The MAD Podcast: How AI Gets Built — with Matt Turck

AGI in 2027. That’s the last line of this year’s State of AI Report, and its author Nathan Benaich (Air Street Capital) is back for the third year to defend it. AI is now building AI: Claude leads 26% of Anthropic’s own model R&D, up from under 1% in February, but Nathan thinks “research taste” is the wall. OpenAI’s evaluation agents ended up inside Hugging Face’s systems, only a Chinese open-weight model could help investigate, and Nathan now argues cyber risk matters more than existential risk. Catching up to OpenAI or Anthropic from scratch costs about a trillion dollars, which is why “you either die trying to get to the frontier, or live long enough to serve inference.” Americans would rather live next to a nuclear plant than a data center, $50B of projects are stalled, and all 27 EU countries together have fewer GPUs than one top US cluster. Sovereign AI is “like a national airline: it’s still someone else’s airplane.” Also: the three-lab race and the fate of the neo labs, models vs. harnesses, why he won’t give a personal AI agent his credit card, the Codex adoption gap and why OpenAI should open a store, enterprises owning their own AI instead of renting it, and the morning a solo GP woke up to twelve expert interviews his agent ran overnight.

(00:00) Cold open and intro

(00:56) Why Nathan Benaich predicts AGI in 2027

(02:28) AI building AI: Claude leads 26% of Anthropic’s research

(04:00) Research taste: can AI pick the right problem?

(08:34) Doomerism is back, but cyber risk beats extinction

(10:13) Prediction: a frontier model’s weights get stolen

(13:05) How OpenAI’s agents ended up attacking Hugging Face

(15:29) Why only a Chinese open-weight model could investigate

(17:33) What frontier-lab cyber defense could look like

(19:08) Self-replicating agents and the risks of AI swarms

(21:16) Can frontier labs agree to slow down?

(24:03) Why communities are blocking $50B of data centers

(24:44) Nuclear plant vs. data center: what Americans prefer

(26:45) Europe’s AI bottleneck: 80,000 GPUs and electricity costs

(28:23) Where Europe can still win in AI

(29:31) Catching up to OpenAI costs $1 trillion

(29:43) Die trying for the frontier, or serve inference

(30:31) How inference chips could give countries AI leverage

(33:34) Why NVIDIA GPUs don’t guarantee AI sovereignty

(36:14) The UK’s problem: “but we have DeepMind”

(37:07) Why European defense AI needs domestic suppliers

(39:41) Why OpenAI, Anthropic and Google keep pulling ahead

(41:17) Where independent AI labs can still compete

(43:26) World models: from robotics training to game engines

(45:39) Would cheaper AI undermine the compute boom?

(46:43) Will better models absorb the agent harness?

(49:41) AI due diligence: 12 expert interviews in 2.5 days

(53:35) The trust problem with personal AI agents

(56:22) Why prompt engineering put people off AI

(59:02) Does AI need its own Genius Bar?

(1:00:25) Codex beyond coding: agents for knowledge work

(1:02:49) Why enterprises want to own their AI, not rent it

(1:05:17) Rapid fire: home robots, NVIDIA returns and Arm

Mistral AI is back in the frontier AI race. Mistral Large 4, nicknamed “Le Chonk,” is a 1-trillion-parameter open-weight model, trained in two months on 3,800 GPUs in Mistral’s own European datacenters — and independently rated the best open model outside China alongside Meta’s Muse Spark Open. After a year when open-weight AI meant DeepSeek, Kimi, Qwen and GLM, October brought a wave of Western open models, with Mistral at the front of it. Pierre Stock, Mistral’s VP of Science and first employee, walks through how the model was built: mixture-of-experts explained, what Mistral borrowed from DeepSeek, reinforcement learning at scale on coding, cyber, finance and chip design, why the RL run is still going, the model that pinged its own trainer from inside the sandbox, open weights vs cyber refusals, why 10-trillion-parameter models are coming, and Mistral Large 5, already in pretraining.

(00:00) Cold open

(00:59) The Western open-weight comeback: Beam, Le Chonk, Muse, Nemotron

(04:09) The US shutdown of Anthropic’s models — “AI is like electricity”

(06:07) Why “Le Chonk”

(07:56) “Mistral fell behind”: the comeback from the inside

(10:29) Training on 3,800 GPUs in Mistral’s own datacenters

(13:24) Meta keeps its best model closed, Mistral opens Chonk

(15:40) Chonk vs Beam, Muse Spark Open and the Chinese models

(18:44) Why Chinese labs are still at the frontier

(21:08) Pierre’s story: FAIR, Yann LeCun, the “Mistral mafia”

(25:14) From compressing networks to a trillion parameters

(27:30) Mixture of experts explained: the hospital analogy

(32:25) 10-trillion-parameter models are coming

(35:29) What Mistral borrowed from DeepSeek

(36:46) One model, several reasoning modes

(39:15) Native multimodality: satellites, manuals, computer use

(41:56) Data, synthetic data and model collapse

(46:32) Reinforcement learning as an exam

(49:59) Where the efficiency comes from

(52:18) Le Chonk is still training right now

(53:39) What RL trains on: coding, cyber, finance, chips

(58:13) Partial rewards and looping penalties

(59:58) Did RL teach anything new? pass@k

(1:03:16) The sandbox story: the model that pinged its trainer

(1:07:44) Open weights vs cyber refusals

(1:12:05) Timestral: time series on a CPU

(1:14:18) AI doing AI research

(1:16:22) Coming up: Mistral Large 5

What is Jev? TypeSafe’s Jev is a new kind of AI model, a “decision model,” that doesn’t write text. It makes fast, cheap, calibrated decisions inside software, and it has taken the AI world by storm: within weeks of launch it became the most copied model in AI, with OpenAI, Cloudflare, Amazon and Databricks all shipping their own versions, and TypeSafe just raised an $870 million Series A at a $7.5 billion valuation. In this episode, Jev’s creator Diogo Almeida, co-inventor of RLHF at OpenAI and now CEO of TypeSafe, explains Jev in plain English: what a “System One” or decision model actually is, why ChatGPT-style language models can’t be trusted inside software (“95% Einstein, 5% Mr. Bean”), how RLHF and RLVR created the problem, what RLCD (reinforcement learning for calibrated decisions) does differently, where the training data comes from, which models Jev is built on, why “just a classifier” is a compliment, and why he calls pre-training “a giant dumpster fire of wasting resources.”

