The continent lost the sky

Hadrien arrived in Ukraine in the first week of March 2022, a week after the full-scale invasion. What he saw was a kind of war NATO had forgotten how to fight: “a high and long intensity warfare,” not the small expeditionary operations in Africa and the Middle East where the West always owned the air. For the first time in modern European history, he argues, “Ukraine, but also NATO, doesn’t have the upper hand on the air superiority.” The weapon that took it away is almost insultingly cheap. As Hadrien traces its lineage, the Shahed-136 is a slow, low-flying design first drawn up by German engineers, then copied in turn by the Israelis, the Iranians, the Russians, and now the Americans. “Nothing changes through the years,” he said. The economics did.

Most of the time, it misses

I asked him to walk through the data loop, the thing every speaker that day had described from behind a laptop: get representative data, run the model, grade the output, go again. Alta Ares runs it in a field at night. “Most of the time it doesn’t hit,” he said, and that is the whole problem. There is rarely internet at the edge and never a data center, so everything runs on “super compressed” models on small GPUs inside the interceptor. A mission has three phases. Radar detection sorts friendly from hostile. Then an interceptor flies into a kill zone roughly two kilometers wide, where an RGB and an infrared camera hunt the sky for a single pixel that is genuinely hard to find, because the feed is low quality and the airspace is being jammed. Then a human, still in the loop, commits to the intercept. One target is tractable. A swarm is not.

The loop only tightens if you close it fast. Alta Ares keeps around twenty engineers in Ukraine, as close to the front as Hadrien can put them, and combat-tests “every week or every two weeks.” The hardware turns over constantly; the vision models update remotely as more data comes back. They do not call the things drones. “We call them munitions, because it doesn’t come back,” he said. There is an 800-gram warhead in the nose. “I don’t recommend them to come back.”

What no simulation can reproduce

Simulation, he allowed, is good for a proof of concept and little more, and the reason is human rather than technical. The engineer who runs a test slept well, had breakfast, had a good coffee, “almost as good as the tea in England.” The operator who fires the system in Ukraine is on their fifth coffee “that doesn’t wake you up anymore,” at 3 a.m., freezing, running on adrenaline, and far less trained. No amount of lab time closes that gap, and no simulator renders the adversary. The Shaheds his team intercepts have started carrying rear-facing sensors that trigger evasive maneuvers when an interceptor closes in. Others fly in a mesh, so when one is shot down another inherits the firing solution. Much of the underlying technology, he said, is Chinese, routed to Russia and Iran. The lesson Hadrien draws is that classical computer vision is not enough against an opponent that keeps learning, and the only thing that trains a system to beat it is real data from real intercepts, which no simulation can manufacture.

Quantity is the quality

So why isn’t modern air defense everywhere? Because NATO and Europe bought small numbers of exquisite systems on the assumption that they would always own the air, and that assumption is gone. “The new reality is that the quantity is the quality,” he said. A sophisticated interceptor fielded in low numbers gets outgunned. Israel’s Iron Dome works because Israel is small to defend, and even so, by his account, “Israel used 35% of the world’s stock of Patriot missiles in only the first 10 days” of the war. Interception is never a sure thing: “never 100%,” he said, somewhere between 25% and 75% on the best systems. The fix is as industrial as it is technical, and the budgets are still pointed the wrong way. France is spending more, he noted, but just committed 10 billion euros to an aircraft carrier that, he argued, takes a decade to build and that a hundred cheap unmanned boats could sink. He put the politics of it in one line: “during peace time, you have time, but you don’t have money. And during war times, you don’t have time, but you have money.”

The talent turn toward European defense

Talent has become his strange dividend of the moment. “One of our best talent acquisitors, his name is Donald Trump,” he said. When JD Vance told the Munich Security Conference that American and European interests had diverged, well-qualified engineers at US companies in London started leaving to build European defense instead. What he sells them is accountability and proximity: write something in the morning, watch it deployed on the front by night. Alta Ares does no offensive weapons, “so far.” And in his telling, this work is the precondition for everyone else’s. The clinical trials, the healthcare, the AI-for-good that had filled the rest of the day are “possible because we live in a country at peace.” Bomb the data center and none of it ships.

Pressed by an audience member to separate the founder from the person, he couldn’t, and didn’t try. He reached for the old Roman maxim, si vis pacem, para bellum: if you want peace, prepare for war. Europe stopped preparing, and “the use of force is being democratized” faster than its institutions have noticed. You can annex territory now, he said, and no one does anything about it. He was honest about the discomfort of his own business: “I would sleep way better if tomorrow the war in Ukraine stops.” But he doesn’t think it will, and his case for rearming rests on a grim simplicity. The people on the other side keep telling us what they intend, so “we should listen to them, because at least they are being transparent.” Freedom doesn’t come for free.

That was once a hard sell. When Odyssey pitched its seed round, he says, “basically no one understood world models,” and convincing people was “an uphill battle.” The category has since shifted underneath him: the argument flipped around November 2025, money and researchers flooded in, and at NeurIPS 2025 world models were, by his count, the field’s number one or two theme. Hawke came to the problem from self-driving, as a founding engineer at Wayve teaching cars to drive end-to-end - one of the few arenas, he says, where frontier AI has really been tested against reality.

Not every world model is a world model

The term is loose enough to mean almost anything, so Hawke was careful to pin it down. Odyssey uses it in the precise sense inherited from model-based reinforcement learning: a learned transition-dynamics model, the thing David Ha and Jürgen Schmidhuber named in their 2018 “World Models” paper. In plain terms, a model that learns how the world evolves and then samples possible futures one step at a time, conditioned on the actions you feed it. That is distinct from “spatial intelligence,” which models the appearance and structure of a scene (the pitch at World Labs and others); from behavior models, the decision-making brain of a self-driving car or a robot, which Hawke thinks is better named as such; and from the “proxy world models” some researchers argue are already latent inside LLMs. Odyssey’s version is what he calls a general-purpose neural simulator: an interactive stream of pixels that models physics, that you can talk to and that talks back, and that you can reach into and change. “I have many questions about what this means for future products,” he admitted. “I have very few good answers.”

Generality wins

Two principles guide what Odyssey builds, both carried out of self-driving. First, end-to-end learning, a simple model trained purely from data, “almost always wins.” It was an unpopular position in 2018 and is now the default in autonomy and in language alike. Second, “generality wins”: narrow models built for a single vertical rarely keep their lead for long. He couldn’t point to durable legal-specific foundation models, for instance, because “Claude just got better.” The ambition that follows is unsubtle. Odyssey wants to build “the GPT-3 of world models” - the InstructGPT-style moment when a category of model stops being a research demo and starts generating real commercial demand. Odyssey’s research splits into four problems, each with a shipped model against it.

Pixels that keep going…with sound too!

The foundation is autoregressive interactive pixels, embodied in Odyssey-2. A standard video model returns a fixed clip; this one streams frame after frame in real time, which is much harder, because error compounds as the model generates in sequence. It is also interactive, absorbing text prompts mid-stream and adjusting to them. Scaling helps in the usual way - a one-billion to a fifteen-billion-parameter jump lifts the benchmarks - but the real difficulty is keeping the interaction open-ended instead of narrowing it to a single domain.

In May 2026, Odyssey released Starchild-1, named for 2001: A Space Odyssey, which generates pixels and audio jointly in one coherent stream rather than dubbing sound onto finished video. It was harder than expected, Hawke said; to his knowledge no one else had shown one publicly. The obstacle is a clash of timescales - a single predicted video frame spans only “half a phoneme” of audio - so keeping the two coherent took a custom KV-cache design running on two clocks at once.

Shared state, not stitched video

The third problem is multiplayer: shared state across world models, which matters as much for a cell of robots working one environment as it does for a game. Odyssey’s Agora-1 demonstrates it, and Hawke ran it live. The audience pointed their phones at a QR code and played a fully generated game of GoldenEye, streamed off H100s “probably in Spain.” Under the hood it borrows a game engine’s split between rendering and simulation, except both halves are learned - a neural simulator and a neural renderer, trained together. People had said multiplayer world models “weren’t possible,” he noted, “and we felt it was worth disproving.”

Learning by being broken

The fourth strand carries the freshest idea in the talk. Almost everyone who pairs agents with world models uses the agent to get smarter inside the model. Odyssey’s PROWL turns that around: it uses a reinforcement-learning agent to improve the model itself. The reasoning is “garbage in, garbage out” - “your agent will never really be better than the quality of your learned environment model” - and yet, Hawke argued, almost no one had bothered to fix the model rather than the agent. PROWL drops an agent into a learned version of Minecraft and rewards it for finding the world model’s failure modes, then folds those failures into a curriculum that patches them. The wrinkle is that reinforcement learning is very good at cheating: left unleashed, the agent would “sit on the spot, spin the camera around at very high speed” to stall the model’s learning, so the team added a KL “leash” to keep it close to an agent that actually plays the game. The payoff was a clear lift in the base model’s performance - a world model that gets better by being systematically broken.

Still the GPT-2 era

For all of that, Hawke was disciplined about where the work sits. World models, he said, are at “the GPT-2 era”: the pre-ChatGPT stage where the outputs already look promising and the first commercial experiments are forming - he name-checked Jasper, the early GPT-3 copywriting startup - but mass adoption is still ahead. That is the gap between today and the “GPT-3 of world models” he wants to build. The honesty extends to limits. Asked from the floor whether a world model would have to encode general relativity and quantum mechanics, he declined the bait. Visual data is “by far the highest volume” and the right place to start, and these models earn their keep where conventional simulation struggles: for a precise definition of turbulent flow, “use CFD”; for crowds, contact, the messy texture of a scene, the neural simulator wins. Biology and harder physics are wanted, but, he conceded, nascent.

The economics are gentler than the hardware implies. The multiplayer demo ran on a single H100, “let’s call it $1.50 an hour” - close enough to a Netflix subscription that, by his estimate, $30 a month buys ten to fifteen hours of generated play.

Where this goes, on his telling, is convergence: as a visual world model takes on audio and, eventually, text, it meets the language models arriving from the other direction through VLMs, and at some point the two merge. None of that is solved, and he was candid that no one yet knows how. But the direction is set, and the market has caught up to the thesis that once needed an uphill seed pitch. On June 17, 2026, days after RAAIS, Odyssey raised a $310M Series B at a $1.45B valuation, led by Natural Capital with Amazon, AMD Ventures and GV. Air Street backed the company’s seed in 2024, on the bet that learning the world directly would become its own category of model. The GPT-2 era doesn’t last long.

Read more!

A science, not an art

Ted is firm on the framing. These systems are complex, he says, but they are understandable. The catch is that understanding them demands “epistemic humility,” because there are always too many variables in flight and “running a controlled experiment is hard.” The skill he prizes most is knowing the limits of your own evidence: “knowing what an experiment tells you and what it doesn’t tell you is important.” That sounds modest. In a field that markets every result as a breakthrough, it is closer to radical. His own path ran through linguistics and neuroscience, but the thing that carried over, he’s clear, wasn’t brain-inspired architecture - it was the scientific method itself: rigor, skepticism, doubting your own results.

The job also changed his relationship to curiosity. In a PhD you pull on a thread because it’s interesting. In a frontier lab the question is colder: what is the cost-benefit, and is the answer going to be interesting but “ultimately less useful to making Claude smarter”? Experiments that survive that test get fed up to the people deciding the big training runs, where the point is to cut uncertainty before committing the compute.

The metric that matters is the counterfactual

I put a striking set of numbers to him, drawn from Anthropic’s recent essay When AI builds itself: more than 80% of the code merged into Anthropic’s codebase is now authored by Claude, a typical engineer merges eight times as many lines of code per day as in 2024, and on one kernel-optimization task Claude’s speedup climbed from roughly 3x to 52x in under a year. The essay’s own line was that “we have not yet seen that curve bend.” So what, I asked, is the honest measure of AI actually accelerating?

Not benchmarks, Ted said. The real test is “what changes the counterfactual” - if you took Claude away and coded alone, what would you do differently? Concretely, it reduces to the “number of human interventions that are required”: how often a person has to step in, how often the model’s first pass is the one you accept. By that measure the line has moved fast. He pointed to Opus 4.5 in November as a jump, and the Mythos-class models in February as a bigger one. The shift is trust: the models still make mistakes and you still have to check them, “but you can trust them a lot more than you could before.” In the labs, people have already stopped supervising every step - they run in bypass mode and let the agent work.

When bigger models get cheaper

Does the future belong to one large model doing everything, or to companies decomposing tasks across smaller, cheaper ones? Ted leans hard toward the former, and the argument is economic, not sentimental. He cited Noam Brown at OpenAI, who plots test-time-compute curves with cost on the x-axis: a bigger model that reaches an equally good answer in far fewer tokens can come out cheaper than a small one grinding away. “It could be more cost-effective to just ask the bigger model.” His blunter version: “people underrate the value of having a really smart model to ask questions to.” The counter-pressure is real - at Ramp, staff are nudged away from using Opus 4.8 to write emails when Sonnet will do the job - but his bet is that raw intelligence keeps winning on cost as well as quality.

The next axis is taste

The frontier the conversation kept returning to was judgment. The same essay claimed models now pick the better next research step about 64% of the time, up from 51% in November. How do you build a system with taste? Here Ted turned characteristically vague - he wouldn’t say what goes into it - but he would say that when his team started using Mythos this year, it felt like it had better research taste than anything before it. With one caveat: still “not as good as your average researcher.” Average researcher where, I asked. “Maybe at Anthropic.” A high bar to be measured against, and one the models are now climbing.

Safety is a capability, not a tax on it

Ask what safety research has ever done for product quality and Ted reaches for cars. Seat belts, crumple zones, airbags, an oven that won’t catch fire - guardrails are what “make the product usable” at all, and a product nobody can safely use generates no feedback to improve it. The clearest case built the whole category: reinforcement learning from human feedback began as a safety project - stopping a chatbot from spewing garbage - and turned out to be the thing that made chatbots good. “Alignment and safety really go hand in hand with capabilities,” he said. Indeed, reinforcement learning from human feedback draws its roots back to a 2017 collaboration between OpenAI and DeepMind:

One encouraging trend I put to him is that bigger models are turning out easier to align, not harder. Ted’s own framing was more guarded - it has “gone different from how many safety researchers expected,” he said - but his read is that “people feel pretty good about the alignment situation right now,” tempered by the obvious caveat that “we don’t know if we will cross some capability threshold and it’ll suddenly flip.” Hence the case for “healthy apprehension.”

