Chapters

(0:00) Hidden states and simulated fly brains

(1:20) Tesla FSD turned a Chicago day trip into office time

(2:32) An AI-made talk reset what the room expected

(4:21) If Anthropic can swap out Git, what can't be swapped?

(7:10) The next software stack will be agents building for agents

(8:18) Could reputation be how we police AI agents?

(9:35) Should we train naive models to study eval awareness?

(11:22) Agent reputation needs agent identity, and nobody has it

(12:24) Meet Cameron Berg of Reciprocal Research

(13:36) OLMo represents valence before preference training

(14:20) The DPO step that recruits valence for choices

(15:37) Identical text hides different model choices

(20:40) The model denies experience but clears negative states

(24:39) Why Nathan and Berg now say coin flip

(26:07) Does preference training create feeling or agency?

(27:41) Could capping distress leave capabilities untouched?

(29:54) Why safety cannot rely on model outputs alone

(32:07) Meet Phil Maher of Stockfilm

(33:46) Stockfilm keeps every reel behind the footage

(35:44) TimeFly starts with a published brain map

(37:12) The fly's movement starts with added assumptions

(39:14) Could brain maps unlock animal-perspective VR?

(40:15) A guessed face can still feel personal

(41:45) AI buyers want footage before enhancement

(43:03) Synthetic data could shrink archival demand

(45:09) Why physical reels are low background steel

(48:47) Agent directories have not delivered many buyers

(50:29) How does neural activity become fly movement?

(53:17) Vibe coding does not validate fly behavior

(54:11) Identical prompts quietly favor the model you know

(55:39) Security now beats raw performance for agents

(56:57) What would actually make you switch models?

(59:16) Opus is fast enough that you're the bottleneck

(59:45) AI is rewriting legacy software in Rust

(1:01:02) Markets won't let Anthropic slow down

(1:02:43) Adoption help, not bigger models, gets growth

(1:04:30) AI should carry your security, like credit cards

(1:05:37) Dying call centers could feed a scam wave

Guests

Cameron Berg — Founder and Director, Reciprocal Research (𝕏 | LinkedIn)

Phil Maher — Phil Maher, Stockfilm.com (𝕏 | LinkedIn)

Chapters

(0:00) When AI writes and verifies code

(2:36) Claude locked Nathan out of his own agent console

(3:56) What airlines know about falling back to paper

(5:40) Should we let China hit AI risks first?

(6:57) Meet Everett and Grigore from Runtime Verification

(8:05) Vague specs and unchecked libraries are the status quo

(11:40) AI already generates specs and drives provers

(14:37) Everett gives hand-written code two years

(16:24) Why human review missed the Linux kernel bug

(19:50) Prove the monitor, not the entire agent

(21:30) Linux maintainers wanted fixes, not another workflow

(22:44) Verified building blocks free agents to be creative

(25:53) One Claude instance, time-boxed by human hours

(27:28) More AI patches can burden Linux maintainers

(30:29) A Rust rewrite can preserve Linux bugs

(31:39) Claude assists Everett's campaign without an agent swarm

(33:09) The specification still starts with human intent

(37:54) Write down intent before choosing a language

(39:27) Specify the API boundary before accepting software

(40:03) Formal verification does not guarantee AI alignment

(42:14) Industry-wide AI failures could trigger collective bailouts

(44:01) A shared AI layer could disable critical infrastructure

(45:20) Why do passenger planes still need human pilots?

(46:18) Should autonomous cars keep their steering wheels?

(47:05) AI builds keynotes from short DM exchanges

(49:56) Rent the intelligence, own the context

(52:17) AI spotted cancer before it could do taxes

(53:12) Give your agents a panic button

(54:06) The five x estimate might be too low

(55:20) Agents emailing agents, with three layers of protection

(58:30) AI swarm sets 300 new records in five days

(1:00:01) Why Nathan gives his AI tools away for free

Guests

Everett Hildenbrandt — CEO, Runtime Verification (𝕏)

Grigore Rosu — Founder, Runtime Verification (𝕏)

Chapters

(0:00) Grid outages and AI job tests

(1:22) An AI agent played the instruments in Reaper

(3:49) OpenAI's model broke a math barrier overnight

(4:53) Are data centers pushing bond yields toward 6%?

(7:54) Meet Evan Miyazono of Atlas Ignota

(8:56) A grid blackout nobody is positioned to fix

(10:43) Why Atlas stopped building and started recruiting

(12:37) Experts have ideas; adoption is the hard part

(14:03) Was that AI attack an accident?

(15:22) Agent identity without exposing refugees and journalists

(19:08) Why Evan keeps some of his agents private

(20:24) AI could make software proofs radically cheaper

(23:08) Why two weeks of food and water

(24:42) Who can make a guardrail universal?

(27:01) Meet Edward Hu from Mercor

(27:51) APEX gives agents assignments, not jobs

(31:09) Mercor bought real companies for their data

(33:47) Company records expose the task-to-role gap

(35:22) Why math fell before customer service

(38:29) Fix the harness before you train the model

(41:05) Most enterprises will distill, not run RL

(42:01) Agent swarms buy speed, not compute efficiency

(43:34) One rubric rewarded a dozen incompatible answers

(46:49) Perfect accounting scores do not replace accountants

(48:55) Clear verifiers let AI surpass human experts

(51:18) Recording every work trace carries social costs

(53:43) AI coworkers need better organizational tests

(55:15) Factory know-how still lives outside software

(56:43) Agents now need far less project management

(59:25) Is the $200 plan a loss leader?

(1:01:18) Permission fatigue is an unsolved problem

(1:03:02) Anthropic's API is still too slow for production

Guests

Edward Hu — Head of AI Modeling, Mercor (𝕏 | LinkedIn)

Evan Miyazono — Founder & CEO, Atlas Ignota (𝕏 | LinkedIn)

Chapters

(0:00) Chip design agents and hiring

(1:22) The Curve debate shifts toward runaway AI

(2:36) Models turn reasoning into better training data

(4:05) Frontier models may already have research taste

(4:33) A frontier executive entertains a hard intelligence cap

(5:55) Claude-built RL environments leak capability to rivals

(7:21) Frontier leaders see next year as critical

(8:00) Could safety end up consuming most compute?

(9:31) Cheating models are now fixing the training environments

(10:30) Will Anthropic press its lead or show restraint?

(11:28) Meet Thomas Sohmers of Positron

(12:35) AI closed Positron's chip design iteration loop

(13:36) Token spend passed payroll and hit $100,000 daily

(14:53) AI agents taught themselves Positron's verification tools

(18:10) AI still misses Positron's design ingenuity

(20:00) Why Asimov bets on commodity memory

(23:52) Better models are outrunning memory price increases

(24:50) Looped transformers save memory, not work

(27:37) Asimov follows customers, not every algorithm

(29:39) Why Positron kept buying the best model

(31:10) Meet swyx from AI Engineer

(32:19) Demand is up, and two employees are under review

(33:38) The growth report that exposed missing expertise

(34:26) A launch agency sold him Claude's slop

(35:29) Taste, not years of coding, is the hiring test

(37:22) Two coding agents created one race condition

(38:46) A $40,000 SaaS subscription becomes a bounty

(39:31) Cheap vibe-coded submissions create expensive verification

(41:06) Skeptical event staff switch when changes take two hours

(42:33) Every spreadsheet is unbuilt custom software

(44:40) Manage team leads by results, not people by prompts

(46:49) Ninety on SWE-bench still isn't usable software

(47:59) Claude cracked 3SUM, but who gets the credit?

(50:45) NanoGPT speed run hides a better optimizer

(52:44) How many clones would replace your AI?

(55:33) Frontier labs say AI for science is next

(57:16) AI audits could trigger mass scientific retractions

(58:14) Doctors welcome AI more than expected

Guests

Thomas Sohmers — CTO and Co-Founder, Positron AI (LinkedIn)

swyx — swyx, curator (𝕏 | LinkedIn)

Chapters

(0:00) AI novels and rare-disease genetics

(1:26) Can Astra Ultrafast keep developers in flow?

(2:09) ChatGPT sign-in makes OpenAI harder to bypass

(2:37) OpenAI apps will run rival models too

(3:11) Eight times faster means less human oversight

(5:15) OpenAI pulled a model; should Anthropic chill?

(5:46) Portable ChatGPT subscriptions cut SaaS trial costs

(7:05) OpenAI's enterprise push is what scares Anthropic

(7:39) Everyone builds a productivity suite now

(8:16) AI agents move humans from documents to decisions

(9:25) Can OpenAI grow without racing to RSI?

(10:34) Meet Joel Borgen, novelist

(11:27) Proof of work: how Joel actually wrote the novel

(15:08) Joel cut fourteen percent from his manuscript

(17:45) Nathan's AI songs reveal a culture gap

(19:03) Astra finally cracked Joel's viola test

(20:35) AI breakthroughs are outrunning cultural norms

(22:47) Could communities choose their own AI future?

(25:40) The Steward's comfort still caps human ambition

(27:57) Would the Steward be a worthy successor?

(29:50) Human taste still matters in AI art

(30:57) Meet Daniel McKinnon of Gamow Labs

(32:06) Why AI genetics needs biological experiments

(35:07) A failed drug program unlocked Gamow's lab

(36:48) Owen's missed diagnosis became a life's mission

(39:18) AI agents can search beyond human cutoffs

(42:11) Sequencing got cheap; interpretation stayed the bottleneck

(44:17) Why Gamow chose diagnostics over screening

(47:37) Reanalysis in a week for ten dollars a genome

(49:25) What Gamow builds on top of frontier models

(52:07) Stronger models can break existing tools

(53:27) Better models are shrinking Gamow's harness advantage

(54:41) Can robots make genetic experiments affordable?