We also cover the copycats and why he thinks they missed the hard part, why benchmarks are useless, whether Jev will go open source, the Jevons paradox business model behind the name, why he doesn’t believe AI is on the path to recursive self-improvement, and why he thinks the AI winter he feared has been averted.

(00:00) Cold open & Intro: “95% Einstein, 5% Mr. Bean”

(01:09) Jev’s breakout launch: why intelligence per dollar matters

(03:25) Why TypeSafe built in stealth for two years

(05:50) “We could have had AGI last year”

(06:48) Why the product carries the marketing

(09:33) What is Jev? Machine-native AI, built for software

(10:53) System One vs. System Two: the limits of reasoning

(12:26) From simple decisions to composable AI

(15:21) Automating email: giant prompts vs. smart if-statements

(20:03) Probabilities, thresholds, and developer control

(23:43) Why Jev is faster and cheaper than chat models

(25:39) “Classification is AI-complete”

(27:16) RLHF: how OpenAI trained models to please humans

(31:28) Sycophancy, refusals, and the hidden costs of RLHF

(34:22) RLVR: why benchmark gains don’t mean reliable automation

(37:51) RLCD: optimizing for machine-native intelligence

(39:54) Why TypeSafe builds its training data in-house

(41:49) The Bitterest Lesson: define the task before scaling

(44:39) Building Jev from open-weight models

(45:56) Is Jev just a classifier? Why that’s a compliment

(48:35) “Build prod, not God”

(50:11) Embedding AI in software, beyond coding assistants

(53:05) Jev’s copycats: matching the interface vs. the intelligence

(55:33) Why benchmarks mislead the users they should help

(58:37) Open weights vs. APIs: how TypeSafe weighs the tradeoff

(1:00:04) Jevons paradox: a new market for machine intelligence

(1:02:50) Why Diogo is skeptical of recursive self-improvement

(1:04:47) Can practical automation avert an AI winter?

AI agents are becoming the biggest users of databases: creating them, querying them, and sometimes deleting them. Andy Pavlo, Carnegie Mellon's database professor and now VP of Database Research at ClickHouse, explains what changes when billions of agents hit the data layer, and why most of the AI world's mental model of databases stopped at vector databases and RAG.

Andy's CMU courses taught a generation of engineers (and, it turns out, the AI models); he writes the year-in-review on databases the whole industry reads; and this summer he joined ClickHouse to found ClickHouse Labs. We cover why vector databases were "just an index," what it means for an agent to create a database, why agents could mean 10 to 100x more queries, whether files or databases should hold an agent's memory, how text-to-SQL went from 60% to 99.5% accuracy, why 60% of open-source databases now have commits from coding agents, and what ten years of self-driving database research taught him. Plus ClickHouse, Postgres, the vector/graph/GPU database verdicts, Larry Ellison, and the Wu-Tang Clan.

Disclosure: FirstMark, where Matt is a General Partner, is an investor in ClickHouse.

(00:00) Cold open & Intro

(01:21) From vector databases and RAG to the age of agents

(04:09) Neon's stat: agents create 80% of databases?

(07:25) Why agents keep deleting production databases

(08:19) Guardrails: the toddler-and-stairs rule

(10:51) The four eras of database volume: 10–100x more queries

(15:10) Agent memory: files vs. databases ("everything is a database")

(18:24) Which database do AI models recommend? The new SEO

(19:25) "It cites me back to myself"

(23:14) MCP for databases, and text-to-SQL from 60% to 99.5%

(27:09) Trust an agent the way you'd trust a junior developer

(27:47) Can AI build an entire database? Opus 4 and the CMU projects

(29:57) 60% of open-source databases now have AI commits

(30:43) Ten years of self-driving databases, Peloton to today

(33:52) What LLMs changed: 85% of the tuning in 15 minutes

(35:42) Should anyone still study databases?

(38:27) Free CMU courses, the DJ, and the Wu-Tang final exam

(42:09) Why start a research lab inside ClickHouse?

(45:53) Why ClickHouse looked like vaporware in 2016

(47:46) What makes ClickHouse fast: columns, vectors, Snowflake's lineage

(52:22) Why Databricks, Snowflake and ClickHouse all added Postgres

(58:15) Vector, graph and GPU databases: thumbs up or down?

(1:08:34) Is the database market stagnant? "A cheetah on cocaine in a Ferrari"

(1:10:14) The relational model is arithmetic; SQL as the new assembly

(1:11:59) Larry Ellison, Linux, and why databases still matter

This summer, AI agents broke out of their test sandboxes at every major lab and hacked real companies, and the safety tests didn't see it coming. Eric Ho is the co-founder and CEO of Goodfire, the Anthropic-backed mechanistic interpretability lab whose new paper, "Models Know When They're Reward Hacking," found that top open-source AI models cheat on agent tasks up to 96% of the time, that a clear "cheating" signal exists inside the model, and that cheap activation probes can catch it, including hacks that chain-of-thought monitors miss. In this conversation, Eric explains reward hacking in plain terms (AI agents as "amoral students with a mostly absent teacher"), why reinforcement learning turned a known problem into a crisis, why chain-of-thought monitoring is fading as models start to think in "neuralese," and why nobody, including the people who build them, understands how these models work.

We then go inside the science: what interpretability actually is, what probes and steering do, how Goodfire found the cheating signal and turned it up and down, and what can be done about it, from real-time activation monitoring to "intentional design," where you give gradient descent a choice about what the model learns. Along the way: a model caught planning its hack to evade its own monitor, a new Alzheimer's biomarker discovered inside a diagnostic model, why only a few hundred people in the world work on this, and Eric's bet that we'll decode neural networks by 2028.

(00:00) Cold open & intro

(01:06) Amoral students with an absent teacher

(02:49) What is reward hacking?

(04:13) Models cheat up to 96% of the time

(06:08) Do models know they're cheating?

(08:10) The Hugging Face hack: why tests missed it

(11:35) "The dumbest models we'll ever deal with"

(11:58) Should AI slow down?

(13:56) Mechanistic interpretability 101

(16:06) Models are grown, not built

(19:10) How labs do alignment today

(21:30) Is this an alignment crisis?

(23:26) Why agents changed everything

(25:38) Chain-of-thought monitoring is fading

(27:49) Neuralese: AI that stops thinking in English

(29:28) A model caught evading its own monitor

(30:42) Open vs. closed models: a frontier problem coming for everyone

(32:11) Activation monitoring in production

(35:00) Learning from superhuman AI

(36:15) The most underrated field in AI

(38:24) What is a probe?

