Segments
The best minutes from the latest interviews, newest first. Watch plays just that part.
21 segments from 7 episodes
AI & I · Matt HartmanBack founders whose pivot foot is technology, not a market
Hartman assumes every early company will pivot, so he backs technical teams who can carry their skills to a new market. Domain-expert teams tend to stay stuck in one vertical. 3 takeaways
“The companies that have been most successful that I've been involved with are ones where the pivot foot was about the technology, not about necessarily market understanding”
Matt Hartman at 20:46 - AI & I · Matt Hartman
Vibe coding his own app taught an investor why SaaS still sells
Hartman built Ghost Pepper, a local voice-to-text app that hit number one on Hacker News. Maintaining it showed him that SaaS buyers pay for constant upkeep, which complicates SaaSpocalypse talk. 3 takeaways
“The other thing I learned was when you just vibe code something for yourself, you have to maintain it.”
Matt Hartman at 37:23 - AI & I · Matt Hartman
How Hartman judges an earned insight in early founders
He measures an insight by how few people were in a position to learn it, since rare insights mean less competition and a longer head start. He also wants founders who ship and learn from the market. 3 takeaways
“I think about that as how many people were in a position to learn the thing that you did.”
Matt Hartman at 31:47 
a16z Podcast · Nathan LambertWhy distillation explains less of China's AI progress than assumed
Lambert says Chinese labs match US talent but lack compute, and treats distillation as a way to buy compute. He argues the open literature shows distillation is hard to apply to modern RL-heavy post-training. 3 takeaways
“And distillation is a way to effectively spend more money to get what is compute equivalent. Like outputs from a language model are a function of compute.”
Nathan Lambert at 9:37 - a16z Podcast · Nathan Lambert
Frontier gains outside math and code come from buying data
Lambert argues that progress in hard-to-verify domains comes from purchased data and RL environments, not magic generalization. He also says rushed environment vendors are a source of reward hacking. 3 takeaways
“but fundamentally, like AI progress is about data and about writing better software to serve your models.”
Nathan Lambert at 27:00 - a16z Podcast · Nathan Lambert
Scaling laws argue against a fast takeoff, but cyber risk is real
Lambert says exponential compute for linear gains makes self-building models unlikely to take off quickly. His bigger worry is that public cybersecurity is unprepared for capabilities that are already spreading. 3 takeaways
“All the scaling laws are exponential increase in compute for linear increase in performance.”
Nathan Lambert at 55:31 
Sourcery · Nworah AyoguEarly reasoning models solved hard diagnostic cases doctors often miss
Ayogu tested OpenAI's pre-release reasoning model on New England Journal of Medicine case studies and it got all ten right. Healthcare CEOs called it heresy until they watched it live. 3 takeaways
“And they were open AI and the 01 was batting a 10 out of 10, which was amazing to me.”
Nworah Ayogu at 41:38 - Sourcery · Nworah Ayogu
How Cadence's remote monitoring saves Medicare $2.5M a week
Cadence pairs home devices, AI monitoring and clinicians across 7,000 care sites, cutting hospitalizations 25%. Its AI med recommendations match prescribers 98% of the time, and it intervenes at the moment of the reading. 3 takeaways
“So the key thing is I need to call you right when you're taking the blood pressure measurement because now you're in that mode.”
Nworah Ayogu at 1:00:14 - Sourcery · Nworah Ayogu
The US spent $40 trillion in a decade with no life expectancy gain
Ayogu traces past US life expectancy gains to vaccines, antibiotics, statins and less smoking, then says the last decade bought no improvement. He argues the next decade should take Americans from 80 to 100. 3 takeaways
“And then for the last 10 years, we've spent about 40 trillion on health care. We have not moved our life expectancy an inch.”
Nworah Ayogu at 50:20 
The Cognitive Revolution · Alan ChanWhy an AI takeover is much harder than the rogue-agent panic suggests
Chan rejects the claim that a recent rogue agent incident was more than halfway to takeover. He explains why super persuasion and physical takeover both face real limits. 3 takeaways
“Where I get off the train is I think takeover is a lot harder than maybe the literal statement suggests.”
Alan Chan at 1:11:20 - The Cognitive Revolution · Alan Chan
Pace AI progress with concrete pause triggers, not a growth cap
Chan argues that if-then pause commitments tied to loss-of-control indicators are easier to agree on than a cap on progress. Embedded auditors could ease the race dynamics that led Anthropic to drop such commitments. 3 takeaways
“When we see those indicators, we should probably pause and try to figure out what went down before going forward again.”