Where the leverage is

I closed by asking what the highest-leverage work in 2026 looks like. Ted was self-aware about his own bias and unhedged anyway: if you believe AGI is close, go to a frontier lab, because your lever there is bigger than it is outside one. He counts himself among the converted - “I’m definitely more AGI-pilled than when I joined.” OpenAI and Anthropic are, he pointed out, still small companies where a newcomer can move things, and the window matters. If you want to influence the direction these systems take, “sooner is better than later.”

That includes London. Anthropic’s office here has gone from 15 or 20 people a few years ago to a couple hundred, with whole strands of frontier work - Ted’s among them - run from the UK rather than mirrored from California. “It doesn’t feel like we’re a satellite,” he said.

What the conversation kept circling back to is that the hard part of scaling was never the spending. Anyone can buy more compute; the discipline Ted’s team is building is the other half - knowing what each experiment will and won’t tell you, and turning that compute into capability you can measure and trust.

That story, which Vivek Natarajan told from the RAAIS stage, is the kind of result his team at Google DeepMind has been chasing. Natarajan is a Research Lead there, working at the intersection of AI, science and medicine. When he last spoke at RAAIS a couple of years ago, the state of the art was Med-PaLM, a language model tuned to answer medical exam questions. His pitch this year was more ambitious: that the recipe behind AlphaGo can be turned on science and the clinic, and that the trick is teaching models to stop thinking fast and start thinking slowly.

System one is not enough

The problem with using a chatbot as a scientist, Natarajan argued, is that even reasoning models mostly do “system one style thinking,” quick responses drawn from surface-level pattern matching. Real discovery is the opposite: slow, deliberate, rigorous, the product of chewing on a problem for weeks until the spark comes. He wanted “system two style thinking,” and to get it he reached back into DeepMind’s own history. AlphaGo’s 2016 breakthrough came from self-play and search, with agents playing each other, taking feedback from the environment, and reinforcing what won. AlphaZero then showed the same recipe could scale from zero knowledge to superhuman play in months, limited mainly by compute.

The AI co-scientist generalizes that idea. Instead of agents playing a game, they generate scientific hypotheses, then critique, debate and refine them over hours and days, what Natarajan calls a “generate, debate and evolve ideas loop.” Borrowing from AlphaStar, DeepMind’s StarCraft system, the team added tournaments: a ranking agent stages pairwise debates between hypotheses, scores them against a rubric derived from the scientist’s stated goal, and assigns Elo ratings, so only the strongest ideas reach the human. Because the debates run in natural language, they can be summarized and fed back into the agents’ context, which is what makes the system self-improving. It also lets the system signal its own uncertainty, what he called “epistemic humility,” which matters when the scarce resource you are spending is a scientist’s time.

The whole project nearly didn’t happen. The idea came from Gary Peltz, a Stanford geneticist who, after one of Natarajan’s lectures, suggested that a model trained on scientific text might generate hypotheses for the causes of rare disease. Most of the team thought it was too early. They did it anyway. Sometimes, as Natarajan put it, you “jump off the cliff, and then you figure out how to build an airplane on the way down.”

From hypothesis to organoid

The Penadés result, run with collaborators at Imperial College on antimicrobial resistance, was the moment the team realized they were onto something. But the more telling cases are the ones that ended in a wet lab. Physician-scientists at Houston Methodist used the co-scientist to find drug-repurposing candidates and combination therapies for acute myeloid leukemia. Peltz’s own lab pointed it at liver fibrosis, a disease with few treatments, and tested its picks in human liver organoids. One candidate, vorinostat, not only showed anti-fibrotic activity but cut TGF-beta-induced chromatin damage by over 91%, a hint of regeneration. The interesting part is that vorinostat is an FDA-approved cancer drug, exactly the kind of cross-field connection a liver specialist might never make, and the system surfaced it because it could read broadly while the human judged what mattered. Natarajan called this complementary intelligence, and it is the honest version of the pitch: the machine goes wide, the scientist goes deep.

He kept the limits in view. The system itself is general-purpose, he stressed, with nothing in the scaffolding specific to biology; the specialization comes from the tools it reaches for at runtime. But asked where it fails, he was candid that the wins so far have been in biology, where a mass of unread literature hides real signal. Chemistry is harder, and fields like mathematics and physics, which reward narrow depth-first reasoning over wide reading, harder still. The common thread is not biology or medicine specifically, but a way of turning compute into disciplined deliberation, then putting the result back in front of expert humans.

Manufacturing medical experience

The other half of the talk was about access. World-class medicine, Natarajan said, is “pretty much a geographic and socioeconomic lottery,” and his team’s second mission is to close that gap. The vehicle is AMIE, a diagnostic dialogue system he co-leads with Alan Karthikesalingam, a vascular surgeon still practicing in the NHS. Asked once how to choose between two doctors, Karthikesalingam told him to “always go with the one who has more gray hair.” There is no substitute for experience, so the team manufactured it. Using the same self-play machinery, AMIE ran consultations against synthetic patients and a critic, refining itself over millions of simulated dialogues. A human doctor might see 10,000 to 50,000 patients in a career; AMIE has already run hundreds of millions of conversations, “building up the world’s most experienced doctor,” with the heavy caveat that it all happens in simulation.

It is starting to pay off. In evaluations published in Nature, AMIE matched or beat physicians in simulated consultations with patient actors, on diagnosis and, more pointedly, on rapport, empathy and relationship building. “I’m kind of sorry about the doctors and humans in your life,” Natarajan deadpanned to the unsurprised. But this is not about replacement, he insisted: “the story is still about augmentation.” A companion study found that general physicians given complex diagnostic puzzles did significantly better with the AI as a thinking partner than working alone or with standard tools like web search. And in an early, supervised feasibility study with Beth Israel Deaconess in Boston, where patients spoke to the AI before an urgent-care visit under physician oversight, zero safety stops were required under the study’s predefined criteria, patient trust in AI rose after the interaction, and the system’s pre-visit diagnoses held up against the attending physicians, all without the benefit of lab tests.

A third person in the room

For most of modern medicine, Natarajan closed, the core unit of care has been a dyad: the doctor and the patient. His bet is that it is becoming a triad, the doctor, the patient and the AI, with the machine as a teammate rather than a tool. That is the idea behind the team’s next effort, an AI co-clinician. It is a tidy frame, and the evidence on stage made it land harder than it would have a year ago. The deeper claim running under both halves of the talk is that the self-play recipe which once mastered a board game can now give scientists and clinicians a new kind of thinking partner: one that searches widely, argues with itself, and hands humans better starting points. The wet labs and the early clinical studies are starting to agree.

At this year’s RAAIS, Roberta Raileanu set out why that ceiling exists and what it would take to lift it. Raileanu leads the open-endedness team at Google DeepMind and was previously at Meta. Her talk laid out a recipe for superhuman scientific discovery: a general system that makes groundbreaking discoveries across domains faster than people can. Three ingredients hold it together, but underneath all three sits one problem. We are good at searching for anything we can measure. We do not yet know how to measure what makes a discovery good.

The plateau before the breakthrough

The past two years delivered a real proof of concept. In 2024, Sakana AI wired LLM agents into a loop that generates a hypothesis, implements it, runs an experiment, and iterates, producing its first machine-written papers. The bar has risen since: a fully AI-generated paper has passed peer review at a workshop attached to a top machine learning conference, and a wave of startups now aims to automate research outright.

Proof of concept is not parity, however. Put the best agents head to head with human experts on the same open problems and the agents improve early, then stall. They are good at variations and combinations of known methods, and weak at what defines real research: exploring unfamiliar paths and making the conceptual leaps that change a field. Scale up compute and time and the human line keeps rising while the model line flattens.

The reason to think the ceiling can move is breadth. These models train on a far wider cross-domain corpus than any scientist can absorb, and can search for connections across more fields than any specialist holds in working memory.

The lesson of Move 37

Raileanu’s first ingredient is to treat discovery as a reinforcement learning problem. An agent acts in an environment, observes what happens, and learns from the feedback, which is not far from how a scientist forms an idea, tests it, and revises. The appeal is specific: as long as you can measure progress with a reward, the agent is free to find any solution that earns it, including one no human would think to try.

The proof is a decade old. When DeepMind’s AlphaGo played Lee Sedol, its move 37 was so counterintuitive that no human would have played it, and it won the game. But Go is a closed world with a clean reward: a move wins or it does not. Move 37 shows what optimization can do once the objective is given. In science the objective is not given. Deciding what counts as progress on an open question is the actual work, and it is the part no reward function hands you.

To study this inside AI research itself, her team built MLGym, a sandbox where an LLM agent runs shell commands, edits files, and runs experiments across tasks from language modeling to game theory. Even a year ago, simple setups could self-improve against a benchmark, but only by tuning hyperparameters and swapping architectures, not by inventing a method a human expert would adopt. That gap, between optimization and originality, is the rest of the talk.

Why greatness cannot be planned

Most breakthroughs, Raileanu argued, are not solutions to known problems. They involve finding the right problem, and “innovation is rarely this linear process from A to B.” Try to build a personal computer in the 1800s and you would not get there by scaling up the abacus; you would need the vacuum tube, which was invented to amplify radio signals.

She took the frame from Kenneth Stanley and Joel Lehman’s “Why Greatness Cannot Be Planned,” and put its claim on the screen: “No prerequisite to any major invention was invented with that invention in mind.” Optimize too narrowly for an objective and you skip the stepping stones that lead to it. Machine learning has won by hill climbing toward benchmarks, and that has carried the field far. But a hill climber only ever reaches the top of the hill it started on.

The fix is to widen the search. Borrowing from evolutionary methods, you hold a population of candidate solutions, mutate them, and select which to keep. The usual fitness function rewards performance alone. Raileanu’s argument is to score for what scientists actually value too: novelty, diversity, interestingness. This is where the signal problem surfaces in the open, because none of those is easy to measure. Her stopgap is to let an LLM judge what a person would find interesting, which at least keeps ideas that are not useful yet but might combine into something later.

Her team’s Rainbow Teaming did this for AI safety, generating diverse jailbreak prompts across a grid of risk categories and attack styles, then reusing what worked in one cell to seed another. Train on the result and the model gets measurably harder to break. The same machinery, she suggested, should carry over to ideas and methods, where a solution built for one field can matter in a completely different one.

Optimizing discovery itself

The third ingredient is to stop optimizing discovery on a single task and optimize the process of discovery itself. DeepMind has trained RL agents across huge numbers of simulated environments and watched them adapt to new ones far faster than before, in some cases faster than humans. If that transfers to research, an agent could generate its own tasks and get better at discovering, not just at one discovery.

To make that studyable, her team built DiscoBench, a framework that procedurally generates AI research tasks: more than 400 million of them, across problems like language modeling and image classification, with an agent free to target a loss function, an optimizer, or an architecture. The early signal is encouraging, with more and more diverse training tasks improving performance on held-out problems the agent has never seen.

The complementary curve

Stack the three together and you have the recipe: reinforcement learning to discover better solutions where progress can be measured, divergent search to find new problems rather than climb known ones, and meta-learning to speed up the whole process on problems no one has posed yet.

The bet underneath it is complementarity. Humans go deep in one field; a model reaches across many at once. The line worth chasing is neither the human curve nor the machine curve, but the one above both, where the two discover what neither would alone. Yet all three ingredients run back into the same wall. We have good algorithms for search once we know what to reward, and we still cannot reward novelty, a promising dead end, or taste. “The key is, do you have the right signal?” Raileanu asked. The search is the easy part. The missing piece is the signal.

You’ll now find updated data for AI research papers using NVIDIA, TPUs, Apple, Huawei, AMD, ASICs, FPGAs, and AI semi startups. We’ve also expanded the infrastructure side of the index to 1 July 2026: A100, Hopper, standalone Blackwell, Grace-Blackwell, and a new demand-side view of frontier-lab contracted compute in gigawatts. Each chart can now be downloaded, shared, and embedded.

A few notes upfront: The 2026 citation figures use real counts through June 1, 2026 plus a volume-adjusted projection for the rest of the year. Year-over-year deltas are calculated against final normalized 2025 counts, not last year's mid-year 2025 projection. GPU-count charts show NVIDIA data-center GPUs by owner/operator, split into Deployed, Installing, and Announced. Grace-Hopper and Grace-Blackwell parts are counted by GPU dies, so one NVL72 rack equals 72 GPUs. Tenants are not double-counted against the operators whose clusters they use.

Subscribe now

The breather was short

Last year’s update asked whether 2025 was the first real slowdown in open AI compute citations after six years of growth. With final normalized 2025 counts now in hand, and 2026 projected from counts through June 1, the answer looks clearer: 2025 was a pause, not a rollover.

Across the tracked accelerator categories, the 2026 projection reaches 49,339 chip-citation counts, up 10.7% year-over-year and just above the 2024 peak. NVIDIA remains the default: its chips appear in 44,715 of those counts, up 10.9% year-over-year and about 91% of the tracked total.

That does not mean frontier labs have suddenly become more transparent. The largest model developers still publish less of their best work, and many papers built on managed APIs or shared cloud services do not specify the underlying silicon at all. But the open literature has not stopped reflecting hardware diffusion. The 2025 dip looks more like publication-cycle timing, API abstraction, and a quiet year between hardware waves than a collapse in compute usage.

The more interesting finding is the relative lack of upheaval. A lot of companies, labs, and governments are trying to challenge NVIDIA. In the open research literature, they are still mostly not succeeding. There is movement in AMD, Huawei, and Apple, but NVIDIA remains the language researchers use when they describe their compute.

Indeed, AMD citations nearly doubled, from 251 in 2025 to 472 in 2026. Huawei Ascend 910 rose 56%, from 137 to 213. Apple moved from 741 to 998, overtaking AMD to become the most-cited non-NVIDIA, non-Google accelerator in the open literature, which means consumer Macs now out-cite the leading silicon challenger. That likely reflects the spread of local inference and developer workflows rather than frontier training. TPUs, meanwhile, declined 7% in the open paper data, despite Google’s obvious importance to frontier AI compute. That tension is a useful reminder that paper citations are a leading indicator for some forms of adoption, but a poor measure of private, API-mediated, or closed-lab usage.

Subscribe now

NVIDIA is still king, but the kingdom is changing shape

The NVIDIA chart now approximates their product-cycle chart.

