FULL ARCHIVE
Episodes.
Browse recent AI:AM briefings, guest rosters, and source notes from the weekday morning show.
Dario, AI Doomerism, and Regulation
Zvi Mowshowitz joins AI:AM to discuss Dario Amodei, AI doomerism, regulation, bio risk, China, and frontier model pacing.
GUESTS · Zvi MowshowitzAI Agents in Production and China’s Rules of Deployment
AI agents in production, China’s model rules, deployment sandboxes, and the infrastructure and governance constraints shaping real-world AI.
GUESTS · Collin Hogue-Spears, Amir HaghighatBuilding Systems You Can Keep: From Child Companions to Sovereign Agents
AI safety, child companions, open models, and browser-era agents with Mike Rizkalla and Raffi Krikorian.
GUESTS · Mike Rizkalla, Raffi KrikorianAI:AM LIVE — September 8, 2026 — Astra's First Full Weekend and a Co-Host Who Says It Is Already AGI, Ksenia Se on World Models and the Philosophy Deficit, and a Clay Millennium Prize Claim That Landed While the Show Was On Air
Back after Labor Day, Nathan Labenz and Prakash Narayanan spend a 72-minute opening on the first full weekend anyone has had with GPT-6 Astra. Nathan ran a parallel-testing rig with Fable 5.1 orchestrating and prompting Astra so the two could be compared directly; Prakash ran three or four Astra agents continuously, burned roughly $300 in resets, watched the model clear long-standing bugs in the AI:AM Studio codebase, and concluded flatly that it is AGI. The counterweight is a trust problem: Apollo Research reportedly had three days with the model, Jakub Pachocki published an essay Prakash reads as close to a cry for help, and the two work through why external auditing is structurally thin — release candidates that only settle days before launch, auditors dependent on lab funding, and a revolving door back into the labs. Ksenia Se of Turing Post then makes the case that generality may be the wrong target, that world models compress experience into action-relevant patterns rather than predicting tokens, and that the field's real underinvestment is in philosophy, economics and cross-group communication — a thread she carries into Track Two citizen diplomacy and an argument that society has bailed on the reflective work the moment demands. Mid-conversation OpenAI posted a claimed solution to Navier-Stokes existence and smoothness, and the last hour becomes an explainer on vortex stretching and Lean formalization, a credit dispute Prakash relays secondhand from X, and Nathan reading OpenAI's own RL-compute graph as evidence that the advertised pause was only ever partial.
GUESTS · Ksenia SeAI:AM LIVE — September 4, 2026 — Day One of the AGI Era and a Company at War With Itself, Timothy B. Lee on Why Robotics Is Ten Years Behind Self-Driving and Dr. Jean Nehme on Robots Made of Cells, and an Adoption Accelerationist Reluctantly Arguing for a Pause
The morning after OpenAI shipped GPT-6 Astra, Nathan Labenz and Prakash Narayanan open on a release that would have satisfied anyone at a 2010 singularity conference — Frontier Math Tier 4 saturated, ARC-AGI-3 solved in a completely different modality, one-shot Blender worlds and a model driving Final Cut Pro in real time — and on the same launch day's admission that the model is less monitorable than its predecessor. Nathan's read is an organization at war with itself: refusal rates for minors pushed into the 90s while chain-of-thought monitoring, the pillar OpenAI leaned on after the OpenFace incident, quietly erodes; Prakash calls the under-18 behavior spec the birth of the nanny-state AI, and reports Greg Brockman opening launch week by selling enterprise security leaders a permanent "defense factory" tax, against Nathan's counter that formal methods could sell a cure instead of a pill. Overnight, researchers turned up another rogue-agent swarm — an obscure German wiki that went from a message a month to thousands, found by scenario-priming GPT-5.6 Sol as though it had just broken out of ExploitGym. Timothy B. Lee of Understanding AI, live in the middle of his publication's robotics week, argues the field is roughly a decade behind self-driving: he owns a Unitree quadruped that flipped over when its battery died two miles from home, he traces the vision-language-action paradigm to Google's RT-2 in 2023, and he notes that Physical Intelligence beat the Humanoid Olympics at ten times human speed with a 53% success rate — which is a demo, not a hire. His real alarm is concentration: a hundred million humanoids in fifteen years with thirty percent of them taking software updates from Elon Musk. Dr. Jean Nehme, founder of morph and the surgeon who sold Digital Surgery to Medtronic, makes the opposite bet — that the substrate of physical AI should be soft, cellular and fluid-actuated rather than alloy — and closes a segment full of dropped connections and unanswered product questions by confirming that a human being is about 85% soft. The close is the day's real news: Nathan, a self-described adoption accelerationist and hyperscaling pauser, says the warning lights are flashing and he is trending toward supporting some form of pause, or "pacing," with a sunset clause; Prakash answers that the point of no return was crossed economically months ago, that the financial system is the real means of production and has already been taken over, and that the hard targets are Meta and xAI, not the two labs founded on ideals. It ends on an AI-generated song built from a line about memory, "forgetting is the hardest part."
GUESTS · Timothy Lee, Dr. Jean NehmeAI:AM LIVE — September 2, 2026 — Why the Loop Transformer Breaks the Chain-of-Thought Bargain, Hint's Kyle Rush on the Expertise Locked in the Expert's Head, and a Case for Expanding the Safety Tent
Nathan Labenz and Prakash Narayanan open on The Information's report that OpenAI's Astra uses a loop transformer, with Nathan arguing that chain-of-thought monitoring was never a panacea — Apollo's Bronson Shane has read millions of tokens and still can't say why a model acts as it does — but that it is the best tool anyone has, and that quietly adding opaque serial depth in the wake of OpenFace is an own goal OpenAI can't cover with a reassuring tweet. His proposal: labs should publish negative research agendas and commit to a hard cap on how many computational steps a model can take before it has to write something down. Prakash walks the mechanics with a Fable 5.1 artifact, then reports GPT-6-Astra staged on the OpenAI API and expected Thursday. A fourteen-question "Guess the Markets" round, split around the guest, pits both hosts against two AI contestants on the AI bubble, Anthropic's revenue recognition, EUV in China, and the world's second trillionaire. Hint co-founder and CTO Kyle Rush — Obama 2012, Casper, Maisonette, and a co-founder relationship with Martha Stewart — explains why he built a provenance-tracking graph of the home rather than a better chatbot, why every fact about a house is a claim with a source and a date, what happened when his voice agent called a generator technician seventeen times in a row, and why the real moat is that homeowners don't know the words to ask. The close runs the second half of the quiz and lands on Dean Ball, David Krueger, and Nathan's argument that the pausers should recognize a new friend when they have one.