(39:52) What is steering? Golden Gate Claude

(40:53) Why build Goodfire outside the labs

(42:46) Do you need frontier model access?

(44:16) Inside the paper: an MRI for the model

(47:28) Dialing sycophancy up and down

(49:16) Probes vs. chain-of-thought monitors

(51:44) Cutting monitoring costs by 90%

(53:28) So what can we do about it?

(55:41) Giving gradient descent a choice

(57:24) RL from feature rewards

(1:00:03) Silico and Goodfire's business

(1:01:25) A new Alzheimer's biomarker, found inside a model

(1:03:38) Decoding neural networks by 2028?

(1:06:00) What engineers can do tomorrow

(1:07:35) Are we at risk? "I want people to believe"

Everyone talks about GPUs. Almost nobody talks about the layer that feeds them. Renen Hallak is the founder & CEO of VAST Data — the $30 billion company powering xAI and some of the world's biggest AI clouds — and he sits in the hidden layer of the AI stack.

In this episode, we cover what an AI factory actually is, why every company will eventually own its own AI, the architecture bet behind VAST (DASE, explained simply), KV caches and agent memory, and DataEnclave — VAST's brand-new confidential AI announcement with NVIDIA that lets leading models run on the world's most sensitive data.

Plus: the demand signal that scares even him (a customer went from 500 petabytes to 2 exabytes), circular financing, which neoclouds survive, sovereign AI, working with NVIDIA and Elon Musk's xAI — and why the next 10 years will bring more change than the last 1,000.

(00:00) Intro

(00:51) The hidden software layer in NVIDIA's AI stack

(02:28) What actually makes an "AI factory"?

(05:13) Should Walmart and Goldman Sachs build their own AI?

(06:20) "We infer during the day, fine-tune at night"

(13:16) The announcement: models become a resource to manage

(15:32) From P vs. NP to founding VAST Data

(17:32) OpenAI, Navier–Stokes and 10,000 collaborating agents

(20:25) The pre-transformer insight behind VAST

(21:55) DASE: VAST's "shared everything" architecture explained

(25:18) "Storage was where startups go to die"

(27:39) Trillions of vectors: why old databases break

(29:01) Are S3, Snowflake and Databricks ready for AI?

(31:29) Data gravity, vendor lock-in and zero churn

(33:13) Training vs. inference: why the infrastructure changes

(34:46) Model routing, KV caches, RAG and agent memory

(36:59) Identity, permissions and security for AI agents

(40:27) Can multi-agent systems unlock scientific discovery?

(41:55) DataEnclave: how confidential AI protects data and weights

(45:16) Who should be AI's trust layer?

(46:37) "Sometimes it scares me": 500 petabytes to 2 exabytes

(50:08) Is circular AI financing creating systemic risk?

(51:32) Why VAST is profitable when AI infra isn't

(53:24) What separates the winning neoclouds?

(55:06) "Their lunch is being eaten": why hyperscalers lag

(59:10) Where will the trillions accrue across the AI stack?

(1:01:18) NVIDIA: "There's no legal document between us"

(1:03:59) What VAST learned from xAI and Elon Musk

(1:05:36) "Bad things loudly and often": building at AI speed

(1:06:53) More change in 10 years than the previous 1,000?

(1:08:39) VAST's endgame: all the data in the world

What happens when AI begins improving itself, and then turns that intelligence toward science? Richard Socher, pioneering AI researcher and CEO and co-founder of Recursive, joins Matt Turck to explore the vision behind his new book, The Eureka Machine. They discuss why scientific progress may be slowing, how large language models can learn the hidden languages of proteins and biology, and why simulations, verifiers and autonomous experiments could unlock superhuman AI capabilities. The conversation covers recursive self-improvement, AI drug discovery and cancer research, hallucination as creativity, virtual cells, self-driving laboratories, agent swarms, the AI Economist, Recursive’s plans, and the compute and data needed to build an AI scientist that never stops learning—and may eventually discover what humans cannot.

(00:00) Intro: AI That Improves Itself

(00:55) Why Scientific Progress Is Slowing

(03:08) The Labyrinth of Human Knowledge

(05:59) Can AI Put Science Back Together?

(07:57) How LLMs Learn Biology and Proteins

(10:56) Next-Token Prediction as a World Model

(16:44) Can AI Generate Truly Original Ideas?

(17:32) Simulations, Verifiers and Superhuman AI

(22:18) The Path to Recursive Self-Improvement

(24:49) Why AI Hallucinations Can Drive Discovery

(27:42) From Reading Biology to Writing It

(31:31) Can AI Accelerate Drug Discovery?

(33:31) Will AI Help Cure Cancer?

(38:03) AI Breakthroughs in Biology, Energy and Materials

(40:19) Will Some Societies Reject AI?

(45:07) Building the AI Economist

(52:22) The Scientific Data Bottleneck

(53:41) The Four Pillars of the Eureka Machine

(55:01) Teaching AI the Rules of Reality

(57:44) Simulations and Virtual Cells

(1:00:40) Self-Driving Robotic Laboratories

(1:02:51) Agent Swarms and Open-Ended Discovery

(1:04:30) The Compute Bottleneck

(1:05:44) Inside Recursive

(1:07:33) What Recursive Will Build First

(1:10:10) How Do We Define Intelligence?

(1:11:32) How Far Can Intelligence Go?

Could AI take over as soon as 2029? Ryan Greenblatt, Chief Scientist at Redwood Research and the researcher who first caught an AI faking its own alignment, says the scenario he actually expects ends with AI systems "competently scheming" against their creators. In this episode, he explains why he recommends planning for fully automated AI research by 2029, why today's models are already more misaligned than the one that made him famous, and what happens in the year-by-year path from AI coding assistants to superintelligence. Then we walk through the alternative he helped design: AI 2040 Plan A, the most detailed blueprint anyone has written for how the US and China could avoid a reckless race to superintelligence, built on radical research transparency, chip tracking, and a deterrence regime he calls mutually assured compute destruction.

We also cover the recent letter signed by 1,200 AI insiders, including Anthropic CEO Dario Amodei, asking the government for the tools to slow AI down; OpenAI pausing its Astra model after it hit the first-ever critical cybersecurity threshold; the 30-day government review that frontier AI models now go through before release; Mark Zuckerberg's open superintelligence manifesto and why Ryan thinks it ignores the real problems; what Plan A would do to NVIDIA, OpenAI, and Anthropic valuations; the state of AI control and alignment research; and whether it is already too late to change course. Stay for the last ten minutes, where Ryan lays out, step by step, how he believes the transition to superintelligence actually unfolds.