A100 citations are basically flat at 15,327, up 0.4% year-over-year. The A100 is no longer where the growth is. H100/H200 citations have more than doubled to 9,823, up 111% year-over-year, as the 2024 and 2025 Hopper build-out finally works its way into papers. Blackwell-family mentions, tracked as B100/B200/B300 in the Zeta Alpha data, are still small at 902, but are up 4.5x year-over-year.

Meanwhile the older stack continues to drain out: V100 citations are down 34%, RTX 3090 down 23%, P100 down 36%, and K80 is now barely visible. The RTX 4090 is still useful in the academic long tail, up 9% to 6,557 citations, while the new 50-series cards are appearing quickly from a small base.

Startup silicon is no longer one story

Groq is now the most-cited startup chip, up 49% to 264, and in December, NVIDIA acquired it. The fastest way to dent NVIDIA’s share turned out to be getting bought by NVIDIA. Cerebras is second at 242, up 8%. SambaNova rose to 52, Cambricon to 33, while Graphcore fell to 40 and Habana to 24.

That is not yet a market-share story. Startup chip citations remain tiny next to NVIDIA, and papers using startup silicon still often include authors from the chip company itself. But it does show the category splitting into different jobs. Groq is showing up around low-latency inference, Cerebras around large-scale wafer-scale systems, and the older acquired or de-emphasized platforms are fading from view.

The right conclusion is not “startup chips are breaking CUDA.” They are not. It is that the category is specializing into niches, inference latency and wafer-scale, while NVIDIA keeps the general case.

Hopper is the installed base

The second half of the index is not a citation tracker at all. The cluster charts show how much of the AI build-out has already moved from research procurement into industrial infrastructure.

Across the tracked Hopper systems, we count 460,904 deployed H100, H200, and GH200 GPUs, plus 1,328 installing. That is more than 11x the tracked A100 total of 41,208. The A100 chart now reads like a legacy fleet chart; Hopper is the live installed base.

xAI Colossus 1 is the largest tracked Hopper deployment at 200,000 GPUs, and by itself is almost five times the entire tracked A100 set. Tesla Cortex follows at roughly 66,000 H100-equivalent GPUs, then Meta’s GenAI clusters at 49,152, a CoreWeave H200 cluster at 42,000, Voltage Park at 24,000, and Germany’s JUPITER Booster at 23,536 GH200.

There is a geopolitical point hiding in the table. National HPC systems are exact and visible, which makes them easy to count. Private fleets are estimated, harder to observe, and much larger. Europe now has serious machines in JUPITER, Alps, Isambard-AI, Leonardo, MareNostrum 5, and Jean Zay. But the largest private AI clusters are operating at a scale national supercomputing programs mostly do not match.

Subscribe now

Blackwell is mostly pipeline

Standalone B200/B300 deployments in the index total 20,024 GPUs, led by Deutsche Telekom’s Munich Industrial AI Cloud at 10,000, then SoftBank’s DGX SuperPOD and Korea’s Naver Cloud at 4,000 each, E2E Networks in India at 1,024, and SK Telecom’s Haein cluster at around 1,000. IREN’s 50,000 B300 order sits in Announced.

Grace-Blackwell is where the real pipeline sits. The index tracks 100,128 deployed GB200/GB300 GPUs, 308,640 installing, and 1.66 million announced. In other words, just under 5% of the tracked Grace-Blackwell pipeline is deployed, and roughly 80% is still announced. The deployed number crossed 100,000 this month as the first sizable GB300 systems came online, including CoreWeave’s 8,192-GPU cluster, which posted MLPerf training results in June, and Mistral’s 13,800-GPU Bruno in France moving from installing to live. It is a small share of the pipeline, but it is the first month the deployed bar moved on the strength of Grace-Blackwell rather than Grace-Hopper.

The largest announced and installing programs still look more like sovereign or hyperscale industrial projects than normal data-center procurement: HUMAIN in Saudi Arabia at up to 600,000 GB300, Stargate Abilene at a 450,000 GPU target, South Korea’s national 260,000 Blackwell program, Nscale’s 200,000 GB300 commitment to Microsoft, xAI Colossus 2 at roughly 110,000 installing, Argonne Solstice and Equinox at 110,000 combined, and Stargate Norway at 100,000 announced.

This is why “GPU count” is becoming an incomplete question. The constraint has moved outward to power, land, cooling, interconnect, permitting, debt, and offtake. A GPU order is not a cluster. A cluster is not always available capacity. And contracted capacity is not the same as a model training run.

The demand side is now measured in gigawatts

The frontier-lab contracted compute chart looks at the other side of the market: not who owns a cluster, but which labs have contracted capacity in disclosed gigawatt terms.

For GW-disclosed deals, OpenAI is now at 26 GW of contracted compute: 10 GW of NVIDIA systems and 16 GW of non-NVIDIA, the latter split between a 6 GW AMD MI450 commitment and 10 GW of Broadcom custom silicon. Anthropic is at 6 GW total, split between 1 GW of NVIDIA and 5 GW of non-NVIDIA capacity.

On disclosed gigawatts, then, OpenAI’s non-NVIDIA book is now larger than its NVIDIA book, and larger than Anthropic’s entire portfolio. It is worth being careful about what that does and does not say. NVIDIA is still OpenAI’s largest single-vendor commitment and its fastest path to scale; the non-NVIDIA figure is spread across AMD and a custom Broadcom part that has not yet shipped at volume. Anthropic’s non-NVIDIA number is, if anything, understated, because its AWS Trainium and Google TPU commitments are only partly disclosed in gigawatts and are therefore not fully charted.

Still, the direction is clear. Frontier labs are no longer choosing a single chip vendor. They are assembling compute portfolios: NVIDIA for the broadest software ecosystem and fastest path to scale, TPUs and Trainium for strategic supply and cost control, custom silicon for future leverage, and neoclouds or sovereign projects when hyperscaler capacity is not enough.

“CUDA is still king” remains true in the papers. But in frontier-lab procurement, the question is becoming broader: who can turn contracted gigawatts into reliable, liquid-cooled, networked, usable intelligence infrastructure?

Subscribe now

Looking ahead

Several known unknowns will shape the next Compute Index update:

• How quickly announced Grace-Blackwell projects become deployed clusters rather than press releases, now that the deployed bar has finally started to move.

• Whether OpenAI’s Stargate program, HUMAIN, South Korea’s national Blackwell program, and Nscale’s Microsoft commitments hit their published timelines.

• Whether AMD MI300/MI350, Huawei Ascend, TPUs, and Trainium show up more visibly in open paper metadata, or remain hidden behind closed-lab and cloud-abstraction layers.

• Whether Blackwell appears in the literature in late 2026 the way Hopper appears in the 2026 data now.

• Whether OpenAI’s and Anthropic’s non-NVIDIA books convert from contracted gigawatts into deployed, usable capacity, or stay on paper the way the Grace-Blackwell pipeline mostly has.

• Whether national AI factories can close any of the gap with private frontier-lab infrastructure.

The headline from v7 is simple: the open literature has rebounded and NVIDIA remains the default; the challenger story is real but still small in the papers; Hopper is the installed base while Grace-Blackwell has only just begun to land; and on the demand side, even OpenAI’s disclosed book is now majority non-NVIDIA, even as most of that capacity is still a contract rather than a cluster.

See the live charts here: www.stateof.ai/compute.

A few notes

• We take the view that usage of chips in AI research papers by early adopters is a leading indicator of broader industry usage, but not a complete measure of closed-lab or API-mediated compute.

• Paper-citation counts are unchanged from v6. Zeta Alpha’s open-source AI paper index refreshes annually; 2026 figures are real counts through June 1, 2026 plus volume-adjusted full-year forecasts, and year-over-year comparisons use final normalized 2025 counts as the baseline. This update refreshes the infrastructure charts to July 1, 2026.

• GPU-count charts use public disclosures, operator materials, EuroHPC, Top500, SemiAnalysis, The Next Platform, Data Center Dynamics, company filings, and NVIDIA materials. Large private-company figures are best-available estimates; national HPC figures are exact where published.

• Announced capacity is labeled separately from deployed and installing capacity. We do not invent cloud instance counts where vendors do not disclose them.

  Subscribe now

Angelos leads ElevenLabs’ speech-to-text and text-to-speech teams and built its Scribe and Scribe Real-time transcription systems. The Scribe V2 models he shipped this past year rank, he says, as the most accurate transcription models on most popular benchmarks. It gives him a particular vantage on the problem, since voice models live or die on latency and cost at scale. It is also the second year running ElevenLabs has taken the RAAIS stage; in 2025 its CEO, Mati Staniszewski, spoke on the voice frontier. This year, we dove into the engine room.

What a token actually costs

Every optimization starts with knowing what you are paying for. For the autoregressive transformers behind most popular LLMs, a token’s cost reduces to two bottlenecks: compute, how fast the GPU does the matrix multiplications, and memory bandwidth, how fast the GPU’s VRAM can load the model’s weights and its KV cache. Generation runs in two phases. A prefill step reads the whole prompt and fills the KV cache, the model’s working state, and is compute-heavy. A decode step then emits tokens one at a time, each conditioned on the last, and is memory-heavy. The KV cache is what lets the model reuse that prefill instead of recomputing it for every new token, and at scale it is the thing that hurts: a hundred concurrent requests need a hundred separate caches resident in memory. Size is not destiny either. Angelos noted that Qwen 3’s cache costs almost three times as much per token as Qwen 2.5’s despite near-identical parameter counts, so two models of the same size can cost wildly different amounts to run.

Stop letting the GPU sit idle

The first and biggest single win is batching. GPUs are excellent at parallel work and poor at sequential work, and the dominant cost in decoding is loading the model weights, which can be shared across every request in a batch rather than reloaded for each. Naive batching groups requests once and then waits for the slowest one to finish while the GPU idles. Continuous batching fixes that: it batches at the level of each decode or prefill step, so a new request can join a GPU already mid-flight on others.

Shrink the weights, then the cache

With the GPU busy, the constraint becomes memory, so the next moves all reduce it. Quantization comes first. Models are usually trained at BF16, sixteen bits per weight, which is more precision than they need; dropping the weights to FP8 roughly halves their footprint with near-lossless accuracy, given H100-class hardware and a little quantize-aware training, which injects noise into the gradients so the model learns to tolerate the lower precision. That buys headroom for more cache and lifts throughput to twenty users per GPU. The more aggressive options exist too: int4 is lossy but useful on-device, and MXFP4 reaches four bits but only on Blackwell and newer.

Speculative decoding comes next. A cheap draft model proposes tokens and the big model verifies them in a single forward pass, accepting the run until the two disagree. It only pays off when the models agree often, which they frequently do not, so in practice it is used less than its reputation suggests; applied here it nudges the running total from twenty to twenty-eight users per GPU. The more popular cousin is multi-token prediction, where the same model wears extra prediction heads and drafts several tokens itself, with no second model to host. It earns its keep with two or more heads, and it doubles as a training signal: teaching a model to anticipate several moves ahead, like a chess player, tends to make it more stable and sometimes faster to learn. Most big labs use it, and it lands in the same place, around twenty-eight users per GPU.

The largest gain is also the riskiest. The KV cache holds far less redundant capacity than the weights do, so compressing it is genuinely lossy. Angelos was candid about this from his own testing: the much-discussed Google method TurboQuant was announced as lossless but, in his experience, proved lossy in practice, because any change to the cache is hard for the model to recover from. The fix is again on the training side: distill the model so it grows accustomed to a lower-precision FP8 cache, and you keep most of the accuracy while shrinking the cache 2.5x. That single step lifts throughput from twenty-eight to seventy users per GPU - seventy times what the same hardware served at the start.

Where the frontier labs go

Seventy is what disciplined engineering gets you. Going further means changing the architecture itself, and here the labs are placing different bets. DeepSeek’s multi-head latent attention squeezes each token’s key-value pair into a small latent rather than storing it in full, which both speeds inference and stretches context toward a million tokens; it was one of the more copied ideas after DeepSeek-R1. Qwen swaps standard quadratic attention for a linear network on every other layer, cheaper and longer-context, at some cost to quality. NVIDIA goes furthest, replacing the transformer on a fraction of its layers with state-space models that scale linearly and compute faster, keeping enough transformer layers to hold accuracy up. With architecture-level changes like these, the ladder reaches roughly a hundred and forty users per GPU.

Nothing here is free

Angelos was careful not to oversell any of it. Every technique on the ladder carries a cost. Batching adds latency and runs into a memory ceiling. FP8 quantization takes a small quality hit without the extra training. Speculative decoding needs access to weights and a training pipeline to work well. KV cache compression is the one most likely to degrade output, so the real question is how much degradation you can absorb rather than whether you can avoid it. He was blunter still about the gap between papers and production: many compression methods that report no loss of accuracy were tuned on a handful of benchmarks, and scaled to millions of users they can simply fall apart. You often only find out which ones once they are popular enough to be stress-tested in the wild. Which technique pays depends entirely on the workload.

The reason any of this matters beyond the engineering reached the room through a question from the floor: today’s token prices are subsidized, by one audience estimate a factor of ten to forty. Angelos’s hope is that optimization, not subsidy, eventually closes that gap. The largest models, he said, have to be subsidized to make economic sense, but he expects smaller, Sonnet-class models to become good enough for nearly all everyday use, at margins that actually work. He already sees the shape of it in agent systems: route each request to the smallest model that can handle it, and reserve the expensive one for planning.

This week STARK, the German defense company building exactly these systems, raised €500 million led by Sequoia and Founders Fund. I first met STARK’s CEO and Founder, Uwe Horstmann, back at Project A in 2016, years before any of this. Air Street Capital invested because we believe STARK is emerging as a German neoprime: a new prime contractor for unmanned strike systems, built software-first and able to manufacture at the scale a real war demands.

STARK builds autonomous platforms, fast

STARK’s flagship platform, Virtus, is a loitering munition: a drone that flies to a contested area, waits, finds a target, and strikes it. It is not a reconnaissance drone, though it can return and land like one, and it is not a cruise missile. It sits between the two, cheap enough to expend, autonomous enough to find a target on its own, and simple enough to build at scale. Around it STARK is building a family of effectors - Gambit, a man-portable short-range munition, and Cascade, a tube-launched one - alongside Vanta, an unmanned surface vessel that takes the same idea to sea.

