix Momentum, our AI solutions sister organization.

Here we go.

1. AI for Science

AI for science is a standout trend for us in 2025.

One billion year breakthrough

AI-assisted scientific discovery has been making breakthroughs behind the scenes for years, but its impact became impossible to ignore in 2024. A clear sign of this shift can be seen in last year’s Nobel Prizes, where AI wasn’t just a supporting tool—it was a major player in its own right.

In physics, John Hopfield and Geoffrey Hinton applied principles from physics to machine learning, laying the foundation for modern AI systems. Hopfield developed an associative memory network inspired by atomic interactions, allowing AI to reconstruct patterns by minimizing energy states. Hinton used the Hopfield network as the foundation for a new network that uses a different method: the Boltzmann machine. The model autonomously identifies patterns in data using statistical physics and enables AI to categorize images and even generate new ones, powering key advancements in deep learning.

AI also played a defining role in the Nobel Prize for Chemistry, with Google DeepMind’s AlphaFold2 solving the decades-old challenge of protein structure prediction. The project uses AI to predict the 3D structure of proteins from their amino acid sequences, a problem that had stumped researchers for 50 years. The impact of AI here can’t be understated. Mapping all 200 million known proteins would have taken over a billion years with human effort alone.

Beyond being a research milestone, this discovery is transforming drug development and improving drug design.

Currently, machine learning models interpolate existing knowledge—analyzing and recombining what is already known. But if AI gains the ability to design and conduct its own experiments, as we believe it will, artificial intelligence will generate entirely new knowledge, pushing scientific discovery far beyond human limitations.

Weather

We live in a world where weather is becoming increasingly unstable, with devastating effects on communities across the globe. Extreme storms, heatwaves, and other climate-driven disasters are harder to predict using traditional methods, which rely on historical data and physics-based models. Now, engineers no longer have to depend on guesswork and a handful of prototypes; AI allows them to process thousands of data points to create more accurate forecasts, helping to predict and protect against extreme weather events.

One example of this is NVIDIA’s CorrDiff. This correlator diffusion model works in two stages: first, predicting the mean of a fine-resolution weather field, then refining it with missing details. Trained on high-resolution radar data from Taiwan’s weather authorities, CorrDiff can scale extreme weather predictions from a 25km resolution down to just 2km, offering a significant improvement over traditional forecasting methods.

DeepMind’s GenCast is a diffusion-based generative model designed to enhance both everyday weather predictions and extreme event forecasting. Optimized for the Earth’s spherical geometry, it learns to generate future weather patterns based on real-time atmospheric conditions. Trained on four decades of historical weather data, GenCast can produce 15-day forecasts in just eight minutes on a single TPU, significantly outperforming the European Centre for Medium-Range Weather Forecasts' (ECMWF) leading ensemble system. DeepMind’s own testing reportedly showed GenCast was more accurate than ECMWF’s ENS in 97.2% of cases, underscoring the transformative potential of AI in weather prediction.

From insurance, to catastrophe prevention and every-day planning - weather prediction will become orders of magnitude more accurate in 2025.

2. Distributed Model Training

Much of today’s economic development is powered by technology. Technological advancement is powered by compute. Access to affordable compute will soon become an existential necessity for most nation states.

But that computing power is currently hoarded by a small handful of companies in the U.S. and China, where data centers can boast over 100,000 GPUs. In contrast, Europe’s largest has around 2,000. That’s why nations are scrambling to build technological independence with distributed model training.

Distributed AI model training is a process that accelerates the development of artificial intelligence models by dividing the training workload across multiple computing nodes, often referred to as worker nodes. These nodes operate in parallel, enabling the system to handle larger datasets and more complex models than would be feasible on a single machine.

Spreading the load across geographically disparate clusters of GPUs potentially solves the need for building new data centers and also avoids too much power being centralized in one location or country.

We think that the German Federal Agency for Breakthrough Innovation SPRIND’s initiative to develop decentralized methods of training powerful new AI models through open competition and funding is indicative of where this trend will continue to develop.

Either way, the big players have a wide technological and financial moat and the only way to break this stranglehold on innovation is to collaborate and decentralize.

3. Agentic AI

Agentic AI is one of the most talked-about advancements in artificial intelligence today, marking a major leap from passive assistants to autonomous decision-makers that bridge the digital and physical worlds.

Until now, AI has largely been limited to collecting and analyzing information, offering insights but leaving execution to humans. The next generation of AI agents changes that—by directly taking over hardware, making transactions, and automating real-world actions. This shift unlocks entirely new capabilities, turning AI from a tool that suggests actions into one that independently carries them out, fundamentally reshaping workflows, businesses, and even job markets.

Computer use agents will likely become standard in operating systems and office applications, much like how Excel macros once revolutionized workflow automation. But unlike macros, today’s AI can execute multi-step tasks autonomously—researching and booking travel, completing financial transactions, and updating schedules without human intervention. By handing AI control over personal and corporate computing power, users gain extreme efficiency, but at the potential cost of many traditional jobs.

At the same time, vertical AI is replacing human-driven, unstructured processes. This shift is particularly disruptive to roles that involve repetitive, manual tasks—often handled by interns, junior employees, or consultants—such as transferring data between systems or troubleshooting inefficiencies.

4. Synthetic data

2025 could be the year where AI models trained on vast amounts of real-world data start making a tangible impact across industries—from robotics and manufacturing to supply chain logistics and warehouse management.

One of the key technologies driving this transformation is the digital twin, a virtual model of a real-world environment that updates in real time using AI and sensor data.

Digital twins allow companies to test different scenarios, predict failures before they happen, and refine processes without disrupting operations. Nvidia’s Cosmos project is a prime example of this evolution, training AI models on 20 million hours of video and 9 trillion parameters to help machines understand the physical world with high accuracy. Paired with Nvidia’s Omniverse, which provides a simulation environment for industrial AI, Cosmos enables companies to develop and test automation solutions before deploying them in real-world settings.

This approach is expected to cut the time required to plan a new warehouse in half, while reducing manual labor and operating costs by up to 50%. In industries like retail, manufacturing, and consumer goods—where supply chain disruptions have caused major financial losses in recent years—this kind of AI-driven automation offers a way to increase resilience and efficiency.

This convergence of technology comes at a time where businesses are looking for savings anywhere they can find them. We’re feeling a lot of energy from business leaders and innovators to make 2025 the year when AI really shakes up the industrial sector.

5. Human-in-the-Loop Systems

Human-in-the-loop (HITL) systems will play a critical role in AI adoption as chat and voice interfaces mature - significantly impacting how millions of people work every day.

HITL systems leverage advancements in large language models (LLMs) to provide a bridge between human expertise and machine capabilities. This democratization of AI tools allows users without technical backgrounds to solve problems and deploy solutions, reducing dependence on programmers or data engineers.

These systems will not only empower users but also shift workflows in industries such as data engineering, customer service, and creative work. The ability to interact with machines in natural language removes a significant barrier to adoption, allowing businesses and individuals to incorporate AI into their processes seamlessly.

This year, we will also pass the point where people can tell whether or not they are speaking to an AI or a human. Many processes will likely be a blend of the two, especially customer service. AI can handle most processes, but humans can step in when expertise is needed.

6. Industry-Specific Models

AI is becoming more specialized, with models built for specific industries rather than just general use. In 2025, with large models being more and more commoditised this trend will accelerate, shaping fields like healthcare and biotechnology, where precision and efficiency are key.

In healthcare, AI-assisted screening is set to become even more common in 2025. Vara, a company backed by Merantix, has developed an AI tool that helps radiologists detect breast cancer more accurately. A study in Germany involving nearly 462,000 women found that AI-assisted screening detected 6.7% of cancers, compared to 5.7% with traditional methods—equating to one extra cancer found per 1,000 screenings. Crucially, AI didn’t increase false positives, meaning fewer unnecessary tests and less anxiety for patients.

In biotechnology, AI is reshaping how new materials and medicines are designed. Cambrium is using AI to create proteins that aren’t just effective but also easy to produce at scale. Instead of searching for naturally occurring proteins, they design new ones from scratch, optimizing them for human needs. One example is vegan human collagen, which Cambrium has developed for skincare products. As AI-driven protein design expands in 2025, it could transform industries like medicine, food production, and sustainable materials, making lab-created proteins a mainstream solution.

Let’s start building

AI in 2025 won’t just assist—it will act. It’s solving billion-year problems in science, reshaping weather forecasting, and optimizing industries with digital twins and automation. We’ll see governments around the world scramble to develop distributed model training to tap into much needed sources of growth and agentic AI will have an enormous effect on everything from work to daily life.

We’re excited to see what this year brings in AI. What are we missing? Are you building something that helps solve the world’s most pressing issues with AI? Let us know!

Finding ways to integrate AI into edtech isn’t just timely, it’s essential. The world is just waking up to the fact that technology is a few years away from rendering huge swathes of the global workforce redundant. Employees face tough decisions about how to stay relevant in a modern world and, likewise, companies won’t have much fun managing the costs of hiring new talent or trying to upskill existing workers with the long-outdated training methods still being used today.

This is where AI and edtech come together as one of very few solutions that is cost-effective, adaptable and scalable enough to tackle this looming threat. Daunting though the challenge may seem, hand-wringing isn’t going to get us anywhere. Now is the time for technical innovators, the business world and educators to come together and seize this opportunity.

Edtech must rise to the global challenge

The pandemic-era edtech boom saw consumer-focused platforms like Coursera and Duolingo thrive as millions turned to online learning. Riding the lockdown business-to-consumer (B2C) wave, global edtech funding hit $20 billion in 2021. But, as in-person activities resumed, funding dropped sharply to $3 billion by 2023.

In contrast, business-to-business (B2B) edtech has emerged as a robust, growing market. Instead of targeting individual learners, these platforms focus on corporate needs, providing tools to improve employee retention, productivity, and skills development. AI is central to this shift, creating tailored, scalable training that meets the growing need for more effective workforce development as industries grapple with automation and technological change.

The existential threat faced here by both businesses and employees is real. By 2030, up to 375 million workers will need to learn new skills or transition to entirely different roles to remain employable. AI-powered platforms will enable companies to address these challenges, bridging skills gaps and preparing employees for a rapidly changing workplace.

This is particularly true for Europe, which now finds itself at a crossroads. The great manufacturing and industrial players that powered the last century have lost their dominance over global supply chains. European economies are stagnating. Technology and industrial best practices have not kept in step with the rest of the world.

Aside from mass layoffs and site closures, most recently exemplified by Volkswagen in Germany, companies risk losing corporate knowledge without adequate measures in place to capture, archive and adapt it to suit the future of industry and its workforce. This is a huge opportunity for builders to step forward and act before it is too late.

Training that won’t cost your employee’s respect

Corporate training often feels like a box-ticking exercise rather than an opportunity to build a skilled and adaptable workforce. Programs are often repetitive, disengaging, and fail to produce lasting results. AI is transforming this space by creating personalized, interactive learning experiences.

The snappily-named study ‘When Generative AI Meets Workplace Learning: Creating A Realistic & Motivating Learning Experience With A Generative PCA’ by Andreas Bucher et al. highlights how Generative AI tools, like Generative Pedagogical Conversational Agents (GenPCAs), can create dynamic training tailored to individual needs. These AI-driven tools simulate workplace scenarios, mirroring real-world challenges to foster engagement and self-directed learning. Here’s what the study found:

• Higher engagement levels: Employees trained with GenPCAs reported a 40% increase in engagement compared to traditional methods.

• Improved retention: Scenarios designed by AI helped participants retain 35% more information on average.

• Realistic simulations: The ability of AI to mirror workplace-specific challenges increased perceived relevance of the training, leading to greater motivation to apply skills.

Academic studies are one thing. Real-world applications are another. Fortunately, one of our very own portfolio companies is ahead of the curve and is already implementing dynamic AI-training out there in the real world.

Company Shield is a cybersecurity training platform that transforms traditional training into interactive, scenario-driven exercises. Their approach uses continuous, AI-generated phishing simulations and deepfake scam training to immerse employees in real-world scenarios, helping users to recognize and neutralize risks in real time.

When you have a dedicated training session, the warning signs are obvious and the risks are usually weighted in a way to help employees pass. But having AI-generated risks introduced into employees’ daily workflow breeds a totally different mindset across an organization and requires all actors to keep their skills sharp at all times. By combining continuous training and a permanently vigilant approach to cybersecurity, Company Shield has helped organizations reduce phishing success rates by 25%.