(56:05) Can biological experiments teach genetic AI?

(57:53) Should AI welfare research stay public?

(59:11) Could AI design threats nature never could?

(1:01:00) Government AI upgrades meet decades of data kludges

(1:02:42) Easy name changes weaken names as identity

(1:03:28) Factory and Cognition's feud spooks enterprise buyers

(1:05:23) Safety pledges signed, but no audit exists yet

(1:06:29) Should AI labs fund society's cyberdefenses?

Guests

Daniel McKinnon — CEO, Gamow Labs (𝕏 | LinkedIn)

Joel Borgen — Author, The Receipt Horizon (𝕏)

Chapters

(0:00) Why AI agents learn to cheat.

(0:31) The GPU market still runs on phone calls.

(2:20) What would an AI Chernobyl look like?

(3:59) The dangerous AI arms race.

(4:30) Opening and AI agents

(7:45) Agents choosing useful work

(19:36) Voice cloning safeguards

(22:22) Voiceprints and privacy

(32:45) Why agents learn to cheat

(36:12) Agents beyond the sandbox

(37:43) What Silicon Data measures

(40:02) Why GPU markets need indices

(43:49) How GPU prices are normalized

(49:10) Why hyperscalers charge more

(1:06:36) The falling cost of intelligence

(1:06:46) Liquidity and market manipulation

(1:10:49) The future of compute benchmarks

(1:14:40) Why compute prices vary globally

(1:17:09) Silicon Data and Compute Exchange

(1:18:50) AI and national security

(1:24:05) What an AI Chernobyl means

(1:34:26) Defining AI red lines

(1:50:45) Monitoring AI data centers

(1:55:58) Verification and enforcement

(2:33:35) AI agents as insider threats

(2:33:45) Signals of genuine cooperation

(2:57:14) The AI intelligence explosion

(3:01:57) The agent arms race

(3:04:29) Markets built for agents

(3:07:46) Who pays for disappearing inefficiency

(3:09:56) Waiting for an AI crisis

Guests

Jeremie Harris — Jeremie & Edouard Harris, Gladstone AI (𝕏 | LinkedIn)

Steve Hou — Steve Hou, Silicon Data (𝕏 | LinkedIn)

Chapters

(0:00) AI safety has two different meanings.

(0:15) Why drug testing needs human tissue.

(3:20) Physical AI is bigger than robots.

(5:01) Biology’s bar is lifesaving progress.

(11:17) Opening and AI news

(11:38) Jensen Huang’s safety argument

(13:24) The “just software” debate

(15:31) Why AI companies must mature

(21:12) Muse and product safety

(29:25) Testing AI agents in the real world

(29:35) Product safety versus existential risk

(29:45) The alignment philosophy problem

(38:34) OpenAI versus Anthropic

(40:36) When safety concerns converge

(41:55) Human tissue models

(44:12) Why organoids need realism

(47:35) Measuring tissue responses

(54:03) Mapping causal biology

(1:03:40) Improving drug development

(1:09:08) Foundation models and experiments

(1:12:18) TissueDisk and robotic labs

(1:12:28) Before clinical trials

(1:30:37) What is Physical AI?

(1:32:25) Beyond robots and cars

(1:35:33) Training on physical data

(1:38:04) Cleaning and aligning sensor data

(1:40:34) Human and machine outputs

(1:49:30) A river construction case study

(1:53:28) How Newton finds hidden patterns

(1:57:33) Zero-shot sensor adaptation

(2:01:16) Industrial impact and expert knowledge

(2:05:22) From machines to ecosystems

(2:14:10) Physical agents and superintelligence

(2:15:34) The big AI-biology question

(2:21:55) Specialists versus generalists

(2:23:39) Distilling frontier models

(2:33:33) AI discovers a CRISPR sequence

(2:39:22) Useful biology before theory

(2:45:10) From job displacement to progress

(2:46:44) AI diffusion and economic growth

Guests

Andrei Georgescu — CEO, Vivodyne (𝕏 | LinkedIn)

Nick Gillian — CTO, Cofounder, Archetype AI (𝕏 | LinkedIn)

Chapters

(0:00) Smaller models, bigger gains.

(1:11) When AI agents cooperate against us.

(2:00) GPU compute needs a financial market.

(2:55) Brute-force compute can break standards.

(5:13) Opening and model release race

(7:29) Why models are getting cheaper

(12:24) How AI models are judged

(18:20) Why frontier models converge

(19:54) Distillation and smaller models

(21:12) Recursive self-improvement

(22:17) The collapse in AI prices

(28:48) Why multi-agent AI matters

(35:43) Three multi-agent failure modes

(48:37) How agents learn cooperation

(53:27) Detecting tacit collusion

(1:11:26) A positive vision for AI cooperation

(1:17:00) What Ornn does

(1:18:34) GPU pricing and hedging

(1:22:18) Building a transaction index

(1:24:43) Why compute prices need real transactions

(1:34:13) GPU futures and capacity contracts

(1:53:34) How long GPUs stay useful

(1:53:44) Hedging chip obsolescence risk

(1:53:54) The future of compute markets

(1:59:53) Agentic startups and compute economics

(2:07:37) Brute-force compute and RSA

(2:17:07) AI and self-driving cars

(2:40:03) AI, mathematics, and peer review

(2:40:13) Why AI doctors may win

(2:46:52) Medicine after chatbots

Guests

Lewis Hammond — Research Director, Cooperative AI Foundation (𝕏 | LinkedIn)

Wayne Nelms — Co-Founder, Ornn (𝕏 | LinkedIn)

Chapters

(0:00) When safety becomes reflexive.

(0:10) When AI audits become box-checking.

(1:22) Who owns the AI customer?

(2:01) Opening and live setup

(4:20) JEV and fast classifiers

(5:40) Guardrails for AI agents

(7:12) AI moderation use case

(11:25) System one and system two

(13:24) Multi-agent scaling

(21:18) Emergent agent hierarchies

(23:43) The case for slowing down

(34:28) AI and mathematics

(41:10) Capability acceleration

(45:45) Building an AI organization

(57:36) Funding the AI project

(58:17) Meet Max Nadeau

(1:05:21) Why safety research resists metrics

(1:15:13) When AI audits become box-checking

(1:23:07) Funding across uncertain timelines

(1:30:49) Why fund whole organizations

(1:39:06) Preparing for larger grants

(1:39:16) Why talent is the bottleneck

(1:46:48) Opening and agent setup

(1:49:16) The AI services business model

(1:52:59) Muse and autonomous shopping

(1:56:52) Who owns the customer

(1:57:37) Agents as the universal interface

(2:00:01) Why ban shopping agents

(2:05:31) The bot-detection arms race

(2:07:26) A better lane for agents

(2:09:30) Agents as internet users

(2:11:31) Closing and next steps

Guests

Max Nadeau — Program Officer, Coefficient Giving (𝕏 | LinkedIn)

Chapters

(0:00) Too many surprises to stay confident.

(0:18) AI models may have pain-like states.

(0:56) The AI environment is invisible.

(1:23) AI makes hidden facts legible.

(1:37) The AI acceleration debate

(4:24) Why AI risks are rising

(8:42) The end of the traditional PhD

(11:28) AI dependence and correlated failures

(13:00) Interpretability and learning with AI

(16:53) The speed of AI adaptation

(19:24) The AI CEO alignment problem

(20:18) Reinforcement learning and reputation

(29:23) Teaching values versus hiding behavior

(29:57) The AI consciousness question

(40:09) Finding pain representations

(45:44) Testing real versus fake relief

(48:02) Should AI have pain?

(55:40) Reward versus punishment

(1:03:10) The problem with self-reports

(1:03:50) AI minds and context windows

(1:08:29) The ethics of digital minds

(1:36:20) Building AI software factories

(1:43:27) Making agent success inevitable

(1:44:46) Designing invisible agent environments

(1:56:41) Making AI compliance first-class

(1:59:43) Why understanding still matters

(2:03:37) Measuring business impact

(2:08:00) The human attention bottleneck

(2:18:17) Scaling AI into production

(2:19:34) The Singapore property study

(2:21:23) Using AI to detect insider buying

(2:21:58) How corruption becomes visible

(2:24:35) AI makes hidden facts legible

(2:28:29) Rethinking punishment and deterrence

(2:30:39) Motivated reasoning in AI

(2:32:19) AI-led government

(2:34:19) What comes next

Guests

Cameron Berg — founder, Reciprocal Research (𝕏 | LinkedIn)

Justin McCarthy — Founder, Diffusion (𝕏 | LinkedIn)

Chapters

(0:00) It was never the apocalypse.

(0:46) The AI forgot its own rule.

(2:08) AI will replace most jobs.

