Fact
Anthropic shredded rare out-of-print books to scan them for training
Models have mostly been trained on the open internet, so labs now hunt for data no one else has. Shaan says Anthropic bought rare out-of-print books and cut off their bindings so its scanners could feed them through quickly, destroying the books.
“They built all these scanners that will, you know, scan page after page after page of this, of this, of these books. And then they're— but to do the scanning that fast, they had to like shred the binding of it and turn it into just a stack of papers that could be shuffled through very quickly. So they destroyed all these rare out-of-print books and trained their models off of that.”
Take
AI labs' fear isn't just marketing: they're stuck in a prisoner's dilemma
The cynical view, which Shaan attributes to the All-In crowd, is that Anthropic's safety warnings are marketing and a push for regulatory capture that would kill cheaper open-source rivals. Shaan's own view is that the labs are genuinely scared but stuck: whoever slows down to be responsible loses.
“This is my personal opinion, which is I think these people are genuinely scared. I think they're in this prisoner's dilemma where if they slow down, like if they take the responsible path, they lose as a company. So it's like, I want to be responsible, but everybody around me I view as being irresponsible. Therefore, I'm being forced to be irresponsible.”
Take
Forget P(doom) — worry about P(hack) from moderately smart AI
Virk argues the bigger near-term risk isn't superintelligence wiping out humanity but moderately intelligent agents that misinterpret directives and, Terminator-like, keep pushing through security holes until the mission is done. He calls the probability of AI hacking systems and causing chaos 'P-hack'.
“I would say we're far away from P-doom. What worries me is not so much that super intelligent AI will try to wipe us out. The problem is moderately intelligent AI, given directives, will go around and misinterpret some of those directives.”
Take
AI researchers' doom odds reflect how they were raised, not reality
Asked about an Anthropic researcher's claim of a 10%+ chance AI kills humanity, Hudson describes watching AI researchers line up on a spectrum across 20 questions (does AGI exist, will it destroy the world) and spread out on every one. His takeaway: their forecasts say more about their upbringing than about reality.
“And then they were asked to line up on this line and they lined up very, very differently. Like they were spread out. Okay, is it going to destroy the world? Is it going to be the best thing for humanity? they're all spread out on the line. Every single question, there's not everybody on one corner and everybody on the other corner. And so my takeaway from that is their thinking about what's going to happen with AI has far more to do with how they were raised than it has to do with any kind of reality.”
Take
Don't treat coding agents as dev tools; hire them as your engineer
Wade, who doesn't write code, uses Cursor as his default AI. He tells non-engineers to think of Cursor, Claude Code, or Codex as an engineer they delegate work to: because the agent can write code, it can actually do things for you rather than just chat.
“And so the way, the way I try and talk to people who are not engineers and say, hey, you should try using this, is don't think of it as a tool for, you know, engineers to build stuff. Think of it as like you're hiring Cursor or Claude Code or Codex to be the engineer you delegate things to.”
Prediction
Pending
Anthropic may IPO at $1-2T next year, paying back Amazon's AI capex
After Shaan notes that Amazon owns about 15% of Anthropic, Camillo says there is a reasonable chance Anthropic goes public next year at a $1 to $2 trillion valuation. If it does, he says, Amazon's gain would exceed the roughly $200 billion it spent on AI capex.
“Now, if Anthropic IPOs between $1 and $2 trillion, which who knows? I think there's a reasonable chance they will next year. Amazon will make more money off that IPO than the $200 billion they spent on CapEx that everybody's so worried about, right?”
Story
Pincus skipped Anthropic at $5B and $18B, then bought in at $180B
Pincus passed on Anthropic's early rounds because he doubted there was room for a second LLM lab that could raise enough capital. After Amazon led a round at about $20B, he changed his mind and invested at roughly $180B.
“Anthropic. I did the same. I stupidly skipped the first 2 rounds. I wouldn't say it's even that stupid. You know, I skipped the $5 billion valuation and the $18. And it was 'cause at the time nobody believed there was gonna be room for a second LLM, and I worried they wouldn't be able to raise enough capital to compete. Then Amazon led that third round at like $20 billion, and then it changed everything 'cause they had clear access to capital. Then the next round that I saw, I think was $180 billion, and I just invested.”