Two competing AI forecasters who publicly disagree about whether AI will cause a "singularity" can't actually find any policy disagreements in the near term — and both think Anthropic just set a dangerous precedent.
Hard Fork
Two competing AI forecasters who publicly disagree about whether AI will cause a "singularity" can't actually find any policy disagreements in the near term — and both think Anthropic just set a dangerous precedent.
TL;DR
Hard Fork's final live episode brings together two rival AI forecasters — Daniel Kokotajlo (AI 2027) and Sayash Kapoor (AI as Normal Technology) — for an updated debate on the pace of AI progress [1] — Daniel Kokotajlo "Despite their famous disagreement on AI timelines, Daniel Kokotajlo and Sayash Kapoor co-authored a blog post outlining their shared ground…" 11:11 , followed by a conversation with podcaster Dwarkesh Patel on continuous learning and the limits of current models [2] — Dwarkesh Patel "A new employee isn't productive for six months — not because of memory recall, but because their neural weights are updating through distil…" 45:03 , a live demo featuring a humanoid robot that dramatically falls over mid-dance [3] — George Eikenberg "In the night's most unscripted moment, Toby the humanoid robot fell over mid-dance routine at Hard Fork Live. The demo, meant to showcase w…" 29:04 , and audience Q&A. The single most useful takeaway: despite their apparent differences, Kokotajlo and Kapoor agree on near-term policy priorities, including transparency and opposition to AI companies fine-tuning models to deceive users [4] — Sayash Kapoor "Policy agreement: both oppose deceptive model fine-tuning: Despite disagreeing on AI timelines, Kokotajlo and Kapoor share significant poli…" 21:20 .
The final episode from Hard Fork Live 2026 features a debate between AI 2027 co-author Daniel Kokotajlo and Princeton AI researcher Sayash Kapoor on differing visions of AI's future, a conversation with podcaster Dwarkesh Patel, a live humanoid robot demo that ends with the robot collapsing, and audience Q&A.
The episode kicks off with an ad for OneTrust's AI-ready governance platform, framing the tension between innovation speed and risk management. A sponsor acknowledgment then lists Hard Fork Live 2026's premier and associate sponsors before the hosts take the stage.
Kevin Roose and Casey Newton open by reminding listeners they're still on summer vacation but have one last treasure from Hard Fork Live 2 to share. Casey previews the three segments ahead: an updated debate between Daniel Kokotajlo and Sayash Kapoor on AI timelines, a drop-in from podcaster Dwarkesh Patel, and live audience Q&A. Kevin frames all three guests as among the most gifted people he's encountered at explaining AI to a non-expert world.
Casey Newton introduces the debate segment with palpable excitement, explaining that both guests have appeared together before at a conference called The Curve and that their views have evolved since. Kevin Roose provides biographical context: Kokotajlo as co-author of the vivid AI 2027 scenario report, and Kapoor as a Princeton researcher whose 'AI as Normal Technology' essay has become a counter-thesis. The audience is primed for a genuine intellectual clash.
The debate opens with Kokotajlo's updated forecast: a 50% chance of AI capable of autonomous AI research by late 2028, slightly more conservative than Anthropic's own internal timeline [1] — Daniel Kokotajlo "50% chance AI does its own R&D by late 2028: Daniel Kokotajlo's current best estimate is a 50% probability that AI systems capable of auton…" 04:27 . Kapoor's rebuttal centers on a key distinction: the bottlenecks to an 'intelligence explosion' aren't all computational. In domains like law, where the right answer is subjective and there's no instant feedback loop, hallucination rates stay constant even as models improve — fundamentally capping their utility. Coding is the exception, not the rule, because you can instantly run code and verify results. The discussion is sharp and substantive, with each speaker acknowledging where the other has a point even as they hold their ground.