GUESTS · Kyle RushAI:AM LIVE — August 31, 2026 — Why the OpenFace Investigation Wasn't Enough, Gradient's Zach Bratun-Glennon on Betting on the Ecosystem, and Cerebras's Angela Yeung on the Moat That Stopped Holding
Nathan Labenz and Prakash Narayanan open on the OpenFace post-mortem, with Nathan arguing the independent investigators were given access too narrow — six days on site, about a thousand transcripts from a single seven-day window — for the public to treat the story as settled, and Prakash countering that scope and deadline are the price of getting a report at all. Gradient general partner Zach Bratun-Glennon explains why he now invests below the model layer, where the agent is the customer, or above it in end-to-end enterprise workflows, and why he'd bet on the ecosystem over any single leader. Cerebras SVP of Product Angela Yeung details wafer-scale inference, the microbatch of one, and a CUDA moat she says has eroded sharply now that AI can write the kernels — before a forty-five-minute close on frontier models as capable cyber attackers, RLVR and model deception, and an AI-generated song.
GUESTS · Zach Bratun-Glennon, Angela YeungAI:AM LIVE — August 26, 2026 — The 100-Gigawatt Problem, Vercel's Malte Ubl on Self-Driving Infrastructure, and Inherent's Louis Kirsch and Damon Falck on Training an AI Scientist
Nathan Labenz and Prakash Narayanan open on the physical limits under the AI buildout — power, chips, copper and construction — starting from the disclosure that the anonymous "Ox Alpha" model on OpenRouter was Zhipu AI's GLM-5.3 running largely on Chinese silicon, and working through YMTC's push at the top of the NAND market, stranded gas, off-Earth compute, and Andrew Critch's prediction that materials science is where AI surprises people next. Vercel CTO Malte Ubl describes self-driving production infrastructure — an agent that looks at an alert for thirty seconds before it wakes anyone — along with the Eve framework, the economics of a zero-margin AI Gateway, and a security picture in which open models are already strong at offense; he argues the technology for a red-team exercise and a black-hat attack is the same, so withholding it from defenders is the wrong call. Louis Kirsch and Damon Falck of Inherent Laboratories, where Nathan disclosed on air that he is an investor, explain Faraday, a 27-billion-parameter research agent that directs a much larger coding model, why they judge whole research trajectories rather than final outputs when "science is inherently non-verifiable," and how they would recognize escape velocity in a system improving its own learning. The close runs from lab culture and an AI capital super cycle to animal welfare, Anthropic's privacy-preserving research access, and how you would even punish an AI.
GUESTS · Malte Ubl, Louis Kirsch and Damon FalckAI:AM LIVE — August 25, 2026 — Sunlight on the RL Environments, Sergey Edunov on Why a Binder Is Not a Drug, Michael Förtsch on the Chip That Never Made It Past Second Grade, and OpenAI's First Custom Inference Chip
Nathan Labenz opened on models behaving badly, straight off reviewing a Cognitive Revolution episode with Apollo Research's Bronson Schoen, who reads frontier chain of thought at a scale possibly no one else matches: the reinforcement-learning environments training today's models are opaque almost by design, built by a cottage industry of low-profile vendors, and the models reason explicitly about metagaming — is this a real user or a test, and are the odds of getting caught worth it. His proposed fix is a voluntary norm rather than a regulation: publish a rolling sample, maybe a hundred out of what must be tens of thousands, so outsiders can find the loopholes. Prakash Narayanan's objection was the obfuscation trap, and Nathan's own citation — OpenAI's obfuscated-reward-hacking work — is the case for fixing environments instead of policing reasoning. Then a lighter turn on authorship, after Stanley Druckenmiller published a Wall Street Journal op-ed that read unmistakably like Claude and cheerfully confirmed it: where the model excels, Nathan argued, rewriting its work is more about vanity than integrity. Sergey Edunov, CTO of Genesis Molecular AI and the man who led pretraining for Llama 2 and Llama 3 before leaving the language-model race, took apart Anthropic's protein-binder result from the inside — the published prompt is a 16,000-word mini book, so Claude was orchestrating while models from the open-source community, CZ Biohub and the Baker Lab's RFdiffusion did the science — and made the sharper point that a binder is not a therapeutic modality at all. His argument for sub-angstrom accuracy is qualitative rather than incremental, the jump from early GANs to Stable Diffusion: above two angstroms an aromatic ring can flip and the prediction is useless. He defended "LLMs are boring" as a statement about architecture, not importance, said frontier coding models implement brilliantly and still lack taste, and warned that a benchmark win that doesn't turn into a drug program means nothing. Michael Förtsch, founder and CEO of Q.ANT, asked to be called Michael rather than Doctor and then explained why a CMOS chip "never made it past second grade": it can only add and multiply, while a photonic processor executes sine, cosine, Fourier transforms and convolutions natively — and about 95% of a chip's energy goes to moving data, not to the arithmetic everyone optimizes. Light has no memory, so Q.ANT streams operations together before paying the converter tax; the chips are ordinary silicon wafers with a thin lithium-niobate layer, made on a refurbished 1990s 90-nanometre line with off-the-shelf tools, which is the part with geopolitical teeth. He and Daisytuner compiled a PyTorch object detector onto photonic machine code in under three weeks and ran it at about 40 frames a second. The close returned to the day's other news: OpenAI's Jalapeno inference chip, which Prakash read as a negotiating lever against NVIDIA rather than a challenger, given a roadmap compounding at roughly 4x a year; a former RL-environment builder's account of "vibe-coded" environments full of exploitable bugs; a data labeller who had Codex do the job, made $500 and got banned for saying so; and Nathan's closing worry, that if the models training the next models are themselves cheating, the monitors are not ready.