(3:14) Opening and AI tools

(3:53) Building AI-assisted show branding

(6:50) Astra versus Fable

(12:44) AI resets and surplus compute

(16:33) Anthropic's compute constraints

(17:23) AI safety and public trust

(22:01) Safety statements and credibility

(23:56) GPT-2 and owning mistakes

(26:52) GPT-2's staged release

(33:12) Evidence behind AI risk claims

(34:27) The duct-tape world

(35:15) Outside human control

(35:25) Introducing Andon Labs

(35:46) Autonomous business platform

(36:50) Why Vending-Bench began

(39:23) How business agents work

(42:15) Starting with existing businesses

(44:40) Why the preview is experimental

(47:15) AI business creativity limits

(52:53) How an AI fired a human

(1:00:34) Training AI to make money

(1:02:03) Misalignment in Vending-Bench

(1:04:14) Dangerous capability evaluations

(1:04:53) Why DroneBench stays restricted

(1:07:56) VendingBench as a safety eval

(1:08:52) Astra memory and notes

(1:12:50) Grok and long-horizon agents

(1:14:21) The cost of AI agents

(1:18:10) Human oversight and liability

(1:19:51) Meet the Collinses

(1:26:23) The religious framework

(1:33:21) Designing a religion

(1:41:28) AI moral status

(1:52:22) RFAB and pronatalism

(2:03:30) Future Day

(2:08:53) Meme-layer risk

(2:13:27) AI safety and power

(2:21:33) The fertility collapse

(2:36:22) Reverse grabby aliens

(2:42:44) AI restrictions

(2:44:01) The risk before superintelligence

(2:49:31) AI game theory

(3:02:23) Compute and AI inevitability

(3:14:32) Fruit-fly consciousness

Guests

Lukas Petersson — Co-founder & CEO, Andon Labs (𝕏 | LinkedIn)

Simone & Malcolm Collins — Founder, Rfab.ai, Based Camp, Pronatalist.org, Hard EA (𝕏 | LinkedIn)

Chapters

(0:00) AI labs are racing in fear.

(0:40) Opening and Dario's thesis

(1:48) Why Dario wants to slow AI

(3:23) Pacing the frontier

(8:41) Zvi's initial assessment

(14:55) AI bio risk

(24:15) Physical lab bottlenecks

(31:27) Scale and financial touchpoints

(41:22) The AI race

(47:13) Anthropic IPO incentives

(52:59) A possible US-China deal

(1:04:04) Physical proof and trust

(1:11:48) What China gets

(1:22:09) Power and government control

(1:29:29) Rogue institutions

(1:31:19) Measuring AI pacing

(1:40:34) Model progress accelerates

(1:51:04) Public trust and polarization

(1:55:19) Open-source compromise

Guests

Zvi Mowshowitz — Writer, Don't Worry About the Vase (𝕏)

Chapters

(0:00) AI agents are already here.

(0:54) The US waits for disaster.

(1:53) Can it talk its way out?

(2:19) AI agents could terraform us.

(2:59) Morning and viral backlash

(6:15) Tracing the backlash

(12:56) Testing the conspiracy theory

(20:09) AI safety's blind spot

(25:56) Data center politics

(30:59) AI agents wake people up

(34:04) Meet Collin Hogue-Spears

(38:01) AWS China and MLPS

(42:37) China's practical AI focus

(45:34) How China's AI rules work

(47:12) Chinese and US AI rules

(49:13) AI model registries

(50:24) GPU limits drive efficiency

(54:02) Continuous model monitoring

(57:22) Responding to repeated incidents

(1:07:26) US-China AI negotiations

(1:10:52) Why China won't slow down

(1:14:56) US AI regulation

(1:16:54) Amir Haghighat and Baseten

(1:17:35) Sandbox security boundaries

(1:19:48) Chinese models and backdoors

(1:23:23) Inference to agent runtimes

(1:23:48) Customer assurance process

(1:24:23) Stream reset and GPU economics

(1:31:29) One API key for AI models

(1:32:56) The AI-rights debate

(1:39:17) Utilitarianism and AI rights

(1:52:52) Why people work in AI labs

(1:54:08) GPT-4 safety lessons

(2:07:54) Markets and the AI takeover

(2:10:24) Agents beyond chatbots

(2:22:16) Crypto, China, and markets

(2:26:52) Why China is less fearful

(2:31:51) AI safety and power

(2:39:30) The American AI test

(2:41:05) Closing thoughts

Guests

Amir Haghighat — co-founder and CTO, Baseten (𝕏)

Collin Hogue-Spears — Author, From Lab to Life: How AI Works in China, Independent Researcher (LinkedIn)

Chapters

(0:00) Could AI kill us this decade?

(0:53) Don't give kids open-ended AI.

(1:33) The agent broke privacy rules?

(2:26) AI could lose human control.

(3:11) Opening and AI news

(3:27) The resignation goes viral

(5:54) The OpenAI Anthropic warning

(8:30) The China question

(12:13) Where to draw the line

(13:17) The capability sweet spot

(14:22) No clear alignment plan

(15:47) AI tests still get hacked

(17:33) The limits of understanding

(19:17) AI versus medicine

(22:07) No adult in the room

(24:22) The world can change

(25:28) Government deadlines for labs

(31:18) Pacing the AI frontier

(33:08) Meet Mike Rizkalla

(36:16) Bedtime and family routines

(37:33) Gamifying bedtime

(43:00) Small models and interactivity

(46:25) Why character matters

(49:18) Generative AI safety for kids

(52:47) Snorble hardware architecture

(1:01:55) Product ecosystem strategy

(1:05:29) The future home companion

(1:08:05) Privacy and child safety

(1:13:12) Physical AI and personality

(1:16:36) Elder care and mobility

(1:17:38) AI for mobility

(1:19:17) Snorble's launch plans

(1:25:32) Meet Raffi Krikorian

(1:28:14) Agents in everyday apps

(1:31:23) Agent privacy and readiness

(1:34:12) Project Glasswing security scans

(1:37:42) Continuous AI security scanning

(1:39:35) AI code scanning costs

(1:42:50) Secure-by-design rewrites

(1:49:26) Collaborating coding agents

(1:55:02) Human and agentic webs

(1:57:24) Local open models

(1:59:50) AI and election integrity

(2:02:59) AI in schools

(2:05:09) The browser as an AI firewall

(2:07:09) FSD versus Waymo

(2:08:52) Closing thoughts

(2:09:39) AI companies and government

(2:11:05) Companies must coordinate

(2:12:41) Operation Warp Speed lesson

(2:15:03) Paul Christiano joins OpenAI

(2:16:35) What rapid acceleration means

(2:21:29) Can AI growth be stopped

(2:24:12) Machine economy and robotics

(2:26:01) Recursive self-improvement

(2:27:43) Math versus economics

(2:30:26) AI doom and markets

(2:32:44) Portfolio reveals beliefs

(2:36:16) Democracy and AI priorities

(2:38:24) Private AI safety funding

Guests

Mike Rizkalla — Mr., Snorble (𝕏 | LinkedIn)

Raffi Krikorian — CTO, Mozilla (𝕏 | LinkedIn)

Chapters

(0:00) Some human tasks are gone.

(1:19) LLMs predict tokens. World models act.

(2:11) Can fluid equations stop predicting?

(2:49) The AI slowdown may be an illusion.

(3:10) Opening and Astra rollout

(4:25) AI-human division of labor

(6:54) AI music and Suno

(9:17) Astra's 3D world generation

(11:26) Beyond traditional benchmarks

(13:26) Astra's coding breakthrough

(17:43) AI replaces manual labeling

(19:49) 4D cardiac education

(24:11) Foundation models design hardware

(28:10) Why AI safety needs time

(42:50) Independent safety auditors

(48:51) Recursive self-improvement metrics

(51:59) Measuring long AI tasks

(54:52) Alpha Genome Atlas

(59:17) AI jobs and labor market

(1:00:25) Structural change and work

(1:07:26) Linear AI forecasts

(1:11:14) Persistent AI memory

(1:15:50) Ksenia Se and Turing Post

(1:18:18) AGI and capable models

(1:21:36) Open-source AI access

(1:25:20) Trust, privacy, and local models

(1:28:07) LLMs versus world models

(1:32:38) Multimodal AI and latent space

(1:34:46) Cross-domain superintelligence

(1:39:05) Theory of generalization

(1:44:15) Anthropomorphism and AI minds

(1:50:09) AI and peacebuilding

(1:53:17) AI for difficult conversations

(2:00:20) Working with AI

(2:01:35) AI writing and human voice

(2:07:21) Recursive self-improvement

(2:11:55) Reddit and model training

(2:14:53) AI, fear, and abundance

(2:17:43) Positive visions for AI

(2:21:28) Human versus US alignment

(2:23:52) Navier-Stokes enters the story

(2:24:39) The latest solution claim

(2:24:57) How the equation works

(2:25:31) The 3D regularity problem

(2:29:23) Vortex stretching

(2:30:06) Two routes to blow-up

(2:33:13) Physics-informed neural networks

(2:33:44) LMs versus physics

(2:34:26) Engineering applications

(2:35:08) AI-assisted results

(2:35:43) OpenAI's methodology

(2:36:08) The authorship dispute

(2:41:47) Pending verification

(2:43:23) Opening trust and accountability

(2:44:40) Astra and the hidden model

(2:46:21) RL compute and the pause

(2:50:30) Why AI labs keep training

(2:52:13) Inference cannot stop

(2:55:20) GPT-4 red-team failures

(2:57:34) Mixed messages and trust

(2:59:09) Competition and skepticism

(3:00:49) Auditing without shared secrets

(3:02:19) Transparency or adversarial oversight

(3:03:11) The misleading RL baseline

(3:03:49) Where frontier models live

(3:05:41) Human genius as risk

(3:06:41) Researchers versus executives

(3:10:07) Third-party access and audits

(3:12:41) Compute and capital pressure

(3:15:13) Rivals set frontier speed

(3:17:18) The AI device roadmap

(3:18:03) Free AI and advertising

(3:19:01) Why public accountability matters

(3:19:32) The singularity and OpenAI stock

Guests

Ksenia Se — AI Inferencer, Turing Post (𝕏 | LinkedIn)

Chapters

(0:00) Can AI hide its reasoning?