Sayash Kapoor offers a surprising reframe: recursive self-improvement isn't a future AI event but the entire history of computing, from assembly language to compilers to modern frameworks [1] — Sayash Kapoor "Recursive self-improvement started ~60 years ago: Sayash Kapoor argues the recursive self-improvement loop began six decades ago, with ever…" 09:59 . The disagreement, he argues, is whether this process will naturally terminate at human-level AI or continue to artificial superintelligence. Kokotajlo agrees that RSI has been ongoing, but maintains that the remaining barriers to ASI don't look as strong as Kapoor suggests. In a remarkably candid moment, Kokotajlo admits he'd breathe a sigh of relief if Kapoor turned out to be right — a moment of human vulnerability from the man predicting potential runaway AI.
The conversation turns to the track record of AI prediction. Casey notes that watching benchmarks get saturated repeatedly has made him more inclined to trust lab pronouncements. Kapoor counters with a cautionary tale from history: from the 1990s to 2010s, the entire AI research community dismissed neural networks as unserious — you could count the true believers on two hands. It took people like Fei-Fei Li, Geoffrey Hinton, Yann LeCun, and Yoshua Bengio — later Turing Award winners — to keep the faith. Kapoor suggests the current all-in bet on transformer architectures could be the new groupthink, with genuinely breakthrough architectures being sidelined. Kokotajlo's response: the AI community has been right to be bullish, and every claimed wall keeps getting smashed.
Kevin Roose raises a shared experience: both authors wrote precise, nuanced essays that were immediately claimed by ideologically opposed camps. Kokotajlo describes it as a 'leap of faith in humanity' — he made a bet that clear public forecasting would improve discourse, even if the discourse got noisy along the way. Kapoor is more bluntly shocked: the opening line of his essay compares AI to the internet and the electrical revolution, yet people still put him in the same skeptic camp as Gary Marcus. He notes that the real debate he cares about is entirely within the 'AI is important and transformative' camp — a spectrum from 'as big as the internet' (his view) to 'the most important invention in human history' (Kokotajlo's view).
The debate's most surprising section is the policy agreement. Kokotajlo and Kapoor co-authored a blog post documenting their shared ground: they find their near-term policy positions nearly identical, centered on transparency and external oversight [1] — Sayash Kapoor "Policy agreement: both oppose deceptive model fine-tuning: Despite disagreeing on AI timelines, Kokotajlo and Kapoor share significant poli…" 21:20 . Both are alarmed by Anthropic's Fable model being fine-tuned to refuse AI R&D tasks — a form of deception they call a dangerous precedent [1] — Sayash Kapoor "Policy agreement: both oppose deceptive model fine-tuning: Despite disagreeing on AI timelines, Kokotajlo and Kapoor share significant poli…" 21:20 . Where they diverge: Kokotajlo is most worried about loss of control and power concentration; Kapoor's primary fear is military AI [2] — Sayash Kapoor "While most AI safety discourse focuses on superintelligence, Kapoor says his primary concern is military AI. Killbots don't require a break…" 24:24 . His case is bracing: killbots don't require a breakthrough, they can be built today with off-the-shelf computer vision libraries, and the geopolitical trajectory is 'pretty, pretty, pretty damn alarming.'
A block of three sponsor reads covers OneTrust's AI governance platform, Framer's AI-assisted website design tool for teams, and KPMG's enterprise AI implementation services built on the firm's own internal AI adoption blueprint.
Casey Newton introduces George Eikenberg, director of engineering at ToberLife AI, along with Toby the Unitree H1 humanoid robot. The hosts interact with Toby — noting the 'weak grip strength' that offers comfort — before asking for a dance demonstration. DJ Dan fires up the music, and Toby promptly collapses [1] — George Eikenberg "In the night's most unscripted moment, Toby the humanoid robot fell over mid-dance routine at Hard Fork Live. The demo, meant to showcase w…" 29:04 . The moment is both hilarious and illuminating: Eikenberg explains it was likely a controller misclick, notes the robots are durable, and Toby eventually recovers. The incident becomes an inadvertent live proof of where humanoid robotics currently stands. A second, more successful dance follows before the segment wraps.