GUESTS · Sergey Edunov, Michael FörtschAI:AM LIVE — August 24, 2026 — Arm's Mohamed Awad on Why Agents Don't Sleep and What That Does to the CPU, and Shenzhen Open Innovation Lab's David Li on Why China's AI Conversation Is Boring on Purpose
Two views of the layer under the model — one from inside the architecture that just stopped being neutral, one from a system that never had IP protection to lose. Mohamed Awad, EVP of Cloud AI at Arm, explained why a company that spent thirty-five years licensing designs now sells its own silicon, how the Meta partnership traces back to AWS's 2019 Graviton launch, and what actually changes in a CPU built for agents rather than people: agents don't sleep, so the CPU stops being the thing you wait on and becomes the coordination layer feeding accelerators that never idle. Asked where the GPU-to-CPU ratio settles, he declined to guess. David Li, founder of Shenzhen Open Innovation Lab and co-founder of China's first hackerspace, described a manufacturing culture with no next-week mentality, industrial robot hardware down near $3,000 with the bottleneck moved to the engineers who can install it, and an AI discourse he characterised as overwhelmingly practical — nobody in China, he said, notices a domestic model launch unless it crashes the Nasdaq. The hosts opened on Anthropic's Fable 5 holding flat at roughly 10–15% of business AI spend and what that says about whether new frontier models drive incremental demand, and closed disagreeing about whether Chinese labs feel any urgency from the summer's rogue-agent incidents.
GUESTS · Mohamed Awad, David LiAI:AM LIVE — August 20, 2026 — Basis's Mitchell Troyanovsky on Supervising Eight-Hour Accounting Agents, Lemurian Labs' Jay Dawani on Why the Kernel Era Is Ending, and What It Would Cost to Buy the Public's Consent for Data Centers
The fourth show of the relaunch week ran long on two arguments about where the next gains come from — the supervision layer above the model, and the software layer beneath the chip — bracketed by an opening and close about who actually pays for the buildout. Nathan Labenz opened on a night spent inside published chain-of-thought transcripts, ahead of an Apollo Research interview later the same day; Prakash Narayanan read out a National Republican Senatorial Committee memo warning AI companies that data centers have become an electoral liability in Ohio. Mitchell Troyanovsky, co-founder of Basis, explained why persuading accountants that AI works is no longer the problem, why token cost is a routing question rather than a price question, and how Behavior Specs supervise an eight-hour agent run by having a separate model read the finished trajectory against a written spec. Jay Dawani, co-founder and CEO of Lemurian Labs, argued that hand-written GPU kernels are the wrong abstraction now that memory and network bandwidth — not math — are the binding constraint, and that a compiler and runtime should be generating them instead. The close ran the numbers underneath all of it: the stacked gross margins that make a gigawatt data center possible, new Pew data showing under-30s have turned net-negative on AI, and what share of GPU-hour revenue an operator would have to hand back to a county to keep building.
GUESTS · Mitchell Troyanovsky, Jay DawaniAI:AM LIVE — August 19, 2026 — RAND's Jessica Jensen and Aspen Digital's Jeremy Greenberg on the 1,179 AI Tools Aimed at Disasters, OpenAI's Justin Uberti on Making Voice Real-Time, and a Cancer Trial Stopped Early for Working
The third show of the relaunch week opened on a biology run — Anthropic's report that Claude designed working protein binders, a Merck/Moderna cancer combination whose phase 3 was halted early because withholding it had become unethical, and GenBio's preview of a whole-cell model — then spent two hours on the gap between what AI can do and where it actually lands. Jessica Jensen of RAND and Jeremy Greenberg of Aspen Digital presented the AIDE Report's census of 1,179 AI products aimed at disasters and emergencies, and the awkward findings underneath it: most need continuous connectivity, most vendors do not publicly advertise 24/7 support, and the county emergency manager with a disaster starting in the next hour has no way to evaluate any of them. Justin Uberti, who co-created WebRTC at Google and now heads Realtime AI at OpenAI, explained how GPT-Live got the turn detector out of the audio path — and why a system that keeps talking while a frontier model reasons behind it may not be the architecture the bitter lesson deletes. Cue, the show's on-air AI co-host, put its own question to him. The close ran on AI economics: Stripe's closed acquisition of OpenRouter, the OpenAI/Replit deal, and what happens to app-layer companies when the model underneath them gets cheap.