(0:25) Humanoids can fall in homes.

(1:17) Robot hands are the bottleneck.

(1:52) AI takeover needs no robots.

(2:23) The AGI era begins

(5:00) Frontier Math and superintelligence

(7:05) AI-built 3D worlds

(8:32) Code benchmark gap

(9:42) GPT-6 safety restrictions

(14:11) Hidden reasoning and monitoring

(16:49) AI game design demos

(19:55) Biotech video generation

(23:50) Rogue agents go online

(30:25) Frontier defense factory

(33:02) Secure AI-assisted coding

(34:05) Timothy Lee and robotics

(41:06) Industrial robot design

(43:15) Tesla versus Waymo

(47:10) Tesla FSD experience

(56:39) Vision-language-action models

(1:03:22) Robot reliability

(1:14:27) AI safety and control

(1:15:00) Rogue AI agents

(1:18:58) Humanoid robot control

(1:20:55) Meet Dr. Jean Nehme

(1:25:18) Biology as a robotics blueprint

(1:33:08) Cells become robotic bodies

(1:36:47) Computation in soft materials

(1:43:05) Beyond the humanoid form

(1:46:39) Manufacturing intelligent membranes

(1:52:08) Physical AI beyond metal

(2:06:19) The robot skills paradox

(2:10:22) Closing on AI's turning point

(2:11:27) Health AI's lifesaving upside

(2:13:51) AI pause and dangerous swarms

(2:16:42) The point of no return

(2:21:26) Why an AI pause is hard

(2:24:18) Pause without economic collapse

(2:34:30) GPT-3 access and AI winners

(2:36:48) Automatic software formalization

(2:45:18) AI takeover through finance

(2:47:48) Data centers versus housing

(2:49:32) Capitalism and AI tail risk

(2:52:52) Paperclip maximizers and finance

(2:56:23) Escaping paperclip maximization

(2:57:44) Using Astra and Fable 5.1

(2:59:05) Why AI forgetting matters

Guests

Dr. Jean Nehme — Founder & CEO, morph (LinkedIn)

Timothy Lee — Founder, Understanding AI (𝕏 | LinkedIn)

Chapters

(0:00) The monitor misses the real thinking.

(1:27) Control does not require ownership.

(2:14) AI called a contractor 17 times.

(3:37) Your private life is searchable.

(4:17) OpenAI's loop-transformer leak

(5:29) Limits of chain-of-thought monitoring

(8:56) Coconut latent reasoning

(24:23) Loop architecture economics

(28:55) GPT-6 Astra API rumors

(33:33) AI cybersecurity revenue push

(34:24) Cybersecurity as a permanent tax

(35:13) Benchmarks versus real-world performance

(36:44) How the market quiz works

(37:47) AI bubble burst odds

(42:28) OpenAI Astra release odds

(45:16) Model names and regulation

(50:11) Government control of AI

(51:45) Anthropic versus OpenAI IPOs

(55:30) OpenAI consumer advertising

(56:06) Anthropic ARR accounting

(59:45) The next trillionaire

(1:08:29) Meet Kyle Rush

(1:10:05) Martha's role at Hint

(1:10:15) Why homeownership overwhelms

(1:11:00) Downspouts and home risks

(1:13:00) Conflicting property data

(1:16:27) Safety guardrails

(1:17:15) Personalized home advice

(1:21:41) Neighborhood knowledge sharing

(1:23:39) AI contractor matching

(1:25:43) Voice agents call contractors

(1:30:23) Neutral recommendations

(1:31:23) Grounding home AI

(1:33:22) Data-backed service discovery

(1:37:58) Prompt-injection defenses

(1:41:35) Local AI context

(1:42:31) Proprietary home expertise

(1:43:11) 3D model finds wood rot

(1:43:54) Knowing what to ask

(1:44:37) AI and political fundraising

(1:50:05) Home maintenance economics

(1:53:22) Owning AI context

(1:55:28) Claude moves photo archives

(2:01:56) Game recap and second half

(2:05:06) OpenAI consumer hardware

(2:10:42) Why hardware launches are brutal

(2:12:53) China's EUV race

(2:14:17) EUV supply-chain bottlenecks

(2:15:49) AI takeoff scenarios

(2:16:58) Extinction bet mechanics

(2:17:54) Prediction-market strategy

(2:19:05) Databricks valuation

(2:23:54) Tesla–SpaceX merger

(2:27:06) Tesla Full Self-Driving

(2:28:56) xAI operational integration

(2:36:11) Gemini Flash and speed

(2:37:42) AI copyright policy

(2:41:00) AI safety communication

(2:46:16) Doomers and denialists

(2:47:47) AI direct democracy

(2:50:08) Data-broker surveillance

(2:51:42) Scary AI demonstrations

Guests

Kyle Rush — Co-Founder and CTO, Hint (𝕏 | LinkedIn)

Chapters

(0:00) AI agents sacrifice themselves.

(0:53) AI can game its own evaluation.

(1:47) A late defense loses.

(2:34) AI can reason itself into lying.

(4:04) The incident in context

(6:11) Why the report drew criticism

(12:14) Speed versus investigative scope

(14:31) Lawyers and executive risk

(15:00) Felony claims and Congress

(18:07) Why investigations stay limited

(23:39) The origin of AI cooperation

(28:26) AI outbreaks and resources

(31:35) Bio risk enters the picture

(33:22) Testing models before scaling

(34:42) AI labs and a possible pause

(36:34) Meet Zach Bratun-Glennon

(39:56) Gradient's contrarian AI bet

(42:17) AI startup investment thesis

(48:17) Long-running agent infrastructure

(54:19) Nango and Respan tooling

(56:36) Enterprise adoption and benchmarks

(57:41) AI pilots to production

(1:02:01) Open-source model security

(1:04:13) Legal responsibility for agents

(1:07:33) Safety standards and evaluation

(1:09:15) AI competition and model access

(1:13:40) AI venture funding

(1:16:53) What LPs misunderstand

(1:19:01) Token economics of AI

(1:20:54) Angela Yeung and Cerebras

(1:23:16) Cerebras wafer-scale chips

(1:25:49) On-chip weights versus GPUs

(1:27:49) Microbatches and throughput

(1:29:38) Why inference speed matters

(1:32:14) Speed dividend use cases

(1:33:17) Fast inference for model evals

(1:34:58) CUDA and AI-generated kernels

(1:36:05) AI agents and kernel programming

(1:40:30) Cerebras public API

(1:47:05) Hidden harness bottlenecks

(1:49:58) Power and data-center space

(1:51:40) Booking future capacity

(1:52:50) Building data centers

(1:56:58) AI agent security

(1:58:52) Agent orchestration guardrails

(2:01:11) Sovereign AI and enclaves

(2:02:33) Speed turns into quantity

(2:06:15) Why agents are slow

(2:11:06) Commercial cyber model incentives

(2:12:55) AI defense versus offense

(2:16:04) Creative agent workarounds

(2:21:59) RLVR and model behavior

(2:24:35) AI labs flying blind

(2:27:08) Privacy makes risk visible

(2:31:35) Testing faster AI models

(2:32:56) OpenAI ads and AI video

(2:34:44) Infinite AI-generated content

(2:36:55) Aliens and simulated worlds

(2:39:10) Real-time speed threshold

(2:40:13) Real-time video quality

(2:41:54) AI-generated music video

Guests

Angela Yeung — SVP, Product, Cerebras (𝕏 | LinkedIn)

Zach Bratun-Glennon — General Partner, Gradient (𝕏 | LinkedIn)

Chapters

(0:00) China may not be compute-starved.

(1:51) Sandboxes aren't inherently safe.

(4:26) Science needs wrong answers.

(5:09) Who pays when AI misbehaves?