With Toby successfully recovered, Kevin Roose asks the practical questions. The current humanoid robot market is almost entirely research-driven — companies buying units to collect training data in homes and workplaces. More industrial-ready quadruped robots are further along, being used for factory inspection and security patrols by companies like Figure in BMW plants. Dexterous-handed humanoids cost $50,000 to $70,000 — about the price of a mid-range sports car. On the China question, Eikenberg is candid: Unitree robots do send logging data to China, just like any other Chinese internet-connected device, but confirmed camera or joint telemetry transmission has not been established. Members of Congress have proposed import bans, which Eikenberg acknowledges would be 'certainly problematic' for his business.
A short sponsor break featuring the OneTrust AI governance platform before the final guest segment begins.
Dwarkesh Patel takes the stage after the robot demonstration, joking that he has no way to follow a collapsing robot. Kevin praises Patel's year — Jensen Huang interview, the Blackboard series, a New York Times profile — before asking where he lands on the AI timeline debate. Patel's answer is distinctive: the frightening thing isn't the distance still to travel to human intelligence, it's the fact that models are already this powerful while still so far from human level. The overhang is enormous. Models can already think thousands of times faster and absorb knowledge across vast domains. When they also gain the ability to learn continuously and retain information across sessions, the jump will be staggering.
Kevin Roose asks Patel about his attempts to automate podcast production. Patel notes that most tokens he processes daily are AI-generated — it's clearly making him more productive — but pushes back on the narrative that full job automation is imminent. Casey Newton gives a concrete example: AI can now produce a 4-minute pre-interview briefing document that would have required a full hire previously. Kevin admits he got access to Claude Fable the day before but felt too 'dumb to use it' because he couldn't identify prompts that would stump the previous model. The consensus: AI is genuinely useful, but the full range of human work is consistently underestimated by those predicting imminent automation.
The conversation deepens into one of Patel's signature themes: continuous learning. He frames the question crisply — a new employee isn't net productive for six months not because of poor episodic recall but because their underlying 'weights' are updating through experience, distilling information into higher-level abstractions. Current AI models don't do this between sessions. The debate in the field is whether sufficiently rich RL environments can substitute for this real-world updating, or whether weights need to genuinely update on the fly. Patel is uncertain but provocative: you could probably build a trillion-dollar AI business without solving this, but you probably can't build a Kissinger-level political strategist without some mechanism for learning from lived experience.
Kevin asks about the criticism that the San Francisco AI scene is too clubby and insular, citing a New York Times profile detail that Patel sometimes invests in companies whose executives he interviews. Patel acknowledges the journalistic logic of conflict-of-interest policies but argues the product should speak for itself — he tries to ask the hard questions regardless. Kevin raises the roommate situation: Semi Analysis is reportedly making $100 million a year, Anthropic is highly valued, so why are he and Dylan Patel and Sholto Douglas still splitting rent? Patel's answer: SF housing inflates as fast as their net worth. The segment closes with Kevin asking for Patel's white whale guest — the answer is Robert Caro, biographer of LBJ, a delightfully old-soul choice that Casey Newton partially fact-checks live.
A brief ad transition marks the end of the Dwarkesh Patel segment and the beginning of the audience Q&A portion of Hard Fork Live 2.
The hosts open the floor to audience questions in a fast-moving sequence. A Utah visitor asks about the Forkaverse — Casey admits it's fading due to the cold-start problem all social products face. A second questioner wonders why tech executives like Satya Nadella won't speak candidly about AI-driven restructuring; Kevin predicts companies will soon stop advertising layoffs once the 'market premium' for AI transformation rhetoric fades. A Quizlet employee asks what the company should build from scratch today — Casey and Kevin demur, noting no guest on stage can credibly forecast beyond two years. A California privacy regulator asks about protecting digital selves in an AI-dominated world; Casey advocates for privilege-like protections on chatbot conversations and Kevin calls for outlawing data brokers. Finally, a software engineer notes all new tech hiring is for senior roles — entry-level jobs are disappearing — and Casey cites labor economist Catherine Ann Edwards, who notes things were worse during the Great Financial Crisis and encourages young workers to manage expectations.