Mass you can afford to build

For decades, Western firepower meant a small number of exquisite platforms, each costing millions and far too precious to lose. Ukraine inverted that. The decisive weapon turned out to be the cheap one you can field by the thousand and expend without flinching. A Virtus assembles in roughly ten minutes, and Germany has already put STARK on a framework worth up to €2.8 billion to supply the Bundeswehr, a landmark deal for a new defense company.

Once the weapon is cheap, the constraint moves to manufacturing and a robust supply chain. More than 80% of this round goes into manufacturing and R&D, and STARK is standing up production lines across Germany, the UK, and Ukraine. In a war of attrition, throughput is the moat: the side that can keep replacing what it loses sets the tempo. STARK builds on civilian supply chains and simple lines that stand up in weeks and replicate from city to city, which is both a way to scale and a way to survive being targeted. The defense companies that win this decade will treat the production line as the product.

Lastly, with a growing portfolio of autonomous products, STARK needs software to tie them all together: Minerva. The same software that swarms Virtus in the air coordinates the unmanned boats at sea, navigates when GPS is jammed, and plugs into NATO battle-management. Each platform improves with a software update rather than a new airframe, and every deployment teaches the whole fleet.

Why STARK

Uwe spent a decade as a general partner at Project A, one of Europe’s most active defense investors and an early backer of Quantum Systems, before leaving to run STARK. He has assembled the four branches this company needs at once. To build at scale: Martin Rost, sixteen years at Zalando running a roughly €10 billion unit, and Johannes Schaback, a repeat founder who was CTO of SumUp. To sell into defense: Jan-Patrick Helmsen, former CEO of Rheinmetall's weapons-and-munitions business. To win the politics: Johannes Arlt, until this year a member of the Bundestag's defense committee. And to learn from the war as it is fought: a Ukraine team running an R&D and production hub in Kyiv that turns frontline feedback into design changes.

Regular readers will be familiar with our defense thesis at Air Street Capital. We recently led the Series A for Alta Ares because Europe has to build its own air defense shield. In our letter from this year’s Munich Security Conference, we argued that Europe must move from emergency buying to structural production capacity, and that when a government contracts a domestic firm it confers the industrial gravity that pulls in private capital and localizes supply chains. STARK is what answering that call looks like: sovereign manufacturing, software-defined systems, and a founder building for the war that is actually being fought. The mass that wins the next conflict will be cheap, built at home, and improved in software, and we’re here to stand behind the people making it happen.

The difficulty is in the control plane around it: one gateway, one governance layer, measurable fallbacks, cost controls, layered human review, and a way to run all of it inside a regulated bank that serves more than 70 million customers across over 40 countries. Revolut ships more than 200 products and has handled over a trillion dollars in transactions, with a machine learning model now in the path of almost every one of them. The leverage, in Donets’s telling, has moved from the model to everything around it.

Four constituencies, one bottleneck

For years, Revolut’s AI was classical machine learning: fraud and transaction models shipped through three internal libraries, one each for training, serving, and performance monitoring. Then, in 2022, the ground moved. Vendors began exposing large models behind an API, and suddenly you did not have to train anything to build something. Generative use cases started growing exponentially while the classical models kept running underneath.

Donets spent as much time on the people problem this created as on the technical one. Four internal groups pull in different directions: researchers who want compute and freedom to explore; builders who want one common API and to ship today; operators who want predictability, rollbacks, and cost under control; and a compliance function that owns human-in-the-loop controls, security audit, and data sovereignty. Left to themselves, every product team solves the same problems its own way, and governance fragments into tribal knowledge spread across hundreds of teams. That is expensive, and it does not scale.

Govern the use case, build one gateway

Rather than govern each model one by one, Revolut made two moves that changed the shape of the problem. The first shifted the unit of governance to the AI use case, a move that lines up with the EU AI Act’s use-case-based view of risk, so that one set of risks, budgets, and rules can cover several models at once and match policy to context. The second put a single gateway at the center of the company, with the governance layer on top of it, rather than shipping capability as libraries each team installs for itself.

But there’s a tradeoff: whereas libraries push reliability onto whichever product team owns the service, a central gateway makes one team responsible for everyone. Even so, the cost of improving a library means cutting a release, then persuading hundreds of busy teams to upgrade and absorb breaking changes they never wanted. With one gateway, the central team ships the improvement once and every product inherits it at, in Donets’s phrase, “zero effort.” Compliance and monitoring move to the same place. As a result, Revolut runs roughly twice as many generative use cases as classical ML ones, all off that single platform.

What breaks when the model is someone else’s

Once you are renting frontier models rather than training your own, you inherit failure modes you do not control. Pay-as-you-go providers run at around 98.5% uptime, which sounds high until you count the hours of dead service it implies each month for a scaled, global business. So Revolut wires a fallback chain into every generative product: if the primary model degrades or stops responding, traffic rolls to the next, and the next. Slightly degraded service beats no service at all.

Subtler, and more painful, was a failure they could not see at all. Because the platform watched only inputs and outputs at the interface, a model buried in the fallback chain quietly stopped working and nobody noticed. “Everything was fine, uptime was high enough, but the model itself was not functional,” Donets said. Or, as one of his slides put it: without per-model visibility, a model doing nothing looks exactly like one that works.

Money was the other lesson, and an easier one to swallow. Teams reach for the newest and most expensive model by reflex, but most workloads are over-provisioned, and right-sizing the model to the task cuts cost by as much as eight times with no loss in quality. Donets’s rule: do not default to the newest model in production. Measure first, then use the smallest model that clears the bar.

A note on the org chart

Underneath the platform sits an org chart doing as much of the work as the code. Revolut is flat and built as a matrix: AI engineers are embedded in product teams, each staffed to ship end to end, with a functional line back to the platform. Standards and tooling flow down; field requirements flow up to Donets’s central group, which sets direction and pushes compliance rules out. He called the product teams “our forward-deployed engineers,” the mechanism by which one team’s hard-won experience becomes everyone’s. The architecture, as one slide noted, ends up shaped like the org chart.

From Rita to AIR

Where all of this lands is a single product that has been running for years. It began as Rita, a support chatbot built on intent models and pre-filled scenarios - the “slot machine” era - that frustrated as often as it helped. In 2022 the team tested large models, Bloom and BloomZ at 175 billion parameters, and found they worked. The first thing they put into production was mundane: paraphrasing a multi-screen FAQ into a short, relevant answer. LLM-based Rita reached production in Q2 2023, then rolled out country by country, Europe first and Japan the hardest, finishing around Q1 2025.

Voice came next, and it runs on a simple pipeline: audio is transcribed, a small LLM decides whether to answer directly or hand off to the full multilingual chatbot, and an end-to-end response comes back in under two seconds. It now runs in 20 countries, handles around 25,000 calls a month, and resolves a customer’s problem roughly eight times faster than a human agent. AIR, the latest layer, followed in Q2 2025 and pulls in transactional data: it can break down your spending, propose hotels inside a budget computed from your own history, or explain why a stock is moving. Across the arc from the old chatbot to today, the share of cases resolved without a human climbed from 17% to 80%, Net Promoter Score went from low to high, and the financial impact, Donets said, ran into double-digit millions of pounds. AIR began rolling out in the UK in April 2026, where Revolut says it has 13 million customers.

Where human oversight is mandatory

Holding all of it up is the monitoring layer. Revolut stores every input and output and runs a panel of LLM “judges” against live traffic - one dedicated to hallucination - currently 9 to 12 mandatory metrics and rising, backed by human review teams that sample chats and transcripts, and by the blunt signal of Twitter and Reddit when something goes badly wrong.

Above all of it sits a hard line: no decision that can change someone’s life is made by an AI system. That position got tested in the room. An audience member who works on regulated healthcare AI pushed back - in his field, he said, “humans make that process unsafe,” so an AI judge might be the safer choice. Donets gave ground on the evidence, agreeing that machines “provide more stable and better help to users,” but not on the principle: the critical calls still do not go to the model. Asked how soon that might change, he did not hedge: “This year, definitely no.”

The frontier gets the headlines, but shipping AI inside a regulated bank across 40 countries is won or lost on the plumbing beneath it: one gateway, the right unit of governance, a fallback for when the vendor fails, and a human who still gets the last word.

A decade ago, the AI community was alight with enthusiasm over the first generative models for images, Generative Adversarial Networks. At the time, the model's outputs looked more like a Microsoft Paint attempt than anything close to photorealism. Ten years later, Latent Diffusion Models, and the scaled-up, improved architectures since built by the team at Black Forest Labs, have ushered in a level of photorealism arguably indistinguishable from reality, were it not for the unreal scenes:

The same trajectory from “pixelated if you’re lucky” to “I can’t tell the difference between reality and AI” is happening in the domain of world models, and Odyssey is leading the frontier.

Last week, Odyssey announced that it has raised a $310M Series B at a $1.45B valuation, led by Natural Capital, with participation from Amazon, GV, AMD Ventures, EQT, IQT, and others. We wrote the first check into Odyssey's seed in late 2023.

On a personal note, I have known both founders far longer than Odyssey has existed: Oliver Cameron from his years building Voyage and then Cruise (the first self-driving car I experienced thanks to him!), and Jeff Hawke from the founding team at Wayve, where I was involved from day 1. I invited Oliver to speak at RAAIS 2023 in London, which created an opportunity for the two of them to spend significant time together in person. They started Odyssey later that year.

This round is a good moment to explain what the team has actually been building. The answer is more interesting than “AI video.”

From videos to worlds

A world model is not a video generator. Sora, Veo, Kling, and their successors take a prompt and render a fixed clip. The output can be beautiful, but the future is locked from the outset of generation. As a consumer of the video, your only mode of interaction is to watch.

By contrast, a world model is closer to a simulator than a renderer. It predicts the next state of an environment from the past and from whatever a participant does next. It accepts input mid-rollout. It holds a persistent state, a memory of the world that can be acted on and changed. Just as next-token prediction enabled language modeling, Odyssey is betting that next-state prediction, at enough scale, forces a model to learn physics, because a model that has not learned physical regularities drifts into nonsense over a long rollout.

The path to frontier world models

Scale is necessary, but not sufficient. The bottleneck is experience, and over the past year Odyssey has attacked it from three directions: 1) making each moment richer, 2) making experience shared and persistent for people and agents, and 3) teaching models to generate their own.

First, richer experience. Most world models are mute, which is not only bizarre to experience but means that a rich information stream is discarded. Sound is where collisions, distance, intent, rhythm, and emotion live. Nothing is in the intellect, as the line goes, that was not first in the senses. As such, Odyssey’s Starchild-1 generates synchronized audio and video in real time, while responding continuously to streaming text, speech, and action. This is the first real-time multimodal world model.

While this sounds logical, the hard part is that audio and video move on different clocks. A small error in one modality can corrupt the other during a long rollout. Starchild-1’s approach is to let each run on its own clock while staying synchronized, turning a bidirectional audio-video foundation model into a causal, real-time world model.

Second, shared experience. Agora-1 is a multi-agent world model that decouples simulation from rendering. One function evolves a shared world state from player actions, while another renders consistent views of that state from independent viewpoints. The result behaves like a game engine with no hand-coded engine underneath: a shared state the model maintains for every participant, which can be edited into new levels while the dynamics hold.

At RAAIS in London on June 12, Jeff Hawke, Odyssey's CTO, ran a live session of an Agora-1-generated GoldenEye death match, with every frame conjured on the fly. Attendees could join the game and play one another in real time.

Beyond games, robots need shared environments before they touch the real world. Agents need places to collide, coordinate, compete, and fail. Simulators need to cover worlds that have never existed. Odyssey is pursuing all of these directions with design partners it will name in time.

Play GoldenEye here!

Third, self-generated experience. PROWL is a reinforcement learning agent rewarded for breaking the world model: freezing a waterfall, losing a crosshair, popping geometry under the camera, ignoring a control input, or collapsing through a hard scene transition. Through active exploration, an agent can therefore improve a world model.

Readers may recall OpenAI’s “Faulty reward functions in the wild” from 2016, where an RL agent steering a boat learned to rack up points by spinning in circles instead of finishing the race. RL agents are unreasonably good at finding the cracks in a system. PROWL points that talent at the world model itself: the agent hunts a weakness, the model trains it away, and the agent comes back for a harder one. It is a way to manufacture the experience these models are short of.

Why Odyssey

When we first invested, we wrote that great research is never enough on its own. The teams that win pair it with execution, product taste, and a feel for where the customer needs to go next.

Oliver Cameron and Jeff Hawke came out of self-driving, where the job is to break an impossible physical AI problem into tractable pieces, simulate the world, make the model robust and generalizable. Odyssey’s shipping cadence since their seed round continues to accelerate: Odyssey-2 Max for scale and physics, Starchild-1 for multimodal grounding, Agora-1 for shared state, and PROWL for closed-loop improvement.

Odyssey is betting that the next leap in machine intelligence comes from systems that build worlds, act inside them, and learn how reality behaves. If that is right, the lead in robotics, agents, and simulation goes to whoever can generate the richest experience fastest. We wrote the first check because we think Oliver, Jeff, and the team are the ones who will.

Join the team!

Guilherme Penedo and Hynek Kydlíček built that data. Over roughly three years at Hugging Face they shipped FineWeb, FineWeb 2, FinePDFs, and FineTranslations, the reference datasets a generation of open models trained on.

The duo have now left to do the same thing for robots with Macrodata. As a FineWeb fan, I’m excited to share that Air Street Capital led their $4M pre-seed alongside a group of angels from the leading AI labs, and today the company comes out of stealth.

Why now

Physical AI is the field’s next scaling story. Jensen Huang calls this “the ChatGPT moment for physical AI,” and the money has followed: robotics drew record venture funding in 2025, and 2026 is on track to dwarf it, with a cluster of companies building robot brains and bodies now carrying multi-billion-dollar valuations - Figure at around $39 billion, Skild around $14 billion, Physical Intelligence reportedly raising near $11 billion. The model side has caught up to the ambition, with vision-language-action models that fold perception and control into one system and world models that let a policy be tested in simulation before it touches hardware.

Every one of those companies needs the same thing to keep scaling: large amounts of well-prepared real-world data. Macrodata does not have to pick which of them wins, it refines the data all of them depend on. And that layer barely exists today. In autonomous driving, models trained on tens of thousands of hours of fleet data already generalize to cities they have never seen; most of robotics has no equivalent.