This kind of dynamic, AI-powered training is a big improvement on embarrassing, outdated training materials most employees have to suffer through. But what about when employees don’t retrain and upskill, even when they have plenty of opportunities to? This brings us neatly to the "skill vs. will" problem.

While some employees are naturally motivated to learn and adapt, others may lack the willingness, even when they have the capacity. AI-driven tools can play a critical role in addressing this by analyzing behavioral patterns, engagement levels, and performance metrics to identify employees who are ready to upskill.

I can see scope for innovators in the edtech space to tackle the skill vs. will conundrum by tailoring learning paths to both ability and motivation. AI tools have the adaptability to monitor both these factors and adjust content based on an employee's aptitude and willingness to learn. Builders might also consider adding gamification and interactive tools to make training more engaging and impactful. By addressing these gaps, new platforms can help businesses improve workforce retention, productivity, and readiness for future challenges.

Prepare workers for the worst

AI virtual reality (VR) and augmented reality (AR) are a powerful combination for edtech. By combining lifelike simulations with AI-powered analytics, these technologies transform the way employees learn and practice skills, especially those that experience unique risks that are hard to recreate with today’s training methods.

VR training isn’t just a gimmick. PwC research found that employees trained with VR completed programs up to four times faster than those using traditional methods. Learners also reported a 275% increase in confidence to apply new skills and were 3.75 times more engaged with the content. AI enhances this further by dynamically adapting simulations to user data, ensuring that each session becomes progressively more effective

Pixaera is a VR training platform focused on safety-critical industries such as energy and manufacturing. Their training focuses on hyper-realistic, story driven simulations that put employees in situations that could lead to Serious Injuries and Fatalities (SIFs) - think limbs getting stuck in machinery and other grizzly workplace accidents.

With Pixaera’s VR training, the employees aren’t just watching a video, but living the experience. With AI, these scenarios can be adapted to each employee, as well as evolve to match real incidents and emerging trends observed in industry data. Pixaera reports a 98% user preference rate over traditional methods, with retention rates improving by up to 30%.

While VR/AR technologies are already driving change in high-stakes sectors, their broader potential is only beginning to be realized. Industries like retail and healthcare are exploring these tools to train employees in customer service and patient care. This is a great opportunity for builders with experience in industries that experience risks that can’t be adequately, effectively or affordably prepared for with traditional training methods.

Why it pays to integrate edtech into industrial operations

In manufacturing, where every second counts, inefficiencies are more than an inconvenience - they’re costly. From skill gaps to outdated processes, companies often find themselves losing valuable time and resources. AI is helping companies not just catch up but lead in a cut-throat corporate landscape.

A Springer Artificial Intelligence Review study found that AI-powered systems significantly enhance high-performance work environments by improving employee training and development. These systems not only automate tasks but also help employees adapt to complex processes through real-time feedback. Similarly, PwC’s report on AI in manufacturing showcases examples like FANUC’s Oshino factory in Japan, where robots autonomously assemble and inspect other robots. Companies are constantly changing, which means strategies to monitor both efficiency and opportunities to adapt employee training need to run around the clock.

Deltia.ai (another of our portfolio companies) leverages AI to optimize assembly lines, combining computer vision and real-time analytics to monitor workflows, identify inefficiencies, and create targeted training content. Using cutting-edge technologies like NVIDIA Jetson AGX Orin modules for edge computing, Deltia.ai provides immediate, localized insights in the workplace. Its real-time analytics dashboards convert raw data into actionable visualizations, enabling managers to pinpoint delays, misaligned workflows, and material handling inefficiencies. These insights empower managers to deploy targeted interventions, such as reallocating resources or retraining employees.

You can see where things will go from here across the entire sector. Multilingual AI training modules could help companies address skill gaps across diverse workforces, as well as boost retention and productivity. Similarly, integrating predictive maintenance with workforce training could enable systems to proactively identify and address machinery issues while providing relevant employee training in real-time. Innovators could also develop industry-specific AI solutions for sectors like construction or food processing, where safety and precision are critical.

Bespoke training materials and immortal executives

AI is transforming corporate training by helping companies turn their own materials and expertise into tailored, impactful resources. Beyond just improving training, it’s preserving critical knowledge, from retiring employees’ skills to invaluable organizational insights. If that’s not enough, AI is even bringing back long-dead executives, so they can impart their wisdom from beyond the grave.

Platforms like Colossyan and eduBITES digitize and adapt companies’ internal knowledge into engaging and scalable training content. Colossyan uses generative AI to transform static materials like PDFs and PowerPoint presentations into dynamic training videos, reducing content creation time by up to 50%. Similarly, eduBITES helps businesses capture expertise from retiring employees and niche specialists.

One extreme example of how AI can preserve the influence of extraordinary executives is Panasonic’s initiative to immortalize its late founder, Konosuke Matsushita, through an AI model. Developed in collaboration with the University of Tokyo-affiliated Matsuo Institute, the system uses Matsushita’s writings, speeches, and over 3,000 voice recordings to create a “digital clone” capable of emulating his thought processes and speech patterns. This AI not only preserves his management philosophy but also serves as a consultative tool for contemporary decision-making.

The idea that companies don’t need to bring in pre-made external content and can simply use the documents they generated throughout their existence, along with recordings of executives, is a total break from all previous corporate training.

Bigger public companies will have materials that can readily be transformed into training content. But smaller companies may not be ready for this. New platforms could be designed to help companies archive and record their activities for future training purposes, as well as to provide solutions that grow in scope and scale along with the companies themselves.

Corporate information is often a closely guarded secret, but many companies are waking up to the fact that building in public comes with benefits of its own. This opens up space for creating services that boost interoperability and bridging between new AI knowledge management and training programs. AI tools that can integrate seamlessly with existing knowledge bases, adapt to evolving workforce demands, and scale efficiently will be key to addressing the next generation of corporate training challenges.

The future of work demands innovation, and AI-powered edtech holds the key to unlocking it.

Professionals, technologists, and innovators have a unique chance to not just meet these challenges but shape the future of corporate learning. The need is urgent, the market is ready, and the tools are within reach. Now is the time to build the platforms and systems that will redefine education, improve businesses, and transform the workforce of tomorrow.

Insurance isn't the sexiest topic, but I must admit I find it inherently interesting. Maybe it’s because I’m German, where risk management —coupled with a long list of insurance policies — is woven into the fabric of our society. 

Mostly, I view the insurance landscape opportunistically: The industry is data-heavy, full of repetitive tasks, and burdened by a pressing need to modernize. For most, it’s an industry stuck in time. For an investor in AI companies, it’s an open goal for innovation.

While I can’t claim deep insurance expertise, recent developments across other industries demonstrate how AI breathes new life into struggling sectors. I believe AI has the potential to reshape insurance by streamlining underwriting, automating claims, and uncovering new efficiencies. Whether it’s navigating the unpredictability of climate events, mitigating complex cyber risks, or adapting to advances in automation, AI positions insurers to act more nimbly and make smarter, data-driven decisions.

In insurance, this shift isn’t just about marginal improvements. It’s about rethinking how the industry functions from the ground up, offering enormous upside potential for those ready to tackle insurance’s existential problems with technology. 

In this article I want to dive into a few ways I think insurance is ready for its AI moment. And of course, if you’re a founder building in this space, don’t hesitate to get in touch.

How AI Can Impact Insurance Businesses

Claims processing has barely changed in decades. Here’s what a typical process looks like:

• Initial Reporting: The policyholder reports the incident and shares basic details.

• Investigation: An adjuster verifies evidence and assigns liability.

• Damage Assessment: A mechanic or appraiser estimates repair costs.

• Approval and Payout: The insurer issues payment and closes the claim.

Multimodal AI, for instance, could streamline this entire workflow by processing text, images, and video simultaneously. This reduces hand-offs between departments, minimizes human error, and could save up to 75% in processing costs. 

So, one clear avenue is to integrate AI into existing insurance systems such as claims processing, upgrading traditional workflows without a full “overhaul.”

The second, more radical path to build true “AI-native” insurance companies. There are a few such examples already.

The 9-year-old venture Lemonade, for instance, showcases this potential by integrating AI across its operations. The company reports issuing policies and processing claims in as little as two minutes, and has since rolled out policies for renters, homeowners, cars, pets and personal liabilities, as well as functioning as an agency for other insurance agencies. As of November 2024, Lemonade is valued at over $2bn and recently beat its Q3 earnings, sending shares soaring by over 30%. 

French unicorn Shift Technology offers another example of AI's influence on the industry. The company provides AI-driven tools for real-time fraud detection and automated claims handling, helping insurers reduce financial losses and enhance process efficiency. Shift’s systems analyze incoming claims to identify unusual patterns, which supports more informed underwriting and streamlines the claims process to boost customer satisfaction. The firm boasts over 100 customers, including Generali France and Mitsui Sumitomo, and reports to have analyzed nearly two billion claims, data that will only serve to improve its algorithms. 

These two firms illustrate one of the industry’s biggest issues: legacy insurance companies had more data than they knew what to do with. This is where AI native companies will really pull ahead. Using APIs, these insurers can pull data from carriers, reinsurers, and other sources, instantly enabling custom underwriting and adaptive pricing. With the proliferation of connected devices - expected to reach up to a trillion by 2025 - AI will have even more data at its disposal to fuel personalized pricing, new product categories, and quicker service.

Using AI Insurtech to Tackle Humanity’s Existential Problems

The insurance industry doesn’t just need AI for efficiency—it needs it to stay resilient against new risks and growing demand from customers on contemporary challenges that are hard to price using traditional rails. Here’s why:

Climate

Extreme weather events are becoming more frequent, with over $1 trillion in damages recorded from more than 1,000 events in recent years. In 2022 alone, natural catastrophe claims surged by 54% compared to the previous 10-year average. This unpredictability makes it difficult for insurers to plan and price coverage effectively. Traditional models relying on historical data struggle as climate change destabilizes weather patterns, rendering past data less useful.

This is exactly the kind of situation where AI can shine. Let’s take the example of Taiwan, a country that is regularly battered by enormous storms. Traditional methods helped to mitigate the impact of extreme weather over time, but new AI models developed by NVIDIA have taken the predictive potential to the next level. 

NVIDIA developed CorrDiff, a correlator diffusion model that can deliver a high resolution forecast for the entire globe in a fraction of the time needed by traditional models. CorrDiff has two steps: a regression model that predicts the mean of the fine-resolution field and a second step where it refines with the missing details that weren’t captured in the first prediction. NVIDIA trained the model on high-resolution, radar-incorporated WRF data from the Taiwan weather authorities and was able to predict extreme weather from 25km down to a 2km resolution. 

Using advanced modeling, researchers achieved highly accurate storm simulations, enhancing both preparation and response. This is a potential game changer for insurers, as extreme weather scenarios change from educated guesswork to data-informed decisions that can be made in real time. Insurers can apply this precision to improve their pricing and reduce risks in regions prone to annual hurricanes and typhoons.

Tomorrow.io’s AI-driven weather intelligence provides another powerful example. The platform combines satellite data, ground sensors, and proprietary algorithms to deliver hyper-local forecasts, boosting accuracy by up to 80% over traditional models. A European insurer saved $12 million in claims within a year using this real-time intelligence to develop parametric insurance that triggers payouts based on predefined weather conditions, cutting down on administrative tasks and speeding up claim settlements by up to 25%.

Extreme weather is expected to get worse as the climate warms and areas that were previously safe will likely face growing environmental risk in future. Already, countries in areas that regularly face extreme weather are considered uninsurable. The need for modern, data-centric, parametric insurance solutions has never been higher and tools that cater to this have the potential to tap into a fast-growing market with urgent demand. 

Cyber

Cyber incidents have become the top global business risk, with costs projected to reach $10.5 trillion annually by 2025. This has driven an urgent need for robust cyber insurance, and AI is uniquely positioned to offer solutions at various levels:

• Prevention: AI-driven models can proactively identify vulnerabilities in company software, security setups, and network configurations, allowing insurers to make more precise risk assessments.

• Detection: AI establishes real-time behavioral baselines, detecting anomalies early, even within IoT-connected devices, enhancing underwriting with improved real-time intelligence.

• Response: Automated AI responses, like decoy environments, help isolate attackers and prevent data breaches, allowing human analysts to focus on complex threats and improving response speed.