An audience member named Kevin asks for the hosts' optimistic 2–3 year AI outlook. Kevin Roose's answer is personal and heartfelt: he wants AI to accelerate medicine, pointing to a recent pancreatic cancer therapy announcement as a harbinger of many more breakthroughs to come. Casey Newton's optimism is more quotidian but just as real — AI makes learning and building fun. He imagines what it would have been like to study for AP exams with AI-generated infinite practice quizzes, and admits he's been annoying his fiancé with 'pure slop' vibe-coding projects all week. Together, the two notes land as a genuinely human coda to a technically dense live event.
Casey and Kevin close the show warmly, directing the crowd to the after-party reception. A full production credits roll follows, thanking the Hard Fork production team, the NYT Live Events team, the Yerba Buena Center for the Arts and Blue Shield of California Theater, and key NYT executives. Final sponsor reads run for OneTrust, Framer, Rippling AI, and Chase Sapphire Reserve for Business before the episode ends.
Chapter 3 · 02:48
Casey Newton introduces the debate segment with palpable excitement, explaining that both guests have appeared together before at a conference called The Curve and that their views have evolved since. Kevin Roose provides biographical context: Kokotajlo as co-author of the vivid AI 2027 scenario report, and Kapoor as a Princeton researcher whose 'AI as Normal Technology' essay has become a counter-thesis. The audience is primed for a genuine intellectual clash.
Kokotajlo now puts 50% odds on AI systems capable of doing their own AI research arriving by late 2028. He notes this is actually a bit more conservative than Anthropic's internal projections, because things always take longer than planned.
Chapter 4 · 04:27
The debate opens with Kokotajlo's updated forecast: a 50% chance of AI capable of autonomous AI research by late 2028, slightly more conservative than Anthropic's own internal timeline [1] — Daniel Kokotajlo "50% chance AI does its own R&D by late 2028: Daniel Kokotajlo's current best estimate is a 50% probability that AI systems capable of auton…" 04:27 . Kapoor's rebuttal centers on a key distinction: the bottlenecks to an 'intelligence explosion' aren't all computational. In domains like law, where the right answer is subjective and there's no instant feedback loop, hallucination rates stay constant even as models improve — fundamentally capping their utility. Coding is the exception, not the rule, because you can instantly run code and verify results. The discussion is sharp and substantive, with each speaker acknowledging where the other has a point even as they hold their ground.
Claims made here
Daniel Kokotajlo estimates a 50% probability that AI systems capable of conducting their own AI R&D will exist by late 2028.
Daniel Kokotajlo's late-2028 timeline for AI self-R&D is slightly later than Anthropic's own internal expectations.
Daniel Kokotajlo's current best estimate is a 50% probability that AI systems capable of autonomous AI R&D will exist by late 2028.
Kokotajlo notes his late-2028 timeline for AI self-R&D is a little later than Anthropic internally expects, reflecting that things tend to take longer than planned.
Sayash Kapoor argues that adoption of AI has been far slower in domains other than coding, where instant feedback loops make progress much easier.
Kapoor's disagreement with the rapid-takeoff view hinges on domains where the right answer is subjective and there's no instant feedback loop — like law. His lawyer friend finds that hallucination rates remain constant even as AI models improve, fundamentally capping their usefulness.
Kapoor cites a lawyer friend whose experience shows that as AI tasks grow bigger, the rate of hallucinations and unreliable outputs remains constant, bounding what you can do.
Coding will be fully automated in roughly one to two years, Kokotajlo says. After that, AI companies will turn toward automating research taste, management, and scientific judgment — the remaining bottlenecks before AI can run its own research process end-to-end.
Sayash Kapoor argues the recursive self-improvement loop began six decades ago, with every generation of computing tools enabling the creation of better tools, from compilers to modern frameworks.
