Skip to main content

All In Podcast Episode 282 Recap and Notes July 24 2026

July 24 2026

  • Chinese AI models catching up to US frontier models
    • Kimi K3 seems awesome

    • Moonshot AI seems really good

    • Michael Kratsios alleges distillation

    • Reports that Trump White House might be divided on banning Chinese (or any) open source models

    • Sacks predicted this in E275 - May 2026

    • Any threat is used to take a shot at open source models

      • Try to enter predicate facts into the public record to justify a push to ban these later on
    • Sacks thinks it would be a tragic mistake for US government to ban open source

    • If Anthropic was sincere about their distillation concerns, then they would push for a ban on Chinese models distilling American models, not to hurt American developers’ access to models

      • Why aren’t they stopping it if its happening at “industrial scale”?
    • Instead of KYCing their customers like Chamath suggested they are asking to ban their competitors

    • Hilarious hypocrisy of these labs - especially Anthropic - when they have already paid fines for “distilling/learning” from books, internet, blogs, others’ code, etc

    • Seems to be more of a valuation preservation game by the frontier labs - because it seems the models are getting commoditized faster than anticipated

    • Seems value is more at the application layer and at infrastructure layer

    • If US government intervenes, Chamath thinks it will tank the stock market

      • Seems American enterprises will pay a token tax if they are forced into an AI duopoly - forced into an island of just 2 models
    • If US bans open source that only applies here - rest of the world will run Chinese models for much cheaper and US will lose the AI race

    • Distillation issue - are they taking copies of copyrighted software or are they looking at the output - looking at the output of a product and tuning your product in response is NOT copyright infringement

    • The Chinese models are not - at least right now - being accused of stealing weights or software - so its not so pernicious

    • OpenAI is saying the same thing - in the NYT case issue - they are taking NYTimes output and training their model - if Anthropic and OpenAI have a problem - they should enforce their terms of service

    • Open source is responsible for the internet - AI can benefit the same way - the whole world economy will benefit

    • The value capture is diffused

    • Chinese models doing great


    • Its almost like the web browser in the 90s and how it won’t get commoditized

    • Frontier models are growing exponentially and not being slowed down by the open source models

    • Jason thinks their IPOs are going to get derailed

    • Anthropic is being unseemly - most successful tech company in history - is racing to the government to protect themselves from their future competitors

    • This is like flopping in the NBA

    • China has 8x the electricity production and 20x the manufacturing capacity compared to US - by compressing knowledge/services economy which is US’s advantage they are commoditizing US’s advantage - creation/usage of bits gets commoditized and they are left with their advantage of creation/usage of molecules

    • Most open source can do 95% of all tasks

  • Anthropic lawsuit settlement - $1.5B
    • JCal taking victory lap (called it in 2023 and 2024)
    • But the nuance here is they didn’t pay for even 1 book (that would have been fair use) - instead they apparently “stole” the content from libgen
    • The hypocrisy here is they think its fair use to train their models and derive their own weights - even against the wishes of the authors - and they think the only content that should be protected is their own output
    • They didn’t even claim initially that the Chinese companies were committing IP theft - they claimed a national security threat because those models don’t have guardrails
      • If they do then they will lose all these court cases
    • They might have made a fatal mistake by arguing that distillation is IP theft
      • Hoisted by their own petard
  • Google and Tesla earnings vs AI capex
    • Google incredible earnings - Google Cloud doing very well but both Google and Tesla reported negative cash growth (first time for Google) so their prices dropped
    • Google’s 20+ year average return on invested capital since going public is 32% - Chamath thinks they will do great
    • The best thing that can happen to them is 500 different models proliferate and they work with all of them
    • Google will spend 20% of US military budget in cash this year - Ryan Petersen tweet
    • Google also own a good chunk of SpaceX and Anthropic
    • GCP might be the best place to capture value from AI because of all the context/state they maintain about your business
    • Should Apple get in on this game? They could buy a neo-cloud
  • Socialism corner
    • John Adams quote about how sacred private property rights - this is fundamental to the founding of America
    • Private property rights are the foundation of liberty in America
    • All anarchies end in tyranny
    • Always starts with the framing as “you are bad” to the private property owner
    • The other issue is that the problem is the delinquent tenants who are hurting the ability of the landlord to service the building - it affects the neighbors even more - old people who can’t afford to go anywhere, poor people are hurt
    • Rich liberals can afford to have luxury beliefs about public spaces because they don’t live in those spaces
    • If you want low rent, you need to have more units, if you want more units, you need faster permits
    • In Argentina they got rid of rent control and rents went down
    • Socialists never learn
    • Now they want to ban credit and background checks on tenants and banning evictions - so what can the landlord do?
    • So what incentive does the renter have if he’s not going to get evicted
    • NYC already has 50000 ghost apartments

Comments

Popular posts from this blog

All In Podcast Episode 271 Recap and Notes May 1 2026

May 1 2026 Fun with names OpenAI revenue and growth issues Apparently missed user target, revenue targets (internal) Reported issues between Altman and Friar - prompted joint statement maybe They are almost trading at the price equal to their spending commitments On the other hand, GPT-5.5 is doing very well and it seems their compute commitments may be saving the day for the coding use case in Enterprise The key chokepoint is power - they don’t have enough to get tokens out Too much red tape 40% of data center jobs have been canceled recently - why won’t that negative trend continue? This will benefit the Hyperscalers - Oracle, Microsoft, Meta, Google, Amazon This could help Elon with xAI/SpaceX/Cursor 4 : 2 : 1 typical market share of a mature market Cool MIT paper shows we could reduce inference cost by 10x with no loss of accuracy This could increase usage Maybe Google has cut into OpenAI’s consumer market share Clearly these compan...

What's your favourite Iron Maiden song?

Part of the NWOBHM ( ignoramuses can go ahead and google that :P ), Iron Maiden have rocked for a long time, with many brilliant compositions (yes compositions, as intricately crafted as any symphony you philistines - the cadence of mcbrain's drums, the richness of dickinson's voice, steve harris's incomparable lyrics and I will stop before waxing lyrical about the guitars of Dave Murray, Adrian Smith and Janick Gers....man!). So what is your favourite Iron Maiden song? I ask because frankly I don't think I have one, and I doubt if its possible for me to single any one out. (get it -- "single"? I use it while referring to songs? mmmpff :D, k, never mind) (wrt the inaugural blog post which was in keeping with the spirit of this webpage, I thought I might as well clear the air regarding the only thing meaningful in the whole piece - the title )

Mxyztplk

kljfkldsajfks djfsdfklj23kj2kl3j423kl j4l23k jrkl;qjr;lwqej klqwjdklasdjklj234kj23kjsa kljdaskd'asdasd sad w 23 424k 23l4k23lk 4l23 4l32 kweerkwqe23 14 wqd asda d 234 2134 23 42343rldla;dkl;sakdl;kasl;dkas;ldk;asdklkasdklal23l;k4klsaldksaldksl;adk