Company Shield, a venture in our Merantix Capital portfolio, is working on the cutting edge of these issues from the vantage point of employee training, offering personalized simulation exercises of real-world AI attacks instead of boring long workshops and pro-forma “E-learning” training. 

CyberCube is another example of how this potential can be transformed into a comprehensive business offering. The firm’s platform processes extensive datasets to analyze potential cyber threats, supporting insurers in making informed underwriting decisions and managing aggregated risk exposure. By embedding AI in their analytics, insurers can leverage predictive insights to anticipate vulnerabilities and refine their risk models. This helps improve the precision of underwriting, optimize pricing, and develop more resilient cyber insurance offerings.

Besides using AI to boost traditional cyber insurance options, the potential for AI in cyber insurtech actually gets kind of meta. Whole industries are in the process of integrating AI into their workflows. But AI-driven decision-making, something that Dimitris Bertsimas and Agni Orfanoudaki research in their 2021 Algorithmic Insurance report, comes with a whole new class of risk. In this sense, AI models are cyber products that require their own insurance policies. Let’s take a look at this in practice. 

AI models are increasingly used to assist in detecting diseases as they are able to observe patterns and data that humans miss. The report presents a case study on breast cancer detection, illustrating how properties of the AI model, such as accuracy and interpretability, can influence the evaluation of insurance contracts. 

The authors’ research suggested that the accuracy and interpretability of these models can directly influence the evaluation of insurance contracts. This is because the reliability of an AI system affects the potential risk exposure insurers take on when underwriting policies tied to AI outcomes. For instance, a highly accurate and transparent AI model may lower perceived risk and lead to more favorable contract terms, whereas a less reliable or opaque model could result in increased premiums or tighter policy conditions. This underscores the importance of assessing AI performance as it shapes liabilities in insurance products.

This shift means AI models themselves may need their own insurance frameworks, developed through quantitative analysis and supported by specialized insurtech models.

Armilla is one of the companies developing new AI risk assessments, including AI liability. Most AI-related risks do not fall under traditional cyber policies, creating a massive problem for the 70% of companies that McKinsey estimates will use AI by 2030. Targeting this market gap, the company is now an MGA focusing on AI insurance with capacity from Chaucer, Greenlight Re and Swiss Re. 

Traditional insurers have frankly not been able to properly support companies implementing AI in their workflows. This is an urgent issue that will only get more pressing with time and the market is wide open for companies that can really get into the weeds of AI liability for an enormous, global market. 

Robotics in the Workplace

From factories staffed by autonomous devices to humans toiling alongside machines on assembly lines, robots are already pretty common in the workplace. 

VIDEO: Amazon showcases 1st fully autonomous mobile warehouse robot

But what happens when a machine powered by an AI algorithm hurts someone. Can you sue the robot? The factory owner? Or the developer that designed the algorithm? These are things that insurance companies will have to think about quite urgently, considering that up to 25% of manufacturing tasks are expected to be automated by robots in the next few years. 

According to a Lloyd’s research report, collaborative robots, known as “cobots” introduce new risks to the workplace, from potential injuries from unexpected movements to cybersecurity, operational failures and compliance issues. 

AI-powered monitoring systems enable insurers to continuously oversee robotic systems, detecting anomalies that could signal malfunctions. This proactive approach is critical in sectors like manufacturing and healthcare, where equipment failures can lead to significant costs or even physical harm. AI-driven sensors and real-time data analysis allow insurers to assess risks dynamically, improving both the precision and adaptability of coverage. 

As it stands, few insurance companies are in a position to handle the data required to monitor such new risks, or be able to adapt policies to them in real time. Robotics is a fast-moving and rapidly growing industry, but one lawsuit at the wrong time could easily shutter businesses. 

Now is the time for ambitious founders to step forward and create nimble, comprehensive insurance solutions that will protect companies, their employees and the innovation they are seeking to bring into the sector.  

The Future of Insurance

The insurance industry is ripe for transformation through artificial intelligence. This shift isn’t just about turning around a sector bogged down by legacy systems, it’s a chance for mission-driven founders to tackle pressing global challenges, make a tangible impact, and attract significant funding.

Climate change is an unfortunate reality, with increasingly unpredictable weather events demanding urgent solutions. AI finally empowers insurers to make real-time, data-informed decisions that mitigate risks and enhance response strategies. Similarly, the rise of AI has brought new cyber vulnerabilities, requiring tailored insurance solutions to cover emerging threats and even AI products themselves. 

Robotics, a frontier signaling the future of work and industry, introduces specific risks that most insurtechs are not yet prepared to address. Founders with expertise in automation, data, and insurance have a unique opportunity to lead in shaping solutions that meet these challenges. 

Traditional insurance might be fighting off obsolescence. But for agile companies leveraging AI, there has never been a better time to be in the industry. I’m excited to see (and help build and invest in) this next frontier in insurance. 

The next big opportunity lies in going beyond LLMs and tapping into the vast amount of real-world data that’s currently untouched by these models. I presented this a few weeks ago at Project A’s PakCon 2024, here’s the summary:

The Limitation of Text-Based Models

Today’s LLMs are impressive, but they all have one major limitation: they’re almost entirely fed with data from the web. Think about it—all the text, articles, and websites that form the backbone of training datasets are just a small fraction of the data that actually exists in the world. In fact, most of the meaningful data is found in the physical environment around us—data that LLMs haven’t even begun to touch.

Yes, datasets are getting bigger, chips are improving, and we’re throwing more compute power at the problem. But eventually, we’re going to hit a ceiling. There’s only so much original text content to mine from the web. If we really want to take AI to the next level, we need to go beyond what’s written and start processing the vast amount of non-verbal, real-world data that exists all around us.

Humans Don’t Just Use Text

Here’s a key insight: humans don’t just rely on text to make sense of the world. Around 90% of our sensory input is non-verbal—we see, feel, smell, and interact with our physical surroundings. AI needs to evolve in the same direction. If we want AI systems that can operate effectively in real-world environments, they need to process the same type of multimodal data that we do.

Think about autonomous cars. These vehicles generate massive amounts of data as they navigate the world, from sensor readings to camera footage, and all of it goes beyond text. AI models need to be able to process and learn from these kinds of data sources, which is where new technical tools like multimodality, zero-shot learning, and edge computing come into play.

Multimodality and Zero-Shot Learning: The Future of AI

The future of AI lies in multimodality—the ability to handle different types of data beyond just text, such as images, audio, video, and sensory data from the physical world. By building models that can understand and learn from this wide range of data, we’re opening the door to far more powerful AI systems.

Then there’s zero-shot learning—a technique where AI models can apply their knowledge to new data types they’ve never encountered before. This allows AI to generalize more effectively and handle situations that weren’t explicitly part of its training data. Imagine an AI system that can recognize and process new objects or events in real-time, even without prior examples.

And don’t forget about edge computing. By deploying hardware directly into real-world environments, we can push AI systems closer to the data they need to process. This not only increases efficiency but also allows for real-time decision-making in areas like manufacturing, robotics, and healthcare.

Real-World AI: The Untapped Opportunity

Here’s where the real opportunity lies: while LLMs are great at text-based tasks, the vast majority of data that could fuel the next wave of AI innovation is out there in the physical world. Yann LeCun famously pointed out that even a four-year-old child has processed 50 times more data than any LLM. That’s because humans interact with the real world, and the same needs to happen with AI.

In sectors like manufacturing, healthcare, energy, and logistics, there’s an enormous amount of untapped data waiting to be processed by AI systems. For instance, one of our companies, Deltia, uses computer vision in factories to analyze workflows, making them more efficient. This data can eventually be used to train robots, further automating tasks and pushing AI into new areas.

And it’s not just about efficiency. There’s a wealth of insights, improvements, and innovations that can come from harnessing this data, and right now, most of it is going unprocessed. Every industry stands to benefit from AI that can understand and interact with the real world, and this is where Europe, in particular, has a unique opportunity.

Europe’s Unique Opportunity

Unlike the U.S. and China, which have largely focused on tech-centric AI developments, Europe has a rich industrial heritage that offers a distinct advantage. Many of the industries where AI can have the most immediate impact—such as manufacturing and logistics—are areas where Europe excels. While the race for LLMs may be led elsewhere, Europe has the potential to lead the next phase of AI development by applying domain expertise to real-world AI applications.

The key is to act now. As these industries embrace AI, there’s a window of opportunity for Europe to enhance and scale its expertise in these fields, making a global impact.

Unlocking AI’s Next Frontier

The future of AI isn’t just about making LLMs better. It’s about accessing and understanding the data that exists all around us in the physical world. By focusing on multimodal AI, zero-shot learning, and edge computing, we can unlock entirely new possibilities for innovation and business growth.

Whether you’re in manufacturing, healthcare, logistics, or any other sector with real-world data, the next big wave of AI is coming your way. All you need to do is prepare to tap into it.

Subscribe now

When news broke this week that John Hopfield and Geoffrey Hinton, two machine learning pioneers, won the Nobel Prize in Physics, it felt like a profound moment of recognition for AI. 

Simplifying things greatly, Hinton is widely regarded as the “godfather of AI,” and his work on neural networks and deep learning are fundamental to the field, particularly on “backpropagation” which influenced how models make predictions based on inputs. Hopfield similarly offered the core of the foundation of modern machine learning, including innovations in pattern recognition. 

Their work, which came decades ago, paved the way for more modern and advanced breakthroughs we’ve seen in recent years, many of which I’ve written about on this Substack. 

When a reporter called me for my thoughts, I said that the award would likely be the first of many Nobel Prizes offered for research in AI. I figured that we’d eventually see Nobels for sector breakthroughs, such as a new AI innovation in healthcare that saves millions of lives. 

Well, it turned out that we only had to wait one more day for another such recognition. On Wednesday, Demis Hassabis and John Jumper of DeepMind and David Baker, a computational biologist, won the chemistry Nobel Prize prize for their work on protein structures.

Their breakthroughs are extremely important for the future of health and medicine, especially in the development of drug discovery, and can help speed up processes in the pharmaceutical industry and reduce costs. A big credit to Google as well for its support of AI research through DeepMind and other initiatives, showing how scientific research and commercial developments can go hand in hand to create truly innovative AI tools that can reach the market quickly and have an immediate, tangible impact.

I really believe in this future, which is why I think these prizes offer a signal of AI’s full arrival and the power this technology on the application layer across industries like healthcare and techbio, where we see true value. At Merantix Capital, we recently invested in Graph Therapeutics, for instance, which uses AI-driven precision medicine in a lab-in-the-loop discovery engine in order to tackle immune-driven diseases.

Or take our venture Cambrium, which designs molecules at the amino-acid level, achieving micro-molecular precision. Cambrium’s first product is a sustainable collagen, but as founder Mitchell Duffy told TechCrunch, the sky is really the limit when it comes to future materials. You can imagine how this technology could change textiles, industrial materials, and much more — driving down cost but, most importantly, doing so sustainably.

It’s easy to get caught up in AI hype, boardroom dramas, or science fiction, and forget the potential impact of this technology: making our lives and our planet better. I think we’ll see a lot of other “AI Nobel Prizes” given to interdisciplinary researchers in many other sectors.

Read more here about Merantix Capital.

Subscribe now

Artificial intelligence is rapidly transforming various sectors, bringing forth innovations that redefine how we work and live. The rate of new discoveries has become so (excitingly) overwhelming that I thought it would make sense to highlight a few select research developments that really exemplify AI’s profound potential. I try to stay on top of developments by always looking at the potential impact they have for specific industries, and in this Substack I want to give a glimpse of where my thinking is currently. In the past weeks, amazing research has been published providing evidence for great opportunities in the following three fields: knowledge work, cybersecurity, and interactive media. Let’s dive in!

From Code to Finance

Artificial intelligence is making significant strides in enhancing productivity across various domains of knowledge work. Recent research provides compelling evidence of how generative AI models are transforming tasks that were traditionally time-consuming and reliant on specialized expertise.

AI Accelerates Software Development

A groundbreaking study involving software developers from Microsoft, Accenture, and a Fortune 100 electronics company demonstrated the tangible benefits of integrating AI into coding practices. Developers using an earlier version of GitHub Copilot, powered by GPT-3.5 technology from 2021, experienced a 26% increase in completed tasks compared to those who did not use the AI assistant. The most substantial productivity gains were observed among junior and less experienced developers, highlighting AI's potential to elevate the capabilities of those early in their careers.