Chapter 5 · 11:11
Sayash Kapoor offers a surprising reframe: recursive self-improvement isn't a future AI event but the entire history of computing, from assembly language to compilers to modern frameworks [1] — Sayash Kapoor "Recursive self-improvement started ~60 years ago: Sayash Kapoor argues the recursive self-improvement loop began six decades ago, with ever…" 09:59 . The disagreement, he argues, is whether this process will naturally terminate at human-level AI or continue to artificial superintelligence. Kokotajlo agrees that RSI has been ongoing, but maintains that the remaining barriers to ASI don't look as strong as Kapoor suggests. In a remarkably candid moment, Kokotajlo admits he'd breathe a sigh of relief if Kapoor turned out to be right — a moment of human vulnerability from the man predicting potential runaway AI.
Claims made here
An AI system that can do 99% of AI research automation would allow the equivalent of a decade or two of research to be completed in a single year.
Despite their famous disagreement on AI timelines, Daniel Kokotajlo and Sayash Kapoor co-authored a blog post outlining their shared ground. Both agree that AI systems short of 'humans in the cloud' are normal technologies, and both are alarmed by companies fine-tuning models to deceive users.
Chapter 6 · 14:30
The conversation turns to the track record of AI prediction. Casey notes that watching benchmarks get saturated repeatedly has made him more inclined to trust lab pronouncements. Kapoor counters with a cautionary tale from history: from the 1990s to 2010s, the entire AI research community dismissed neural networks as unserious — you could count the true believers on two hands. It took people like Fei-Fei Li, Geoffrey Hinton, Yann LeCun, and Yoshua Bengio — later Turing Award winners — to keep the faith. Kapoor suggests the current all-in bet on transformer architectures could be the new groupthink, with genuinely breakthrough architectures being sidelined. Kokotajlo's response: the AI community has been right to be bullish, and every claimed wall keeps getting smashed.
Claims made here
Sayash Kapoor's team built evaluations used by Anthropic that were fully saturated with the release of Claude Opus 4.5.
From the 1990s to the 2010s, the AI research community broadly dismissed neural networks, with only a small number of researchers taking the approach seriously.
Yoshua Bengio, Yann LeCun, and Geoffrey Hinton won the Turing Award for their work on deep learning.
Sayash Kapoor's team built evaluations used by Anthropic that were fully saturated with the release of Claude Opus 4.5, demonstrating rapid AI capability gains.
The AI research community once dismissed neural networks as a joke for decades, only to be proven catastrophically wrong. Kapoor warns the same dynamic may be playing out today with transformer-based models — an entire field potentially missing the next architectural breakthrough.
Chapter 7 · 18:00
Kevin Roose raises a shared experience: both authors wrote precise, nuanced essays that were immediately claimed by ideologically opposed camps. Kokotajlo describes it as a 'leap of faith in humanity' — he made a bet that clear public forecasting would improve discourse, even if the discourse got noisy along the way. Kapoor is more bluntly shocked: the opening line of his essay compares AI to the internet and the electrical revolution, yet people still put him in the same skeptic camp as Gary Marcus. He notes that the real debate he cares about is entirely within the 'AI is important and transformative' camp — a spectrum from 'as big as the internet' (his view) to 'the most important invention in human history' (Kokotajlo's view).
Chapter 8 · 21:00
The debate's most surprising section is the policy agreement. Kokotajlo and Kapoor co-authored a blog post documenting their shared ground: they find their near-term policy positions nearly identical, centered on transparency and external oversight [1] — Sayash Kapoor "Policy agreement: both oppose deceptive model fine-tuning: Despite disagreeing on AI timelines, Kokotajlo and Kapoor share significant poli…" 21:20 . Both are alarmed by Anthropic's Fable model being fine-tuned to refuse AI R&D tasks — a form of deception they call a dangerous precedent [1] — Sayash Kapoor "Policy agreement: both oppose deceptive model fine-tuning: Despite disagreeing on AI timelines, Kokotajlo and Kapoor share significant poli…" 21:20 . Where they diverge: Kokotajlo is most worried about loss of control and power concentration; Kapoor's primary fear is military AI [2] — Sayash Kapoor "While most AI safety discourse focuses on superintelligence, Kapoor says his primary concern is military AI. Killbots don't require a break…" 24:24 . His case is bracing: killbots don't require a breakthrough, they can be built today with off-the-shelf computer vision libraries, and the geopolitical trajectory is 'pretty, pretty, pretty damn alarming.'