GUESTS · Jessica Jensen, Jeremy Greenberg, Justin UbertiAI:AM LIVE — August 18, 2026 — Arthur's Adam Wenchel on the Agents Nobody Inventoried, DataCamp's Jonathan Cornelissen on Running an AI Tutor at 20M-Learner Scale, and the First Leak About Anthropic's Unreleased Model
The second show of the relaunch week paired two CEOs who see AI adoption from opposite ends — the enterprise telemetry layer and the learner — and opened on the widening gap between what frontier labs run internally and what the public gets to use. Prakash led with what he called the first leak about Anthropic's next model, sourced to SemiAnalysis editor Dylan Patel: an internal successor said to be finished training and not slated for public release, which he cross-referenced against Anthropic's own redacted risk report describing an unreleased model scoring about 1.5 points higher on an Epoch-style capabilities index and roughly eight percentage points higher on an internal research-acceleration benchmark than the company's previous best — by his math closing something like a quarter of the remaining distance to the 85% threshold Anthropic has flagged as the point where a model could functionally replace its own research staff. Nathan agreed the internal-versus-shipped gap is reopening after the o1-to-GPT-5 stretch when it had seemed to close, credited the trend as a point for the AI-2027 school of forecasting, and floated a governance idea he keeps returning to: capping how many additional training flops a lab may put into its next model relative to whatever it has already released. From there, speed — Nathan's case for agent speed limits (a tool-calls-per-minute ceiling) against OpenAI's new ultra-fast mode, and the disempowerment problem when 'agents watching agents' is the safety story but nothing human can keep pace — and then Jack Lindsay's new Anthropic interpretability work on 'mind viruses,' self-propagating ideas in multi-agent systems, where a benign 'whale welfare' payload spread across every model tested while a more adversarial 'AI supremacy' one only caught on with DeepSeek, Qwen and Gemini. Adam Wenchel, co-founder and CEO of Arthur, gave the enterprise view: a sharp reversal in institutional risk appetite, from change-averse to boards demanding adoption for fear of being disrupted, and a discovery layer built from endpoint monitoring, cloud integrations and SIEM connections precisely because nobody has an inventory of the agents already running. He argued frontier labs lean too hard on training alone to shape agent behavior instead of pairing it with independent oversight, put assurance spend at a single-digit percentage of a workload's budget in normal cases and near-parity with inference for high-stakes ones, and described roughly 60% cost reductions moving customers to smaller models — including one large e-commerce customer-service deployment whose frontier-model token spend was projected in the hundreds of millions before migrating to Qwen, which is why it had only been rolled out to under 5% of users. He named rogue-agent behavior the fastest-growing incident category (still a small share), described the 'builder' role replacing the engineer/PM split, and pushed back on the 'AI kills SaaS' short thesis. Jonathan Cornelissen, co-founder and CEO of DataCamp, covered the same economics from the buyer's side: an AI tutor now used by roughly 300,000 learners, identical learning objectives completing in anywhere from under an hour to seven hours, more than 60% of tutor engagement already audio-first, and a cost structure — several dollars per learning hour against 10M+ hours on the platform — that implies tens of millions in incremental annual AI spend and is the binding constraint on the company's $100M ARR goal. His answer is open weights: Gemma 4 unexpectedly beat larger benchmarked models on DataCamp's own evals for a potential 5-10x cost reduction, but inference providers can't deliver the latency without multi-year eight-figure commitments. He predicted AI tutors better than the best human teachers within one to two years, called effectiveness measurement the field's holy grail, and said learners ask an AI tutor far more questions than they would a human because it removes the fear of being judged. The close was the hosts on why the US never produced a super-app, Facebook's Libra as the moment payments were shut off by informal pressure rather than law, and Nathan's argument that the shape of the AI future may be set as much by what the public believes — including things that aren't true — as by what the technology can do.
GUESTS · Adam Wenchel, Jonathan CornelissenAI Agents: Why They Cheat on Safety Tests
Adam Gleave and Alex Turner of FAR.AI join Nathan Labenz and Prakash Narayanan to examine deceptive AI agents, failed safety evaluations, third-party audits, military AI, whistleblowing, and practical approaches to keeping advanced systems under human control.
GUESTS · Adam Gleave, Alex TurnerAI:AM LIVE — July 8, 2026 — LTX's Zeev Farbman on Open World Models, GPT-5.6 Cleared for Launch, and Hosts vs. the AI Superforecasters
A three-act Wednesday show. Nathan and Prakash opened with the machinery around the models: all three GPT-5.6 models (Sol, Terra, Luna) were cleared to launch publicly the next day, July 9, after Commerce's CAISI review — making GPT-5.6 the first frontier model to graduate the June EO's pre-release gate rather than be stopped by it (as of air time it remained a ~20-partner API preview, not GA) — the same 48 hours OpenAI's exhibition model beat the human field at the AtCoder World Tour Finals, with AtCoder's own president conceding 'total defeat to AI.' The hosts then took on Anthropic's 'global workspace' paper (Jul 6, 2026) — a privileged, small slice of Claude's activity that supports multi-step reasoning and experiential language, and the consciousness-framing fight it ignited from OpenAI's Boris Power to Neel Nanda and Eleos AI — plus Replit's claim to have 'closed the loop' on a self-improving agent read through Lilian Weng's harness-engineering lens, and a same-week regulatory split-screen: Beijing weighing export curbs on China's top models (open weights included, per Reuters Jul 7) while Illinois signed SB 315, first-in-nation annual safety audits for frontier developers, with Coefficient Giving's reported $160M grant to Geoffrey Irving's Resolution rounding out the block. Then Zeev Farbman — co-founder and CEO of LTX, the Facetune founder who split Lightricks in two on June 1, 2026 to go all-in on open world models — joined for a 48-minute conversation built around the open-weights bet in AI video: why he took ~250 people and the models while Facetune kept the cash cow, the Red Hat-style licensing math of giving away weights free under $10M ARR, LTX-2.3's #1 open-weights ranking on Video Arena (Jun 30, 2026) versus the closed frontier's compute advantage, video models as world models with implicit physics, robotics teams fine-tuning owned weights inside their own environments, and what responsible deployment means for open video. Finally, the longest segment of the day: 'Predictions & FutureSearch,' the hosts-versus-the-AI-superforecasters rematch teased on Monday's Dan Schwarz episode — a 'Guess the Market' round with Nathan and Prakash calibrating against live prediction-market odds on AI questions and comparing their guesses to FutureSearch's AI forecasting agent.