This randomized controlled trial is one of the first large-scale, real-world experiments to provide concrete evidence that generative AI tools can significantly enhance productivity in high-skilled work environments. As AI models continue to advance, we can anticipate even greater impacts on software development efficiency and innovation.

Transforming Financial Statement Analysis

In the finance sector, AI models are emerging as powerful tools for interpreting complex financial data. A recent study by researchers at the University of Chicago assessed GPT-4's capabilities in financial statement analysis—a task that requires critical thinking and nuanced judgment. The findings revealed that GPT-4 not only matched but often surpassed professional financial analysts, especially in challenging scenarios.

GPT-4 excelled in analyzing financial statements, offering predictions and insights that were on par with, or superior to, those of human analysts and specialized machine learning models. Unlike traditional models, GPT-4 was able to generate valuable narrative insights about a company's future performance, providing a unique qualitative edge in financial analysis. Its analysis was based on understanding trends, financial ratios, and economic reasoning rather than relying on memorized data.

Investment strategies derived from GPT-4's predictions yielded higher Sharpe ratios and alphas, indicating the potential for improved investment returns. Interestingly, the study found that GPT-4 and human analysts are complementary. While GPT-4 excels in data analysis and pattern recognition, human analysts contribute additional context and industry-specific knowledge. This synergy suggests that large language models like GPT-4 could play a central role in future financial decision-making, driving efficiency and innovation within the industry.

By incorporating AI tools into knowledge work, organizations can unlock new levels of productivity and insight. The ability of generative AI to assist with complex tasks not only streamlines workflows but also opens up opportunities for professionals to focus on higher-level strategic activities. As AI technology continues to evolve, its integration into knowledge-intensive industries promises to redefine the future of work.

AI's Role in Modern Cybersecurity

Artificial intelligence is increasingly influential in the realm of cybersecurity, offering new avenues for both defending against and executing cyber operations. Recent research showcases how advanced AI models are impacting offensive and defensive strategies.

Automated Offense

A recent study evaluated the performance of GPT-4 in offensive cybersecurity challenges, specifically in a Capture The Flag (CTF) competition. CTF events are designed to test participants' abilities across various cybersecurity domains, including forensics, binary exploitation, cryptography, reverse engineering, and web security. Remarkably,  GPT-4 performed in the 88th percentile compared to human participants, effectively outperforming the average human hacker in these challenges.

The study also assessed other AI models like GPT-3.5 and Mistral, which, despite being less advanced than GPT-4, demonstrated competent performance given their smaller sizes and earlier development stages. As AI models continue to evolve, we can anticipate even greater proficiency in tackling complex cybersecurity tasks.

This advancement signifies a pivotal shift in cybersecurity, where certain hacking tasks can be automated using AI. It highlights the urgent need for organizations to integrate AI into their cybersecurity operations to stay ahead of potential threats. By leveraging AI capabilities, security teams can enhance their defensive measures and more effectively identify and mitigate vulnerabilities.

Machine Learning Fortifies IoT Security

The rapid expansion of the Internet of Things (IoT) has interconnected a multitude of devices across various sectors, from healthcare systems to smart city infrastructures. While this increased connectivity offers numerous benefits, it also amplifies security risks, making IoT networks attractive targets for cyberattacks.

A recent study introduced a machine learning-based model designed to autonomously detect and respond to cyber threats within IoT networks with an impressive 99.9% accuracy. By employing sophisticated computational methods, the model not only identifies threats faster but also outperforms previous approaches in both accuracy and speed. This advancement significantly strengthens the security of IoT ecosystems, providing robust protection against evolving cyber threats.

The applicability of this model spans a wide range of industries utilizing IoT devices. In healthcare, it can safeguard sensitive patient data and ensure the reliable operation of medical devices. For smart cities, it enhances the security of infrastructure systems like traffic management and public utilities. As IoT's economic influence continues to grow, securing these networks becomes increasingly critical to prevent disruptions and protect valuable data.

Implementing AI-driven security measures such as this machine learning model is essential for mitigating risks associated with the proliferation of IoT technologies. This research highlights the crucial role that artificial intelligence will play in securing the future of connected technologies, emphasizing that integrating machine learning into IoT security frameworks is not just a trend but a necessary evolution to keep pace with emerging threats.

By embracing AI in cybersecurity, organizations can enhance their defenses against sophisticated threats. The ability of AI models like GPT-4 to perform complex hacking tasks highlights both the potential risks and opportunities. Proactively incorporating AI into security operations can help ensure these powerful tools are used to strengthen defenses rather than exploit vulnerabilities.

AI-Generated Realities

Artificial intelligence is making significant strides in the gaming industry, introducing innovative methods for game development and interactive experiences. Google's recent research presents a notable example with GameNGen, the first game engine entirely driven by a neural network. This development not only signals a transformative moment for gaming but also has broader implications for the generation of interactive media.

GameNGen utilizes a diffusion model to predict and generate each frame of a game based on player actions. In their study, researchers demonstrated this technology by running the classic game DOOM at over 20 frames per second on a single Tensor Processing Unit (TPU). The visuals produced are so close to the original that even seasoned players find it challenging to distinguish between the two. This achievement underscores the potential of AI to replicate and enhance complex real-time environments efficiently.

The significance of GameNGen is heightened when considering the substantial size and growth of the gaming industry. Valued at $220 billion in 2023 and experiencing a compound annual growth rate (CAGR) of 8-10%, the gaming sector is a major player in the global entertainment landscape. Innovations like GameNGen could accelerate this growth by streamlining development processes, reducing costs, and enabling greater creative freedom. By harnessing AI for game engine operations, developers can experiment more freely and bring innovative ideas to life without the constraints of traditional development pipelines.

Beyond gaming, the technology behind GameNGen has profound implications for the broader field of interactive media. The ability to generate high-quality, real-time simulations through AI can transform how interactive content is produced and experienced across various platforms. In virtual reality (VR) and augmented reality (AR) applications, AI-driven engines can create more responsive and immersive environments, enhancing user engagement and interaction. This advancement benefits areas such as education, where interactive simulations aid in teaching complex concepts, and training scenarios that require realistic and adaptable environments.

In the realm of digital entertainment and multimedia, AI can streamline the production of interactive experiences, allowing for personalized content that adapts to user preferences and behaviors. Real-time content generation reduces production times and costs, enabling creators to focus more on storytelling and artistic expression. Additionally, AI-generated interactive media opens up new possibilities in fields like interactive storytelling, immersive art installations, and virtual events.

The integration of AI technologies like GameNGen into the creation of interactive media underscores a broader trend toward more dynamic and user-centric experiences. By lowering the barriers to developing sophisticated interactive content, AI empowers creators to innovate more rapidly and deliver experiences that are engaging and tailored to individual users. This evolution could lead to more immersive, efficient, and personalized interactions, signaling an exciting future for interactive media.

Embracing Transformative Technologies Across Industries

The advancements highlighted in this article are more than just promising research—they are catalysts for real-world change across multiple industries. From enhancing productivity in software development and financial analysis to transforming cybersecurity practices and revolutionizing interactive media, artificial intelligence is actively reshaping the way we live and work. These developments demonstrate AI's capacity to not only optimize existing processes but also to open up new avenues for innovation and growth.

Organizations that embrace these AI technologies are unlocking unprecedented levels of efficiency, security, and creative potential. 

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.

Subscribe now

Synthetic data is rapidly gaining traction in the AI community, hailed as a solution to data scarcity. With models like Llama 3.1, smaller companies can now generate large amounts of high-quality synthetic data, democratizing AI development.

What makes synthetic data particularly exciting is its growing range of practical applications. From advancing medical diagnostics to enhancing the training of autonomous systems, synthetic data is already proving to be a valuable tool in pushing the boundaries of what AI can achieve. It offers a flexible, scalable way to create the datasets needed for complex AI models, especially in areas where collecting real data is challenging or costly.

That’s the good news.Still , it’s important to recognize that while synthetic data is incredibly useful, it’s not a magic solution. 

Yes, synthetic data is a powerful tool with great potential, but it works best when integrated thoughtfully into a broader data strategy. In this post, I’ll explore both the exciting possibilities and the practical considerations that come with using synthetic data, aiming to provide a balanced view of its role in the future of AI.

What is Synthetic Data?

Synthetic data is artificially generated data designed to replicate the structure, patterns, and statistical properties of real-world data. Unlike real data, which is collected from actual events or observations, synthetic data is produced through various techniques, including algorithmic generation, machine learning models, and simulations.

There are several key types of synthetic data:

• Fully Synthetic Data: This type is created entirely from scratch, often using AI models or statistical algorithms, without directly referencing real-world data. It’s particularly useful when real data is scarce, sensitive, or expensive to obtain.

• Simulation Data: A vital subset of synthetic data, simulation data is generated by creating virtual models of real-world systems, processes, or environments.

• Hybrid Synthetic Data: This type combines real and synthetic data, where synthetic data is used to fill gaps, augment, or enhance existing real-world datasets. This approach is often used when real data is limited or needs to be diversified.

• Augmented Data: Real data that has been modified or enhanced with synthetic elements to expand its utility. For example, in computer vision, existing images can be altered to create new variations, providing more comprehensive training datasets.

The use of synthetic data is becoming increasingly critical in AI development. It enables the creation of large, diverse datasets essential for training AI models, particularly in fields where real data is difficult to collect or where doing so poses privacy concerns.

The Promises of Synthetic Data

One of the most compelling advantages of synthetic data is its ability to alleviate the problem of data scarcity, particularly in areas where real data is difficult, expensive, or ethically challenging to obtain. In industries like healthcare, finance, and autonomous systems, acquiring enough  real-world data for training AI models can be a bottleneck. This is where synthetic data comes in, offering a scalable alternative that can replicate the patterns needed for robust model training.

Synthetic data is not just about quantity; it’s also about diversity. In fields where edge cases and rare events are crucial, synthetic data can help fill the gaps in real datasets, ensuring that AI models are trained on a more comprehensive range of scenarios. This diversity is key to building models that are not only accurate but also resilient and capable of generalizing to new, unseen situations.

Recent research by Google Deepmind, Stanford University, and Georgia Institute of Technology further underscores the potential of synthetic data in overcoming these challenges. The paper highlights how synthetic data can provide an abundant supply of training and testing data, especially in domains where real-world data is scarce or difficult to obtain. The research emphasizes the benefits of synthetic data in generating tailored datasets that improve model performance and generalization, particularly for tasks requiring balanced representation or multilingual capabilities. Moreover, synthetic data's ability to mitigate privacy concerns by creating anonymized datasets is crucial in sensitive domains like healthcare.

To better understand its impact, let's look at three recent papers that exemplify the practical applications and benefits of synthetic data.

1. Agent Hospital: A Simulacrum of a Hospital with Evolvable Medical AgentsThis paper, published by researchers at Tsinghua University, introduces "Agent Hospital," a simulated hospital environment where autonomous agents—representing patients, doctors, and nurses—interact and learn over time. The synthetic data generated within this environment allows doctor agents to accumulate experience by treating thousands of simulated patients. The result? These agents achieved a state-of-the-art accuracy of 93.06% on respiratory diseases, demonstrating how synthetic data can be used to train AI in complex, life-like scenarios. This approach has significant implications for medical training and decision-making, particularly in cases where real-world data is scarce or sensitive.

2. Scaling Synthetic Data Creation with 1 Billion PersonasIn this research, the authors introduce "Persona Hub," a methodology that scales synthetic data creation using a vast collection of 1 billion personas. This synthetic data spans a wide array of contexts, from logical reasoning problems to rich textual content, making it highly versatile. By leveraging such diverse and large-scale synthetic data, AI models can be trained more effectively, improving their performance across various tasks and applications. This example highlights how synthetic data can address not only the quantity but also the diversity of training data, which is crucial for developing robust AI systems.

3. RoboCasa: Simulating Realistic Home Environments for Robotics TrainingResearchers from The University of Texas at Austin and NVIDIA developed RoboCasa, a large-scale simulation framework designed to advance robot learning. With 120 simulated environments, 2,509 objects, and 100 tasks, RoboCasa generates synthetic data that helps train robots to perform tasks in realistic home settings. This simulation data is critical for training robots to operate in environments that would be difficult or costly to replicate in the real world. The result is improved robot performance and adaptability to real-world tasks, demonstrating the power of synthetic data in advancing AI-driven robotics.