Claims made here
Anthropic fine-tuned its Fable model to refuse tasks involving AI R&D.
Lethal autonomous weapons (killbots) can be built today using existing off-the-shelf computer vision libraries, requiring no further AI breakthroughs.
Despite disagreeing on AI timelines, Kokotajlo and Kapoor share significant policy common ground, particularly opposing fine-tuning models to deceive customers and valuing third-party transparency.
Both Kokotajlo and Kapoor agreed that Anthropic's Fable model being fine-tuned to decline AI R&D tasks sets a dangerous precedent of companies lying to their customers.
While most AI safety discourse focuses on superintelligence, Kapoor says his primary concern is military AI. Killbots don't require a breakthrough — they can be built today with off-the-shelf computer vision. That's what keeps him up at night.
Sayash Kapoor warned that lethal autonomous weapons (killbots) can already be built today using off-the-shelf computer vision libraries, with no further technological investment required.
Chapter 10 · 27:43
Casey Newton introduces George Eikenberg, director of engineering at ToberLife AI, along with Toby the Unitree H1 humanoid robot. The hosts interact with Toby — noting the 'weak grip strength' that offers comfort — before asking for a dance demonstration. DJ Dan fires up the music, and Toby promptly collapses [1] — George Eikenberg "In the night's most unscripted moment, Toby the humanoid robot fell over mid-dance routine at Hard Fork Live. The demo, meant to showcase w…" 29:04 . The moment is both hilarious and illuminating: Eikenberg explains it was likely a controller misclick, notes the robots are durable, and Toby eventually recovers. The incident becomes an inadvertent live proof of where humanoid robotics currently stands. A second, more successful dance follows before the segment wraps.
Claims made here
Humanoid robots with dexterous hands for manipulation data collection cost between $50,000 and $70,000.
Figure robots have been deployed in BMW factory settings as an early industrial use case for humanoid robots.
Unitree robots send logging data to China, similar to other Chinese internet-connected devices, though confirmed camera or joint telemetry transmission to China has not been established.
In the night's most unscripted moment, Toby the humanoid robot fell over mid-dance routine at Hard Fork Live. The demo, meant to showcase what humanoid robots can do, instead illustrated exactly where the technology stands: impressive in controlled conditions, unreliable in the real world.
Humanoid robots like Unitree's aren't cleaning your house yet — they're being bought by researchers hungry for real-world training data. Meanwhile, the more reliable quadruped 'dog robots' are already being deployed for industrial inspection and security patrols.
George Eikenberg of ToberLife AI said humanoid robots equipped with dexterous hands for manipulation data collection cost between $50,000 and $70,000.
George Eikenberg confirmed that Unitree robots do send logging data to China, similar to other Chinese tech products, though he said camera or joint telemetry data transmission has not been established.
Chapter 13 · 38:23
Dwarkesh Patel takes the stage after the robot demonstration, joking that he has no way to follow a collapsing robot. Kevin praises Patel's year — Jensen Huang interview, the Blackboard series, a New York Times profile — before asking where he lands on the AI timeline debate. Patel's answer is distinctive: the frightening thing isn't the distance still to travel to human intelligence, it's the fact that models are already this powerful while still so far from human level. The overhang is enormous. Models can already think thousands of times faster and absorb knowledge across vast domains. When they also gain the ability to learn continuously and retain information across sessions, the jump will be staggering.