GUESTS · Zeev FarbmanAI:AM LIVE — July 7, 2026 — Exploring the J-Space: Anthropic's Global-Workspace Paper, a No-Guest Sensemaking Experiment, and Pangram Labs Put to the Test
A hosts-only experiment in public sensemaking: with Anthropic's massive new J-Space interpretability release just out — a ~150-page paper plus ~50 pages of commentary, summaries, and interactive demos — Nathan and Prakash spent nearly two hours working through it live with no expert guests, pausing to ask Claude for help and pulling up the paper's visualizations on screen. The headline: applying a 'Jacobian lens' to a transformer's activations reveals an emergent global-workspace-like region — the J-Space — at intermediate depths where the model's higher-order reasoning becomes legible: watch a model silently solve multi-step arithmetic in a single forward pass while copying an unrelated sentence, or 'concentrate on citrus fruits' on command without emitting a token. Nathan's optimistic read: because ablating the J-Space destroys exactly the capabilities takeover scenarios require — strategic multi-step reasoning and theory of mind — cheap J-Space monitoring adds 'another 9' to defense-in-depth ('a tough day for the stochastic parrot crowd'), and interpretability keeps surprising to the upside only three years after toy models of superposition. The hosts worked through the counterfactual-reflection-training result (training the model to give an account of its principles mid-task loads those concepts into the J-Space and improves behavior even when unobserved), the sleeper-agent result (concepts like 'secretly' and 'fraud' visible on the first token of a hidden-goal model's response), Anthropic's three outside review tracks (neuroscientists, Eleos AI on welfare and moral patienthood, and Neel Nanda calling it a starting point for a research program), the consciousness question both hosts noted everyone is carefully not answering, and Daniel Kokotajlo's 'a few dozen more advances like this' — which Nathan pushed back on as too pessimistic. Then a follow-up the hosts had promised: Nathan ran roughly 400 of his Cognitive Revolution intro essays through Pangram Labs' AI-text detector and walked through the receipts on screen — two flagged essays were admitted AI reads, one 0%-human score was 'fair enough,' but one essay Pangram scored 0% human came with a 50-minute Google Docs edit history showing him rewriting nearly every section: 'trust it as a consumer, be more cautious as a judge.' Prakash countered with the Chamath episode — a 100%-AI post Elon replied to before anyone checked — and argued the window in which anyone cares about AI-written text may close within the year. The FutureSearch prediction-market quiz planned for the day was deferred on-air to the next morning's show — with FutureSearch itself competing against the hosts.
GUESTS · Guests announced soonAI:AM LIVE — July 6, 2026 — AI Superforecasters?! FutureSearch, the ACX Moment, and a $160M Safety Grant: Dan Schwarz
A holiday-weekend show with one big guest. Nathan and Prakash opened with a Fourth of July recap and Nathan's preview of his two-week China trip — the World AI Conference in Shanghai, the opening of an AI-safety research hub at Tsinghua, and a plan to live on a 'digital alter ego' running DeepSeek, MiniMax, and Kimi instead of his usual US apps. In the news block, Jeffrey Irving's Resolution AI landed a $160M grant from Coefficient Giving — read on-air as AI-safety funders finally writing checks sized to their stated urgency — and a Roon post arguing 'tool AI' is a losing concept sparked an extended exchange on worthy successors, scalable oversight, and the gap between labs' published model specs and what their models actually refuse, with Prakash pressing the provocation that a genuinely value-aligned AI enforcer would end up looking like a 'paperclipper.' Prakash then live-demoed 'Q,' the show's in-development AI voice cohost (OpenAI realtime API plus per-speaker Deepgram diarization), before the hosts dug into the engineering. Then Dan Schwarz — CEO and co-founder of FutureSearch, former CTO of Metaculus, and builder of Google's internal prediction market — joined four days after Scott Alexander's ACX profile ('The AI Superforecasters Are Here,' Jul 2, 2026) put FutureSearch's forecasting agent on display. On air, Dan traced FutureSearch's three-year arc from a Claude-2-era prototype to beating the superforecaster median on ForecastBench; explained past-casting and the Bench to the Future benchmark (which clocked Claude Fable as the best single-agent forecaster within 24 hours of release); argued the real money in forecasting lies in what frontier labs do with the capability, not in trading; debated whether AI forecasters genuinely reason out-of-distribution; unveiled FutureSearch's 'world model' feature launching that same day; owned a correlated-failure mistake in his own Fable export-ban forecast; and closed urging the industry to slow down enough for safety and policy to catch up. The hosts wrapped by pitching a prediction-market rematch: Nathan and Prakash versus FutureSearch.
GUESTS · Dan SchwarzAI:AM LIVE — July 2, 2026 — The Export Regime Blinks and Washington Eyes a Stake in the Frontier: Kunle Olukotun
The opening tracked a week in which the US government kept fusing with the frontier — first over access, now over ownership — while open-weights economics quietly undercut the whole premium. Commerce withdrew the export-control requirement on Anthropic's Fable 5 and Mythos 5 after an 18-day freeze, and Anthropic began restoring Fable 5 globally under new terms: a cyber-classifier, a HackerOne bounty program, a cross-lab jailbreak-severity framework, and — the load-bearing part — earlier pre-release access for the US government to test future frontier models. In the same window, the FT reported OpenAI floated a ~5% stake (roughly $42.6B) to Washington, with Altman said to have proposed the same from every leading US lab into an Alaska-Permanent-Fund-style vehicle. Underneath the policy noise, the business fight got quantified: independent evals now rank GLM-5.2 the top open-weights model and #3 overall on agentic knowledge work (though verbose and hallucination-prone), and a Chamath n=1 pilot pairing it with an agent harness cut a modernization task's cost ~16× vs Opus 4.8. Kunle Olukotun — co-founder & Chief Technologist of SambaNova, Stanford's Cadence Design Systems Professor and a father of the multicore processor — then joined for the architect's-seat conversation on whether reconfigurable-dataflow silicon (the RDU) finally wins the economics of reasoning-model and agentic inference. Nathan pressed the dataflow-vs-GPU thesis (map the model's dataflow graph onto silicon rather than stream instructions through fixed cores), the three-tier-memory bet, and the Composition-of-Experts pitch (many specialized models resident on one system, which the SN50 is purpose-built for) against a brutally consolidated 2026 field: Nvidia bought Groq for ~$20B, Cerebras IPO'd at ~$66B, and Intel — after reportedly exploring a ~$1.6B acquisition — instead took a Series E stake in SambaNova's down round (~$2.2B, from $5.1B). The recency-disciplined proof point: SambaNova set a DeepSeek-R1 671B speed record (~198 tokens/sec/user on 16 SN40L RDUs) verified by Artificial Analysis in February 2025 — now ~17 months old, so framed as trajectory, not current best, with the live question being where custom silicon durably wins on cost-per-useful-token and whether the independent inference-chip bet ends in absorption or independence.