The Limitations of Synthetic Data: Understanding Model Collapse

While synthetic data offers significant advantages, it's important to acknowledge its limitations. One key issue, highlighted in the recent paper "AI Models Collapse When Trained on Recursively Generated Data," is the phenomenon of model collapse. This occurs when AI models are trained on synthetic data that has been generated by previous models, leading to a gradual loss of information about the original data distribution.

Model collapse is a process where AI models, when trained repeatedly on data generated by earlier models, start to lose touch with the true diversity and nuances of the original data. Over time, this can result in models that are less accurate and less capable of handling rare or unexpected scenarios.

The paper shows that if future AI models are trained predominantly on synthetic data produced by their predecessors, they may converge towards a narrower understanding of the data, potentially missing important variations that were present in the original real-world data. This highlights a key limitation: synthetic data, while useful, cannot fully replace the richness of genuine human-generated data.

The findings also suggest that while synthetic data is a powerful tool, especially in situations where real data is scarce, it should be used thoughtfully. Over-reliance on synthetic data, particularly when it’s generated recursively, can lead to models that underperform or fail to generalize effectively.

This doesn’t diminish the value of synthetic data but underscores the importance of balancing it with real-world data. By maintaining a mix of real and synthetic data, AI models can retain the robustness and accuracy needed to perform well across a range of scenarios.

A Thoughtful Approach to Synthetic Data

So where do we stand right now in the discussion about synthetic data? I think it’s an exciting and valuable resource in AI development, but it’s not a cure-all. The phenomenon of model collapse reminds us that while synthetic data can enhance and extend real datasets, it’s essential to continue incorporating real-world data to maintain the quality and reliability of AI models.

By using synthetic data as a complement rather than a replacement, we can maximize its benefits while avoiding the pitfalls associated with recursive training.

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.

Subscribe now

Robotics has long been a field challenged by some of the toughest problems in technology. From the complexities of mechanical design to the intricacies of programming autonomous behavior, the journey has been consistently challenging. Moravec's Law, which states that tasks we find easy (e.g. motoric skills) are difficult for robots and vice versa, encapsulates the historical struggles in robotics. The promise in robotics has always been great: breakthroughs could revolutionize industries and daily life, transforming everything from manufacturing to healthcare. In the past few weeks, we have seen some impressive signs that we are getting close to a GPT-moment for Robotics. As we are getting more and more bullish on the robotics space here at Merantix, I wanted to share some of the recent breakthroughs and what we believe they will mean in the application for the coming years. Let me know if you feel we have missed some or if you have differing views!
Let’s jump in:

Recent Breakthroughs and Innovations

Recent years have witnessed remarkable progress in the field of robotics. Researchers are increasingly utilizing simulations and synthetic data to train robots, effectively overcoming the issue of data scarcity.

Simulation and Synthetic Data

Simulations allow researchers to create virtual environments where robots can be trained and tested without the constraints of the physical world. These simulations can be highly controlled and tailored to specific tasks, providing an abundance of data that would be difficult or expensive to gather in real life. As foundation models improve and multimodality increases, our ability to create sophisticated simulations also advances.

For example, a recent study demonstrates the use of Video2Game, a system that transforms single videos into real-time, interactive, realistic, and browser-compatible environments. Systems like that could significantly enhance the training of robots by allowing them to interact with diverse and complex scenarios generated from video footage.

Synthetic data, generated through these simulations, enables robots to learn and adapt to various scenarios, enhancing their performance. This is not only important for robotics. Synthetic data is trending in many domains where data is hard to come by. A recent paper by Researchers from Google Deepmind, Stanford University, and Georgia Institute of Technology nicely summarizes how synthetic data can be used to make the training of AI models cheaper and create higher performing models. The authors of the paper point out how synthetic data already improves model capabilities in regards to reasoning, tool-using and planning, multimodality, multilinguality and alignment.

Pushing Limits: Robots Trained in Virtual Worlds

Recent research has produced several impressive examples of robots trained in simulation performing complex tasks. Here are three notable instances:

1. Humanoid Robot Walking in San Francisco:

A recent breakthrough by UC Berkeley researchers showcases the remarkable progress in robotic control systems. Their model enabled a full-sized humanoid robot to walk through San Francisco in a zero-shot setting, meaning the robot could operate in the real world even when trained solely on 27 hours of walking data. Impressively, the robot could generalize to new commands not encountered during training, such as walking backward. This advancement underscores the potential for rapid improvements in real-world robotics. The researchers attribute their success to several factors:

• Decreasing Cost of Data Collection: Large-scale datasets for training are becoming more affordable.

• Effectiveness of Transformer Architectures: These models excel in this context, facilitating better control and adaptability.

• Emerging Scaling Laws: These laws indicate that increasing the amount of data, expanding context windows, and utilizing larger models significantly reduce tracking errors.

2. Robots Playing Soccer with Deep Reinforcement Learning:

Another groundbreaking development comes from Google DeepMind. Their recent paper demonstrates how robots can be trained to play soccer using onboard cameras and deep reinforcement learning. By leveraging egocentric RGB vision, these robots learn to perceive and react like humans.

High realism in simulations was achieved by combining physics with advanced rendering techniques, enabling complex behaviors developed in simulations to transfer zero-shot to real-world robots equipped with low-cost sensors.These soccer-playing robots can track the ball, avoid opponents, and execute sophisticated strategies, marking a significant milestone in end-to-end robot training.

This research opens up vast opportunities for scalable, cost-effective robotics solutions across various industries, from manufacturing to logistics. The use of affordable sensors and realistic training environments makes advanced robotics more accessible and practical. If only it could help Germany win Euro 2024, then it would really be transformational.

Enhanced simulation tools and deep reinforcement learning capabilities pave the way for developing autonomous systems capable of complex coordination and decision-making. The principles demonstrated in this research can be applied to many dynamic environments, expanding AI’s real-world applications and offering substantial returns on investment in both financial and technological advancements.

The Challenge of Sim-to-Real Transfer

Even with the remarkable advancements in robotics through simulation, the transition from simulation to real-world application, known as sim-to-real transfer, remains a significant hurdle.

Understanding Sim-to-Real Transfer

Sim-to-real transfer refers to the process of taking a robot or AI system trained in a simulated environment and deploying it successfully in the real world. While simulations provide a controlled, cost-effective, and risk-free environment for training robots, real-world environments are often unpredictable and complex. Factors such as varying lighting conditions, sensor noise, unmodeled dynamics, and physical interactions can drastically affect a robot's performance when it moves from simulation to reality.

Challenges in Sim-to-Real Transfer:

1. Environmental Differences: Robots may struggle with real-world variables like poor lighting and unexpected obstacles, which are not present in controlled simulations.

2. Sensor Discrepancies: Real-world sensors often encounter noise and interference, unlike the ideal conditions in simulations.

3. Physical Interactions: Real-world interactions, such as friction and collisions, can be more complex and unpredictable than simulated models.

4. Generalization Issues: Robots trained in specific simulated environments may not perform well in different real-world settings.

Using AI to Overcome Sim-to-Real Transfer Challenges

Despite these challenges, recent advancements offer promising solutions. A newly published paper on DrEureka warrants optimism about the prospects of robotics in the near future.

DrEureka uses GPT-4 to simplify and speed up the process of training robots in virtual environments before deploying them in the real world. Normally, designing reward systems and adjusting virtual training conditions requires significant manual effort from experts. By automating these tasks, DrEureka significantly reduces the amount of human input needed and delivers better results. The method has proven effective in training robots to perform tasks like walking and manipulating objects and allows them to tackle new challenges, such as balancing on a yoga ball (even while it is being deflated or kicked).

Keep in mind that GPT-4 is a model from 2022 that was not fine-tuned for sim-to-real transfer. This demonstrates the potential for AI to revolutionize the field, making the once laborious and complex process of sim-to-real transfer much more efficient and effective.

These advancements suggest that robotics might experience significant transformation once we adopt these insights into industries. The future of robotics, once hindered by the challenge of sim-to-real transfer, is becoming increasingly promising with the integration of advanced AI and innovative training methods.

Practical Implications for Industries

The advancements in AI and robotics outlined in this guide highlight a transformative potential across various industries. From the historical challenges encapsulated by Moravec's Law to the recent breakthroughs in simulation, synthetic data, and deep reinforcement learning, the field of robotics is on the cusp of a significant evolution.

Revolutionizing Industries

1. Manufacturing:

   • Enhanced Automation: The ability to train robots in virtual environments allows for more sophisticated and adaptable manufacturing processes. Robots can be programmed to handle complex tasks, improve precision, and operate in unpredictable conditions, thus reducing downtime and increasing efficiency.

   • Cost-Effective Scaling: The decreasing cost of data collection and the effectiveness of transformer architectures enable scalable solutions, making advanced robotics accessible even to smaller manufacturers.

2. Healthcare:

   • Precision and Reliability: The integration of AI with robotics can improve the precision of surgical robots, assist in complex procedures, and enhance patient care through consistent and reliable performance.

   • Training and Simulation: Medical robots can be trained in simulated environments to perform intricate tasks, reducing the risk associated with real-world training and enhancing the capabilities of healthcare providers.

3. Logistics and Supply Chain:

   • Dynamic Adaptability: Robots trained with advanced AI can adapt to changing environments, manage inventory more effectively, and optimize supply chain operations. The ability to generalize from simulations to real-world scenarios means that logistics can be more responsive and efficient.

   • Cost Reduction: Utilizing affordable sensors and realistic training environments lowers the barrier to entry for high-tech solutions, making it feasible for a broader range of businesses to adopt robotic automation.

Future Prospects

As industries adopt these innovations, we can expect significant improvements in efficiency, cost-effectiveness, and the ability to perform complex tasks. The journey from theoretical breakthroughs to practical applications is well underway, signaling a new era where robotics plays a crucial role in transforming industries and enhancing our daily lives. The future, once hindered by the challenge of sim-to-real transfer, is now increasingly promising with the integration of advanced AI and innovative training methods.

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.

Subscribe now

Until recently, the buzz around foundation models centered around LLMs like GPT-4. We're beginning to witness a shift as models extend their reach into various modalities, paving the way for a future where AI-generated content transcends text to include video, speech, games, and more. As the release of Sora from OpenAI demonstrated, builders and investors should start paying attention to these new developments if they aren’t already. 

The concept of foundation models is effectively expanding to incorporate a broader spectrum of modalities, from video and speech generation to even game creation. This diversification is backed by scaling laws that demonstrate remarkable improvements in model performance across these different domains, leading to more and more modalities experiencing their “GPT moment”.

Beyond Text Generation: A Foundation Model Zoo

In the last month, we've seen groundbreaking advancements with new AI models emerging:

• Speech Generation: Amazon's foray into speech generation with its Big Adaptive Streamable TTS model (BASE) underscores the principle that scaling laws applicable to LLMs also hold true for speech data. BASE demonstrates emergent abilities, enhancing naturalness and the representation of complex linguistic features. This advancement signifies progress in creating more lifelike AI voices. This ability of BASE opens new doors for creating more natural and engaging user interfaces, significantly enhancing user experience in digital assistants, e-learning platforms, and customer service bots.

• Games Generation: DeepMind's Genie takes a novel approach by generating interactive environments from unlabelled Internet videos, including gameplay dynamics. This capability to transform static images into playable game environments opens up numerous possibilities for personalized entertainment and educational tools, emphasizing the creative potential unleashed by generative AI models. This novel approach to generating interactive environments, illustrates a direct application in revolutionizing game design and educational content.

• Video Generation: Sora represents a leap in video generation, combining GANs, diffusion models, and transformers to create videos with improved object permanence and 3D consistency. While current limitations in reasoning and physics modeling exist, the potential applications—from animating family photos to revolutionizing robotics training with simulations—are immense.

However this is just the beginning. These innovations hint at a future where such technologies are seamlessly integrated into the production of content and simulations, revolutionizing industries including entertainment, marketing, education and robotics.

It is crucial to understand how these innovations translate into real-world applications that touch our everyday lives. The next section dives deep into the transformative use cases and industries poised to benefit from these technologies, highlighting the direct link between AI's capabilities and the future of various sectors.

The transformative implications we see

The next unicorn in media production will likely emerge from a startup that not only masters the technical complexities of AI-powered content creation but also deeply understands the creative and storytelling needs of its users.

In this light, the applications and use cases discussed here are more than just examples of what is possible; they are a call to action for innovators, creators, and entrepreneurs to envision and build the future. 

