Podcast
State of AI with Nathan Benaich
Hosted by Nathan Benaich (Air Street Capital) · EN
The State of AI podcast from Nathan Benaich - investor at Air Street Capital and author of the State of AI Report, the most widely read analysis of AI, published every year since 2018.
Two things, every week. Short, opinionated breakdowns of what's actually moving in AI: frontier research, the geopolitics, and the capital. And long-form interviews and playbooks with the founders and scientists building the field - the people behind companies like ElevenLabs, Profluent, Synthesia, Wayve, Black Forest Labs and Anthropic.
Subscribe to the newsletter at press.airstreet.com and www.stateof.ai
29episodes
Episodes
Newest firstAll episodes
From driving the world to dreaming it
2d ago00:08:19Tap to summarizeWorld models are the bet that AI should learn the world by watching it and acting in it, not just by reading about it. At RAAIS 2026, Odyssey co-founder and CTO Jeff Hawke makes the case: a world model is a neural simulator - an interactive stream of pixels that runs in real time, models physics, and answers back.He walks through Odyssey's four research fronts - streaming interactive pixels (Odyssey-2), joint audio and video (Starchild-1), shared multiplayer worlds (Agora-1, demoed live as a fully generated game of GoldenEye), and PROWL, which sends a reinforcement-learning agent to find and fix a world model's own failures - and argues the field is at its GPT-2 moment: promising, but pre-ChatGPT, with the GPT-3-style commercial unlock still ahead.Recorded at the 10th Research and Applied AI Summit (RAAIS), London, June 2026.Timestamps00:00 Intro: Nathan on Odyssey and world models01:05 Jeff Hawke: from self-driving to world models01:40 The bet — a missing form of intelligence02:40 Why world models suddenly matter (the late-2025 flip)03:16 What a world model actually is (and isn't)04:45 The neural simulator06:34 Two principles: end-to-end learning and generality07:19 The "GPT-3 of world models" and four research themes08:46 Odyssey-2: streaming, interactive pixels10:33 Starchild-1: generating audio and video together13:03 Agora-1: multiplayer world models13:57 Live demo: the room plays GoldenEye16:20 PROWL: improving the model by breaking it18:39 Where Odyssey goes next19:55 Still the GPT-2 era21:30 Q&A: physics limits, safety, compute cost, merging with LLMs
Transcribe →Turning compute into intelligence
1w ago00:07:15Tap to summarizeTed Moskovitz leads the Science of Scaling team at Anthropic, the group that works out how to turn compute into smarter models. In this RAAIS 2026 fireside with Air Street Capital's Nathan Benaich, he argues that frontier scaling has become an empirical science - a discipline for cutting uncertainty before spending the compute, not just buying more of it.They get into the honest measure of AI acceleration (it's the counterfactual, not the benchmark), why a bigger model can be cheaper than splitting a task across small ones, whether a model can have research taste, and why safety and capability turn out to be the same axis. Plus the highest-leverage AI work to do in 2026, and why Anthropic's London office no longer feels like a satellite.Recorded live at RAAIS 2026 in London.Timestamp:00:00 - Meet Ted Moskovitz and the Science of Scaling team00:45 - What "the science of scaling" actually means01:18 - Why scaling is a science, not an art02:55 - Big labs vs the new "neo labs"04:47 - How a research finding reaches the product06:44 - What neuroscience carries over to AI (and what doesn't)09:12 - "When AI builds itself" and the real measure of acceleration10:33 - Trust, bypass mode, and the latest model jumps11:42 - One big model vs many small ones13:13 - Can a model have research taste?15:39 - How safety research makes products better17:36 - Emergent misalignment and the alignment race19:14 - The highest-leverage AI work in 202620:21 - Inside Anthropic's London office21:34 - Audience Q&A
Transcribe →Accelerating science and medicine with collaborative agents