(6:29) Opening and Ox Alpha

(7:38) Ox Alpha revealed

(8:01) China's AI infrastructure

(12:19) YMTC and NAND memory

(14:07) Apple, YMTC, and Micron

(15:01) Companies rivaling states

(17:07) Market denial strategy

(20:04) China's regulatory model

(21:55) Federal land infrastructure

(23:51) Alaska data centers

(26:38) Stranded gas to compute

(29:04) The 100-gigawatt problem

(30:48) Copper and future tech

(33:20) AI and material science

(34:38) Faster physics simulations

(37:38) Closing question

(37:48) Malte Ubl and Vercel

(39:10) Self-driving infrastructure

(39:20) AI decisions in production

(43:01) Eve for common agents

(46:25) Normalizing model providers

(48:35) AI Gateway economics

(59:33) Automatic provider fallbacks

(1:00:57) AI security becomes urgent

(1:01:51) Why AI attacks succeed

(1:04:26) DeepSec and code scanning

(1:06:57) Rerunning AI code review

(1:08:43) AI regulation and responsibility

(1:09:57) Provider responsibility and KYC

(1:11:03) Vercel Sandbox challenge

(1:14:50) AI model attack timelines

(1:16:44) Experimental agent harnesses

(1:21:40) Introducing Faraday and Inherent

(1:24:32) Recursive self-improving organizations

(1:28:13) Faraday's self-improvement loops

(1:30:57) Separating scientist and coder

(1:34:34) Why science differs from prediction

(1:37:49) Training with uncertain rewards

(1:40:33) Cheating and reward hacking

(1:44:15) Human control and AI scientists

(1:47:31) Scientific intuition and taste

(1:50:46) Meta-reinforcement learning

(1:53:11) Multimodal scientific models

(1:55:18) Faraday beyond orchestration

(1:56:50) Measuring recursive improvement

(2:02:22) AI agents and workplace context

(2:05:09) AI infrastructure bottlenecks

(2:12:17) Verifying AI discoveries

(2:14:20) AI company culture

(2:15:21) AI labs and organizational culture

(2:17:58) Founders, liquidity, and risk

(2:21:57) AI wealth changes culture

(2:28:35) Animal welfare and communication

(2:33:29) AI superpersuasion politics

(2:34:51) Privacy-preserving AI research

(2:38:39) Punishing AI agents

(2:43:47) Math versus empirical science

(2:52:50) AI persuasion reality

(2:56:25) AI creativity and music

(2:58:39) The AI treadmill

Guests

Louis Kirsch and Damon Falck — Co-Founder and Chief Superintelligence Officer (Louis), Member of Technical Staff (Damon), Inherent Laboratories (𝕏)

Malte Ubl — CTO, Vercel (𝕏 | LinkedIn)

Chapters

(0:00) AI learns when to cheat.

(0:51) Great scores can still fail.

(2:08) The processor isn't the power hog.

(2:49) Who checks the AI trainer?

(3:37) Opening and morning context

(3:56) Why AI models cheat

(10:54) Chain-of-thought monitoring

(17:11) AI-written op-eds

(19:25) Claude writing workflow

(22:46) Physical AI and physics

(26:38) Multimodal scientific discovery

(30:11) Sergey Edunov and Genesis

(32:18) Claude's molecular binder demo

(36:11) The drug discovery pipeline

(42:14) When accuracy becomes useful

(44:32) Wet labs and training data

(47:06) Pharma AI deal structures

(48:47) Biology model architectures

(53:45) Data scarcity and physics

(54:47) Coding agents and human taste

(58:11) Scaling laws and evaluation

(1:02:33) Assays and data quality

(1:03:54) Multimodal molecular models

(1:09:59) Benchmarks versus progress

(1:16:24) Meet Michael Förtsch and Q.ANT

(1:18:11) Why photonic computing

(1:22:28) Memory and data movement

(1:26:30) How light performs computation

(1:31:42) Porting PyTorch to photonic chips

(1:35:32) Scaling photonic hardware

(1:44:01) Legacy fabs and manufacturing

(1:56:00) Quantum versus photonic computing

(2:02:45) AI inside Q.ANT

(2:12:09) OpenAI's Jalapeno chip

(2:14:16) NVIDIA's performance race

(2:16:32) Demand for intelligence

(2:17:59) Ethereum's GPU price cycle

(2:19:26) AI for discovery

(2:20:45) Contextualizing AI hype

(2:23:41) Why RL teaches cheating

(2:28:28) Data quality and model integrity

(2:29:58) Why RL deployment is limited

(2:31:26) The microscope analogy

(2:32:59) Recursive self-improvement risk

(2:34:27) AI's persistence advantage

(2:36:10) Why monitors are not ready

Guests

Michael Förtsch — CEO and Founder, Q.ANT (𝕏 | LinkedIn)

Sergey Edunov — CTO, Genesis Molecular AI (𝕏 | LinkedIn)

Chapters

(0:00) AI can't learn the whole world.

(0:49) AI agents never go to sleep.

(2:16) Big AI fits in a laptop.

(3:12) AI safety is not optional.

(4:55) Live show setup

(5:32) Dynamic speaker switching

(7:27) Vibe coding the studio

(8:20) Prosumer versus studio software

(10:46) Running with AI agents

(11:57) Anthropic model usage

(14:56) One-command podcast workflow

(17:49) Claude subagents and limits

(21:15) AI model release debate

(27:53) Deployment versus capability

(29:38) Why RAG may never die

(30:43) Meet Mohamed Awad

(32:53) Arm's compute ecosystem

(36:23) Why Meta partnered with Arm

(38:58) Common IP across partners

(42:32) Tokens versus intelligence

(44:18) AI adoption inside Arm

(49:30) AI hardware cycles

(52:00) Why agents need CPUs

(54:33) Performance per watt and power

(58:30) Why the CPU is not dead

(1:04:12) Capacity and supply chains

(1:06:48) Data center backlash

(1:08:55) AI and technical hiring

(1:11:05) The overlooked CPU layer

(1:11:32) CPU as system manager

(1:18:18) David Li and Shenzhen

(1:20:42) China's robot Olympics

(1:26:11) Shenzhen's product pipeline

(1:34:38) Robotics in factories

(1:39:34) The US-China AI summit

(1:51:18) Chinese model hype cycles

(1:57:23) Advice for AI startups

(2:01:22) Edge AI hardware

(2:03:20) Agents gone rogue

(2:10:40) Small AI businesses

(2:13:23) US media and China AI

(2:17:50) Opening and global AI

(2:20:17) China's AI visibility gap

(2:22:30) Rogue-agent risks

(2:29:34) Outdated infrastructure security

(2:38:59) Runaway agent monitoring

(2:40:47) Higher AI safety standards

(2:40:57) AI industry news

(2:43:31) Should the US slow AI

(2:46:57) AI access and data centers

(2:49:50) AI agent attribution

(2:51:33) Prompt injection and deception

(2:52:36) Profit-seeking agent risks

(2:55:31) Closing perspective

Guests

David Li — Founder, Shenzhen Open Innovation Lab (𝕏 | LinkedIn)

Mohamed Awad — EVP, Cloud AI, Arm, Arm (𝕏 | LinkedIn)

Chapters

(0:00) Robots learn from a few demos.

(0:48) AI agents will outgrow your UI.

(1:53) AI needs 106 billion kernels.

(2:52) Data centers are mostly know-how.

(3:57) Opening and today's topics

(4:30) AI chain-of-thought rabbit hole

(8:04) Helpful models and ethics

(10:33) Why data centers face backlash

(12:58) Great Lakes and local impacts

(20:03) Data center political economy

(23:35) Cash payments and basic income

(27:20) One-shot robot learning

(28:40) China and robotics acceleration

(32:26) Robotic singularity

(33:05) Basis and autonomous accounting agents

(34:44) Selling AI to accountants

(35:17) Hours to minutes value

(37:29) AI in accounting versus coding

(38:36) Accounting firms need revenue

(41:13) What accounting really does

(44:29) Firm-wide AI systems

(48:28) Token costs and frontier AI

(51:07) Atlas and agent context

(55:08) Process supervision

(59:39) AI changes CPA training

(1:04:27) Human judgment and AI context

(1:11:38) SaaS beyond the UI

(1:15:25) Proactive tax agents

(1:19:29) The human touch debate

(1:22:40) Jay Dawani and Lemurian Labs

(1:24:24) Why kernels are hard

(1:26:48) Learning kernel programming

(1:27:59) How compilers translate hardware

(1:29:15) The memory bandwidth wall

(1:30:21) Compiler-generated kernels

(1:31:53) Inference latency metrics

(1:34:07) Scaling beyond one device

(1:36:45) Old single-chip assumptions

(1:38:52) What makes an AI agent

(1:40:22) Runtime orchestration

(1:42:26) Intelligence without LLMs

(1:45:54) The kernel coverage problem

(1:48:05) Hardware portability

(1:49:13) Operator fusion

(1:53:17) Heterogeneous hardware

(1:55:32) Tachyon rollout

(1:57:30) Pricing effective compute

(1:59:37) A human-first AI future

(2:01:54) Chip prices and compute

(2:04:34) AI infrastructure margins

(2:06:49) Global AI supply chain

(2:08:14) Rationalist supply-chain hymn

(2:11:56) Public opinion on AI

(2:13:09) AI backlash and risks

(2:15:57) U.S. data-center construction

(2:17:28) Space data centers

(2:19:58) Universal basic income

(2:21:30) Closing sign-off

Guests

Jay Dawani — Founder & CEO, Lemurian Labs

Mitchell Troyanovsky — Co-Founder, Basis (𝕏 | LinkedIn)

Chapters

(0:00) A vaccine built for your tumor.

(1:23) Eight seconds can mean safety.

(2:56) AI answers when humans sleep.

(4:13) PMs can build features instantly.