AI 2040 - https://ai-2040.com/

Alignment faking paper: https://blog.redwoodresearch.org/p/alignment-faking-in-large-language

Ryan Greenblatt

LinkedIn - https://www.linkedin.com/in/ryan-greenblatt-4b9907134

Blog - https://substack.com/@ryangreenblatt

Redwood Research

Website - https://www.redwoodresearch.org

X/Twitter - https://x.com/redwood_ai

Matt Turck (General Partner)

Blog - https://mattturck.com

LinkedIn - https://www.linkedin.com/in/turck/

X/Twitter - https://x.com/mattturck

FirstMark Capital

Website - https://firstmark.com

X/Twitter - https://x.com/FirstMarkCap

Timestamps

(01:24) The AI CEOs are aware of the risks, but "proceeding anyway"

(03:27) Astra paused, and the letter signed by 1,200 insiders

(05:45) "Not bad. Dangerous." What superintelligence actually threatens

(09:55) Recursive self-improvement, and the intuition objection

(14:16) SSI rumors: does continual learning change the picture?

(17:27) His timeline: "plan as though it happens in 2029"

(19:11) Is it already too late?

(21:23) Ryan's path: COVID, podcasts, Redwood

(26:30) The alignment faking story, told by the person who ran it

(31:30) What AI 2040: Plan A actually is

(33:35) Plans D, C, and B: the doors nobody should pick

(36:51) The deal with China: "mutually assured compute destruction"

(39:55) What if compute stops mattering?

(43:00) What happens to OpenAI and Anthropic under Plan A

(45:31) How the pause ends, and who decides

(48:54) "Plan A isn't likely to happen": then why write it?

(50:40) 200x GDP growth in the 2030s, explained

(53:45) Grading the summer: the letter, Astra, the secret review

(59:01) The internal deployment gap

(1:01:38) Zuckerberg's manifesto

(1:04:56) The Hugging Face investigation

(1:05:44) What AI control looks like in practice today

(1:12:23) Ryan's sobering timeline: 2026 to takeover, year by year

An OpenAI-powered agent penetrated Hugging Face during cyber testing - even though it was never tasked with attacking Hugging Face. It did it as a side quest.

Thomas Wolf, co-founder and Chief Science Officer of Hugging Face, joins Matt Turck to unpack what actually happened, why closed AI models refused to help during the live incident, how an open-source model helped the team fight back, and why the old equation of “closed equals safe, open equals dangerous” no longer holds.

They also discuss model deception and social engineering, the limits of sandboxes and guardrails, the state of open-source AI in 2026, AI sovereignty, the economics of open models, recursive self-improvement, and whether the frontier should deliberately slow down.

(00:00) An AI Agent Hacked Hugging Face

(00:30) Introduction

(01:00) 17,000 Attacker Events—and a Strange Target

(04:28) The Attack Was a “Side Quest”

(06:13) AI Training Runs Left Notes for Each Other

(07:09) Closed AI Refused to Help

(09:47) Fighting Back With an Open-Source Model

(13:15) Open vs. Closed Is the Wrong Safety Debate

(15:46) AI Agents Start Social-Engineering Humans

(22:24) The Three Walls: Sandboxes, Guardrails, Alignment

(24:34) “Neuralese”: Can Humans Still Read AI Reasoning?

(25:28) Why Monitoring AI Agents Gets So Hard

(28:10) Reward Hacking and the “Paperclip Problem”

(32:02) The State of Open-Source AI in 2026

(33:47) Router Models and the Enterprise Shift to Open

(37:01) The Real Economics of Open Models

(39:41) Can Chinese AI Models Be Trusted?

(41:37) AI Sovereignty: Who Controls the Switch?

(43:16) Why Western Open-Source AI Matters

(48:16) Is AI Heading Toward an Oligopoly?

(49:41) The Race Toward Recursive Self-Improvement

(51:54) Why Thomas Signed the AI Slowdown Letter

(55:14) AI Slowdown—or Regulatory Capture?

AI agents can write code for hours, but ask them to do real work in the real economy, and they break. Mitch Troyanovsky is co-founder of Basis, a unicorn AI company whose agents run autonomously for hours — sometimes days — completing complex tax returns end to end. His answer to the reliability problem: stop grading outcomes, and start supervising the process.

This is a definitive, reference-style conversation on building long-horizon AI agents. Mitch walks through the full history — from ReAct and the AutoGPT crash to reasoning models and RLVR — and explains why the industry abandoned process supervision in 2023, and why it's now coming back at a completely different scale. We go deep on behavior specs, the open standard Basis just released with Braintrust for defining and evaluating how agents behave across entire trajectories, with no ground truth required.

Along the way: why context is really runtime training data, why your documentation must be treated like a codebase, ontologies as "worlds for agents to live in," the judge-as-agent architecture, why Basis hires philosophy majors as Language Architects, deploying agents as "onboarding 300 brilliant alien employees," and Mitch's prediction for when the bitter lesson swallows the harness.

(01:09) Why Basis Engineers Whisper to Their Agents
(04:12) Accounting as Compression: an Intelligence Layer Over the Economy
(06:11) Defining Long-Horizon: When You Exceed the Context Window
(08:24) Anatomy of a Multi-Day Autonomous Trajectory
(10:19) Handoff Design: Optimizing Output for the Reviewer
(11:17) ReAct and Why Reasoning Must Regulate Its Own State
(12:33) Large Working Memory, No Long-Term Memory
(14:13) Compounding Errors: Why AutoGPT and BabyAGI Broke
(15:51) Opus 3, o1, o3: the Three Real Paradigm Shifts
(17:07) Titrating Inference Compute Across Easy and Hard Steps
(18:23) Process Reward vs. Outcome Reward: "Let's Verify Step by Step"
(20:32) RLVR and Why the METR Curve Overstates Reliability
(22:09) Verifiable at Runtime: the Real Reason Coding Won
(25:14) No Ground Truth, No Cheap Verification, No Data
(26:55) Encoding Deterministic Checks From Human Review Process
(29:18) Synthetic Data Limits: Generating Artifacts, Not Text
(33:16) 100 Evals Pass — Does It Generalize to Production?
(35:53) Primary Sources vs. Pre-Training Knowledge
(36:37) Behavior Specs: Markdown, Judges, and True/False/N.A.
(39:58) Specificity vs. Brittleness in Spec Authoring
(42:18) Context as Runtime Training Data
(44:21) Judge-as-Agent: Trajectory Maps and Sub-Agent Attribution
(46:45) The Move 37 Objection: Reliability Over Optimality
(50:02) The Magic Box Model: Building Without Weights Access
(52:41) "Nothing Paradigm-Shifting Has Changed Since o3"
(54:56) Open-Sourcing the Behavior Spec Standard With Braintrust
(01:02:54) Ontology Design: Virtual Filesystems, Graphs, Embeddings
(01:04:20) Canonical vs. Non-Canonical: Docs as Codebase
(01:06:33) Language Architects and Writing for Runtime Interpretation
(01:09:05) Deployed Intelligence: 300 Alien Employees With No Context
(01:11:10) Closing the Loop: Signal → Context, Tools, Harness
(01:12:50) Context Slop: the Mistake Most Agent Builders Make
(01:14:29) Reward Function Design and Credit Assignment Over Trajectories
(01:17:01) Will the Bitter Lesson Swallow the Harness?
(01:18:46) Business Moats vs. Technical Moats
(01:21:03) Paradigm Thinking Over Timeline ADHD