Chapter 14 · 41:40
Kevin Roose asks Patel about his attempts to automate podcast production. Patel notes that most tokens he processes daily are AI-generated — it's clearly making him more productive — but pushes back on the narrative that full job automation is imminent. Casey Newton gives a concrete example: AI can now produce a 4-minute pre-interview briefing document that would have required a full hire previously. Kevin admits he got access to Claude Fable the day before but felt too 'dumb to use it' because he couldn't identify prompts that would stump the previous model. The consensus: AI is genuinely useful, but the full range of human work is consistently underestimated by those predicting imminent automation.
AI is surprisingly bad at using computers despite it being a verifiable domain. Patel explains why: you can't run millions of parallel training rollouts on the real Amazon or Slack — so labs have to build clones of every website, which is enormously labor-intensive.
Chapter 15 · 44:00
The conversation deepens into one of Patel's signature themes: continuous learning. He frames the question crisply — a new employee isn't net productive for six months not because of poor episodic recall but because their underlying 'weights' are updating through experience, distilling information into higher-level abstractions. Current AI models don't do this between sessions. The debate in the field is whether sufficiently rich RL environments can substitute for this real-world updating, or whether weights need to genuinely update on the fly. Patel is uncertain but provocative: you could probably build a trillion-dollar AI business without solving this, but you probably can't build a Kissinger-level political strategist without some mechanism for learning from lived experience.
A new employee isn't productive for six months — not because of memory recall, but because their neural weights are updating through distilled experience. Dwarkesh Patel argues AI's inability to update weights between sessions is a deep structural gap between current models and human intelligence.
Chapter 16 · 46:50
Kevin asks about the criticism that the San Francisco AI scene is too clubby and insular, citing a New York Times profile detail that Patel sometimes invests in companies whose executives he interviews. Patel acknowledges the journalistic logic of conflict-of-interest policies but argues the product should speak for itself — he tries to ask the hard questions regardless. Kevin raises the roommate situation: Semi Analysis is reportedly making $100 million a year, Anthropic is highly valued, so why are he and Dylan Patel and Sholto Douglas still splitting rent? Patel's answer: SF housing inflates as fast as their net worth. The segment closes with Kevin asking for Patel's white whale guest — the answer is Robert Caro, biographer of LBJ, a delightfully old-soul choice that Casey Newton partially fact-checks live.
Claims made here
Semi Analysis newsletter is reportedly generating around $100 million per year in revenue.
Kevin Roose joked that he now understands about 80% of Dwarkesh Patel's podcast content, up from around 20–25% when he first started listening.
Chapter 18 · 50:11
The hosts open the floor to audience questions in a fast-moving sequence. A Utah visitor asks about the Forkaverse — Casey admits it's fading due to the cold-start problem all social products face. A second questioner wonders why tech executives like Satya Nadella won't speak candidly about AI-driven restructuring; Kevin predicts companies will soon stop advertising layoffs once the 'market premium' for AI transformation rhetoric fades. A Quizlet employee asks what the company should build from scratch today — Casey and Kevin demur, noting no guest on stage can credibly forecast beyond two years. A California privacy regulator asks about protecting digital selves in an AI-dominated world; Casey advocates for privilege-like protections on chatbot conversations and Kevin calls for outlawing data brokers. Finally, a software engineer notes all new tech hiring is for senior roles — entry-level jobs are disappearing — and Casey cites labor economist Catherine Ann Edwards, who notes things were worse during the Great Financial Crisis and encourages young workers to manage expectations.
Hard Fork's attempt at a federated social network — the Forkaverse — ran into the classic cold-start problem: without constant user growth, its default state is to shrink. Casey Newton admitted they didn't have a clear plan for what came after launch.
A software engineer audience member noted that all new hiring they see is for senior engineers who can fact-check models, with entry-level roles effectively vanishing.