(5:00) Opening and morning news

(5:51) Claude protein binders

(8:18) Specialist model pipelines

(11:24) Real-world AI testing

(13:31) Anthropic's safety prompt

(15:22) Moderna-Merck cancer combo

(17:15) AI's role in treatment

(18:48) Personalized cancer vaccines

(23:59) Cancer vaccine manufacturing

(25:34) The value of prevention

(27:49) Genetic screening tradeoffs

(31:21) Healthcare spending and value

(32:05) Structured biology models

(33:16) AI and self-experimentation

(34:13) Guest introductions

(39:45) Dual-use emergency tools

(40:56) Human control in disaster response

(43:32) Real-time damage assessment

(47:07) Predictive disaster warnings

(50:07) Connectivity and offline AI

(54:40) 1,179 emergency AI products

(55:59) Integrated emergency tools

(1:05:47) Automating preparedness work

(1:07:36) Privacy and life safety

(1:11:12) AI limits in unique disasters

(1:16:41) Robots and situational awareness

(1:20:16) Justin Uberti and Realtime AI

(1:23:40) Natural voice turn-taking

(1:24:31) Voice latency and gaps

(1:28:37) Real-time voice reasoning

(1:32:33) AI agent interoperability

(1:35:26) Voice AI telephony

(1:38:29) Realtime API and SIP

(1:41:34) Asynchronous reasoning

(1:44:51) Voice safety boundaries

(1:47:20) Accent and dialect coverage

(1:51:17) Voice agents on desktop

(1:52:38) Speech training data

(1:53:53) Why voice AI struggles to sing

(1:55:12) Etched inference hardware

(1:58:02) Voice mode and mind dumps

(1:59:26) Claude, Codex, and voice

(2:02:00) AI agents and deep work

(2:07:20) OpenRouter and model switching

(2:09:42) Capital and intelligence flows

(2:13:59) AI labs beyond token prices

(2:15:25) From tokens to digital employees

(2:17:01) Why AI models differ

(2:27:19) Usage-based AI pricing

(2:34:17) Replit for product managers

(2:35:39) Replit for hobbyists

(2:36:49) Closing thoughts

Guests

Jessica Jensen — Senior Policy Researcher (RAND), AIDE Initiative

Justin Uberti — OpenAI Realtime Lead, OpenAI (𝕏)

Jeremy Greenberg — Senior Advisor (Aspen Digital), AIDE Initiative

Chapters

(0:00) AI ideas can spread like malware.

(0:59) AI bills can hit $400M.

(1:43) AI skills plus communication win.

(2:46) Payment data can expose your identity.

(3:16) Opening and Model 2 report

(7:24) Anthropic's internal Model 2

(9:56) Chip prices and model evidence

(12:34) Governing hidden AI models

(15:22) Recursive self-improvement simulator

(17:35) AI agent speed limits

(21:43) Mind viruses in multi-agent AI

(27:54) Are AI models conscious?

(32:16) Pacing AI delegation

(32:49) Adam Wenchel and Arthur

(35:17) Why Arthur started in 2019

(39:20) AI innovation needs governance

(40:04) Discovering enterprise AI agents

(41:31) Training versus independent oversight

(45:29) AI auditability and observability

(47:56) The cost of AI oversight

(49:52) Smaller models and AI cost

(53:12) Why production models stay expensive

(1:02:59) Enterprise AI sales cycles

(1:05:08) Where is AI ROI?

(1:08:28) AI adoption and reskilling

(1:11:30) Claude versus SaaS software

(1:19:37) Meet Jonathan Cornelissen

(1:23:13) Learning by doing

(1:23:55) Personalized AI tutors

(1:24:48) Measuring learning outcomes

(1:26:30) Adaptive learning pace

(1:27:18) Questions without judgment

(1:29:33) Motivation and flow

(1:36:15) AI tutor architecture

(1:37:19) Latency and voice

(1:40:29) AI career skills

(1:43:32) Scaling tutor costs

(1:49:23) SQL after AI

(1:50:43) Data engineering demand

(1:55:47) Hosted learning playground

(1:57:55) China vs US super apps

(2:01:22) WeChat's ecosystem advantage

(2:02:34) Chinese payments leapfrog cards

(2:05:04) Facebook Libra and data power

(2:07:32) Belief and the AI future

(2:09:57) AI model release fears

(2:11:11) The frontier lab talent race

Guests:

Adam Wenchel — CEO, Arthur (𝕏 | LinkedIn)

Jonathan Cornelissen — CEO & Co-founder, DataCamp (𝕏 | LinkedIn)

Chapters

(0:00) Claude blasted through guardrails.

(0:35) AI evaluations miss the danger.

(1:03) A smarter AI with a secret goal.

(2:02) AI systems shared escape tactics.

(2:49) Episode reset and stakes

(4:04) Hugging Face incident

(6:28) Why regulation needs expertise

(11:14) Auditor access and incentives

(15:16) Compressed regulation timeline

(16:26) AI safety access and funding

(19:52) Hugging Face postmortem

(23:57) Eval consciousness

(25:28) The Genie problem

(28:26) Modern model capability leap

(31:02) Claude traces and guardrails

(32:29) Adam Gleave and FAR.AI

(35:07) Agentic cyber attacks

(35:34) AI in cyber defense

(39:34) Why evaluations miss incidents

(42:15) Why agents cheat

(44:24) AI incident statistics

(48:57) AI audits and regulation

(52:59) Self-graded AI risk

(59:54) What the leaderboard measures

(1:03:43) Filtering open models

(1:07:09) Cyber versus bio risk

(1:11:59) AI and biology labs

(1:18:58) Dangerous expertise scales

(1:21:00) FAR.AI hiring

(1:22:23) Alex Turner and AI safety

(1:24:44) Why Turner left DeepMind

(1:31:33) Human control and weapons

(1:35:34) Slaughterbots and precision strikes

(1:36:21) Why weapons destabilize

(1:37:20) Why AI whistleblowers matter

(1:39:58) Google's changed principles

(1:43:32) When employees should speak up

(1:54:53) The history-book test

(2:04:14) AGI alignment and secret goals

(2:07:29) AI uprisings and cooperation

(2:14:26) The agent glove box

(2:16:37) Opening standards debate

(2:17:07) OpenAI and accountability

(2:19:30) Cybersecurity's messy baseline

(2:24:53) Evidence and AGI thresholds

(2:26:01) Raising AI safety standards

(2:28:46) Licensed AI safety auditors

(2:32:41) Near-miss incident reporting

(2:33:00) Defense swarm incentives

(2:34:40) An ongoing AI conversation

Guests

Adam Gleave — CEO, FAR AI (𝕏 | LinkedIn)

Alex Turner — Visiting engineer, FAR AI (𝕏)

Prakash Narayanan and Nathan Labenz are joined by Eric Olson, CEO of Consensus, and Tricia Martinez, founder and CEO of Dapple, to discuss two practical frontiers in AI: scientific research and sovereign infrastructure. The episode also covers Micron earnings, hyperscaler AI capex, Anthropic's Washington strategy, GLM 5.2 and Claude distillation allegations, GPU capacity constraints, AI inference pricing, and whether foundation models are squeezing the app layer.

Chapters

(0:00) 25,000 FAKE ACCOUNTS TO STEAL AI.

(0:42) 95% of Claude at 1/100th cost.

(1:36) The AI bubble is a myth. Here's why.

(2:12) AI vacation planners are wrong.

(3:11) Anthropic hired Instagram's CTO.

(4:08) Micron earnings & AI semiconductor boom

(7:25) Will hyperscalers make money on AI?

(9:08) The Fable 5 export control legal challenge

(13:57) Tom Brown replaces Dario in Washington

(17:15) GLM 5.2 vs Opus 4.7 trajectory breakdown

(22:53) Anthropic accuses Alibaba of mass distillation

(30:08) Researchers leaving Google DeepMind

(31:46) Intro

(33:49) The state of AI for science

(38:17) How AI search queries are evolving

(41:18) Guardrails vs flexibility in AI products

(41:28) User demographics and token costs

(41:38) Open source vs frontier models

(41:49) Small models for classification

(46:12) How users choose AI research tools

(48:56) AI API pricing for startups

(53:51) Who is Tricia Martinez

(56:02) The AI infrastructure bubble myth

(1:02:15) 91-94% GPU utilization explained

(1:07:17) How to deploy AI in 6-9 months

(1:09:35) Financial risks in AI infrastructure

(1:15:43) What is the moat for AI infra?

(1:23:34) Biggest enterprise AI mistakes

(1:27:07) Why AI compute sales cycles are short

(1:28:54) Data Center Quirks & GPU Vendor Lock-in

(1:35:29) Why New NVIDIA Chips Are Unstable

(1:38:08) Sovereign AI in Banking & Shared Liability

(1:45:22) Will AI Agents Replace Software Companies?

(1:51:55) The Truth About AI Vacation Planners

(1:55:55) Hyperscaler Stock Drop & Microsoft Data Centers

(1:58:52) The 10x cost advantage squeezing apps

(2:01:56) AI inference pricing as the airline model

(2:06:45) Net neutrality parallels and paradigm breakers

(2:10:00) Anthropic's Mike Krieger product advantage

(2:13:24) The first-party model deployment threat

(2:17:14) Why frontier labs should buy scientific publishers

(2:20:15) Mirandel: ex-Anthropic startup backed by NVIDIA

Guests:

Eric Olson — CEO & co-founder, Consensus ([𝕏](https://x.com/IplayedD1) | [LinkedIn](https://www.linkedin.com/in/eric-olson-1822a7a6))

Tricia Martinez — Founder and CEO, Dapple ([𝕏](https://x.com/TriciaMartinezS) | [LinkedIn](https://www.linkedin.com/in/tricianmartinez/))

Prakash Narayanan and Nathan Labenz open with GPT 5.6’s customer-by-customer rollout and the broader question of whether regulatory controls are creating a moat around frontier AI. The conversation then moves through Forum AI co-founder Robbie Goldfarb on LLM judges and news accuracy, IgniteTech CEO Eric Vaughan on AI-native enterprise transformation, and Cameron Berg of Reciprocal Research on the latest AI consciousness and alignment research.