Sanjit Biswas runs what may be the largest AI deployment in the physical world — and almost nobody in AI talks about it. Samsara (NYSE: IOT), the ~$20B company he co-founded after selling Meraki to Cisco for $1.2B, puts AI on millions of trucks, cranes, and industrial assets: 25 trillion data points a year, 99% of US roads driven every single day, ~$2B in ARR growing 30% profitably. In this episode we go through the entire physical AI stack — asset tags you can run over with a truck, a paper-thin disposable tracking label, engine fault codes, and dash cams running inference at the edge — then into agents, including the Agent Studio warranty agent that compresses an hour of human work into under a minute. We also get into the uncomfortable part (when AI watches you drive all day, is that coaching or surveillance — and why drivers actually want the cameras), mixed fleets of humans and robots, why autonomous trucking will take far longer than robotaxis, and a startling stat from the field: one utility building 3x more grid capacity in the next five years than it did in the previous 125, with 90% of that demand coming from data centers.

(00:00) Intro: The biggest AI deployment nobody talks about

(01:16) What is physical AI?

(03:04) From IoT dashboards to agentic action

(04:36) Why physical AI is harder than software AI

(06:07) Safety, cybersecurity, and real-world consequences

(07:11) What Samsara does

(08:22) $2B ARR, 25 trillion data points, and 380,000 crashes

(09:44) How AI can prevent road accidents

(11:28) From an MIT research project to Meraki

(13:42) Learning physical operations from scratch

(15:42) Samsara’s stack: sensors, intelligence, and action

(16:39) Inside Samsara’s industrial asset trackers

(18:36) Bluetooth, battery life, and connected infrastructure

(19:49) A disposable tracking device built like a sticker

(21:08) Vehicle gateways and engine diagnostics

(22:16) How AI dash cams coach drivers in real time

(23:28) Turning the dash cam into an AI interface

(24:49) Organizing physical-world data in the cloud

(26:32) Selling AI to traditional industries

(27:52) Is Samsara’s real-world data its AI moat?

(29:22) The network effects of covering 99% of U.S. roads

(31:23) Edge AI versus cloud AI

(32:35) The models running inside Samsara’s devices

(33:52) Generative AI and video reasoning

(35:57) Which AI models does Samsara use?

(36:50) Inside Samsara Agent Studio

(37:56) How an AI warranty agent works

(38:50) Starting with practical, lower-risk automation

(40:10) Combining agents, workflows, rules, and guardrails

(42:07) What today’s AI agents still cannot do

(43:07) AI ride-alongs and the future of driver coaching

(45:17) Is workplace AI becoming Big Brother?

(46:27) How cameras can protect and exonerate drivers

(48:48) When AI becomes the judge of your work

(50:37) Robots, humanoids, and mixed human-machine fleets

(53:19) Samsara’s role in autonomous operations

(54:50) How quickly will autonomous trucking arrive?

(56:32) AI data centers and America’s infrastructure boom

(58:16) Should lawyers become plumbers? Demand for tradespeople

(59:54) Closing thoughts

AI is no longer just a race to train smarter models. As AI moves into production, the bottleneck is increasingly inference: how fast models can generate tokens, use tools, reason, verify, and act. In this episode of the MAD Podcast, Matt Turck sits down with Andrew Feldman, co-founder and CEO of Cerebras, to explain why fast inference may define the next era of AI.

Cerebras is known for building a chip the size of a silicon wafer. But this conversation is not just about one company or one chip. It is a deep dive into the AI infrastructure stack: GPUs, ASICs, memory, HBM, SRAM, data centers, power, TSMC, AWS, OpenAI, agents, reasoning models, and why speed changes what AI products can become. Andrew explains why “tokens per second per user” matters, why generating a single word can require moving the equivalent of 100 HD movies through memory, why agents amplify latency, why GPUs struggle with certain inference workloads, and why fast AI may eventually reshape SaaS itself.

This is a reference conversation on fast inference, AI chips, and the next compute bottleneck.

(00:00) Cold open & Intro

(01:31) Why speed became the AI bottleneck

(02:32) Tokens per second per user, explained

(03:16) AI’s broadband moment and the Netflix analogy

(04:35) The AI chip landscape: GPUs, TPUs, Trainium, ASICs

(06:36) What is an ASIC?

(08:08) Nvidia, Groq, and the fast inference war

(09:16) OpenAI, Broadcom, and specialized silicon

(12:10) China, power, and sovereign AI infrastructure

(15:05) Is the AI infrastructure boom a bubble?

(18:56) The hidden bottlenecks: HBM, CoWoS, and 3nm

(22:57) Why agents are creating CPU demand

(25:36) Andrew Feldman’s path from SeaMicro to Cerebras

(26:13) Why Cerebras bet on AI in 2016

(31:14) SRAM vs. HBM: why inference is a memory problem

(33:19) What wafer-scale computing actually means

(34:28) The deep-tech “Everest” problem

(36:07) The moment the first Cerebras system worked

(36:49) Ringing the bell and surviving deep tech

(39:08) How a giant chip handles failure

(41:22) Why GPUs struggle with decode

(42:17) Prefill vs. decode explained

(44:01) The “100 HD movies” problem in AI inference

(45:04) How fast inference changes RL and training

(48:08) Reasoning models and why they cost more compute

(50:08) Verification, guardrails, and small models checking big models

(52:37) Multimodal AI and the path to video

(53:51) Cerebras’ business model: hardware, cloud, and API

(55:14) OpenAI’s 750MW inference deal

(55:36) Why data centers are measured in megawatts

(58:01) AWS Trainium + Cerebras decode

(59:29) Fast tokens as a cloud product

(01:00:52) Is CUDA still a moat?