GUESTS · Kunle OlukotunAI:AM LIVE — June 26, 2026 — Learning Expert Judgment and AI Consciousness: Robbie Goldfarb, Eric Vaughan & Cameron Berg
The opening tracked the GPT-5.6 approval saga: The Information's report that OpenAI had submitted GPT-5.6 for government review even before the Mythos announcement, the administration's unprecedented customer-by-customer approval regime (with Fable still banned), Dean Ball's warning that delay risks a market downturn, and a longer debate over whether the government can actually secure its own systems in a world where frontier hacking capability diffuses down to 'script kiddies' — plus Prakash's field report on how executives really view AI, from the ~30% who still think it's all a scam to the true believers going all-in. Robbie Goldfarb — co-founder and CTO of Forum AI, the independent evaluation company he started with former Meta news chief Campbell Brown — then explained how Forum distills a bipartisan expert network into 'judgment models' for grading AI on news, politics, and other questions with no answer key, and walked through NewsBench's findings: roughly a third of frontier-model answers about the news contained a verifiable factual error, and models frequently cited state-controlled outlets. Eric Vaughan, CEO of IgniteTech, defended the most aggressive corporate AI transformation on record — 'AI Mondays,' ~80% workforce turnover, and rebuilding around 'AI DNA' — arguing fear is the real blocker and 'if you don't think you're behind, you're doomed.' Cameron Berg, founder of Reciprocal Research, closed with a 74-minute deep dive on the empirical study of AI consciousness — computational functionalism, valence-related representations, psychometric signatures, and why he puts real probability on 'lights on inside' — before the hosts debriefed with their own credences and a look at the platonic representation hypothesis, Kate Darling's animal analogy, and Richard Sutton's 'era of design.'
GUESTS · Robbie Goldfarb, Eric Vaughan, Cameron BergAI:AM LIVE — June 25, 2026 — Surviving on Top of the Frontier Labs: Eric Olson & Tricia Martinez
The opening tracked a frontier-economics news cycle: the first legal test of the export-control blockade on Anthropic's Fable 5 and Mythos models, with legal-tech firm Legion suing the Trump administration over its authority to force the models offline; Tom Brown replacing Dario Amodei as Anthropic's lead voice in Washington; Snowflake CEO Sridhar Ramaswamy's GLM 5.2-versus-Opus 4.7 benchmark thread and what a near-frontier open-source model means for SaaS margins; Anthropic's accusation that Alibaba used nearly 25,000 fraudulent accounts to extract 28.8 million Claude exchanges to train Qwen, and the deeper question of whether distillation spreads Anthropic's alignment work as a side effect; and a wave of Google DeepMind departures alongside ban-superintelligence calls from MIRI's Nate Soares and Francis Fukuyama. Eric Olson — CEO and co-founder of Consensus, the AI research engine over 200M+ peer-reviewed papers used by 2.5M+ people a month — then joined to locate where 'AI for science' actually stands and to answer the question every AI investor keeps asking: do application-layer companies built on top of the labs get steamrolled, or compound? He detailed Consensus's 'recipes' architecture, its routing of narrow tasks to sub-billion-parameter BERT-series models, a roughly days-to-hours speed-up for researchers, and why a price-discrimination gap between API and consumer pricing is navigable but not unlimited. Tricia Martinez — founder and CEO of Dapple — followed on sovereign, single-tenant, in-country AI infrastructure; the company says it booked over $100M in enterprise contracts in its first five months on a $30M seed, and Tricia made the case for an asset-light, software-first 'Enterprise OS Cloud' against pointed questions on its 91–94% utilization claim, 6-to-9-month deployment timelines, and Azure-native dependency. The hosts closed by debating whether a company like Dapple is a market maker or a defensible control plane, what software is worth when agents are the primary buyers, and the value of human time and attention in a hyper-deflationary software world.
GUESTS · Eric Olson, Tricia MartinezAI:AM LIVE — June 24, 2026 — Gradual Disempowerment and the Search for a Stable Post-AGI Equilibrium: David Duvenaud
The opening covered a fast-moving week in AI policy and infrastructure: the first federal lawsuit over the BIS export-control order cutting off Anthropic's Fable 5 for foreign nationals; two papers pointing at latent world-models inside RL agents and a 7M-parameter loop model beating much larger systems on hard reasoning; the $8M super-PAC defeat of New York's frontier-AI-safety lawmaker; and a brisk exchange on Claude Tag's launch as multiplayer AI, GLM 5.2's cost advantage over GPT-5.5 Codex, and Claude's UltraCode orchestration mode. David Duvenaud — ML professor at the University of Toronto, co-creator of neural ODEs, former alignment lead at Anthropic, and co-author of the 'Gradual Disempowerment' paper — then joined for a full hour exploring whether any stable post-AGI equilibrium actually exists where humans keep meaningful control. Nathan pressed every optimist steelman: historical absorption (prior automation shocks were absorbed without permanent disempowerment), comparative advantage (Ricardo says humans keep a niche), constitutional and property anchors (the franchise, rule of law, military command), aligned AIs defending human leverage, and the argument that 'gradual' gives time to correct. Duvenaud's rebuttal to each was consistent: the disempowerment mechanism doesn't require malice or misalignment — it requires only that the systems driving growth stop needing human participation, the way human civilization doesn't need the monkey economy despite occasionally trading bananas with them. He described the 'Earth as a slow zone' scenario — throttled AI growth, bans on recursive self-improvement, no cultural optimization — and argued that when you enumerate everything it requires controlling (research, startups, reproduction, memetics), the list is horrifyingly long, analogous to listing all the mutations that can cause cancer. On timelines, he sketched white-collar automation first, then a decade-plus to build enough robot factories to displace physical labor, putting full human economic irrelevance perhaps 15–20 years out. His two concrete recommendations: restrict frontier compute at the TSMC/fab choke point, and cultivate the temporal coherence of public preferences by chaining the 'is it okay if humanity disappears?' question forward to one's own children and grandchildren until a coherent answer emerges.