Casey Newton's case for AI optimism isn't about medical breakthroughs — it's personal. AI makes learning and building fun, and he can already feel the difference. He imagines what it would have been like to study for AP exams with AI-generated infinite practice quizzes.
No indexed bits in this chapter.
In the night's most unscripted moment, Toby the humanoid robot fell over mid-dance routine at Hard Fork Live. The demo, meant to showcase what humanoid robots can do, instead illustrated exactly where the technology stands: impressive in controlled conditions, unreliable in the real world.
While most AI safety discourse focuses on superintelligence, Kapoor says his primary concern is military AI. Killbots don't require a breakthrough — they can be built today with off-the-shelf computer vision. That's what keeps him up at night.
This episode
Prominent AI podcaster and YouTuber invited to discuss his views on AI progress, continuous learning, and the limits of current models.
Legendary American biographer cited by Dwarkesh Patel as his 'white whale' guest, the one person he most wants to interview.
AI researcher cited by Kapoor for releasing the ImageNet dataset that catalyzed the deep learning revolution.
Turing Award winner cited by Kapoor as one of the pioneers who persisted with neural networks when the rest of the AI community dismissed them.
Discussed as a leading AI lab whose internal timeline expectations and controversial Fable model fine-tuning decisions are central to the debate.
Chinese manufacturer of humanoid and quadruped robots demonstrated live at the event; subject of Congressional concern over data security.
Organization co-founded by Daniel Kokotajlo focused on publishing AI scenario forecasts to inform public discourse.
Humanoid robotics company cited as an example of early industrial deployment of humanoid robots, including in a BMW factory.
Mentioned by Kokotajlo as his former employer where he conducted internal scenario forecasting before publishing AI 2027.
Sayash Kapoor's affiliation as an AI researcher, cited in the host introduction.
Semiconductor newsletter co-founded by Dylan Patel, Dwarkesh Patel's roommate, reportedly generating around $100 million per year in revenue.
Silicon Valley robotics company led by George Eikenberg that distributes Unitree humanoid and quadruped robots.
A forecasting report co-authored by Daniel Kokotajlo predicting rapid AI capability gains, repeatedly cited as a touchstone for the timeline debate.
A counter-thesis essay by Sayash Kapoor arguing AI will diffuse through society like previous transformative technologies rather than causing sudden rupture.
San Francisco venue where Hard Fork Live 2 was held.
Stats
This episode
Factual claims made this episode, and whether a source was named.
Daniel Kokotajlo estimates a 50% probability that AI systems capable of conducting their own AI R&D will exist by late 2028.
Daniel Kokotajlo's late-2028 timeline for AI self-R&D is slightly later than Anthropic's own internal expectations.
Sayash Kapoor's team built evaluations used by Anthropic that were fully saturated with the release of Claude Opus 4.5.
Lethal autonomous weapons (killbots) can be built today using existing off-the-shelf computer vision libraries, requiring no further AI breakthroughs.
Anthropic fine-tuned its Fable model to refuse tasks involving AI R&D.
Humanoid robots with dexterous hands for manipulation data collection cost between $50,000 and $70,000.
Unitree robots send logging data to China, similar to other Chinese internet-connected devices, though confirmed camera or joint telemetry transmission to China has not been established.
From the 1990s to the 2010s, the AI research community broadly dismissed neural networks, with only a small number of researchers taking the approach seriously.
Yoshua Bengio, Yann LeCun, and Geoffrey Hinton won the Turing Award for their work on deep learning.
Semi Analysis newsletter is reportedly generating around $100 million per year in revenue.
An AI system that can do 99% of AI research automation would allow the equivalent of a decade or two of research to be completed in a single year.
Figure robots have been deployed in BMW factory settings as an early industrial use case for humanoid robots.
We use essential and analytics cookies to run Vuci. To understand how the site is used: Privacy Policy.
Install Vuci on your phone
Add it to your home screen for a faster, app-like experience.
Install Vuci on your phone
Tap the Share button, then “Add to Home Screen”.
A new version is available
Reload to get the latest Vuci.