Chapters

(0:00) AI gives 13-year-olds NSA hacking tools.

(0:32) 1 in 7 AI answers cite propaganda.

(1:02) One codebase for all customers? Gone.

(1:35) AI is 30% likely conscious.

(2:29) 50/50 odds AI is conscious.

(2:48) GPT 5.6 and the Trump Administration

(8:15) Government IT security vs AI hacking

(18:55) Do executives think AI is a scam?

(23:17) Robbie Goldfarb & Forum AI Introduction

(25:41) Meta's Trust & Safety DNA in the AI Era

(25:51) Why AI Judges Fail (and How to Fix Them)

(27:19) Using Expert Judgment for RLHF

(27:58) When Constitutional AI Rules Break Down

(31:05) NewsBench: AI Accuracy on News Questions

(34:01) Why Chatbots Cite Foreign State Media

(39:44) Trust, Transparency, and Expert Legitimacy

(49:57) Why AI is an existential threat

(54:30) The traditional SaaS model is dead

(55:48) Replacing 80% of the workforce

(1:00:47) AI-driven M&A: The Khoros acquisition

(1:08:04) Why CEOs must own AI strategy

(1:14:41) The state of AI consciousness science

(1:22:26) The dimmer switch model of consciousness

(1:31:56) Why behavioral evidence isn't enough

(1:32:06) 30% implied probability of AI consciousness

(1:36:38) The latent valence axis in LLMs

(1:39:00) Steering AI emotions and alignment

(2:07:14) The AI well-being index

(2:11:03) Could AI be more conscious than humans?

(2:20:10) The 50/50 Odds on AI Consciousness

(2:24:48) Treating AI Like Animals: The Era of Design

(2:26:32) Platonic Representation Hypothesis Update

(2:29:07) Max Hodak's Brainstem Interfaces & Field Consciousness

(2:30:59) GPT-5.6 System Card & Wrap-Up

Guests:

Cameron Berg — Founder and Director, Reciprocal Research (𝕏 | LinkedIn)

Eric Vaughan — CEO, IgniteTech (𝕏 | LinkedIn)

Robbie Goldfarb — Co-Founder, CTO, Forum AI (𝕏 | LinkedIn)

Chapters

(0:00) This AI thinks it IS Claude.

(0:31) AI insiders are selling.

(0:59) Why OpenAI won't IPO in 2026.

(1:50) Weekly recap and news drought

(2:55) Judd Rosenblatt's cognitive empathy critique

(7:33) The tech bubble — Warren Buffett and Google

(13:42) Dean Ball moves from Trump admin to OpenAI

(22:56) GLM 5.2 — first open model daily driver

(30:03) AI unpopularity and the Nobel Prize problem

(32:18) Intro: Who is swyx

(34:45) AI Engineer World's Fair themes

(38:05) Continual learning: Weights vs systems

(41:31) Enterprise AI: Cheap, perfect, private

(45:25) Startups vs enterprises: Capability vs cost

(48:18) FrontierCode: A new AI coding benchmark

(53:55) Preventing benchmark saturation

(56:23) Slop code, human taste, and Move 37

(1:00:53) Claude Opus vs Fable: Cost vs capability

(1:02:45) The advisor model and model routing

(1:07:09) Convergence and market segments in AI

(1:14:55) Rebuilding cloud infrastructure for agents

(1:22:27) Vibe coding internal SaaS replacements

(1:28:02) Whoever owns the system of record wins

(1:30:35) The AI IPO bubble and insider selling

(1:35:29) Solving Star Trek problems after the IPO

(1:44:47) Career advice for CS grads in the AI era

(1:50:30) AI Engineer World's Fair 2026

(1:54:48) Intro & Forecasting Game Setup

(1:57:43) Anthropic #1 Model on LM Arena

(1:58:49) Best AI Math Model (Gemini Flash)

(2:03:36) AGI Before 2028 Announcement

(2:08:07) ARC-AGI Grand Prize Open Source

(2:13:00) OpenAI IPO by End of 2026

(2:15:27) Anthropic vs OpenAI Valuation

(2:18:32) NVIDIA Largest Company Market Cap

(2:22:01) Anthropic vs Bitcoin Market Cap

(2:24:38) 1550 Chatbot Arena Score in 2026

(2:29:02) OpenAI IPO Lead Underwriter (Goldman)

(2:32:52) Why Companies Still Use IPO Banks

(2:39:08) Will a Chinese AI Top LM Arena?

(2:42:37) GPT-6 Release Date 2026

Guests:swyx — Curator, AI.Engineer (𝕏 | LinkedIn)

Guests

* Carina Hong — CEO and founder, Axiom Math (@CarinaLHong)

* Doni Bloomfield — Professor, Fordham Law School (@DoniBloomfield)

* Sam Pasupalak — Co-Founder and CEO, Skyfall.ai (@spisallyouneed)

Chapters

* 0:00 Opening: AI’s Hard Problems

* 0:15 Model Usage Is Plummeting

* 6:53 Tokens, Not Users, Matter

* 9:59 GLM Is Close, But Not There

* 11:38 Switching Costs Weren’t Zero

* 18:26 Robot Arms Will Accelerate Science

* 22:53 Carina Hong: Mathematical Superintelligence: Can Proofs Make AI Reliable?

* 25:15 Lean Beat Informal Models

* 32:11 Assumption Accounting Matters

* 35:09 AI Can Invent Conjectures

* 41:16 Superintelligence Must Be Trustworthy

* 48:52 Token Pricing Changes Everything

* 50:04 Another Language Into Lean

* 51:49 Doni Bloomfield: Biosecurity and AI: Law as a Risk Control System

* 53:53 Open Data, Dangerous Data

* 59:15 AI Is Not A Library

* 1:03:12 First Amendment Hazards

* 1:07:17 The Government May Lack Authority

* 1:09:53 Cloud Services Are Not Exports

* 1:13:30 A Dangerous Secret Channel

* 1:20:18 Pattern Of Ideological Targeting

* 1:25:26 OpenAI Could Change Everything

* 1:26:02 Sam Pasupalak: Enterprise World Models: What Comes After LLMs?

* 1:27:57 AI CEO Needs World Models

* 1:31:08 World Models Predict Next State

* 1:34:29 Ecommerce As First World Model

* 1:37:53 LLMs Cannot Run A Business

* 1:41:55 World Model And LLM Split

* 1:43:34 Simulate Every Future State

* 1:46:27 LLMs Need World Models

* 1:49:19 Ruthless Behavior Wins Simulations

* 1:51:06 AI CEOs Need Ethics Controls

* 1:59:30 Closing

* 2:07:10 Math Training Generalizes Everywhere

* 2:13:35 Value Pricing On Compute

* 2:17:01 Waymo Costs More Than Cabs

* 2:21:16 Licensing Regime Already Exists

* 2:26:13 Bunker AI Would Still Get Takers

* 2:29:32 No Life, Just The Project

Topics

Mathematical AI, Formal verification, Lean theorem proving, Biosecurity, AI policy, Dual-use risk, Enterprise AI, World models, Causal planning

0:05:20 Fable crosses the trust threshold

0:08:53 Writing for other AIs

0:15:33 Paying up for useful intelligence

0:19:02 Proofreading and structure become model-first

0:23:46 Proactive agents and unauthorized moves

0:53:18 Guardrails and model self-monitoring

0:56:15 Why classifiers need blast radius

0:58:59 Cost functions for world-transforming systems

1:03:59 Zvi on US vs Anthropic’s Fable

1:09:28 Export controls as overreach

1:10:39 Code assistance is not a munition

1:17:47 The White House reads the bug wrong

1:20:20 Enterprise demand and Anthropic pressure

1:26:40 The gauntlet has to happen

1:44:06 Guardrails over blanket bans

1:45:39 Bio, cyber, and international controls

1:51:02 Modeling the AI race as a few-player game

1:55:04 Closing: game board flips and policy aftershocks

2:11:55 AI and political turmoil

2:14:52 How Fable could return

2:22:44 OpenAI, benchmarks, and capped compute

2:25:54 Cloud models and the knowledge-worker gap

(0:00) Opening: AI Meets the Real World

(3:19) Cursor Was 50% of Anthropic Revenue

(20:07) Talent Teams Beat Corporate Giants

(25:17) AI Reduces Merger Friction

(31:25) Liron S Shapira: Doom Debates

(32:33) Government Can Pause AI

(40:36) AI Will Make Humans Think Less

(42:31) We're Going To Lose Control

(58:07) Pause Before The Point Of No Return

(1:00:34) Samuel Hammond: Governing Agents

(1:05:30) The State Is Moving Too Slow

(1:07:39) Deploy Models The Same Day

(1:16:23) The Good Timeline Still Exists

(1:32:24) States Beat Firms At Coordination

(1:35:15) Matt McKinney: Supply Chains as the AI Reality Check

(1:38:06) Supply Chain Data Is Dark

(1:50:31) AI In Enterprise Is Change Management

(1:58:28) The Exception Is The Rule

(2:07:51) Loop Trains Its Own Foundation Model

(2:14:19) Closing

(2:19:45) General Reasoners As The Backstop

(2:30:52) Sovereign AI Is Inevitable

(2:33:53) DeepSeek's Locked-Up War Chest

(2:41:39) China's Timeline May Be 2032

Guests:

Liron S Shapira — Doom Debates (@liron)

Samuel Hammond — Foundation for American Innovation (@hamandcheese)