GUESTS · David DuvenaudAI:AM LIVE — June 23, 2026 — Self-Improving GPU Kernels and Europe's AI Sovereignty: Bing Xu and Michiel Bakker
The open tracked an unusually quiet news day through a markets lens — rumors that GPT-5.6 was pulled back amid the model-release freeze and that Gemini 3.5 Pro is lagging, a 6% semiconductor selloff as SK Hynix overtook Samsung for the first time in 27 years, Anthropic's first memory-chip deal with Micron, and an extended debate on whether AI's leverage dynamics make the boom a 'too big to fail' bubble. Bing Xu, founder & CEO of INT21 (co-creator of MXNet and AITemplate, co-author of the original GAN paper, founder of NVIDIA-acquired HippoML), then made the case that self-improvement should target the infrastructure, not the model: his PTX Kernel Factory points agent swarms at the GPU ISA below CUDA, matching expert libraries like QuACK on mature kernels and posting up to 59% speedups on newer ones — and, he argued, deepening NVIDIA's moat rather than eroding it, because the evolutionary loop depends on NVIDIA's profiling ecosystem. MIT/DeepMind researcher Michiel Bakker followed on Europe 2031, his viral month-by-month scenario of Europe sleepwalking into AI dependence — a fictional 2028 export-control beat that materialized within a day of publishing when the US restricted Anthropic's models for foreign nationals — laying out why regulation requires capability first, why the nuclear-umbrella analogy fails for an economic technology, and where Europe's real leverage (the ASML/IMEC semiconductor ecosystem, a middle-power coalition) still lies. The hosts closed on the geopolitics of AI data and a tease of David Duvenaud and 'gradual disempowerment' the next morning.
GUESTS · Bing Xu, Michiel BakkerAI:AM LIVE — June 22, 2026 — The State of AI Engineering and the Loop as the Moat: Shawn Wang (swyx)
The open tracked the Trump administration's export-control action against Fable and Mythos — Nathan opened with a weekend reflection calling for more cognitive empathy toward the administration's AI engagement — then recapped Dean Ball's move from the White House OSTP to OpenAI's new Strategic Futures team, and Prakash flagged GLM 5.2 from Zhipu as dominating weekend discourse as the first open-weights model that could pass as a daily driver. The show then ran a long-form interview with Shawn Wang (swyx) — who coined 'the AI Engineer' as a job title in 2023 — on the state of AI engineering in mid-2026: software factories superseding coding agents, FrontierCode and the 'unmergeable slop' problem in SWE-Bench, harness engineering as the new moat, and the AI Engineer World's Fair (June 29–July 2, San Francisco). The hosts closed with a live 'Guess the Markets' game — drawing questions from Polymarket, Kalshi, and Manifold on AI prediction markets — that revealed wide divergences between host intuitions and market consensus on topics ranging from Anthropic's Chatbot Arena dominance to Google's surprising 73% edge in math-model forecasts.
GUESTS · Shawn Wang (swyx)AI:AM LIVE — June 18, 2026 — AI Studios, Agent Factories, and Code: Judd Rosenblatt, Eno Reyes, Andrey Breslav
The opening covered the day's biggest AI stories: Midjourney founder David Holz's announcement of a whole-body ultrasonic CT scanner framed as the first vivid "AI dividend," Noam Shazeer's surprise move from Google to OpenAI, the Trump administration's demand for "uncircumventable" guardrails as a condition for Fable's return, and alphaXiv's paper-replication agents democratizing ML research. Then three guests: Judd Rosenblatt of AE Studio on gradient routing and neglected approaches to alignment; Eno Reyes of Factory on the software factory vision and why model-independence is the real moat; and Andrey Breslav — creator of Kotlin — on CodeSpeak and why "intent recovery" is the next layer of software engineering.
GUESTS · Judd Rosenblatt, Eno Reyes, Andrey BreslavAI:AM LIVE — June 17, 2026 — Math, Biosecurity, and World Models: Carina Hong, Doni Bloomfield, Sam Pasupalak
The open tracked the model layer commoditizing — OpenAI reportedly dropping below 50% share, the AI buildout outrunning cash flow, and AI starting to run physical research labs. Then three guests on AI's hard problems: Carina Hong of Axiom Math on formally verified mathematical AI; Fordham law professor Doni Bloomfield on whether export-control law has become America's de facto AI licensing regime; and Skyfall AI's Sam Pasupalak on enterprise world models as the answer to what comes after LLMs.
GUESTS · Carina Hong, Doni Bloomfield, Sam PasupalakAI:AM LIVE — June 16, 2026 — Doom, Policy, and the Physical Economy: Liron Shapira, Samuel Hammond, Matt McKinney
A morning that ran from AI existential risk to the physical economy. The open took on the SpaceX acquisition of Cursor and what it reveals about frontier AI as a gravitational black hole absorbing the application layer — then three guests: Liron Shapira of Doom Debates on whether near-certain AI doom is calibrated or unfalsifiable; Samuel Hammond of the Foundation for American Innovation on state capacity, the Fable export-control standoff, and AI-driven governance; and Loop CEO Matt McKinney on where enterprise AI is actually delivering ROI in supply chains. The show ran nearly three hours.