Physical data is messy in ways text never was: large video files, sensors sampling at different rates, actions and language interleaved, and a dozen incompatible formats with no agreed standard. Teams rebuild brittle scripts every time they swap a robot or a sensor.

What Refiner does

Macrodata’s first product, Refiner, is the tooling for that mess. It is an open-source Python library that reads the formats teams actually use - LeRobot, HDF5 (ALOHA, robomimic, LIBERO), Zarr, MCAP, raw video, Hugging Face datasets - and turns raw episodes into training-ready datasets. You compose a pipeline locally, inspect it in a data viewer built for multimodal data (you cannot cat a video in a terminal), then run the exact same code on managed cloud compute when it is time to process at scale.

Along the way it does the work that lifts policy quality: trimming idle motion, annotating subtasks, tracking what the hands did, and scoring trajectories with reward models, with VLMs run in the loop where a model needs to label or judge the data. You pay for the compute by the second. A pipeline that takes eight minutes on a laptop runs in under a minute on five H100s, for about $0.27.

The same craft, a harder problem

The throughline from FineWeb to Refiner is refinement: the unglamorous work of turning raw capture into the precise signal a model learns from. With text, that meant deduplicating and filtering trillions of tokens until what remained was worth training on. With robots, it means unifying formats, trimming, labeling subtasks, and keeping the demonstrations that teach while dropping the ones that do not. It is the same discipline applied to a harder, less mature, more valuable problem. The business mirrors it: an open-source core that becomes the default way teams handle robot data, and metered cloud compute they run it on.

We backed Guilherme and Hynek because they have done this before: the pair built the open data standard for LLMs. The industry knows that every strong model starts with great data. We think the next generation of robots will too.

Get started with Refiner

—

A week ago the United States government ordered Anthropic, the world’s most valuable AI start-up, to switch off its most capable model, Fable, for every foreign national on earth - whether they worked for Anthropic or not. This was not an export ban on a weapon sold to an adversary. It was an instruction to disable a commercial product, four days after its release, after officials acted on a claim - which Anthropic disputed as narrow and unproven - that its safeguards could be jailbroken to expose cyber-offence capabilities.

I have spent my career around this technology, first as a graduate student and for the past decade as an investor. In that time I have watched AI move from recommending films to driving cars, speaking with a human voice and editing the genome. I have also watched the debate about its risks settle on only half the question.

That debate is mostly about capability: how powerful these systems are becoming, and whether one might escape human control. Those are real questions. But they are not the only ones, and the Anthropic episode exposed the half we have neglected: access and control. The most advanced AI is built by a handful of American companies, on American soil, under American law, and what the rest of us are permitted to do with it can change on a Friday afternoon. The risk that matters today is not only that AI goes rogue, but that we do not control access to it at all.

Consider what “renting intelligence” now means in practice. A European hospital triaging scans, a bank screening fraud, a defence ministry planning for a conflict: increasingly each runs on an American AI system that’s governed by its export regime. A single directive in Washington cascades, instantly, through every institution wired to that model. We have built core economic and public infrastructure on a supply that a foreign government can switch off. And while there are open-source alternatives, they’re either Chinese or not at the frontier, and building European infrastructure on Chinese open weights trades one dependency for a thornier one.

And these systems are starting to improve themselves. As they do, AI stops being one industry among many and becomes the input to all the others - writing the code, running the research, designing the products and, increasingly, generating the growth itself. Once intelligence is the engine of an economy, a country without a frontier model of its own does not lose a sector; it loses control of the inputs to everything else, and the independence that depends on them. Worse, the gap compounds: capability that improves itself gets harder to chase with every month it runs ahead. This is not a race Europe can plan to enter in a decade. The window to be a builder rather than a buyer is measured in the time it takes to stand up a cluster, not a career.

This should sting, because Europeans invented much of modern AI. DeepMind was founded in London and sold to Google in 2014, and a great deal of the talent that followed now lives in California. Today Europe faces a company worth almost $1tn and American tech giants spending an estimated $450bn a year on AI infrastructure. Its answer has been the EU AI Act and a capital commitment that is a rounding error by comparison. A single American site, xAI’s Colossus in Memphis, runs more than half a million GPUs. Europe has nothing remotely at that scale. The instinct to govern this technology is right, but we’re off on the ambition by orders of magnitude.

It is fair to object that regulation is itself a form of power. But a rule book is not a substitute for the thing it governs. You cannot regulate, or be cut off from, an industry you do not have.

Europe’s instinct, when it is cut off, is not to build but to ask. We saw it within the week. The G7 convened in Évian and floated a “trusted partners” scheme to win back the access it had just lost, while Emmanuel Macron feted Donald Trump beneath the gilt of Versailles, the palace where France once helped midwife American independence. Two and a half centuries on, the dependency has reversed, and the posture is courtship.

None of this means Europe can match the American frontier dollar for dollar. On today’s capital it cannot, and pretending otherwise only wastes the little it has. But the goal is not parity, it is leverage. A country does not need the best model in the world to be sovereign; it needs a credible one of its own, on its own soil, good enough that being cut off is survivable rather than catastrophic. That is the difference between negotiating your access from dependence and negotiating it with an alternative in hand. The point is not to win the race. It is to make sure no one else can end it for you.

Sovereignty of that kind is something you build, and Europe has done it before. The Financial Conduct Authority’s regulatory sandbox, launched in 2016, let start-ups test products with real customers under supervision instead of waiting years for authorisation. The pro-innovation culture it signalled helped make London the fintech capital of Europe, home to Revolut, Wise and Monzo. Government should be AI’s most demanding early customer rather than writing rules for systems it has only ever imported.

Industry has to stop behaving like a tenant. Too many European companies rent the entire stack from American providers and build a thin product on top. That earns a margin and owns nothing: when the lab that supplies you decides to compete with you, or its government decides to cut you off, you have no ground to stand on. Where it counts, build and hold your own models and compute.

And our universities, which should be the source of all this, still work against it. I have argued in these pages before that Europe’s spinout system is broken, and it remains so. Too many institutions treat the companies their research creates as something to extract value from, rather than as the vehicle through which a discovery reaches the world. The best research should leave the building as a company, in addition to a paper.

We keep framing AI safety and AI ambition as a trade-off, as though a country must choose between governing this technology and building it. It is not a choice. The safest position is not the most heavily regulated one. It is the one where the model runs on your terms, in your jurisdiction, and no one on the far side of an ocean can reach over and turn it off. Right now that finger is not ours. Until it is, every other conversation about AI risk is one we are having on someone else’s permission.

I couldn’t think of anyone better to discuss this with than Ali Madani, the founder and CEO of Profluent. Profluent is an AI lab building frontier models to design proteins, with the goal of taking medicine from discovering molecules nature already made to designing the ones it didn’t. I first read Ali’s ProGen paper back in 2021, DMed him on then Twitter, and wrote the largest first check from Air Street Capital into the company at inception. Last month, Profluent announced a $2.25B deal with Eli Lilly, one of the largest to date between a frontier AI biology lab and big pharma.

We discuss the shift from discovery to design, why Profluent bet sequence-first while others went structure-first, the Lilly deal and large-scale DNA editing, fine-scale base editing, whether LLM-style scaling laws hold for proteins, and much more. You can either watch the interview in full here or on YouTube or read the transcript below.

Timestamp

Timestamp timeline

• 0:00 – Teaser: AI-designed molecules & the $2.25B Lilly deal

• 0:22 – Intros: Nathan Benaich (Air Street Capital) & Ali Madani (Profluent)

• 2:10 – What is Profluent, and why AI matters

• 5:45 – The landscape: readers vs. writers

• 7:45 – Profluent’s edge: 100B+ sequences and a wet lab

• 9:20 – OpenCRISPR and the exponential curve

• 12:55 – Why sequence beats structure

• 14:50 – The Eli Lilly deal and large gene insertion

• 16:20 – Fine-scale vs. large-scale editing

• 18:00 – Why it’s hard: the pre-AI era and the activity/specificity trade-off

• 20:45 – The Verve news, and scaling beyond one-offs

• 23:45 – Rare vs. common disease

• 26:10 – “What do you know that no one else does?”

• 27:40 – bio × AI is an undersaturated field

• 32:40 – When will a top-10 pharma be AI-first?

• 34:30 – Every molecule will be designed with AI

Transcript

Teaser (0:00)

Ali: We’re not a CRISPR company. We’re not a gene editing company. We’re an AI company. $2.25 billion is a great number, but what’s even more exciting than the number itself is the opportunity. This is an example of AI unlocking something, not just accelerating. Within the next two to three years, every single molecule will be designed with AI.

Intros: Nathan Benaich (Air Street Capital) & Ali Madani (Profluent) (0:20)

Nathan: Hey everybody, I’m Nathan Benaich, founder and General Partner of Air Street Capital, a venture capital firm that invests in AI-first companies in the US and Europe. Today I’m really excited to be joined by Ali Madani, founder and CEO of Profluent.

I first came across Ali through his paper ProGen, one of the first protein language models, which we’ll dig into today, back in 2021. After I saw that paper I DM’d him on Twitter and we started a discussion about the future of AI and biology. That led to me writing the biggest first check I’d ever done from Air Street, to be part of Ali’s first round at Profluent.

We’re now a couple of years into the journey. A lot has evolved, both for Profluent and the space overall, and it’s an exciting time, because we just signed a deal with Eli Lilly worth $2.25 billion to develop AI-designed gene editors for therapeutic use. It’s one of the largest deals of its kind in our field and a big moment for frontier AI applied to biology.

So we thought it’d be a good opportunity to take stock of the field: what Profluent is, our mission, why AI matters for protein engineering, and a bit about the Lilly deal that we can publicly discuss. We’ll also get into the difference between fine-scale and large-scale gene editing, the difference between bio models that are readers versus writers, some very recent news from Verve Therapeutics and Lilly, and a few of the recent model releases from other teams we respect. Ali’s one of the few people I know who can credibly speak both to AI as it applies to biology and to how we use it to advance human health.

Before we dive in, it’s worth taking stock of what Profluent is. Ali, can you give us the high-level pitch?

What is Profluent, and why AI matters (2:13)

Ali: Thanks, Nathan. At the highest level, Profluent is an AI lab, and we build foundation models to design proteins.

Nathan: What does that mean? How do we currently design proteins, and why does AI matter here?

Ali: To take one step back: proteins are molecular machines that power everything in human health, disease, sustainability, and the environment. A single protein like Keytruda can generate an incredible amount of revenue, over $30 billion annually. But the way we go about discovery today is really finding a needle in the haystack of nature, and intentional protein design is incredibly hard.

Nathan: So we’re coming from trial and error, and we want to move toward more thoughtful design dictated by the characteristics we actually want in the protein.

Ali: Yes. The mission we’re on is to make biology programmable. That means having levers you can control to design a molecule from scratch based on its intended function. That level of programmability is the moonshot for all of humanity.

Nathan: A quick sidebar: a lot of people are used to prompting ChatGPT or Claude to generate the outputs they want. How does programming a protein look different? How do we actually prompt these models to generate the sequences we care about?

Ali: Great question. Proteins can be represented as a sequence, and they are. A lot of biology is organized that way, whether it’s DNA as a sequence of nucleotides (A, T, C, G) or proteins as a sequence drawn from a standard vocabulary of 20 amino acids.

So we train language models, the same transformer architecture used for text, except the tokens aren’t words or subwords, they’re amino acids, the building blocks of proteins. We train with masked language modeling or next-token-prediction objectives.

Nathan: And what are the labels?

Ali: In the unsupervised setting, where we’re doing pre-training, we just use the sequence information itself. It’s similar to scraping the internet for human-generated text: you know a human wrote it and that it’s useful. For proteins, the analogy is that evolution selected these sequences under selective pressure, so you know a protein existed for a purpose and was functional. We can learn from that in an unsupervised way, then layer in other metadata, organism tags, taxonomic and environmental information, predicted or known structures, and eventually move to the post-training setting where you have actual laboratory measurements of function.

The landscape: readers vs. writers (5:45)

Nathan: It’s interesting to contrast Profluent’s approach, in-house data creation, in-house models, versus using open-source tools, and where the company sits on internal discovery versus partnerships, selling models versus selling drugs. Notably, how does it contrast with Isomorphic Labs, which just raised $2 billion and in many ways is our peer?

Ali: There are a lot of teams exploring biology with AI, and the broader AI-for-science landscape is incredibly exciting. Drug discovery is one of those immediate applications with a large market and large impact.

Isomorphic, for example, is a frontier AI lab born out of the structure-prediction era, from AlphaFold into their latest models. Those are broadly readers of biology that you can use for a variety of applications, and they’re also focused on small molecules and moving into antibodies.

Our heritage, which is important here, is born out of language models, quite literally the same architectures that enabled ChatGPT. Not through a tenuous analogy, but the same architectures behind the commercial interest in ChatGPT or Claude for programming. We use similar architectures and principles, but for proteins. One principle is to keep scaling data and parameters. Another is aligning these models on preference data we derive not just from human feedback but from laboratory feedback.

So one way to think about the field: there are credible players like Isomorphic working on readers of biology, versus writers of biology. We’ve been focused on the generative side, building language models in that paradigm.

Nathan: So we can segment by small molecule versus protein, reader versus writer, and sequence-first versus structure-first.

Ali: Exactly, and we’re firmly in the sequence-first paradigm. That doesn’t mean sequence-only; we can still layer in other information. But we want to capture the vast amount of sequence information available.

Profluent’s edge: 100B+ sequences and a wet lab (7:51)

Ali: We have over 100 billion protein sequences that we’ve curated to be high-fidelity and that we can train models on. That has incredible promise for representation learning, which ultimately lets us design proteins better.

Nathan: And curation means going out into nature scavenging for proteins in esoteric environments? Or is it more number-crunching across databases that aren’t particularly friendly?

Ali: All of the above. It starts out similar to Common Crawl, where the raw internet is available to everyone, but the question is who can actually curate that dataset effectively, access more raw sources, and understand the latent distributions within it. That’s ultimately what creates the winner, and we’ve spent a lot of effort there.

On the post-training side, we built a wet lab from day one. That was intentional. It lets us not just validate the proteins we’ve designed, but also generate supervised datasets, assay labels for a given sequence, that feed back into our models. That’s incredibly important for lifting the capabilities out of these models.

OpenCRISPR and the exponential curve (9:13)

Nathan: One of the cool things you did not long ago was capture a dataset specifically for CRISPRs, the gene-editing tools that transformed modern genetic medicine. Can you talk about what OpenCRISPR is, how you got the data, what it does?