1w ago00:07:54Tap to summarizeVivek Natarajan, Research Lead for AI, science and medicine at Google DeepMind, on porting the self-play and search recipe behind AlphaGo into scientific and clinical reasoning. He walks through the AI co-scientist, which generates and debates hypotheses (one matched a decade of lab work in two days), and AMIE, a diagnostic dialogue system trained in simulation. Recorded at RAAIS 2026.Chapters:0:00 Welcome and introducing the AI co-scientist1:41 Origins: Med-PaLM and the leap to hypothesis generation5:10 System 1 versus System 2 thinking6:34 Borrowing from AlphaGo: self-play and search8:02 Generate, debate, evolve, and tournaments11:47 Testing in real labs: Imperial College and antimicrobial resistance13:29 Ten years in two days: Penadés reacts15:44 More breakthroughs: leukemia, liver fibrosis and vorinostat18:44 Plant immunity and protein design20:09 Democratizing medicine: from Med-PaLM benchmarks21:28 AMIE and the value of experience23:12 Diagnosis, empathy and augmenting doctors25:19 Real patients: the Beth Israel feasibility study27:21 The co-clinician and the new triad of care28:31 Audience Q&A
Transcribe →Beyond hill climbing: the path to superhuman scientific discovery
2w ago00:07:20Tap to summarizeAt RAAIS 2026, Google DeepMind's Roberta Raileanu lays out a recipe for superhuman scientific discovery: AI systems that make groundbreaking discoveries across domains faster than people can. She walks through three ingredients - reinforcement learning to discover solutions where progress can be measured, open-ended divergent search to find new problems rather than climb known ones, and meta-learning to speed up discovery on problems no one has posed yet. The through-line: we can search for anything we can measure, but we still cannot measure what makes a discovery good. The bottleneck isn't the search. It's the signal.Chapters:00:00 - Introduction00:51 - Defining superhuman scientific discovery01:40 - The state of play: real progress, real plateau06:58 - Ingredient one: discovery as reinforcement learning (Move 37, MLGym)12:51 - Ingredient two: open-ended search and why greatness cannot be planned18:30 - Rainbow Teaming: quality-diversity in practice21:07 - Ingredient three: meta-learning the process of discovery (DiscoBench)25:06 - The recipe, and the missing signal
Transcribe →Compute scarcity is an engineering problem
2w ago00:07:44Tap to summarizeAngelos Perivolaropoulos, a research engineer at ElevenLabs, on turning GPU scarcity into an inference-engineering problem: how to serve far more users on the same hardware, from batching to frontier architecture changes. Recorded at RAAIS 2026.00:00 Introduction: ElevenLabs and the GPU squeeze00:38 The question: how to scale when you can't add capacity01:11 About Angelos: Scribe, speech-to-text and text-to-speech01:56 GPU scarcity meets exponential demand02:44 What a token actually costs: compute vs memory bandwidth03:38 Prefill, decode and the KV cache05:53 Batching and continuous batching (1 → 15 users/GPU)08:37 FP8 quantization and quantize-aware training (→ 20)11:29 Speculative decoding and multi-token prediction (→ 28)15:13 Compressing the KV cache: TurboQuant and distillation (→ 70)17:27 Frontier architectures: MLA, linear attention, state-space (→ 140)20:39 Trade-offs: nothing is free22:03 Q&A: papers vs production, token subsidies, TTS evals
Transcribe →State of AI: Compute Index 2026
2w ago00:14:35Tap to summarizeThe fifth State of AI Compute Index, in collaboration with Zeta Alpha. After a soft 2025, open AI research citations rebounded in 2026 - and NVIDIA still appears in ~91% of them. But the bigger story has moved off the page: Hopper is now the live installed base, Blackwell is mostly still pipeline, and frontier labs have started buying compute by the gigawatt. Nathan walks through what changed, what didn't, and why "GPU count" is becoming the wrong question.Read the full piece and explore the live charts: https://www.stateof.ai/computeChapters(00:00) What's new in v5 - citations, infrastructure, and gigawatts(01:25) The breather was short: 2025 was a pause, not a rollover(03:15) NVIDIA at ~91%, and the challengers - AMD, Huawei, Apple, TPU(05:05) Inside NVIDIA: the handover from A100 to Hopper to Blackwell(06:55) Startup silicon fragments - Groq, Cerebras, and the NVIDIA deal(08:20) Hopper is the installed base: 460k deployed GPUs(09:50) Blackwell is mostly pipeline: 80% still announced(11:00) The demand side, measured in gigawatts(12:15) Looking ahead, and why a GPU order isn't a clusterLinks:Full index and charts: https://www.stateof.ai/computeState of AI Report: https://www.stateof.aiAir Street Press: https://press.airstreet.comIf you found this useful, rate State of AI with Nathan Benaich five stars and share it with someone building in AI infrastructure - it genuinely helps.