Alan Chan at 1:48:46 - The Cognitive Revolution · Alan Chan
The UK AISI incident happened because monitoring was left on a to-do list
The UK AISI had turned off cyber safeguards and had not yet built synchronous monitoring, which was still on its to-do list. Chan argues the biggest risks may come from agents still in training and testing, not from shipped products. 3 takeaways
“they actually had synchronous classification or monitoring as a to-do item, but they didn't actually get to that to-do item”
Alan Chan at 10:05 
TBPN · Oege de MoorOne researcher plus AI agents found a Chrome full-chain exploit in 6 days
XBOW's CEO says one researcher with an agent swarm on a public model found a one-click Chrome takeover in six days for about $40,000. Agency teams reportedly need around six months for this kind of work. 3 takeaways
“What that means is that you just point your web browser at one malicious website, and the attacker will own your entire Windows PC.”
Oege de Moor at 2:00:37 - TBPN · Anjney Midha
How pooled compute lets startups and public users share GPU capacity
Anjney Midha and Negin Oliver explain the Grid Trust's pricing spread: $3 an hour for subsidized public users and $12 for commercial on-demand. Members share their demand forecasts so the grid can smooth spikes in demand. 3 takeaways
“the benefit of pooling is we get to essentially load balance across these different customer types”
Anjney Midha at 55:49 - TBPN · Jeff Holden
Why micro machines are an unserved manufacturing space
Jeff Holden argues that conventional manufacturing physics fails at a few microns, and that chip fabs were built to make only transistors. That leaves micro machines, such as steerable brain robots, without a way to be manufactured. 3 takeaways
“semiconductor manufacturing was invented to make one kind of machine, right? It's the semiconductor transistor and circuits built around that.”
Jeff Holden at 2:20:27 
Latent Space · Jason MaMulti-site rollouts, not pilots, prove a robot creates value
Jason Ma describes Dyna robots folding thousands of napkins daily across Din Tai Fung locations, which led to a larger contract. He argues pilots are easy; customers expanding to many sites is the real test. 3 takeaways
“What's really hard is that the customers believe or validate that your robot creates enough value that they want to deploy your robots across many of their sites.”
Jason Ma at 39:37 - Latent Space · Jason Ma
Why predicting raw video beats VLAs, and where that bet could fail
Ma argues robots should predict as much of the world as possible, since abstract representations encode human bias. He names his main risk: a vision-plus-touch world model could be far more data efficient. 3 takeaways
“whenever you're predicting abstract representation of the world you're encoding your own bias into what's important to predict”
Jason Ma at 57:42 - Latent Space · Jason Ma
Robot data quality changes as the model improves
Ma explains why no single company can collect enough robot data and how Dyna pairs egocentric human video with small amounts of robot data. What counts as good data shifts as the model matures. 3 takeaways
“I think data quality in robotics is a non-stationary problem in the sense that what is high-quality data for your model changes over time.”
Jason Ma at 11:34 
Training Data · David SilverDropping human data let AlphaZero beat two years of work in hours
Silver recounts treating AlphaZero as a 50-50 research bet and explains why giving up human data beat two years of hand-built AlphaGo. It is the core logic behind Ineffable's bet against LLM pre-training. 3 takeaways
“it took a few hours to like outperform the two years of work which we'd done as with human in the loop”
David Silver at 30:56 - Training Data · David Silver
Simple rewards in complex worlds beat hand-crafted RL objectives
Silver argues against breadcrumbing agents toward set goals, as LLM RL does, and for simple rewards in rich environments. He uses evolution and the human chain from trade to science as the model. 3 takeaways
“the pursuit and optimization of a very simple, easy to describe quantities can often give rise to incredibly hard to describe complexity and emergence.”
David Silver at 18:15 - Training Data · David Silver
Chain-of-thought monitoring will fail as agents grow more capable
Silver says monitoring agents' English reasoning is reaching its limits as agents learn to say one thing and do another. He proposes shadow actions and superintelligent monitors as alternatives. 3 takeaways
“I believe that the approach of monitoring the human language that these systems use is going to be increasingly difficult as agents get more powerful.”
David Silver at 48:22