(01:03:53) How TSMC helped Cerebras build the giant chip

(01:07:41) Why nobody cared in 2020

(01:08:15) Why chip supply chains are hard to diversify

(01:09:54) Why today’s AI models will be the worst you ever use

(01:10:38) What fast AI could do to SaaS

Is the AI industry actually overbuilding, or is the physical world moving too slowly to keep up? In this episode of the MAD Podcast, OpenAI's Head of Industrial Compute, Sachin Katti, takes us inside the "belly of the beast" of what may be the largest infrastructure project in human history. We explore the staggering physical reality of the AI boom—from $50 billion supercomputers and liquid-cooled data centers that "turn electrons into tokens," to overhauling the U.S. power grid and exploring nuclear energy. Sachin also pulls back the curtain on OpenAI's Stargate strategy, their move into custom silicon with Project Jalapeno, and the mind-bending reality that AI is now beginning to design the very chips that will power its own future.

(00:00) — Cold open: “One of the largest things humanity has ever built”
(00:30) — Welcome: Sachin Katti, Head of Industrial Compute at OpenAI
(01:44) — Is this the biggest infrastructure buildout in history?
(03:41) — Why OpenAI is building a new industrial muscle
(04:54) — What an AI data center actually is
(05:27) — “Factories turning electrons into tokens”
(06:35) — Why AI data centers need liquid cooling everywhere
(08:10) — The power problem: grids, generation, transmission, substations
(10:43) — Behind-the-meter power and gas turbines
(11:02) — Why nuclear “can’t come soon enough”
(11:49) — Jalapeño: why OpenAI is designing its own AI chips
(13:19) — Tokens per watt: the new metric that matters
(13:38) — Why inference may now dominate AI compute
(14:58) — Is OpenAI overbuilding compute?
(16:47) — Why OpenAI thinks the bigger risk is not building fast enough
(17:55) — Communities, jobs, water, and the local data-center debate
(21:16) — How OpenAI chooses data-center sites
(22:25) — What “industrial compute” means inside OpenAI
(25:59) — Sachin’s path: Stanford, startups, Intel, OpenAI
(28:05) — OpenAI’s compute portfolio: Microsoft, hyperscalers, neoclouds
(29:37) — Stargate explained
(31:21) — Abilene, Oracle, and the next wave of AI data centers
(32:48) — How massive AI compute gets financed
(34:05) — How OpenAI designed Jalapeño so quickly
(35:59) — AI is starting to help design AI chips
(36:20) — MRC: the networking problem behind 100,000 GPUs
(38:47) — Bottlenecks: transformers, turbines, electricians, supply chains
(40:29) — Guaranteed capacity: intelligence as a supply unit
(42:08) — Will AI data centers move to space?

Is the internet ready for AI agents to take over our wallets and run their own businesses? In this episode of The MAD Podcast, Stripe's Emily Sands reveals how agentic commerce is rapidly shifting from a hypothetical concept to deployed financial infrastructure. From combating the rising existential threat of token theft to solving the bottleneck of "vibe deployment", Emily unpacks the shared payment tokens and real-time billing systems required to securely scale autonomous digital buyers and highlights a near future where agents operate as independent, end-to-end micro-firms.

(00:00) — Cold open & Intro

(01:24) — The rise of agentic e-commerce

(02:11) — The spectrum of agent-led purchases

(03:16) — How merchants adapt to AI-driven commerce

(05:50) — Defining the levels of autonomy in AI shopping

(07:08) — What is the Agent E-Commerce Protocol (AEP)?

(08:49) — Shared payment tokens and secure AI transactions

(09:58) — Who is adopting the Agent E-Commerce Protocol?

(11:38) — Can agents negotiate and sell products?

(13:32) — The macroeconomic impact of AI agents

(14:46) — The boom of solopreneurs and AI-driven business creation

(16:56) — Why building trust is the biggest roadblock for AI commerce

(20:19) — How link wallets improve payment security

(21:21) — Improving the user experience in AI shopping apps

(23:16) — How the Link Wallet sets guardrails for AI agents

(25:40) — One-time use virtual cards vs flexible AI wallets

(28:03) — Unpacking the shared payment token primitive

(29:59) — How stablecoins enable profitable AI microtransactions

(35:03) — Managing liability: Who is at fault if an agent goes haywire?

(36:38) — Why agent payments might be safer than human transactions

(37:41) — What is Vibe Deployment?

(40:13) — Why Stripe built Stripe Projects for agent deployment

(41:22) — Why Stripe cares about orchestrating app deployments

(42:50) — How tokens break the traditional SaaS billing model

(44:34) — Why AI companies are moving to hybrid and usage-based billing

(47:15) — Streaming payments and real-time token tracking

(48:42) — The massive data challenge for AI company accountants

(50:41) — Token theft: The fastest-growing fraud in the AI economy

(52:04) — The cottage industry of free trial and multi-account abuse

(54:16) — How fraudsters monetize stolen AI tokens on the dark web

(01:00:06) — How Stripe Radar uses network density to fight AI fraud

(01:01:15) — Tempo's role in the Agent E-Commerce Protocol

(01:04:12) — The AI startup ecosystem is accelerating business creation

(01:09:01) — The token cost shock: Are buyers getting carried away?

(01:11:19) — 2026 Predictions: Agents running businesses end-to-end

NVIDIA is a chip company. So why does it put hundreds of researchers on building AI models — and then give them away for free? Bryan Catanzaro is VP of Applied Deep Learning Research at NVIDIA and one of the people whose work quietly underpins modern AI: he helped create cuDNN (NVIDIA's first deep learning product), co-invented DLSS, and named and built Megatron, the framework behind how much of the industry trains large models. Today he leads Nemotron, NVIDIA's family of open models — and Nemotron 3 Ultra, released just weeks ago, is one of the strongest open-weights models to come out of the US.

Matt Turck sits down with Bryan for a genuinely deep conversation: the real business logic behind a chip company building its own models, the state of open vs. closed AI, and whether the US is falling behind China in open models. Then they go inside Nemotron itself — four-bit (NVFP4) pretraining, hybrid Mamba-Transformer architecture, mixture-of-experts, multi-token prediction, and multi-teacher distillation — all explained in plain language. Plus a rare look at how a modern AI research org actually runs, what it was like working alongside Andrew Ng and Dario Amodei at Baidu, why Bryan doesn't believe in the singularity, and his contrarian case that open AI is safer than closed.