Ali: OpenCRISPR was the first demonstration that we could use AI to edit the human genome, to generate molecules that bind to DNA precisely and execute the change you’re seeking, whether a double-stranded break or a precise base edit, an A-to-G edit, for example.

The contrast is with traditional drug discovery, where you pluck something from nature, from bacterial settings, and cram it into a human therapeutic application. Instead, we use a generative model to design a protein from scratch that doesn’t exist in nature and has the intended purpose a clinician or patient would use. It’s gotten a crazy amount of adoption; there’s a voracious appetite for it across industries.

Nathan: Can you give a sense of how hard this was? Is it a landmark moment, or one data point on a steady climb?

Ali: It’s a data point on an exponential curve. When I first started out, we trained our first model, a 1.2-billion-parameter model, the first ProGen model. It was actually the largest model in all the physical sciences at the time, and we didn’t even understand what it was doing or whether it was working.

So we quickly partnered with research labs at UCSF and with biopharma folks and asked a simple question: here are some generated samples, can you test whether these proteins are functional and useful? To our surprise, it worked. The de novo proteins generated by the model had incredibly high hit rates, and their functionality rivaled exemplar proteins that had millions of years of evolution to reach an optimal state. That was one data point on the trajectory.

Nathan: So AI can essentially accelerate evolution to find a much better peak.

Ali: It can ground itself in nature and evolution, then start interpolating and extrapolating. We started with simple monomeric proteins and moved into more complex settings. OpenCRISPR is the next logical step: proteins that are quite large, around 400 amino acids, with multiple domains, protein-protein interactions, large conformational changes, dynamics, and protein-nucleic-acid interactions, protein-guide-RNA and protein-DNA. That full system is truly a molecular machine.

It’s incredibly challenging to build that from first principles, atom by atom. The better approach is information-based: learn from existing examples, uncover the underlying biophysical principles, and generate something new from scratch. Going back to sequence versus structure, this is where sequence-based models really outshine structure-based approaches.

Why sequence beats structure (13:01)

Ali: In the peer review for our Nature paper, a reviewer asked us to baseline against structure-based approaches like ProteinMPNN. We found the language models really outperformed them; the structure-based approaches couldn’t perform at all, because of the complexity involved.

Nathan: What’s the intuition for why structure is less powerful than sequence?

Ali: Because function is complex, and function is what everyone ultimately cares about, whether it’s a patient for therapeutics, a farmer for agriculture, or a consumer for protein-based products. Capturing function can involve many concepts, including dynamics, so it’s not just one structural state. Capturing that sequence-to-function relationship is the most important thing, and we build all of our infrastructure with that in mind.

Nathan: So the problem is that structure captures a protein in one conformation, but it can assume many?

Ali: Yes, these proteins are very dynamic. Structure freezes them in one state, but there are many states they could be in. Sequence is more flexible because everything ultimately arises from sequence, so you get more diversity. Think of disordered proteins or loop-like proteins that have many states and no set conformation, or multi-state proteins. We can bake all of those priors into the model, building toward a broader concept of fitness.

The Eli Lilly deal and large gene insertion (14:53)

Nathan: This led, among other things, to the big Eli Lilly deal, $2.25 billion in milestones, which is pretty epic. Walk us through what the deal means, how it came about, and the plan.

Ali: The number is great, $2.25 billion is a great number, but what’s even more exciting is the opportunity. This is an example of AI unlocking something, not just accelerating.

AI is going to be transformative across many aspects of drug discovery. The easiest value proposition is AI as an accelerant: compressing timelines, making things more efficient. That’s great, and we operate there too and provide value for our partners. But what really excites us is finding unlocks, problems you could not have solved before AI. The specific problem we’re working on with Lilly is large gene insertion: inserting large genes into the genome.

Nathan: What qualifies as a large gene, and what’s the difference with base editing and prime editing? A lot of buzzwords. Can you unpack them?

Ali: We think about two types of effort within gene editing: fine-scale and large-scale.

Fine-scale vs. large-scale editing (16:19)

Ali: Fine-scale editing is like genetic scalpels, the genetic-surgery model, where you perform precise edits of the human genome.

Large-scale editing, which is the subject of the Lilly deal, is the idea of inserting whole kilobase genetic payloads into the human genome. The main challenges are doing that efficiently and effectively, and then specificity. We have examples of proteins called recombinases, like BxB1, that are widely used, but they may not be specific or work well in human cellular contexts. So we have proof points in nature that it’s possible; the grand challenge AI can enable is making it programmatic and controllable.

Nathan: So nature has shown it’s technically possible to snip and stitch large pieces of DNA, potentially an entire gene cassette, in organisms with shorter genes. Now the task is to make it work in human cells?

Ali: Less about bending and molding it, and more about learning the underlying principles of why it occurs, then building it from scratch with AI models.

Why it’s hard: the pre-AI era and the activity/specificity trade-off (18:02)

Ali: Going to your point about base editing, prime editing, and other forms of gene editing: all of those are a pre-AI-era approach of taking something from nature and cobbling it together. It’s worked remarkably well, but it’s the way drug discovery has always operated, find a needle in a haystack, perform random mutagenesis, screen, and hope to find a winner.

Nathan: So this kilobase editing isn’t going to be solved by finding the enzymes that do this in bacteria and then fine-tuning a model to adapt them to human DNA?

Ali: It uses evolutionary information as examples of what has worked, and through that you learn the underlying grammar, similar to what we built on the foundation-model side. The way humans learn to write the next great American novel is by reading other novels, understanding what makes a great one, and then writing our own, as opposed to grammatically copying and pasting from existing novels.

Nathan: Have there been attempts at large gene insertion before, and why have they fallen short?

Ali: There have been attempts. What we find is a big trade-off between activity and specificity. A lot of protein problems have these trade-offs, where you want to optimize multiple properties at once and optimizing even one is hard.

Nathan: So you can either make the scissor really precise about where it snips, or efficient at doing the snip, but not both.

Ali: Exactly, that’s one trade-off we see in recombinases. Navigating that multi-attribute optimization is very difficult.

The Verve news, and scaling beyond one-offs (20:26)

Nathan: There was huge news from Verve, which was also working with Lilly, Lilly acquired the company not long ago. It involved a specific gene tied to high cholesterol and heart disease. They used a base-editing technique in vivo, inside the body, in a handful of patients, and it looks like those patients had their mutation changed and are pretty healthy. Walk us through what this means. Is it as exciting as it sounds, or are there caveats the Twitterverse is missing?

Ali: Even if there are caveats, it’s point-blank insane, in the best way. We should cheerlead these efforts as much as possible, because they can be transformative.

The question we ask is: how do we scale that? How do we make it not a one-off, but use AI to build an engine that enables more and more of these therapeutics, molecules that can become blockbusters down the line?

Nathan: So instead of building tools specific to PCSK9, you could swap in any gene you care about and have off-the-shelf editors, then find patients with those monogenic or more complex diseases and run the same in vivo motion.

Ali: To make it concrete: at ASGCT, the American Society of Gene and Cell Therapy, we announced our ability to 10x the number of mutations and variants we can go after versus state-of-the-art SpCas9-based approaches for base editing. That expands the addressable market, the number of patients and variants you can target. That’s a concrete, non-incremental 10x that AI can deliver, with direct implications for that Verve announcement.

Nathan: Is it a drop-in replacement for what Verve did? Can we now say, if it worked for PCSK9, here’s a 10x version?

Ali: Essentially yes. There’s a payload side to gene editing, and this lets you swap in a different gene editor that’s known to work well, both in silico and validated in experiments, for sites of interest beyond the PCSK9 gene.

Nathan: What about the critique that this only worked on a couple of patients? What should we read into that?

Ali: I’d argue it’s amazing it worked for a couple of patients at all, and let’s see what happens going forward. The ability to have these one-off cures is mind-boggling, that we can go beyond treating disease and symptoms.

Rare vs. common disease (23:52)

Ali: We can go beyond taking pills once a day and worrying about adherence, and instead have one solution, very early on, that can prevent heart disease. That’s incredibly bold, and consistent with the bold bets Lilly is making in obesity and other diseases. And this wasn’t a random shot in the dark; they’ve had successes all along the way, and I’d extrapolate the trajectory beyond this moment.

Nathan: Is it more or less impressive than the curing of baby KJ about a year ago, who had a genetic defect and was treated with a gene editor?

Ali: They’re different use cases. Disease comes in many shapes and forms, rare and common. The baby KJ story is a life-threatening, extreme-need setting: without a liver transplant or some therapy, you die very young. That’s an incredibly powerful use case, because it affects the young, it’s clear death, and there are no other solutions. It’s another form of disease we can tackle with the same underlying technology that AI can scale.

Nathan: What a time to be alive that we have these capabilities.

What do you know that no one else does? (25:52)

Nathan: To tie this together: every drug ever made started in nature, or has been screened to the ends of the earth in pharma. Now we’re inverting that whole system, from discovery to de novo design. What comes next, and where might the field go? You have insight into one of the most exciting companies out there, so, what do you know that the rest of the world doesn’t?

Ali: We need to keep scaling these models. We’re still in early days. If I put it in GPT eras, I feel like we’re in the GPT-1.5 era of the field as a whole, and I want to get us to GPT-3, GPT-4, GPT-5 as soon as possible. I’m impatient to bring the future forward.

That’s not just data scaling, but thinking deeply about inference-time scaling, new model architectures, and incorporating other data. And even though we’re early, it’s pretty incredible that it’s already useful. You can see that with our Lilly deal and across many applications; even the early versions have real utility, and people are willing to bet on them.

Nathan: That might be a big difference from natural-language LLMs, where GPT-1 and GPT-2 were kind of useless economically, entertaining, maybe. Here, companies are staking billions because it already works.

bio × AI is an undersaturated field (27:40)

Nathan: So either this is a domain with lower-hanging fruit, because the industry is more nascent in adopting advanced computation and AI, or AI is a uniquely good interpreter for biology, where any interpretation of what we already have uncovers biologically useful nuggets.

Ali: A connected question we ask internally is: what if this is it? What if you stumble across gold and that was the only application? That’s probably the most unreasonable take; it’s reasonable to expect much more. We see a clear line of sight to unlocking more targets we couldn’t go after before, and many problems that are well-bounded from a science perspective, where the risk is reduced to scientific risk. As scientists, we love those problems, and we feel we’re the best to tackle them.

Nathan: If you had to estimate, how many people work at the frontier of AI and biology versus the frontier of AI generally? What are the relative numbers?

Ali: At least a thousandth, both in compute budget and economic spend, and in number of people. And biology is no less complex than text, and no less impactful, I’d argue more so. We’re totally undersaturated here.

Nathan: But it seems more intimidating for people outside the industry, who think, “I don’t know anything useful about biology, how could my machine-learning skills apply?” Do you have a counter, to build a bigger magnet and pull more people in?

Ali: The proof is in the pudding. We have people who’d never done anything with biology, who built NLP models, and within weeks they sense the usefulness of what they bring. There’s no such thing as a 20-year veteran in using transformer models for protein design; this latest version of AI for biology is new. The intersection of people who can speak both AI, NLP or computer vision, and biology is small, but we see the proof points: you can learn this quickly and provide real value.

Nathan: Can you give a couple of examples of the backgrounds of people who’ve joined and been at the forefront of these papers?

Ali: We have three main pillars at Profluent. The first is machine learning: people from big tech, NLP, computer vision, RL, and computational biophysics backgrounds. The second is data: world-class bioinformatics people who curate the vast datasets, over 100 billion proteins and over 20 trillion tokens, for both pre-training and post-training.

Nathan: So they have taste for the data.

Ali: Absolutely, and taste matters even more in biology, because we don’t natively read and write that language, we don’t speak protein. The bioinformatics element is huge. The third, equal pillar is experimental biology: people from pharma and biotech who understand the domains. And maybe a fourth pillar is our partners, who understand their specific problems and want to take this forward. It takes a village; we humbly go after a central piece of the problem, but advancing it through clinical trials requires partners.

When will a top-10 pharma be AI-first? (32:51)

Nathan: How many years until one of the top-10 biopharma companies is a truly AI-first company like ours?

Ali: I think it’ll come through partnerships. We do what we do best, and partners tell me this directly: they recognize that building the frontier model is what we do best, and they have specific datasets, use cases, and expertise that are complementary. It’s not unfamiliar to pharma, which has long had a symbiotic relationship with biotech, where innovation comes in the form of molecules. There’ll be a similar complement between frontier AI companies, Profluent, Isomorphic, and others, and pharma. That recognition has already happened and seems to have accelerated in the last six months.

Nathan: It’s crazy how fast it’s happened.

Ali: When I trained the first ProGen models and handed sequences to people, the first question was, “Who are you, and what is this alien artifact?” Now the conversation has completely accelerated, and that speed is unprecedented for such large industries.

Every molecule will be designed with AI (34:37)

Ali: On adoption: the way I see it, every drug, every molecule that’s designed is going to use AI, not just AlphaFold for understanding structure, but AI to generate and write the molecule. And not just a percentage; within the next two to three years, every single molecule will be designed with AI.

Nathan: And it goes further into the process, clinical trials, figuring out which patients to enroll, how to monitor response. All of those tasks get fundamentally transformed by AI, especially once big pharma starts treating AI and software as a core part of the product offering rather than just an enabler. Like the shift in financial services, where technology went from “not it” to the product itself. So, if Profluent pulls off its mission, and hopefully the mission keeps expanding, what would that world look like?

Ali: People talk about abundance, and I really believe that, I say it with a straight face. There’s an abundance of problems we can go after and solve with AI, and so many targets we can prosecute. It’s not just compressing timelines or making things more efficient; it’s unlocking new and emergent capabilities from scaling these models, which creates new value.

I’m incredibly bullish, and I say that as a scientist. Profluent wasn’t “let’s start a startup and then figure out the idea.” This was the subject of my research before the company. So I speak from the ground level as a practitioner.

We built foundation models for proteins. We’re not a CRISPR company, we’re not a gene editing company, we’re an AI company for protein design. But the gene editing application is concrete and ambitious, and there’s a future we can point to that motivates us and our partners: imagine a child born with a mutation in their DNA, a genetic disease that, untreated, leads to a life of pain and suffering for them and their family. With our AI, we can design molecules from scratch to correct that disease before it takes hold. That’s an incredibly powerful future, and it’s going to change everything.