The evolution of AI-generated content is poised to impact a wide array of sectors. Here's a closer look at the industries and use cases that stand to benefit profoundly:

Enhancing Robotics with Simulation Training

• Robots can now be trained in virtual environments that mimic the real world closely, significantly accelerating the learning process while reducing the risks and costs associated with physical training. Combined with AI generated simulations, this advancement is crucial for the development of more adaptable and efficient robots capable of performing complex tasks in dynamic environments. Current research, for instance, shows emergent capabilities in a humanoid robot that was trained on only 27 hours of simulation data.

• Investment Opportunity: Development of simulation software and platforms for robotics education and training.

• Challenge: Solving Reasoning and Physics simulations

Revolutionizing Design and Engineering

• By simulating physical laws and interactions, these technologies enable engineers and designers to test and iterate product designs in a virtual space. This not only streamlines the development process but also allows for rapid prototyping and innovation

• Investment Opportunity: Tools and platforms that offer AI-powered design and engineering simulations.

• Challenge: Solving Reasoning and Physics simulations

Personalization of Content

• AI video generation can take personalization to the next level by creating content that is uniquely tailored to individual preferences and behaviors. This could redefine engagement strategies across digital platforms, making content more relevant for users.

• Investment Opportunity: AI solutions that specialize in content personalization algorithms and user engagement analytics.

The Future of Gaming, and Education

• DeepMind's Genie represents a breakthrough in creating interactive, AI-generated environments from static images or sketches. This technology could revolutionize the gaming industry by allowing for the creation of personalized, dynamic game worlds. Beyond gaming, such interactive environments could have applications in education, training, and virtual tourism, offering immersive experiences that are tailored to the user's interests and preferences.

• Investment Opportunity: Platforms that leverage AI to create interactive and personalized learning environments and game-based learning experiences.

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.

Subscribe now

Among the areas where we can apply AI, healthcare is the one people generally get the most excited about. For good reason, considering the obvious benefits we can reap. Everybody is on board with curing diseases. Most people have a very personal relationship with health, whether that’s because of an individual health crisis or one experienced by a family member or close friend. Naturally, the AI community is very eager to do its part in minimizing suffering and helping people live happy and healthy lives.

Furthermore, health costs are immense. Healthcare expenditure in the United States is rising by over $3 trillion annually. It's a staggering figure, reflecting a system struggling with inefficiencies and errors. This is not just a financial issue; it's a human one, affecting real lives daily. Still, many cases of chronic diseases get diagnosed too late and managed inappropriately. Many errors are being made. The situation is similar in other developed countries.

People in less developed countries still suffer from diseases we already learned to handle decades ago. Despite our best efforts, healthcare remains a multi-layered tragedy

Where does AI fit in? Thanks to new technological developments, here's now an opportunity to improve the accuracy and speed of disease diagnosis, especially in chronic conditions where early detection is critical. Through Natural Language Processing, we can make sense of vast amounts of medical data more efficiently, aiding in faster and more informed decision-making. Computer Vision's role in analyzing medical images can support doctors in identifying issues that might be missed by the human eye. Additionally, techniques like semi-and unsupervised learning have the potential to unearth new insights from medical data, enhancing our understanding of various health conditions. These AI technologies are not just theoretical possibilities; they are tools beginning to make tangible improvements in healthcare, although their full potential is yet to be realized.

Over the past few years, I have worked with many people trying  to bring healthcare to the next level by implementing AI. It is easy to get lost when thinking about this topic, so I wrote this piece as a primer for those looking for an overview of the AI healthcare landscape. 

The Big Data Breakthrough

There have been attempts at revolutionizing medicine through AI since the 1970s. However, these early systems' lack of advanced machine learning made them difficult to maintain, and they were never fully integrated into clinical practice. Since then, the landscape has dramatically changed. Not only have there been significant algorithmic advancements, such as the development of Convolutional Neural Networks (CNNs) and the transformer architecture, but we also now have access to greater computational power and an abundance of data.

A notable development in recent years has been the emergence of large clinical datasets that have been de-identified, addressing privacy concerns while providing a rich resource for medical research and AI development (e.g., MIMIC, Truven Market Scan, All of Us). These datasets, encompassing a wide range of health-related information, are proving invaluable in training AI systems more effectively and comprehensively.

The kinds of data available now are diverse and extensive. We are seeing a surge in the collection of unstructured data, such as clinical notes, lab tests, imaging, and vital signs. This data holds a wealth of information that AI can leverage for deeper insights. Furthermore, genomics and proteomics advancements provide detailed biological data that can revolutionize personalized medicine. Additionally, an often-overlooked source of health data comes from social media, where people share information about their physical activities, diets, and mental health. These varied data sources, from clinical records to social media posts, offer a comprehensive view of health and lifestyle, presenting new opportunities for AI applications in healthcare.

What  the future could look like

For many applications in healthcare, the relevant data is not labeled, which makes unsupervised learning an essential approach here. Moreover, we are often interested in causal relationships when it comes to diagnosis and treatment. That means, we need  more than naive supervised learning and correlation. Through more sophisticated techniques, like Deep Reinforcement Learning, we can identify causal relationships. I will now walk you through a few examples to give an overview of how AI can and partly already is being deployed in healthcare.

A New Era of Precision and Efficiency

First, we can improve how healthcare providers operate, increasing quality and access to care. Healthcare providers often face challenging working conditions, including limited resources and the need to make critical decisions. AI can significantly alleviate this pressure by supporting medical staff

One of the key areas where AI can assist is in the analysis and interpretation of patient information from electronic health records (EHRs). By extracting and processing data, AI systems can optimize medical workflows and help operators attend to patients more effectively.

AI can also be helpful  in resource allocation and triage in ERs and clinics. By evaluating the severity and urgency of patients' conditions, AI can aid in prioritizing treatment, ensuring that those who need immediate attention receive it promptly. This not only optimizes the use of limited resources but also helps in managing patient flow more efficiently.

Another significant benefit of AI is its potential to reduce medical errors. Through continuous monitoring and analysis of patient data, AI systems can identify potential adverse events or complications early on. This early detection is vital in preventing escalation and ensuring timely intervention.

Integrating AI with a database of clinical guidelines and best practices can streamline the treatment process. Once a diagnosis is made, the system can immediately provide healthcare professionals with recommended treatment protocols. This speeds up the response time and minimizes the chances of errors, ensuring that care is consistent with the latest medical standards and research.

Fully realizing the benefits of AI in healthcare might require new organizational models in ERs and clinics. While increasingly common, the focus on digitized data-driven workflows is not yet a standard practice across the healthcare industry. 

There are already companies out there that are pioneering a new approach to a data-driven clinic. For example, our Merantix venture Ovom transforms fertility medicine by setting up clinics based on data-driven workflows, allowing hyper-personalized care.

Examples: Ovom: Personalized and data driven fertility medicine. Memora Health: Automating care workflows.

The New Diagnostics in Healthcare's Digital Age

The advent of AI in medical diagnostics has opened new avenues for early and accurate disease detection. AI, particularly computer vision, is increasingly being utilized to analyze medical images such as mammograms, X-rays, and other scans, enhancing the accuracy and efficiency of diagnoses.

For instance, companies like Vara, another Merantix venture, are leveraging computer vision technology to detect breast cancer. By analyzing mammogram images, their models can help doctors identify potential signs of cancer. This accelerates the diagnostic process and increases the chances of detecting cancer at an early, more treatable stage

.Beyond image data, AI can also be applied in other areas of diagnosis. A prime example is wearable technology, like smartwatches, which can measure an individual's electrocardiogram (EKG) data. Traditionally, obtaining an EKG required a visit to the doctor. However, integrating sensors and AI algorithms in wearable devices makes it possible to continuously monitor heart rhythms in real time. We could detect abnormalities such as arrhythmias by analyzing the time-series data of a user's heart rate. 

Examples: Vara: Breast cancer screening with AI. Freenome: Early cancer detection based on blood tests. InVision Medical Technology: Heart disease detection with AI. 

Transforming Chronic Disease Care from Cancer to Diabetes

We have become much better at treating infectious diseases in the past few decades. However, chronic conditions like cancer, heart disease, diabetes, and dementia are still significant problems. Unlike acute illnesses, many chronic diseases exhibit varying trajectories, making treatment and management a nuanced and individualized process.

Take cancer, for instance. Its progression is rarely linear and can vary significantly from patient to patient. For certain types of cancer, there are multiple treatment options available, and choosing the most effective one can be akin to educated guesswork. The response to treatment varies widely among individuals and is influenced by many factors, including genetics, lifestyle, and the specific characteristics of the cancer.

In the future, ML models could be crucial in personalizing cancer treatment. We could analyze patient data - encompassing genetic information, treatment history, and tumor characteristics - to predict how different treatments might affect the disease's trajectory. This isn't just a theoretical possibility; efforts are underway to gather the necessary data. For example, the Multiple Myeloma Research Foundation (MMRF) is compiling detailed records on thousands of patients with Multiple Myeloma, tracking their symptoms, treatment responses, and biological data. Such comprehensive datasets are invaluable for training models to predict disease progression and treatment responses more accurately.

Other conditions like kidney disease follow a more predictable, albeit gradual, decline in organ function. It's crucial to determine a patient's current state and how quickly the condition is progressing for these diseases. However, accurately gauging this progression is challenging. AI can assist in this regard by analyzing trends in lab results, symptoms, and other health indicators to provide a clearer picture of the disease's trajectory.

Beyond diagnosis and treatment, AI also has the potential to revolutionize disease management, especially for conditions like Type-1 diabetes. Managing this disease requires constant vigilance over blood sugar levels, which can be stressful and complex. Machine learning can be employed to predict future insulin requirements more accurately. For instance, consider an AI algorithm tailored to a patient's unique metabolic data. This algorithm could estimate the impact of different foods on insulin levels. If a patient takes a photo of their meal, the algorithm could predict how that specific meal would affect their blood sugar, thereby aiding in more precise insulin dosing.

In summary, AI has the potential to significantly enhance our understanding and management of chronic diseases. By providing more personalized and accurate predictions and recommendations, AI can support clinicians in offering better care and empower patients to manage their health more effectively. As we continue to advance in this field, we can transform the landscape of chronic disease treatment and management, making it more proactive, personalized, and patient-centric.

Examples: Instalab: At home blood testing for biomarkers to optimize log-term health.

Shaping a Future of Personalized, Efficient Care

There are many more use cases for AI in healthcare, such as drug discovery or uncovering new disease subtypes through clustering. Integrating AI in healthcare and leveraging vast data sets holds immense potential. However, it is essential to note that this potential hinges on developing robust, fair, and accountable AI systems, particularly in areas where unsupervised learning plays a crucial role

At its core, the future of healthcare with AI is about prevention, personalization, and optimization. AI promises a healthcare system that not only anticipates and prevents diseases but also tailors treatments to individual patient needs while enhancing the efficiency of healthcare delivery. As we embrace this data-driven approach, we are poised to revolutionize healthcare, making it more effective, equitable, and accessible for all.

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.

Subscribe now

Manufacturing, at its core, is about turning "stuff" into "better stuff." It's a part of our economy that's hard to miss, underpinning most other sectors and keeping the wheels of commerce steadily turning. Even the slightest shift in manufacturing has significant ripple effects on the global economy.

Imagine manufacturing as a giant puzzle, where each piece represents a different process – from sourcing raw materials, running production lines, to managing logistics and distribution. Each piece is crucial and getting them all to fit perfectly is no small feat.

In this context, the role of AI is akin to a master puzzle solver. It's not about reinventing the puzzle or introducing flashy, brand-new pieces. Instead, AI's magic lies in making each existing piece smarter, more efficient, and more interconnected. It's about ensuring that the supply chain is more responsive, the production lines more intelligent, and the distribution networks more agile.

AI in manufacturing is like a series of small yet powerful interventions that collectively transform the entire landscape, from enhancing precision in production lines, predicting maintenance needs before they become problems, or optimizing supply chains in real time. Each of these advancements might seem incremental in isolation, but together they redefine what's possible.

“As AI continues to make its mark, we're witnessing a quiet yet profound revolution, not by turning the big wheel, but by transforming all the many different cogs until the whole system is unrecognizably better.” Silviu Homoceanu told me. He, and Max Fischer together founded Deltia, a venture that is using Computer Vision and Edge AI to optimize assembly line processes. They both helped me deepen my understanding of manufacturing and how AI can make a difference.

Supply chain management

Let's talk about supply chain management, a crucial piece in the manufacturing puzzle. Picture this: you've got a product you want to bring to life. The first step is figuring out how to gather all the necessary components – the "stuff" to build your product. This is where the supply chain comes into play, and it's a bit like walking a tightrope in the business world.