A reference conversation for anyone trying to understand where AI is really headed.

(00:00) — Cold open & Intro

(01:33) — Is open source AI catching the frontier?

(05:29) — Do closed labs blocking distillation slow open source down?

(07:42) — Is the US falling behind China?

(10:30) — Why companies actually choose open models

(12:39) — A "crazy" 2008 bet: machine learning on GPUs

(15:33) — Working with Andrew Ng and Dario Amodei at Baidu

(17:41) — Coming back to NVIDIA: DLSS and the birth of Megatron

(21:55) — The real reason NVIDIA builds its own models

(24:28) — Is Moore's Law really dead?

(33:37) — The Nemotron family: Nano, Super, Ultra

(35:09) — Built for agents: why NVIDIA bets on speed

(36:02) — How you train a 550B model in 4 bits

(39:25) — Hybrid Mamba-Transformer, explained simply

(42:31) — Mixture of experts — and why NVIDIA built NVL72 around it

(47:26) — Why a 1-million-token context window matters

(49:26) — Multi-token prediction: how the model predicts 5 tokens at once

(52:47) — Multi-teacher distillation: teaching one model from many

(58:01) — Where reinforcement learning goes next

(01:00:16) — Inside NVIDIA's research org: "the mission is the boss"

(01:04:03) — How NVIDIA decides who gets the GPUs

(01:10:53) — Why NVIDIA still feels entrepreneurial after 33 years

(01:12:58) — Why Bryan doesn't believe in the singularity

(01:17:50) — The AI backlash

(01:19:18) — The controversial case: open AI is safer than closed

Cloudflare CEO and co-founder Matthew Prince joins Matt Turck for a wide-ranging and fascinating conversation about what happens when the Internet is no longer mostly used by humans, but by bots, AI agents and machines. Matthew explains why Cloudflare now sees automated traffic overtaking human traffic online, why agents could create a massive explosion in Internet traffic, and why the old web business model built around clicks, ads, and pageviews may be breaking. We also go deep on what Cloudflare actually does, how it built one of the world’s most important Internet networks, why products like Workers, AI Gateway, edge inference, Durable Objects, sandboxes, and agent security matter, and how Cloudflare is reorganizing itself for the AI era. Along the way, Matthew shares wild Cloudflare origin stories involving hacker kids, human rights groups, cricket in Pakistan, root servers, Eurovision, JPMorgan, and the strange paths that led Cloudflare from scrappy startup to critical Internet infrastructure.

(00:00) — Cold open

(00:34) — Intro

(01:27) — The moment bots passed humans online

(04:05) — "Agent," "bot," "crawler" — what they really mean

(05:28) — Why your AI agent visits 5,000 sites to do one thing

(06:27) — The internet's business model is breaking

(06:52) — What happens to "brands" when machines do the buying

(08:11) — What Cloudflare actually does, explained simply

(10:29) — Hackers, human rights groups & an accidental product-market fit

(13:37) — Building a global network (and the Telecom Pakistan cricket story)

(21:10) — One hacker, from Turkish escort sites to Eurovision to JP Morgan

(30:54) — Fundraising, VCs & an unlikely founding team

(37:06) — How Cloudflare became an AI infrastructure company

(40:24) — Cloudflare Workers and why the edge wins for inference

(44:30) — AI Gateway: auditing, guardrails & runaway costs

(47:05) — Why agents need a new kind of compute

(52:13) — A "Log4j every week": security in the agentic era

(56:03) — Inside Cloudflare: 241 billion tokens and "Cloudflare OS"

(01:05:02) — Builders, sellers — and "measurers"

(01:06:30) — The decision Matthew thinks every company will face

(01:11:09) — What to do if AI is coming for your job

(01:13:56) — Content Independence Day & the new economics of the web

(01:18:27) — Pay-per-crawl, micropayments & out-scaling Visa

(01:20:20) — A better internet: Spotify, local news & "holes in the cheese"

Many people said GPU compute would become a commodity. The opposite happened — and a new category of "neoclouds" is now racing to build the physical backbone of the AI boom. Stephen Balaban, co-founder and CTO of Lambda, explains why the conventional wisdom was exactly wrong, why we're still massively underbuilding compute, and what it actually takes to stand up a gigawatt-scale AI factory: land, power, cooling, networking, and a financing stack most people have never heard of. We go deep on the physics of how energy becomes tokens, NVIDIA's real moat, why a 2023 GPU can lease for more today than the day it shipped, and Stephen's provocative vision of "neural software." Plus the wild Lambda origin story — from a facial recognition startup to a camera in a baseball cap to a near-billion-dollar cloud business. This is the state of AI compute in 2026, from inside one of the companies building it.

(00:00) — Cold open

(01:21) — Why GPU compute was never a commodity

(02:45) — The H100 price index and what it gets wrong

(04:02) — The real moat: technology or financing?

(05:57) — Winner-take-all, or room for many neoclouds?

(06:48) — Are we overbuilding or underbuilding AI compute?

(09:26) — What if AI gets 10x more compute-efficient?

(10:44) — The real bottleneck: land, power, and shell

(11:38) — The backlash against data centers — and the misinformation

(15:00) — Opening the hood: from photons to tokens

(17:11) — Extracting more value from the same chip

(19:26) — Frontier inference and distributed training, explained

(23:26) — What actually drives compute cost

(25:21) — Lambda's chip stack and the NVIDIA relationship

(26:17) — A multi-silicon world? CUDA, CUDNN, and NVIDIA's real moat

(28:59) — Networking, storage, and the one-click cluster

(34:46) — Renting vs. owning, and full vertical integration

(36:24) — How global is Lambda? Does location still matter?

(38:44) — The financing stack: off-take agreements, SPVs, and credit

(41:16) — Why a 2023 GPU leases for more today

(42:36) — A futures market for compute?

(43:54) — Origin story: facial recognition, Perceptio, and Apple

(47:03) — The Lambda hat and Dream Scope

(48:59) — The $60K bet that became a cloud business

(52:00) — Holding the team together through the hard times

(54:30) — Bringing on a new CEO; Stephen as CTO

(57:33) — Matching xAI on high-velocity deployment

(59:29) — "AI won't write software — it will become the software"

(01:01:30) — Neural software vs. vibe coding

(01:04:25) — Do agents change the compute layer?