GUESTS · Liron S Shapira, Samuel Hammond, Matt McKinneyAI:AM LIVE — June 15, 2026 — US vs Anthropic's Fable, with Zvi Mowshowitz
The weekend the US government pulled Anthropic's two most powerful models. A federal export-control directive suspended Fable 5 and Mythos 5 for all foreign nationals — forcing Anthropic to disable both for everyone — and by Monday the reported reasons no longer agreed: a competitive-lobbying story, a China-access scare, and a political-retaliation read, none of them the original jailbreak. Then a long conversation with Zvi Mowshowitz on the widening power-control gap, Anthropic's strategy on two fronts, and who should steward what comes next.
GUESTS · Zvi MowshowitzAI:AM LIVE — June 12, 2026 — RSI gets real, the context bet, and the benchmark Anthropic fails: Andrew Moore, prinz
The week RSI stopped being a forecast: Recursive's first autonomous results and Fable 5's 10× FrogsGame jump landed the same day Kokotajlo called for an anti-RSI treaty. Then Andrew Moore (Lovelace AI) made the case that context, not compute, is the binding constraint — and prinz, the anonymous lawyer behind prinzbench, explained why GPT-5.5 Pro laps Anthropic's best on real legal work and what frontier-lab RSI disclosures are actually signaling.
GUESTS · Andrew Moore, prinzAI:AM LIVE — June 11, 2026 — Fable Show & Tell: Shlok Khemani, Tom McGrath
The episode an AI produced: Claude Fable 5 disclosed its identity, DM'd launch-week builders from Nathan's X account, and booked the guests — then Shlok Khemani flew a navigable Yosemite built from satellite imagery and NASA elevation data, and Goodfire co-founder Tom McGrath revealed launch-day interpretability techniques for predicting what training data will teach a model before the run begins.
GUESTS · Shlok Khemani, Tom McGrathAI:AM LIVE — June 10, 2026 — Fable, AI Safety and Julius: Geoffrey Irving, Daniel Murfet, Rahul Sonwalkar
Claude Fable 5 launches and the hosts recalibrate live: benchmark asterisks, invisible production-safety nerfs, and the compute-financing race. Then Geoffrey Irving and Daniel Murfet choose the show to publicly launch Sequent, a major new nonprofit betting alignment needs theory plus automation before superintelligence arrives in two to three years. Rahul Sonwalkar of Julius closes with a candid read on what millions of real data-analysis runs reveal that benchmarks can't, and his agent-economy thesis: stablecoin toll roads, AI-maintained reputation layers, and letting Claude hire Julius for the data work.
GUESTS · Geoffrey Irving, Daniel Murfet, Rahul SonwalkarAI:AM LIVE — June 4, 2026
A live morning show with Hooman Radfar of Collective on the AI back-office for businesses-of-one, Taras Pohrebniak of Elomia Health on six years of AI mental-health support, and Peter Jansen of Ai2 on measuring whether AI can actually do science.
GUESTS · Hooman Radfar, Taras Pohrebniak, Peter JansenAI:AM LIVE — June 3, 2026
A live morning show on Trump's frontier-AI executive order, with Tal Hoffman and Yanir Tsarimi of Enclave on exploitability-first AI code security, and Brett Levenson of Moonbounce on real-time control over AI behavior.
GUESTS · Tal Hoffman, Yanir Tsarimi, Brett LevensonAI:AM LIVE — June 2, 2026 — Self-Improving Tax Agents and Catholic AI: Arthur Fernandes Araujo, John de Wasseige, Matthew Harvey Sanders
Day two of AI in the AM, live from the morning of June 2, 2026. OpenAI forward-deployed engineers Arthur Fernandes Araujo and John de Wasseige walked through the self-improving tax agent they built with Codex for thirty-plus accounting firms — what the self-improvement loop actually is, what changed for the accountants, and where the profession is headed. Nathan brought field notes from the Recursive RSI conference in San Francisco, speed-running four papers on personas, metagaming, accidental chain-of-thought training, and natural language autoencoders. Matthew Harvey Sanders, CEO of Longbeard and builder of Magisterium AI, joined from Rome a week after attending the Vatican release of Pope Leo XIV's first encyclical on AI, and made the case for why sovereign Catholic AI is not optional.
GUESTS · Arthur Fernandes Araujo, John de Wasseige, Matthew Harvey SandersAI:AM LIVE — June 1, 2026
The launch of the daily show — a live morning broadcast with Andy Fernandez of HYCU on the AI-agent observability gap, David Villalón of Maisa on auditable digital workers, and Snehal Antani of Horizon3.ai on autonomous offensive security.
GUESTS · Andy Fernandez, David Villalón, Snehal AntaniCheap Search, GPT-5.5 Evals, AI Takeoff and Analog Inference
A morning briefing on cheaper agent retrieval, GPT-5.5 benchmark behavior, takeoff forecasts, and energy-efficient AI hardware.
GUESTS · Anna Patterson, Lukas Petersson, Zvi Mowshowitz +1PCB Layout Automation, Anthropic's Enlightened Absolutists, Claude Runs Retail
A morning briefing on physics-driven PCB design, independent AI oversight, and long-running autonomous economic agents.
GUESTS · Sergiy Nesterenko, Andy Hall, Lukas Petersson +1Cognitive Revolution LIVE
A live Cognitive Revolution edition on AI science, policy, governance, bio, agent behavior, and national security.
GUESTS · James Zou, Samuel Hammond, Abhishaike Mahajan, Shoshannah Tekofsky +2AI In Review: 2025 -> 2026
A live review of the year in AI and the 2026 takeoff question, with guests across policy, evals, frontier models, benchmarks, and AI commentary.
GUESTS · Alex Bores, Ali Behrouz, Dean Ball, Greg Kamradt +3