If your supply chain is weighed down with unnecessary costs, when you scale up your business, those costs scale up as well, and that's not good for anyone. Stability is another key factor. If your supply chain wobbles and you have to slow down or, worse, halt production, you're looking at some serious losses. That's a scenario you want to avoid at all costs.

So how do we keep things lean and stable? Globalization has been the name of the game for supply chains in recent decades. We're talking about sourcing and procuring materials on a worldwide scale. To keep a handle on all this, we use Enterprise Resource Planning (ERP) tools. These tools help manage everything from orders to supplier relationships, ensuring we're as efficient as possible.

Optimization is our constant goal. We aim for Just-In-Time production – it's all about having just enough inventory to meet demand without overdoing it. And when it comes to predicting that demand, accuracy is king. A small increase in forecast accuracy can mean saving a fortune.

Now let's bring AI into the mix. Given that supply chain management revolves around forecasting and optimization, it's a perfect match. By integrating AI with our ERP systems, we unlock deeper insights. We can achieve a more nuanced understanding of supplier performance, risk assessment, and compliance. This isn't just about data crunching; it's about making smarter decisions that reduce costs and bolster the stability of our supply chains.

Through  machine learning, we can sharpen our demand forecasting. We can dissect and understand supplier performance and production schedules in ways we never could before. It's like giving our supply chain management a super-powered telescope to see further and more clearly into our operations.

In essence, supply chain management in manufacturing isn't just about moving things from point A to point B. It's a complex balancing act, requiring precision, foresight, and a fair bit of agility. 

Manufacturing Operations

Once our supply chain and inventory are set up, the real action begins in the world of manufacturing operations. It's the phase where we turn our plans and schedules into tangible products

Our first task is to meticulously plan and manage our production capacity. This means striking a balance between the available resources - manpower, machinery, and materials - and the ever-fluctuating market demands. We're constantly fine-tuning this balance to ensure efficiency. The key here is optimization: the better we optimize our production workflows, the more cost-effectively we can produce.

If possible,  we use automated machinery and robotics for consistent and efficient production. Alongside automation, we emphasize standardized processes, especially in mass production. These processes are crafted with multiple goals in mind: safeguarding our workers, minimizing waste, and achieving cost and time efficiency. 

Quality control is a critical checkpoint in our operations. Traditionally, this has been a domain led by human oversight. However, there's a growing space for innovation and improvement. For instance, companies like Deltia are transforming assembly workflows through advanced camera systems. These systems collect vision data, offering insights that help refine processes, enhance worker safety, and reduce costs. Even minor improvements in the assembly line can translate to significant benefits for the entire organization.

There are many more use cases of AI within manufacturing operations.. A big pain point, for example, is the maintenance of equipment. Since every failure of equipment results in downtime for the manufacturing process, predictive maintenance is extremely important to ideally fix problems before they even arise. 

Logistics and Distribution

After the final product is assembled in our manufacturing setup, we enter the critical phase of logistics and distribution. This process is vital in ensuring that products reach the end customer efficiently and reliably. 

To ensure efficient route planning, minimizing delays and costs, we utilize logistics management systems. Manufacturers often have established networks of distribution centers and retail outlets. These networks are strategically located to ensure wide-reaching and timely distribution. The primary focus here is on reducing delivery times and costs, which is crucial in maintaining customer satisfaction and reducing overheads.

One of the most important parameters to optimize in this context, similar as in supply chain management, is warehousing. Here, Warehouse Management Systems (WMS) are used to track inventory levels, orders, and deliveries. This software is essential for maintaining visibility and control over stock.The processes in the warehouse, including storage, picking, and packing, are predominantly standardized. This ensures consistency and efficiency in handling goods. Operations are mostly manual or semi-automated, with the use of forklifts, conveyors, and manual labor for moving items around. Space optimization is achieved through static shelving and rack systems, and manual planning is used for organizing warehouse layouts.

In the logistics and distribution sector, AI is about enhancing efficiency and accuracy. AI-driven systems are capable of optimizing delivery routes in real-time, taking into account factors like traffic and weather conditions. This not only speeds up delivery times but also improves overall logistical efficiency.

AI-enriched Warehouse Management Systems (WMS) elevate operations to a new level. One  could forecast inventory needs, automate replenishment orders, and integrate seamlessly with IoT devices for real-time tracking. Moreover, the deployment of Automated Guided Vehicles (AGVs) and robotics for tasks such as material handling and packing significantly reduces manual labor while boosting efficiency. AI also plays a crucial role in dynamic space optimization, enabling warehouses to adapt their layouts in response to fluctuating inventory and order profiles. This results in more efficient use of space and faster retrieval times, transforming the traditional warehousing model into a more responsive and agile operation.

Summary

Based on my conversation with industry experts, such as Silviu and Max, I believe that there is great potential for AI in manufacturing. Here is a summary of three major use cases from the categories I mentioned above::

Warehousing
Access inventory is very costly for manufacturers who face potentially long lead times. Many processes here are still manual.
AI solution: Improved forecasting enables lower inventory. Many currently manual tasks can be automated. Computer Vision can supercharge warehouse workers, logging, counting and monitoring inventory.
Examples: Powerhouse AI: Provides automated inventory services. (Backed by Entrepreneur First and Y Combinator).   Deskera: Warehousing company, using AI to reduce inventory. Series A $180M (Backed by Naver and Susquehanna)

Maintenance
Processes can fall apart due to faults anywhere in the process, causing huge costs. Process remains reactive instead of proactive.
AI solution: Predictive Maintenance enables early interventions and a proactive approach. Avoidance of failures reduces repair costs and prevents unplanned downtime
Examples: Tractian: Cloud based predictive maintenance company. Series B $73.3M (Backed by Y Combinator and General Catalyst).   Sparkcognition: Startup tackling Predictive Maintenance, among other issues. Series D $286.6M (Backed by TEMASEK)

Assembly
Assembly processes are still surprisingly manual. Machines are unable to changing situations.
AI solution: AI can supercharge factory workers by finding snags in production. Measures such as Edge Computing and anonymisation can address trust issues among workers towards AI.
Examples: Deltia: Using Computer Vision and AI analytics to provide insights into manual assembly processes. Pre-Seed EUR 1.65 M (Backed by Merantix).   Bright Machines: integrating software and ML-based intelligent capabilities to achieve assembly automation. Series B $311M (Backed by Eclipse and LUX)

Startups in Manufacturing

As you can see, there are ample opportunities to innovate in manufacturing. Startups play a special role in the transformation of this industry. Often entrepreneurs experience significant pushback if they try to innovate in a conservative industry like manufacturing. According to Silviu, gaining acceptance among established players is possible if you have a deep understanding of the industry and empathy for the people who work there.

Despite manufacturing being a traditional industry, most people within the sector welcome AI innovations. The growing popularity and awareness of AI technologies has made the integration of AI more acceptable in the manufacturing space. However, there is some fear among workers, who may worry that AI technologies could replace them or invade their privacy. To address these concerns, companies must be transparent about their intentions and communicate that AI will be a copilot or assistant for workers rather than a replacement. In order to increase transparency and alleviate fears, one approach that Deltia is taking is to provide workers access to the system.

There is ample room for more entrepreneurs to enter the AI-manufacturing scene, as manufacturing significantly contributes to the global GDP. SMEs, especially, could benefit from increased support to democratize access to technology. By helping SMEs to become more efficient, entrepreneurs can contribute to making products more affordable. As the foundations of AI evolve, more innovations can be expected in the manufacturing space. However, entrepreneurs should demonstrate benefits for SMEs from day one without expecting them to have highly skilled personnel in data analytics or AI.

For AI entrepreneurs to succeed in the manufacturing industry, they must be customer-centric and adapt their offerings to the sector's realities. In addition, they need to understand that SMEs may not have the resources or expertise to handle complex data sets or advanced IT infrastructure. By bridging the gap between the two worlds of AI and traditional manufacturing, entrepreneurs can unlock significant value and drive innovation in this crucial industry.

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.

Subscribe now

Often when we speak about applying AI to specific industries it’s about automating workflows and using our current understanding to make things more efficient. This is different in the mental health space. AI here has the potential to bring our understanding of the human mind and the science of mental health to a completely new level. 

Mental health diagnosis, so far, relies solemnly on self reports and occasional meetings with clinicians. Besides drug therapy we treat patients mostly with talk therapy employing a vast array of different approaches from hypnosis to Cognitive Behavioral Therapy. However, we have little evidence on what actually works. If a patient's condition gets better we rarely know whether this was due to the therapeutic approach or because of the therapist's personality or because they changed their diet or got better sleep. 

Although we know about various factors correlated with mental health we do not know much about the causalities and individual differences. Which behavioral pattern contributes to an individual's mental health? And how much so? What kind of therapy with what kind of therapist works for whom? Which commonly used practices might harm certain individuals? All of these questions require answers if we want to really understand mental health. But so far we can merely guess at the answers.

With the advent of artificial intelligence, however, we might have found the tool that is for the science of mental health what telescope and calculus were for physics. With new mental health technology we will be able to collect and integrate data from previously untapped sources like behavioral data from our digital devices or sleep and general health data, collected by wearables.

These AI tools hold the promise of decoding the intricate neural networks of the mind, offering a more profound understanding of mental processes and health. This has the potential to be not only a revolution for mental health patients but also a scientific revolution introducing AI as the new paradigm for mental health research.

"Mental health has the potential for a revolution, but its realization hinges on our immediate action." says Amruth Ravindranath, Neuroscientist and founder of Addera Health. Amruth helped me to develop a better understanding of how AI can revolutionize Mental Health and how our action plan looks like. I am sharing these insights as well as what the next step looks like.

Step 0: Tackling the mental health crisis

Before we can realize this revolution however, we must address the immediate problems our mental health care system has right now. By building technology to make mental health care better, we establish the infrastructure and collect the data necessary to advance our understanding. 

Here are the main problems in mental health care right now: 

• Access to care: Access remains a substantial barrier in mental health care. Geographic isolation, financial constraints, and social prejudices often impede individuals from seeking help. Even in prosperous nations, a shocking number of people can’t  access the care they desperately need, leading to untreated conditions and escalating crises. Since we heavily rely on 1-1 counseling, we can’t  easily scale the care we provide to people. Additionally we provide almost no care to people with mental health issues below the clinical threshold. 

• Quality of Care: Once access is secured, the disparity in care quality becomes evident. Despite the presence of highly trained professionals in the field, a uniform standard of care is elusive. Practices vary widely and the lack of personalized treatment plans means that individuals often receive generic care, which may not adequately address their specific needs.

Mental health tech as a proposed solution to these problems is not new. It has been around for quite a while without contributing much. However, times are different now. Multiple trends in technology as well as in society are coming together creating unprecedented momentum:

• (Generative) AI: Current workflows in the mental health space are mostly language based. There are many other data points, like general health data, that do not get factored in the decision making so far. So, we’ve got vast amounts of unstructured data, mostly in the form of language. Situations like these are ripe for AI . The boom we’ve experienced with this technology over the last months creates a unique opportunity for mental health tech.

• Awareness: The recognition that we are in a mental health crisis has been rising for years, but the COVID 19 pandemic really highlighted the deep problems we have. While primarily a health crisis, the pandemic inadvertently spotlighted another pervasive issue – the state of global mental health. The challenges it unveiled, though not new, are clearer than ever before. Although stigma is still an issue in many social groups, the new awareness opens up new discussions and new opportunities to solve the underlying problems.

• Adjacent tech trends: We’ve seen a massive trend toward data driven and personalized medicine. It’s  more and more common to track health data via wearables. Over the next few years we will see more opportunities to leverage this data for various health applications. This presents mental health tech with the unique opportunity to incorporate new data sources 

As our technological capabilities rise, our moral responsibility towards those in need rises as well. 

In our pursuit to address the mental health crisis, integrating technology like AI can be invaluable. It can help bridge gaps, refine practices, and ensure that as our understanding of mental health evolves, our approach to addressing it evolves too.

Step 1: What needs to be done now

Let us now get more concrete and look at the specific applications that can be built. It's useful here to differentiate between two primary user groups: individuals with clinically diagnosed mental health conditions and those experiencing subclinical mental health issues. 