Nathan: Well, I wish you all the best of success.

Ali: We’ll work on this as hard as we can, and there are many new announcements to come in the forthcoming months, so stay tuned.

Nathan: Hopefully we’ll check in before long and get a temperature check on where you think we are on this exponential curve toward abundance. With that, thank you so much, Ali, and thanks everybody for listening.

Ali: Thanks, Nathan.

Every generation of defense risks building the shield it wishes it had for the last war. France poured concrete into the Maginot Line to stop the invasion it remembered from the First World War. Israel built the Iron Dome to stop barrages of rockets. Both were serious engineering achievements built for threats that moved more slowly than institutions.

That world is gone.

Russia’s war in Ukraine, and Iran’s missile-and-drone onslaught against the UAE and the wider GCC, have exposed the new arithmetic of air defense. Cheap drones, ballistic missiles, cruise missiles, glide bombs, electronic warfare, and massed salvos have changed both the economics and the tempo of the fight. When a cheap drone draws a million-dollar interceptor, the defender may win the intercept and still lose the campaign.

The next shield has to be affordable enough to fire at scale and robust enough to work under jamming. It cannot be static. It must instead evolve with the threat.

That is why Air Street has led the $60M Series A in Alta Ares.

From argument to action

When I wrote in the Financial Times in 2023 that European governments needed to take defense innovation seriously, I meant it as a challenge to governments and the venture industry. Europe had the capital, talent, and technical ambition to build the technologies that safeguard democracy, security, and our way of life. Too often, it chose easier markets, while procurement systems rewarded incumbents built for a slower era.

Since then, the argument has become harder to dismiss. For the last year, I have been looking for the company that could make it real in European air defense: not “AI for defense” slideware, and not a controlled-range demo, but a sovereign, operationally grounded, full-stack company built around a feedback loop from the field.

Alta Ares is that company.

The new air defense stack

Alta Ares is building full-stack, integrated air defense across the entire kill chain: AI-first software, sensors, command-and-control, and effectors built to operate in contested environments against a range of aerial threats.

The company began with software for intelligence, surveillance, and reconnaissance video analysis. Work alongside operators in Ukraine pulled it into the broader air-defense problem: seeing a target, maintaining track, supporting operator decisions, guiding an interceptor, and integrating the result into a system that can actually be fielded.

In many AI applications, failure means a bad answer or a model that needs retraining. In air defense, the object is small, fast, cheap, and often deliberately hard to see. The operator may be tired, cold, under jamming, and making decisions in seconds. A useful AI system cannot live as an analyst beside the workflow. It has to become part of the kill chain itself.

Alta Ares’s products reflect that architecture. Pixel Lock provides onboard computer vision for detection, tracking, and terminal guidance while preserving human control over engagement. Ukrainian drone pilots are already hitting Russian targets from 500km away and as the Financial Times reported, Alta Ares’s terminal guidance software helps interceptors detect and close on Russian drones in the final phase of flight.

Gamma supports autonomous guidance and ISR workflows. X-Lock and Black Bird are both used in the field: X-Lock against short-range one-way attack drone threats, including Shahed-type systems, and Black Bird against faster aerial threats, including cruise missiles and glide bombs.

This isn’t about software grafted onto hardware. Alta Ares develops models, avionics, guidance, operator workflows, and manufacturing around the operational problems faced by the warfighter.

Born from the battlefield

Alta Ares’s most important asset is its feedback loop with the warfighter.

That loop comes from constant in-field deployment. Alta Ares has been active in Ukraine for years and interceptors equipped with Pixel Lock began shooting down Shahed-type drones in 2025. The company has since demonstrated systems with NATO, tested Black Bird in arctic conditions with the Estonian Defense Forces, and is deployed across multiple operational theaters. In rapid succession, Alta Ares has signed large contracts from half a dozen countries across Europe, the Middle East and Asia.

Those milestones matter, but the deeper point is what they make possible. Simulation is useful. Range tests are useful. Neither exposes systems to the full mess of real conflict: electronic warfare, bad weather, changing drone signatures, damaged equipment, uneven training, and the pressure of live operations.

Alta Ares’s data advantage is not a static dataset sitting on a server. It is a relationship with the field: operators use the system, the system sees where it fails, engineers recover the evidence, and the product changes. That is how AI systems become robust in high-consequence environments: not by claiming perfection, but by shortening the distance between failure and adaptation.

That is what the old defense procurement model cannot absorb. As we argued in Bringing Dynamism to European Defense, Europe’s problem has never been talent alone: it is a financial, political, and institutional climate that fails to reward mission-driven defense entrepreneurship. A decade-long program assumes the threat profile will stand still long enough for the program to arrive. In autonomous warfare, that assumption is fatal. The measure of a system is not only how well it works on day one, but how quickly it can be improved for day two.

Why this team

Hadrien Canter’s path into defense technology is not the standard prime-contractor biography. He has deep links with Ukraine for years, and Alta Ares does not feel like a lab looking for a battlefield use case. It feels like a company working backward from the operator: what they can see, what they cannot, how quickly the threat changes, and what a system has to do in the thick of it.

You feel it in the office: the energy, mission, urgency, and seriousness of people building for a live war.

Together with his co-founder Stanislas Walch, Hadrien has recruited software, hardware, government relations and sales talent from Anduril, Helsing, Palantir, Safran, Thales, MBDA, Embraer, and the French Army. They’ve also built an advisory board that includes Philippe Lavigne, former Chief of Staff of the French Air and Space Force and former NATO Supreme Allied Commander Transformation, and General Corentin Lancrenon, a three-star French Army general.

For Air Street, all of these features are critical. We back AI-first companies, but “AI-first” should not mean model-first in isolation. In the physical world, AI advantage usually belongs to the team that owns the loop from data to deployment. Alta Ares has that loop.

Europe must build

Europe now broadly agrees that it needs more defense capability. As I wrote in A letter from the Munich Security Conference, the harder transition is from crisis buying to permanent capacity. The question is whether Europe will build the new capabilities itself, or simply allocate larger budgets to imported systems and slower incumbents. e.

No single company will be Europe’s entire Iron Dome for autonomous air defense. A real adaptive shield will take an ecosystem: sensors, decision systems, interceptors, electronic warfare, command systems, procurement reform, and operators who can field the technology. But every ecosystem needs a company that shows the way.

Europe has written enough policy documents about waking up. It needs companies that jolt us into action.

We believe Alta Ares is that company.

Hadrien is co-founder and CEO of Alta Ares, an AI-first air defense company building an integrated platform for detection, identification, tracking, and interception. Founded in 2024, it works on one of the most demanding problems in applied AI: defending against mass-produced attack drones, cruise missiles, and glide bombs in contested environments, where a system has to perform in seconds, at the edge, and under operational pressure.

Alta Ares began as a software company focused on intelligence, surveillance, and reconnaissance (ISR) video analysis. The feedback loop from Ukraine pushed it into a wider air defense architecture spanning data-fusion software, edge AI, and hardware effectors built to operate from Arctic to desert conditions. Its stack includes Pixel Lock for embedded detection, tracking, and terminal guidance; Gamma for autonomous interceptor guidance; X-Lock for short-range drone interception; and Black Bird for faster aerial threats.

Apply to join RAAIS 2026

Why the air defense gap is so large

Recent salvos over Eastern Europe and the Middle East have exposed a hard truth: legacy air defense systems built to stop fast jets are losing the economics against mass-produced aerial threats. NATO partners increasingly face coordinated waves of one-way attack UAVs paired with cruise missiles and glide bombs. This is an enormous problem that is defining capability gaps in modern defense.

Unlike many AI applications, air defense is not forgiving. The enemy object is small, fast, and often deliberately cheap. The operator may be tired, cold, and working at night. The environment may be jammed. Protecting people, critical infrastructure, and military assets now demands systems engineered from day one for autonomy, modularity, interoperability, and unit-cost discipline.

That is what makes counter-UAS such an important test case for applied AI. There are many hard parts to the problem: recognizing an object in a poor quality video feed, fusing sensor inputs, holding a track, guiding an interceptor, preserving human control over the final engagement decision, and doing all of it inside a system that can be carried, deployed, and iterated quickly. Pixel Lock is Alta Ares’ answer: onboard computer vision that detects, classifies, and tracks targets in real time and supports autonomous terminal guidance while keeping the operator in the loop. Here the AI sits inside the control chain itself, guiding the interceptor rather than only flagging a target for an operator to act on.

The Ukraine feedback loop

Alta Ares’ development is shaped by proximity to the battlefield. Interceptors running Pixel Lock began shooting down Shahed-type drones in November 2025. Hadrien’s public interviews describe an engineering culture built around fast field feedback: simulation helps, but the front line reveals failure modes a lab cannot.

That loop matters because drone warfare is changing faster than long procurement cycles and static product roadmaps can absorb. Threats adapt, operators adapt, and countermeasures adapt in turn. The companies that make a difference in this category are the ones that can move from deployment to model improvement to hardware iteration without treating each step as a separate world.

From software to systems

Alta Ares’ public milestones track a company moving from software into an integrated air defense architecture. In March 2025, NATO Allied Command Transformation named Team Alta Ares the winner of its 15th Innovation Challenge for an “Embedded AI for Recognition, Detection, and Identification” submission focused on glide bombs - a system that detects, identifies, and predicts the trajectory of these low-cost guided munitions from visual and acoustic data.

Later in 2025, Alta Ares demonstrated its drone-interception system to NATO at the DGA missile test site in Biscarrosse. The company calls the configuration a Tactical Protection Dome: radars, interceptor drones, data fusion, and Pixel Lock software.

The most recent milestone came in Estonia. Early in 2026, working with the Estonian Defense Forces and Ukrainian partners, Alta Ares tested Black Bird, its turbojet-powered interceptor, in Arctic conditions. The company reported three consecutive flights, ground temperatures of -17 degrees Celsius and -25 degrees Celsius at altitude, and a top recorded speed of 450 km/h. The trial also validated the less cinematic but more important parts of the system: communication links, antenna performance, live video transmission, and Pixel Lock target detection, tracking, and locking. In parallel, the company has begun mass-producing interceptor drones in France.

Why it matters for RAAIS

The next generation of air defense is being built as layered systems: sensors, command and control, autonomy, and low-cost effectors combined quickly enough to keep pace with changing threats. NATO’s own 2026 work on layered counter-UAS points the same way, treating the challenge as one of integrating sensors, effectors, electronic warfare, command systems, and battlefield lessons into something coherent. Alta Ares is one version of that thesis, built from the edge inward: European, field-informed, and aimed at a class of threats that has already changed the character of modern conflict.

For RAAIS, the interest goes beyond defense. Alta Ares is a working case study in applied AI inside a live operational system, where robustness, cost, latency, and human judgment all bind at once. The same problem shows up across robotics, autonomy, and other high-consequence settings, where a model that performs on a benchmark still has to keep working once it meets conditions that shift under it.

Hadrien’s background

Hadrien’s path into defense technology is unusual. Before Alta Ares, his public background spanned law, Ukraine, and operational fieldwork rather than a conventional defense prime career. He studied at the University of Paris 1 Panthéon-Sorbonne, qualified with the Paris Bar, served as an OSCE international observer around Mariupol in 2019, and worked on humanitarian projects in Eastern Ukraine.

That background shows in the company he has built. Alta Ares designs from the operational problem backward: what the operator sees, what they miss under stress, how fast the threat changes, and what kind of AI stack survives that reality.

Apply to join RAAIS 2026

The harder half of voice AI?

ElevenLabs built its name on synthetic voices that made generated speech sound natural, expressive, and controllable. But the reverse problem also exists: turning messy, real-world speech back into accurate text. For voice agents it is often the part that decides whether the product works in the ears of the human user.

A live agent cannot reason about what it has not heard. It needs a transcript that is fast enough to preserve conversational flow, accurate enough to carry names, numbers, technical terms, and intent, and robust enough to handle accents, background noise, interruptions, and people switching languages mid-sentence. Speech-to-text is a key perception layer for interactive AI systems.

Angelos’ work at ElevenLabs focuses on model quality, inference design, latency budgets, and production reliability.

Two Scribes for two production regimes

Angelos has worked across both of ElevenLabs’ latest transcription models: Scribe v2 and Scribe v2 Realtime.

Scribe v2, launched in January 2026, is optimised for high-accuracy transcription of long and complex recordings: batch transcription, subtitling, captioning, media libraries, training material, compliance workflows, and research audio. These are settings where the model can use broader context, but where errors compound quickly. A missed drug name, a malformed account number, or a confused speaker label can make the downstream transcript much less useful. ElevenLabs built Scribe v2 with production transcription features such as keyterm prompting, entity detection across 56 categories, smart multi-language transcription, speaker diarisation, word-level timestamps, and audio tagging.

On Artificial Analysis’s AA-WER v2.0 benchmark, which combines a held-out voice-agent dataset with cleaned public datasets for parliamentary speech and earnings calls, Scribe v2 led the overall ranking with a 2.3% word error rate. It also led two of the three component datasets, including AA-AgentTalk and Earnings22-Cleaned-AA. That is a useful reminder that “accuracy” is not one thing: the model has to work across short agent-directed speech, formal speech, and long business audio, not just a clean public benchmark.

Scribe v2 Realtime, released in November 2025, solves the same problem under a much tighter constraint. It is built for live agents, meeting assistants, captioning, and conversational interfaces where a transcript that arrives too late is almost as bad as a wrong one. ElevenLabs describes it as delivering live transcription at around 150 milliseconds of latency across more than 90 languages, with features such as automatic language detection, voice activity detection, manual commit control, text conditioning, and predictive transcription for the next words and punctuation. On FLEURS, a multilingual benchmark spanning 30 languages, ElevenLabs reports the lowest word error rate of any low-latency ASR model.

Why latency changes the shape of the problem

For most of the last decade, speech-to-text progress was mainly discussed through benchmark word error rate. That number still matters, but it no longer captures the whole product problem. A transcription model that is accurate after the fact can be excellent for subtitles and useless for a live agent. A real-time model that is fast but unstable can make the agent interrupt, hallucinate intent, or miss the moment to respond.

This is why Scribe v2 and Scribe v2 Realtime are better understood as two parts of the same system-level push rather than a single leaderboard entry. The batch model pushes for the cleanest possible transcript when full context is available. The real-time model asks how much of that accuracy can survive when the system has to stream partial understanding under a human conversational latency budget. In one case the challenge is depth of context. In the other it is speed without collapse.