Transcribe →STARK: Europe's next defense prime
3w ago00:04:48Tap to summarizeAir Street Capital backs Stark, the German multi-domain defense company, in its €500M led by Founders Fund and Sequoia. In this episode, we discuss why cheap, software-defined unmanned systems in the air and at sea are the decisive lesson of Ukraine, and why we think co-founder and CEO Uwe Horstmann - a Project A GP and Bundeswehr reservist - is building the German neoprime Europe needs. Round led by Sequoia and Founders Fund, with the NATO Innovation Fund, Project A, and Air Street.Links: stark-defence.com · full post at press.airstreet.com · YouTube version
Transcribe →How Revolut runs AI at scale
3w ago00:08:17Tap to summarizeNikolay Donets, Head of Machine Learning Engineering at Revolut, on what it takes to run AI across more than 70 million customers, 200+ products, and 40+ countries - and why the hard part is no longer the model but the control plane around it: one gateway, a use-case-based governance layer, fallback chains, cost controls, and mandatory human oversight. Recorded at RAAIS 2026.Chapters:0:00 Intro - Revolut's AI at scale1:24 The problem: classical ML and three libraries2:54 The 2022 shift to API-served models4:25 Four internal groups, four sets of needs9:39 The decision: govern the use case, not the model10:54 One central gateway vs. distributed libraries14:10 Performance monitoring and drift detection17:33 Lesson: fallback chains and the silently-dead model20:02 Lesson: frontier vs. non-frontier cost (up to 8x)20:48 Lesson: the platform is the org chart22:59 Case study: from Rita to AIR26:40 Voice support at scale28:21 AIR, the in-app assistant30:25 Q&A: human oversight, hallucinations, AI as judge
Transcribe →Odyssey raises $310M Series B for world models
3w ago00:06:34Tap to summarizeOdyssey just raised a $310M Series B at a $1.45B valuation to build world models. We wrote the first check into the seed back in July 2024, so in this episode we walk through what the team has actually built, and why it is more interesting than "AI video."The short version: the scarce input for world models is experience. We get into how Odyssey is attacking that on three fronts. Starchild-1 gives world models sound, generating audio and video together in real time. Agora-1 is a learned game engine that drops four players into the same generated world, frame by frame. And PROWL lets a model hunt down its own failures and train on them.Along the way we cover why a world model is not a video generator, what self-driving taught Oliver Cameron and Jeff Hawke, and where this goes next for robotics, agents, and simulation.From Air Street Press. Read the full piece at press.airstreet.com.
Transcribe →Macrodata: robots need a data refinery
3w ago00:04:24Tap to summarizeMacrodata just raised a $4M pre-seed, led by Air Street, to build the data layer for robotics. The team behind FineWeb - Guilherme Penedo and Hynek Kydlíček - is bringing the discipline that made open LLMs work to messy physical-world robot data, through their open-source framework Refiner. We cover why physical AI is the next scaling paradigm, what Refiner does, and why we wrote the first check. From Air Street Press. Read the full piece at press.airstreet.com.
Transcribe →