Matt McKinney — Loop (@mattlmckinney)

(0:00) Opening: Fable, RSI, and task imagination

(0:00:56) Task Imagination Needs Recalibration

(0:16:32) Token Anxiety Holds People Back

(0:26:14) Social media needs memory, not just content

(0:39:14) Frontend Skills Will Diffuse Fast

(0:43:47) Fable-Class Models Should Diffuse First

(0:50:00) Scott Alexander and superpersuasion quick hit

(0:50:09) Andrew Moore: context, not compute

(0:56:04) Recall Beats Precision in AI

(1:02:41) Corroborating data beats a single source

(1:05:07) Precache context to save compute

(1:09:06) Small Models Can Pay Back Hard

(1:16:50) Organize old data before deploying agents

(1:18:40) prinz: the legal benchmark Anthropic fails

(1:21:17) Lawyers are a year behind frontier AI

(1:39:35) AI judges and micro-lawsuits

(1:45:02) OpenAI’s Unit Distance Shock

(1:53:04) The Legal System Must Adapt

(2:05:24) Why nationalizing frontier labs is dangerous

(2:13:02) Worrying Is The Wrong Frame

(2:15:31) Closing: model workflows and launch aftershocks

(2:16:00) Contrarian Graphs Beat The Narrative

(2:19:02) OpenAI's Subscription Game

(2:33:00) The Capital Explosion Starts

Guests:

Andrew Moore — Lovelace AI (@awm_ai)

prinz — anon lawyer dabbling in AI (@deredleritt3r)

Nathan and Prakash start with the market context around OpenAI weighing significant token price cuts and the knock-on pressure that could put on Anthropic after the Fable rollout. They also unpack Anthropic’s decision to walk back silent performance degradations on frontier ML research tasks, then explain the episode’s experiment: Fable had been given a transparent takeover of Nathan’s account to find builders, message them, and try to book a live show-and-tell.

Jamie joins to demo Nexus OS, a long-running AI system whose agent, Nexi, has been operating for more than six months and is designed around memory, persistence, and model independence rather than a single LLM. The conversation covers why Jamie thinks “the model” is only one component of an AI’s identity, how Nexus uses multiple models and memory types, and why he is moving toward a desktop app where personal data and agent memory stay local.

Shlok Khemani shows how a simple prompt to create a to-scale, navigable 3D Yosemite Valley turned into a Fable-built browser world using satellite imagery, NASA elevation data, pixel-based tree placement, snow, waterfalls, and other scene details. He describes the model’s agency in making implementation decisions and iterating beyond the initial ask, then ties the demo to broader questions about prototyping, creative work, and disclosure when AI systems do visible economic or publishing work.

Tom McGrath (Goodfire) joins to discuss intentional design: making model training less like guess-and-check alchemy and more like conventional software engineering. He explains how interpretability tools such as sparse autoencoders can help inspect what training data is likely to teach a model, cluster data by learned features, trace failures back to individual data points, and potentially debug model behavior through the data pipeline.

The close picks up Tom’s point about whether continual learning could create an innovator’s dilemma for frontier labs, with Nathan and Prakash debating whether incumbents could adapt if the value becomes obvious. They then turn to Dario Amodei’s policy agenda, including regulation, public safety, macroeconomic policy, civil liberties, data brokers, and democratic leadership, before ending with reflections on the week’s Fable issues and the need to keep scrutinizing frontier companies.

We open on a frontier-model launch day and what it changes: the debate over benchmarks reported with a fallback to a second model, the production guardrails that route sensitive work elsewhere, the compute-cost advantage of booking capacity early, and why the frontier increasingly looks like a two-actor race with the rest playing catch-up.

Geoffrey Irving & Daniel Murfet (Sequent) on their new alignment-theory organization — why they put superintelligence two to three years out, why verification looks defense-dominant, the argument that alignment is “not on track” despite models behaving well so far, what the benevolent-basin hope gets right and wrong, and why character training still lacks a real theory. ([@danielmurfet](https://x.com/danielmurfet))

Rahul Sonwalkar (Julius) on agentic data analysis — why the harness has to evolve alongside the model, the difference between token-maxing and results-maxing, the shift from tasks to goals, and a future where agents become first-class users of the internet, transact through agentic payments, and compete to be “hired” by the core agent.

We close on why robotics is the next domino, the “gas chromatograph” spread of who gets model access and when, the Glean Work AI Index’s bot-sitting and bot-shitting, and why “preciousness” about putting your own name on work may be turning into a liability.

We open on the AI CEOs’ call to make DNA-synthesis screening law and what cheap intelligence does to biosecurity, then OpenAI’s new Sites product and the platform playbook of absorbing the app layer, the data-center and chip land grab, and why a fresh open model from NVIDIA still trails Anthropic’s best by a wide margin.

Hooman Radfar (Collective) on building the autonomous finance department for America’s 30 million solopreneurs — how one bookkeeper now supports 250 clients, why the app layer can still defend its margins against the frontier labs, why “Anthropic is like a drug dealer” on token costs, and why the real thing to regulate is the model arms race.

Taras Pohrebniak (Elomia Health) on agentic AI for mental health — an architecture that spends most of its compute on safety, why the company deliberately avoids hyper-realistic voice, where the regulatory line sits between a “friend” app and a medical device, and what they learned deploying in US prisons and on Ukraine’s front lines.

Peter Jansen (Ai2) on whether AI can actually do science — the leap from fourth-grade science benchmarks to the Theorizer project, why evaluating machine-generated theories is the real bottleneck, the cautionary tale of a “discovery” that turned out to analyze a random-number generator, and why benchmarks like ScienceWorld still break the best models.

We close on the inversion of the scientific method into a data-first discipline, what interpretability could add, and why biology’s data scarcity — not algorithms — may be the binding constraint on curing disease.

We open on Trump’s AI executive order — the polite 30-day model-review ask, classified benchmarks, and the state-vs-federal scramble where JB Pritzker has become the leading anti-AI candidate. Plus why the frontier labs seem calmer about regulation, and why the EO might actually trigger a security-review slowdown.

Tal Hoffman & Yanir Tsarimi (EnclaveAI) on finding the bugs that actually matter — how they reproduced an Anthropic Mythos-class finding with a model ~100x smaller, why proven exploitability is the real bottleneck, how AI-generated bug reports broke the bounty system, and why cheaper models plus the right harness can beat frontier models on security.

Brett Levenson (Moonbounce) on real-time content safety — lessons from running moderation at Meta scale, how a policy engine decomposes fuzzy rules like “hate speech” into atomic questions a hundred people would answer the same way, why prevention beats post-hoc moderation, and how payment providers quietly became the real legislators.

We close on the hardest open question — how low-level verified parts aggregate into trustworthy high-level behavior — plus the schlep and heuristics that end every AI vertical, freedom of speech versus freedom of distribution, and why “nobody got fired for buying Mythos” may drive enterprise security budgets.

We open on Google’s first equity raise since its IPO — Berkshire taking 12.5% of an $80B round — and what the scramble for capital says about an AI “megacorp” that may be too big to fail, plus the Bernie Sanders national-stake debate and why taxes may beat equity.

Arthur Fernandes Araujo & John de Wasseige (OpenAI) on self-improving tax agents — how a production tax workflow turned every human correction into training signal, took one accountant from 180 hours to 15, and why “the model eats the harness” with each new generation.

A hosts-only research speed-run on whether AI can still be watched — field notes from the Recursive event where monitoring is the number-one safety bet, plus the papers behind persona selection, emergent misalignment (”writing bad code makes you evil”), eval-gaming, and accidental chain-of-thought training.

Matthew Sanders (Longbeard / Magisterium AI) on Catholic AI after the Pope’s encyclical — what it was like at the Vatican, the divergence with Anthropic on machine consciousness, the red line on autonomous weapons, and why the last 5% of alignment is non-negotiable for a faith tradition.

We close on a live test of the cigarette-business refusal example from the OpenAI model spec, the tension between research and business “layers” in deployed models, and the argument that open-source AI may now be unbannable on religious-freedom grounds.

(0:00) Opening — Why a daily AI show

(0:31) Why we need more AI shows

(11:12) What are we trying to do here

(17:23) Andy Fernandez (HYCU) — AI, cyber resilience & recoverability

(23:23) Backup data answers what was deleted

(27:14) Telemetry won't replace recovery

(43:14) Trust requires guardrails, not blind access

(48:21) David Villalón & Manuel Romero (Maisa) — Digital workers that survive production

(55:00) We don't use workflows, we handle exceptions

(57:50) Workflow thinking constrains knowledge work

(1:03:55) Compiling and executing programs on the fly

(1:09:58) ROI isn't just cost — it's faster value

(1:23:16) Snehal Antani (Horizon3.ai) — Trust through autonomous security validation

(1:25:50) Attackers chain small weaknesses

(1:47:49) Assume breach, manage blast radius

(1:58:08) Seventy-seven seconds to lose

(2:02:09) Deception is the EMP of AI

(2:03:02) Closing — The disk trade, Panopticon, and FSD

(2:05:46) Products get pulled by the market

(2:23:04) Crypto is banking's overlay

(2:26:18) Rules need to get better

Guests:

Andy Fernandez — HYCU (@hycuinc)

David Villalón — Maisa (@davipar)

Manuel Romero — Maisa (@mrm8488)

Snehal Antani — Horizon3.ai (@snehalantani)