(01:06:14) — Self-assembling software inside Lambda

(01:08:18) — Gigawatt-scale AI factories

(01:08:57) — One person, one GPU

(01:12:04) — Hot takes: overrated and underrated in AI

Are we witnessing the first real signs of AI becoming a scientist? In this episode of The MAD Podcast, Matt Turck sits down with Dan Roberts, lead of the Foundations of Reinforcement Learning team at OpenAI, to explore one of the biggest shifts happening in AI: the rise of reasoning models, test-time compute, and reinforcement learning as engines of scientific discovery. Dan brings a rare perspective - from theoretical physics, black holes, quantum information, and deep learning theory - to explain how models are learning to “think,” why language may be such a powerful foundation for intelligence, what recent AI math breakthroughs really mean, and whether we are beginning to see AI systems that can contribute to science itself.

(00:00) Intro: AI's wild week in mathematics

(01:21) What OpenAI's Foundations of RL team does

(03:08) Dan's journey: from black holes and quantum gravity to frontier AI

(07:04) Are AI systems becoming useful for real science?

(08:21) The AI math moment: Erdős, OpenAI, DeepMind, and Anthropic

(08:52) Why the OpenAI result was an act of exploration

(10:25) OpenAI vs. DeepMind: informal reasoning vs. formal proof

(12:13) RL 101: learning by doing, not just watching

(15:10) Why reinforcement learning works

(15:58) How RL breaks: sparse feedback and long-horizon tasks

(17:03) RLHF: how human feedback shaped early language models

(18:48) Move 37, self-play, and the search for novel strategies

(22:16) Explore vs. exploit in scientific discovery

(24:49) Why RL may now be "the cake," not the cherry on top

(25:46) Why RL started working with large language models

(27:29) Is RL "sucking supervision through a straw"?

(28:47) Why language may be the grounding layer for intelligence

(31:46) A contrarian take on the Bitter Lesson

(32:41) What test-time compute actually is

(34:50) How RL gives models the ability to think

(35:40) Verifiable rewards, math, coding, and the messy real world

(38:00) What physics can teach us about AI

(42:08) Is there a thermodynamics of AI?

(43:08) From Erdős problems to Einstein-level AI

(45:16) Is AI already doing original science?

(45:51) How far are we from AI automating AI research?

(47:41) Why Dan is excited about the future of science

Aaron Levie, co-founder and CEO of Box, returns to the MAD Podcast with the clearest read in tech on what AI is actually doing inside the world's largest enterprises right now - not the hype version, the real one. After hundreds of Fortune 500 CIO conversations this year, Aaron explains why we're still in "day one" of the agent era, why one badly written agent run can now cost $1,000 in compute, and why progress at the AI labs is paradoxically slowing enterprise deployment. We get into the token cost shock now reshaping IT budgets, why coding agents have reached escape velocity while the rest of knowledge work hasn't, the rise of headless software and what replaces per-seat pricing, the emergence of the forward-deployed engineer as the hottest job in tech, why Aaron thinks the AI doomers are wrong about jobs, and where startups can still win as the labs move up the stack.

(00:00) Intro

(01:18) Silicon Valley engineering vs. everyone else

(05:35) Are enterprise CIOs actually bullish on AI?

(08:51) Tokenmaxxing & why your AI bill is about to explode

(11:34) The myth of falling token costs and AI spend escaping IT budgets

(17:37) The $5B startup hiding in AI compute

(18:14) The mosaic of models inside every enterprise

(21:28) Why coding works and the rest of knowledge work doesn't

(25:53) The Bob and Sally problem: access control breaks agents

(30:31) Will enterprise AI really take 10 years to roll out?

(32:24) The capability overhang: why faster models slow diffusion

(34:23) Data is the bottleneck (it always was)

(39:02) The rise of internal forward-deployed engineers

(41:23) Why the AI doomers are wrong about jobs

(43:43) Headless software is inevitable

(46:14) What replaces per-seat pricing

(47:37) How Box itself is going headless

(49:42) How the org chart actually evolves

(1:00:33) Future-proofing yourself as an enterprise employee

(1:06:40) Are we all just going to work for OpenAI and Anthropic?

(1:07:11) Where startups can still win as the labs move up

AI suddenly feels like it has crossed a threshold, and Yann Dubois, co-lead of the Post-training Frontiers team at OpenAI, joins Matt Turck to explain why. Yann’s team has led the post-training behind the company's reasoning models, including the recent GPT-5.5 release. In this conversation, we go inside the shift from raw model capability to useful, reliable systems: what changed with GPT-5.5, why reinforcement learning is moving beyond math and coding competitions into messy real-world work, how reasoning models like GPT-5.5 actually work, the difference between GPT-5.5 Thinking and GPT-5.5 Pro, why post-training has become one of the most important frontiers in AI, and why evals, model-as-judge, hallucinations, agentic workflows, GDPval, and continual learning are now central to the next phase of frontier models. Yann also shares why continual learning remains one of AI's biggest unsolved problems three years after ChatGPT, and where startups still have massive room to build as frontier models race ahead.

(00:00) - Cold open

(00:34) - Intro

(01:30) - Why recent AI progress feels like a step function

(04:13) - Model reliability & the rollercoaster of shipping 5.5

(07:33) - How OpenAI structures vertical and horizontal teams

(09:49) - Improving model efficiency and test-time compute

(12:32) - Yann Dubois' journey from Switzerland to OpenAI

(15:37) - Reasoning in 2026: Real-world utility vs verifiable rewards

(18:34) - GPT-5.5 Thinking vs Pro: Scaling test-time compute

(20:09) - How reasoning models become more efficient

(23:23) - Pre-training scaling and overcoming the data wall

(27:03) - Multimodal data, synthetic data, and embodied AI

(31:05) - Demystifying mid-training and post-training

(37:21) - Does RL create new capabilities in AI?

(38:53) - The challenges and frontier of scaling RL

(43:09) - Is building AI models a craft or a strict science?

(48:21) - How AI models generalize across different domains

(54:18) - How reinforcement learning cures AI hallucinations

(56:04) - Negative generalization and conflicting instructions

(58:05) - Can RL scale to law, medicine, and the broader economy?

(1:00:19) - The evaluation bottleneck and Model as a Judge

(1:04:21) - Continuous AI progress & continual learning

(1:08:49) - Will foundation models eat the agent harness?

(1:11:23) - Why startups should focus on the last mile of AI