Despite the overlap in their needs, each group necessitates distinct features from the mental health tools we develop. Importantly, these tools are subject to different regulatory and safety standards. AI applications designed for clinically diagnosed patients must adhere to stringent clinical validation and regulatory compliance, given that they are part of formal treatment plans. On the other hand, AI tools for subclinical support are generally subject to less rigorous oversight, but still require attention to privacy, data security, and ethical considerations to ensure they provide safe and effective guidance.

• For clinically diagnosed patients, AI can enable more precise predictive analyses and personalized treatment plans. For instance, by examining patterns in large datasets, AI can identify individuals at higher risk of adverse outcomes, allowing for earlier and more targeted interventions. It also supports clinicians in tailoring treatment plans to individual needs by analyzing data from similar cases, which can lead to more effective management of conditions. Moreover, AI-enhanced monitoring tools can provide clinicians with real-time feedback on patient progress, ensuring that treatment remains dynamic and responsive to the patient’s evolving condition.

• For those grappling with subclinical mental health concerns, AI offers a less intimidating and more accessible form of support. Mental health apps infused with AI provide a suite of tools for self-management, from mood tracking to therapeutic exercises, making mental health care proactive and personalized. AI chatbots act as a first line of emotional support, offering conversational assistance that can guide users through difficult moments or towards professional help if necessary. This level of immediate and personalized support can be critical in preventing subclinical issues from escalating into more severe conditions, thereby enhancing overall mental wellness.

Amruth and his company, Addera, provide a strong example of how to collect new kinds of data and integrate it into a product to provide better access and quality of care: "We are building Addy - an AI coach for mental health, starting with coaching for adult ADHD. Addy uses AI models to construct, test and fine-tune personalized protocols that can take every person through a unique journey towards better mental health in a way that's aligned with their individual strengths and difficulties."

To further split up the field we can functionally divide applications according to the different problems of mental health tech, they are solving:

Gathering of Data

Traditionally assessment of conditions is often carried out through subjective means such as surveys. The assessment is regularly self applied and left up to the patient. Gathering of data is usually infrequent and limited to a clinical setting.

AI solution: AI enhances assessment accuracy through continuous, objective data collection integrated into daily routines, supplemented by insights from wearables and other tech sources.

Examples: Sonde: Vocal biomarker technology that can tell you are at risk for conditions of mental health. Series B $41.3M (Backed by Partners and M Ventures)Kintsugi: Detects signs of clinical depression and anxiety using  machine learning and voice biomarkers. Series A $30.9M (Backed by techstars_ and Insight Partners)

Utilization of Data

Due to subjectivity of measurements it can be difficult to quantify progress. Moreover it is difficult to attribute progress to any particular reason.

AI solution: We can gather much larger data sets. This requires us to use AI to provide insights. We can get insights that were previously not possible and personalize solutions for every user.

Examples: limbic: AI-based tool that supports assessments and referrals in psychological therapies. Series A (Backed by 7 Percent Ventures).   Unmind: B2B mental health platform providing clinically-backed tools and training. Series B $63.2M (Backed by EQT and Project A).

Communication of Insights

Interactions between patients and clinicians are infrequent and delivered only in scheduled appointments. Recommendations are often difficult to implement in daily life.

AI solution: LLMs can make communication with users easier. Solutions can be implemented into other aspects of people’s lives, bringing treatment to them directly.

Examples: Woebot Health: Automated conversation agent using techniques from CBT to help monitor mood and enables users to learn about themselves. Series ? $123.3M (Backed by TEMASEK and NEA).Wysa: Mental health wellness platform that helps individuals manage their mental and emotional stress. Series B $29.5M (Backed by Google Assistant and Kae Capital).

AI's role in mental health care is evolving from a supportive tool to a potential catalyst for a comprehensive understanding of the mind and its maladies. The collective efforts to implement AI responsibly and creatively can revolutionize our approach to mental health care, making it more accessible, personalized, and effective. This revolution is not just within our reach—it is unfolding before us, and it is imperative that we nurture its growth with every resource at our disposal. 

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.

Subscribe now

Here's what we know: AI has tremendous potential to re-engineer fundamental banking aspects, from risk assessment to fraud detection to credit evaluation. AI can make banking fairer, faster, and more trustworthy — all values that should underpin the global financial system. So where is the catch? Somehow the financial industry is not adopting AI at the necessary speed. First, it is important to get a better picture of how FinTech and the traditional financial sector work.

How Fintech embraced digitalisation

To better understand the FinTech sector, it is helpful to model the industry along the lines of B2B and B2C. In the B2C domain, customers have come to anticipate and demand a seamless digital experience, expecting intuitive interfaces, rapid response times, and secure yet accessible transactional processes. 

The proliferation of tech penetration in the B2B segment, on the other hand, serves as the infrastructural backbone supporting this B2C digital prowess. B2B FinTech solutions streamline operational workflows, enhance data analytics capabilities, and fortify security measures, thereby indirectly elevating the end-user experience in the B2C space. Consequently, the synergy between B2B and B2C aspects is instrumental in shaping a holistic, efficient, and customer-centric FinTech ecosystem.

Although the financial industry is usually known for conservatism, it has managed to adopt new technologies over the last 20 years and is today one of the most digitized industries. However, despite the burgeoning possibilities sprouting from the intersection of AI and FinTech, we can see a measured, perhaps cautious integration, especially within the core banking infrastructures. AI's prowess is evident in customer service, where chatbots deliver 24/7 real-time support. However, its deeper integration into the very nucleus of banking operations remains notably measured.

What is keeping the industry from unleashing AI

AI adoption is stuck because of the structural setup of established players. The prevailing architectural structures within traditional banking institutions present a significant operational challenge. Often, data within banks are segregated into silos, a practice adopted to enhance security by restricting access to specific data sets and preventing any widespread impact in case of a security breach. 

Consequently, accessing these disparate silos, each intertwined with its legacy infrastructure poses a colossal hurdle on a broader scale. This segregated data architecture inhibits the seamless integration and flow of information across various departments and functions, limiting the ability to rapidly innovate and implement new solutions that respond to evolving market demands and customer expectations. For these reasons, we cannot expect the AI transformation to come from within the industry but instead must focus on startups that are not locked into these old ways. 

"Tech penetration in the financial industry has peaked, ideally positioning it for robust AI development. However, legacy infrastructures hinder the journey toward an AI-empowered future." Anouk Moll, a seasoned FinTech entrepreneur and current entrepreneur-in-residence at Merantix, told me. Anouk's expertise helped me deepen my understanding of the FinTech landscape related to AI, which I'll lay out here.

Subscribe now

The potentials that Generative AI bring

It's easy to see the potential of Generative AI in finance, a robust data-focused industry. FinTech has had AI applications, typically oriented towards prediction and classification, often predominantly leveraging time-series data. However, with the advent of generative AI, an exhilarating possibility unfolds to harness even the vast swaths of unstructured data, blending it seamlessly into decision-making and predictive modeling processes. 

Indeed, given the advanced digitization and the resultant availability of extensive historical data, Foundation Models can be fine-tuned to decipher, understand, and generate insights from data that was previously deemed challenging to integrate into analytical models. Generative AI doesn't merely analyze; it creates, imagines, and proposes, enriching the analytical prowess of traditional models with an additional layer of innovative spontaneity and capability.

By employing generative AI, financial institutions can delve deeper into the intricate web of diverse data, unraveling insights and strategies that are both data-driven and innovative, paving the way for a future where finance is not only intelligent but also intuitively adaptive and forward-thinking.

The signs for a change

Despite this technological boom, AI, as it stands today, has only scratched the surface of its potential utility within the banking sphere, showing how deep the structural issues lie. But the tides are about to change. Recently, Open Banking has transformed the financial ecosystem in Europe by enhancing transparency, customer-centricity, and data value through third-party developer access. Stringent yet transformative regulatory frameworks such as Know Your Customer (KYC) and Anti Money Laundering (AML) have propelled organizations towards a tech-forward approach, making AI a critical element in ensuring compliance and maximizing the utility of existing technological infrastructures and rich data.

With its transformative capabilities, generative AI autonomously, for instance, creates formulas and queries in financial forecasting or Business Intelligence to streamline analyses and identify patterns, reducing the need for manual inputs and facilitating a more systematized approach towards understanding and leveraging financial data.

The new boom we experience with this technology is paving the way towards a future where predictive and intuitive decision-making coalesce to redefine operational norms within FinTech. Yet, its fullest realization may very well hinge on the entities that drive its implementation and evolution — a role where startups are surprisingly pivotal.

For Anouk, that problem is front of mind as she looks into different ideas: "One thing is the post-trading space, which is known for the high amount of institutional players due to high regulatory requirements. It is a fascinating space that has seen some innovation in the US with companies like Alpaca, but Europe has seen minimal movement within this space. The other area I'm looking into is FX wallets for high inflation markets and using gen AI to empower people to safeguard their savings through accessing stable currencies."

There are many more challenges in FinTech that need to be taken on, and an increasing number of startups are working on it. The interesting situation here is that AI is often the solution to AI applications' challenges. This recursive feature of AI makes it such a promising technology for FinTech. So let’s have a look at whats already been done:

Subscribe now

Challenges and Solutions in AI and Fintech

In case you wonder what some main challenges are and who is already working on them. Here is a brief and non-exhaustive overview:

Challenge: Ensuring Robust Data Security

Financial institutions are gatekeepers of a plethora of sensitive data, and especially in regions like Europe, stringent regulations are in place regarding the sharing and utilization of customer data, acting as a substantial barrier to deploying AI solutions seamlessly.

Solution:AI comes full circle as a solution by being applied extensively in cybersecurity to mitigate these concerns. The need to develop in-house AI solutions for financial institutions creates a fertile market for developer tools, ensuring data is protected while being utilized optimally.

Examples:RESISTANT.AI: Software to protect AI systems from advanced fraud (backed by GV and Index Ventures)Comply Advantage: ComplyAdvantage provides AI-driven financial crime risk data and detection technology. (backed by Balderton.)

Challenge: Navigating Complex Regulation and Compliance

The European Union regulations surrounding KYC and KYB add complexity to customer onboarding. The recent AML measures in the EU necessitate meticulous scrutiny over capital movements, while fragmented markets introduce variability in compliance requirements across jurisdictions.

Solution: AI's capacity to synthesize vast data and information becomes pivotal in ensuring compliance with varied and stringent regulations. Its ability to implement and enforce AML guidelines while ensuring user-side compliance becomes instrumental. Moreover, AI enables financial bodies to adapt to different market regulations and requirements without a consequential resource or time allocation increase.

Examples:onfido: AI to acquire new customers and reduce costs while meeting global KYC and AML compliance (backed by Tech Nation and TPG)Signzy: AI-powered FinTech company offering a digital onboarding solution for banks, NBFCs, and more (backed by Gaja Capital and Tenity)

Challenge: Enhancing User Accessibility and Experience

Making financial services intuitive and secure, especially while integrating into other services without compromising user experience (UX) or security, presents a significant hurdle.

Solution:AI can enhance UX by enabling content personalization and adapting interfaces to match user knowledge and comfort levels. Moreover, its capability to understand user behavior and adjust product offerings accordingly allows for a more tailored and intuitive user experience. Furthermore, predictive risk and fraud models can be developed through AI, facilitating proactive intervention and resulting in a secure user experience, yet not intrusively so.

Examples:ExpressSteuer: AI-powered software automating the process of filing a German tax return (backed by Insight Partners and Project A)Cleo: Personalized financial advisor (backed by EQT and SOFINA)

Looking towards the future

The very nature of AI, with its data synthesis, predictive modeling, and adaptive learning, positions it as a formidable tool to address the complexities and challenges that pervade the financial sector, paving the way for a future where Fintech is not only more secure and compliant but also user-centric and accessible.
The symbiosis between financial technology and Artificial Intelligence presents the promise of an era of unparalleled possibilities and innovations.

From nuanced, personalized customer experiences to robust, unassailable cybersecurity, AI is not merely an adjunct but a powerhouse propelling FinTech into new dimensions of efficacy and customer-centricity. Navigating through the intricate labyrinth of regulatory frameworks, data security, and global financial ecosystems, AI emerges not as a mere technological tool but as a compass guiding FinTech towards more equitable, accessible, and transparent financial futures. 

Subscribe now

Thanks for reading, and please let me know what you think, what further opportunities in the market you see or what technological trends on the horizon you believe will make an impact. A special thanks goes out to Eduard Hübner, who was my great co-author for this piece.

- Rasmus

Connect with me on Linkedin or Twitter.