For RAAIS, that makes Angelos’s work a particularly good example of applied AI becoming harder as it becomes useful. Offline model quality is only the beginning. The real question is whether a research result can be made fast, stable, observable, and cheap enough to sit inside millions of interactions where people do not care about the benchmark. They care whether the agent heard them.

Apply to RAAIS 2026

Angelos’s background

Angelos’s path into speech-to-text runs through systems work, which is part of what makes it interesting. He studied Software Engineering at the University of Glasgow, graduating with First Class Honours in 2020. His master’s project developed a reinforcement-learning-based scheduler for IoT networks, and before ElevenLabs he worked across cloud-native infrastructure and reliability roles at Skyscanner, Ondat, and Beacon Platform. He also contributed to Gentoo’s Portage package manager through Google Summer of Code.

The audio thread appears early. In 2017, his team won the Amazon challenge at the Glasgow University hackathon with Emotionify, an app that combined facial recognition, text-to-speech, and the Spotify API to match music to a user’s mood. He later won the Goldman Sachs and IBM challenges at subsequent Glasgow hackathons, with projects involving speech recognition, text-to-speech, and custom machine-learning models.

He also keeps teaching the fundamentals. At AI Engineer Europe 2026, Angelos ran a workshop called Training an LLM from Scratch, Locally, walking engineers through the practical components of building a small language model on local hardware. That instinct - to understand the whole stack from first principles, then make it work in production - is exactly the one needed for speech-to-text now. Voice AI will not be judged by whether it can speak beautifully in a demo. It will be judged by whether it can listen accurately enough to be trusted.

Short bio

Angelos Perivolaropoulos leads research engineering for speech-to-text at ElevenLabs, where he works across Scribe v2 and Scribe v2 Realtime, the company’s high-accuracy batch transcription and low-latency streaming transcription models. His work sits at the intersection of model development, inference, and production reliability. He studied Software Engineering at the University of Glasgow, graduated with First Class Honours, and previously worked across cloud-native infrastructure and reliability roles at Skyscanner, Ondat, and Beacon Platform.

In this piece, I’ll share why a system that connects research, development and clinical decision-making across the full lifecycle of a drug is critical, and why Perceptic is the company to do it.

Can AI accelerate drug discovery?

Drug discovery is expensive, slow and low-hit-rate. We all know this. What is new is that we now live in an era of AI systems that can consume and reason over the multi-modal evidence drug development actually generates - targets, chemistry, biology, clinical, commercial - and across the sprawl of databases, dashboards, notebooks, departments and geographies it lives in.

Look at where the frontier labs are putting their attention. Anthropic recently added Vas Narasimhan, the CEO of Novartis, to its board, and is openly building Claude into a partner for scientific work. OpenAI’s reasoning model just disproved an 80-year-old Erdős conjecture in discrete geometry, verified by external mathematicians on a problem the field had not cracked since 1946. Science - and biology especially - is the pinnacle AI use case, and the labs are not being subtle about wanting it.

A few months ago I wrote Can AI discover new science?, arguing that AI is now an accelerator of discovery, not just an automator of existing work. But drug development is one of the highest-stakes environments that does not come with clean verifiers: no leaderboard, no public benchmark, no “the model proved the lemma” moment. The contributions only count when they are wired into the workflows where billion-dollar, multi-year decisions get made, with every conclusion traceable back to the evidence that produced it. The labs have given us the substrate: frontier models. Pharma now needs the operating system, and this is where Perceptic comes into play.

The Perceptic operating system

Perceptic is the intelligence layer that connects data, decisions and context across the drug lifecycle, so every insight compounds and every decision is made with the full picture. Three AI applications run on one shared architecture:

• Scout triages external assets, including licensing candidates, competitive programs, pipelines, against the customer’s evolving strategy. Evaluation time has gone from a week to an hour in production while asset screening from hundreds per week to thousands in minutes.

• PercepticOS is the intelligence layer above the customer’s internal tools and data. It is where scientific teams test hypotheses, compare internal evidence against external benchmarks, and build a knowledge base that doesn’t restart with every new project.

• Atlas is the clinical data foundation that recapitulates internal and external trial data, providing the substrate everything else stands on. Live deployments have produced a 50-fold increase in clinical data extractions.

Perceptic in practice

A pharma company evaluating a new therapeutic area starts inside PercepticOS, pressure-testing a hypothesis against the evidence base. That triggers Scout to sweep external assets and rank them against the evolving thesis, in minutes instead of weeks. Candidates feed back into PercepticOS with full context, where Atlas surfaces the trial history, benchmarks and endpoint precedents that determine which assets are tractable.

A customer doesn’t buy three products. They deploy AI workers that learn their organization, their tools, their data, their decisions, and the instance gets more valuable every month. Perceptic follows the drug, not the department.

The perfect team, forged at Palantir

Air Street Capital has invested in many AI-first techbio companies including Profluent, Allcyte (acquired by Exscientia), Valence Discovery (acquired by Recursion), all of which are in the business of discovering novel drugs. Perceptic is the opposite bet: the AI-first software that makes drug developers themselves better and faster.

As an investor I am generally of the view that the money is in the drug, not the tools. Perceptic is the exception. The moment has come to bet on the AI-first software layer, and the right team has shown up to build it.

Two ingredients matter, and they are rarely found together. The first is the DNA of shipping production AI into the hardest enterprise environments. Tilman, Martin and Zaki were core contributors to Palantir’s AIP, the suite designed to securely connect frontier AI with an organization’s internal data and operations. They are operators who learned over a decade what it takes to put production AI inside regulated, data-sensitive, multi-stakeholder enterprises. They know where deployments break, and what the six-month security reviews actually ask for.

The second is deep knowledge of pharma workflows themselves. Frontier models are spiky in their capabilities, and the spikes only line up with value when you graft them onto the proprietary workflow knowledge of the end user. The team’s years inside Palantir’s Life Sciences practice mean they understand the nitty-gritty of bending increasingly capable frontier AI systems into the shape pharma R&D actually requires. That is not something you can buy on the open market.

CSL and multiple top-20 pharma companies that we cannot name publicly trusted that thesis enough to deploy Perceptic before the company came out of stealth - the highest-signal proof point any seed-stage company can have.

What comes next

If Perceptic is right, drug development moves from a 15-year linear bench-to-bedside process to one that runs on always-on infrastructure, where every insight from every team is wired into every subsequent decision. The handoffs stop being where information dies: they become where it accelerates.

That is the category that is forming around Perceptic.

Congratulations to Tilman, Martin, Zaki and the whole Perceptic team. We’re proud to be on the journey.

- Nathan

Join the team!

From clip-generators to simulators

For most of the last three years, generative video has meant clip-makers: models that take a prompt and render a fixed-length, fixed-trajectory output video. Veo, Sora, Kling, and their successors have made the visual fidelity of generated video remarkable. They are also fundamentally offline systems. Once generation begins, the future of the clip is locked.

World models are a different flavor of system. First, they predict the next state of an environment given the past, conditioned on what a participant - such as a person, an agent, a robot - does next. Second, they accept streaming input mid-rollout, and the world responds. Third, they hold a persistent, manipulable state, which makes the model a simulator rather than a render.

So what’s changed more recently? Causal video distillation matured (CausVid, Self-Forcing), and bidirectional joint audio-visual foundation models arrived (Ovi, Veo 3). Odyssey has been heads-down expanding these threads into something interactive.

Two new frontier models

Starchild-1 jointly generates audio and video autoregressively at up to 24 fps, while continuously responding to streaming text, speech, or action input. Odyssey frames the case for sound through Aquinas: “nothing is in the intellect that was not first in the senses”. Pretending the world is silent leaves a large amount of signal - physics, dynamics, intent, emotion - out of the model. Audio and video also evolve at very different temporal resolutions, and small errors in either modality compound during long-horizon rollout.

Starchild-1’s answer is a causal distillation pipeline that adapts Ovi, a bidirectional audio-video foundation model, into a real-time autoregressive one, plus an asynchronous KV-cache architecture that lets the two modalities run on their own clocks without losing synchronization. A single model supports four interaction regimes: interactive world exploration, scripted dialogue control, conversational interaction, and narrator-style companionship. The team is candid about what’s left: scene and acoustic identity still drift over long horizons, and quantitative benchmarks for interactive causal audio-video generation don’t yet exist.

Agora-1 matches up to four players into a shared deathmatch - built on GoldenEye, a game many on the Odyssey team grew up playing (as did I) - and every frame each player sees is generated by Agora-1 while the model maintains a shared world state across all participants.

Prior multi-agent work has tried three paths. Multiverse concatenated player views into a single “split-screen” world. Solaris stacked agents along the sequence dimension of one autoregressive transformer - more robust, but context grows with the number of players, so the approach doesn’t scale linearly. MultiGen maintains an explicit shared world state but doesn’t separate simulation from rendering. Agora-1 decouples the two and learns each as a separate function. One model evolves a shared, manipulable world state from player actions; a second renders consistent views of that state from independent viewpoints. The closest analogue is a modern game engine, only with both halves of the engine learned from data rather than hand-authored.

Learning through discovery

PROWL, released last week, gives world models a way to find their own failure modes and generate training data from them. None of these are products. They are the substrate for a class of interactive system that does not yet exist at scale: games that are generated rather than authored, robots that train in shared synthetic environments before they touch the real world, and foundation agents that grow up inside open-ended simulated worlds rather than on static datasets.

Learn more at RAAIS 2026!

Jeff Hawke, Odyssey’s co-founder and CTO, will go deep on this work at RAAIS 2026 in London on June 12.

Agora-1 can be played at agora.odyssey.ml. The Starchild-1 preview and technical report are live. Odyssey-2 is available via API at developer.odyssey.ml.

The bet behind world models is that the next leap in machine intelligence comes from interacting with a world, not from reading about one. Today, that world has sound and room for more than one.

At RAAIS we have a focus on translating cutting-edge technology and research into production-grade products for real-world problems.

The platform behind production AI at Revolut

Nikolay runs Machine Learning Engineering at Revolut, where his organisation builds the platform that supports every production AI system inside the company - from classical ML for fraud and personalisation, to time-series foundation models, to the voice agents now serving customer support. Revolut has crossed $1.3 trillion in transaction volumes and is the number one finance app in 19 countries; machine learning now sits in the path of millions of financial decisions a day.

The most concrete recent example of that platform in production is the rollout of voice agents across Revolut’s customer service operation, built with ElevenLabs. The system handles live calls in more than 30 languages, resolves tickets in under five minutes - roughly 8x faster than the previous escalation path - with a 99.7% call-handling success rate across more than four million customers in the UK and Europe.

One platform for builders, operators, researchers, and compliance

A central theme in Nikolay’s public work is that the hard problem in production AI is not building a model in isolation. It is building one platform that has to serve builders, operators, researchers, and compliance at the same time - and do so inside a regulated financial product. That framing is especially relevant now, because most organisations have already discovered that strong model performance does not by itself solve deployment. The harder challenge is the infrastructure around the model: evaluation, release discipline, governance, monitoring, and cost control, all without slowing iteration to a crawl.

For a technical audience, this is where a large share of the field’s practical difficulty now sits. As production AI moves into regulated settings - finance, healthcare, public services - the systems around the model have to satisfy operational and supervisory requirements as well as engineering ones. The platform is not separate from the model work. It is what determines whether model progress becomes durable capability inside a real organisation.

Governance as a velocity enabler, not a blocker

Nikolay has publicly outlined a framework for launching GenAI products in 90 days under regulatory constraints, built on three pillars: data lineage (treating compliance data as feature material rather than overhead), continuous delivery with multi-layered validation that goes beyond pass/fail tests, and compliance guardrails plus the documentation needed to defend them. The underlying claim is that governance, designed well, is a velocity enabler - moved into the development environment with clear tiers, predictable review cycles, and regulation treated as a technical requirement with a defined path to production.

As more companies try to support classical ML and generative AI side by side inside regulated products, this is becoming the central question in deployed AI. The bottleneck has shifted out of the model and into the systems that surround it.

Nikolay’s background

Nikolay holds a PhD in engineering from Siberian Transport University, where his thesis applied wavelet transform analysis to damage detection in beam superstructures from the response of traversing vehicles — structural health monitoring for bridges, an early grounding in reliability, monitoring, and operational discipline for critical infrastructure that carries through to his current work. His career has spanned Moscow, St Petersburg, Seoul, Stockholm, Toronto, and now London. He maintains active open-source projects and writes publicly on MLOps, AI governance, and risk in fintech at donets.org.

Apply to RAAIS 2026

Short bio

Nikolay Donets is Head of Machine Learning Engineering at Revolut, where he leads the team that builds the AI platform behind the company’s production models - covering classical ML, time-series foundation models, and the voice agents now serving customers in 30+ languages. He has publicly outlined a 90-day framework for shipping GenAI products under regulatory constraints, built on data lineage, continuous delivery, and compliance guardrails. He holds a PhD in engineering, with earlier work in structural health monitoring and predictive maintenance for critical infrastructure.

I’m excited to bring you the next Air Street NYC AI meetup on 14 May 2026, which brings together New York’s best researchers, founders, and engineers working in AI. Featuring Ramp, Arena Physica and Air Street Capital.

NYC AI brings together New York’s best researchers, founders, engineers, and operators who are building and deploying AI systems. We keep the group deliberately small, curated, and focused on people who are building - not talking about - AI. The goal is to help you learn new best practices, exchange ideas with peers, and meet future collaborators, co-founders, and team members.

At this edition of NYC AI, we’ll cover the following topics:

• Deploying AI inside a high-growth fintech - Seb Goddijn, Product Lead, Internal AI at Ramp

• Physics-aware AI - Pratap Ranade, CEO & Co-Founder of Arena Physica

• State of AI Report 2026 - Nathan Benaich, Air Street Capital

We’ll follow the talks with happy hour drinks, food, and plenty of time to meet people.

Recent meetups have included people from OpenAI, Anthropic, Google DeepMind, Meta, Hugging Face, Runway, Scale AI, Cohere, top labs at Columbia, NYU, Cornell Tech, Princeton, and startups including Ramp, Lumaril, Cursor, Decagon, Sierra, Harvey, Mercor, Granola, and many others.

If you work in research, engineering, product, BD, or you’re a founder, request a spot here:

Request a spot here