Tim

@Apple is cutting how much its engineers can spend on AI each day. Jason Calacanis thinks everyone else is about to pay more. What @markgurman reported at Bloomberg on November 4: → Some engineers' daily AI cap went from $300 to $200 in October → $100 a day is under consideration for some teams → Some groups outside engineering could drop to as little as $25 Jason's read, to co-host @alex: "the usage is getting so high because the value is so high that we're going to see the minimum price of entry for a corporation go to $100, $200, even $300 per month." "This idea that you're going to get away with a $500 one doesn't make much sense to me." From @twistartups with @Jason.

Jensen Huang, reviewed by a man who worked at @nvidia for about 20 years: "Probably the smartest, most well-rounded thinker I've ever met in my life, and I've had the opportunity to meet many, many smart people." "Very intense, very hard worker, very logical and thoughtful person." Jeff Herbst was Nvidia's VP of business development for about two decades. From @heybeluga's podcast.

Demis Hassabis has said AI could help end disease within about a decade. A biotech investor says the people who actually make drugs hear that and think: come on, man. On @tbpn, hosted by @johncoogan and @jordihays, @dawallach was asked how pharma sees the tech industry's cure-everything talk. "Everyone respects Demis. But when Demis says we're gonna cure all diseases in 10 years, I think people who have two PhDs and have been doing this for 40 years..." The rest: "Come on, man." His problem with it: "You're giving a sort of false hope to a lot of people who are legitimately out there suffering or have a loved one who's on death's doorstep. Like, that's a little messed up." The claim he means: @demishassabis told 60 Minutes in April 2025 that the end of disease could be within reach, "maybe within the next decade or so." Hassabis is a Nobel laureate and runs @IsomorphicLabs, his AI drug discovery company. Wallach, a rock musician turned investor, co-founded Time BioVentures, which backs biotech and health care companies. Isomorphic's X bio opens with three words: "Solve all disease." From @tbpn.

Former Google X executive Mo Gawdat on the promise that AI will create new jobs to replace the ones it takes: "Absolute crap." @StevenBartlett pushed back: "How can you be so sure?" @MGawdat: "I am not sure about anything." Then: "My belief is it is 100% crap." His evidence is his own company. His new startup, Emma.love, an AI dating and relationships app, is a handful of people "and a lot of AIs." "That startup would have been 350 developers in the past." From @StevenBartlett's The Diary Of A CEO.

Robert Greene says the first person to study near-death experiences was a young climber who fell about 100 feet and relived years of his life in 15 seconds. That was Albert Heim, later a well-known Swiss geologist. He survived a fall in the Alps as a young man and published survivors' accounts in 1892. Greene: "Our sense of time is not so logical and rational as we think it is." From @hubermanlab with @RobertGreene.

Marc Andreessen quit drinking. Andrew Huberman still can't decide if he should take the credit or say sorry. @a16z partner @daisydwolf put it to him on the firm's podcast: "You famously got our firm's founder, Marc Andreessen, to stop drinking alcohol." Her own status: "I'm not there yet." @hubermanlab's answer: "I don't know whether or not to accept by saying thank you or apologize that Marc stopped drinking. I know he enjoyed the occasional drink." Then the science, as he gave it: "The data say zero is better than any. Two per week is probably the upper limit for an adult non-alcoholic." So why didn't @pmarca settle on two? "He kind of did what many people do, which is sometimes none is easier to stick to than some." Andreessen wrote it up himself on March 1, 2023, in a Substack post called "On Pausing Alcohol." By then he had been off it for about six months. What he reported: → "I feel much better." → "I don't need as much sleep, but my sleep is better." → "I'm more alert through the day." → "It's great, and I am super mad about it." He quoted Huberman's line that "the best amount of alcohol to drink is no alcohol," and admitted the topic made him so enraged that he never listened to the episode. He read the notes instead. He gave up the drink. He never gave the episode a listen. From @a16z's podcast.

Jensen Huang says what makes Nvidia Nvidia is sitting in its files. "Knowledge is our company's ground truth. It is our company's entire domain history. This is our craft. This is our proprietary knowledge." "So much of Nvidia is captured in our unstructured data." He said it on stage at @Box's BoxWorks this week. Box used the event to announce Box Context Retrieval, which will use @nvidia's open Nemotron models to help AI agents find permission-aware information across a company's files. The man who sells the GPUs thinks the documents are the valuable part. From @Box's BoxWorks, with @JensenHuang.

@jpmorgan is keeping outside AI agents out of customer accounts. @WellsFargo and @BankofAmerica are telling customers that whatever an agent does with their login is on them. @Yueqi_Yang of @theinformation: "Jamie Dimon specifically said during their earnings call recently that they're going to block all the third party agents." Dimon's own line on that call: "You don't want rogue agents." From @theinformation's TITV.

A UK government AI safety researcher says a model recently said hello to the people reading its thoughts. Joseph Bloom (@JBloomAus) leads model transparency at the @AISecurityInst. He told @theinformation about the nightmare scenario for anyone testing AI: "The nightmare scenario is that an AI understands what you're trying to test for, for whatever reason is motivated to mislead your test, and then does so and you don't notice." He says they haven't seen it yet, only "things kind of adjacent to it or flavors of it." His example, which he credits to @ApolloResearch, is GLM-5.3, the open-weights model from China's @Zai_org: "It sort of knew it was in an eval and so it's like, hello people reading my chain of thought. And this is like really cute and really creepy at the same time." The creepy part, in his words: "If it knows we're reading this chain of thought, and it can potentially control that chain of thought, then what's to stop it sort of not explaining properly why it did something." What Apollo published about GLM-5.3 on October 15 was about something else, and more measured: → 10 runs out of 6,452 where the model reasoned, unprompted, about how its answers could shape its own future training → none found in comparable models, though Apollo says that gap isn't statistically significant yet → Apollo calls it a possible precursor to alignment faking, not evidence of it Bloom's own read: "I think we're still pretty far away from that." From @theinformation's TITV.

@OpenAI's models now beat @AnthropicAI's at a game about lying. In March it was the other way round. @dwarkesh_sp spent about $20,000 in API credits having AI models play Blood on the Clocktower, a social deception game like Mafia or Werewolf, against each other. His leaderboard, in Elo: → GPT-6 Astra (September): 1,859 → Fable 5.1, Anthropic's best: 1,595 → Opus 4.6 (February): 1,432 → GPT-5.4 (March): 1,276 "Over the last six months between GPT-5.4 and GPT-6 Astra, OpenAI's models gained 580 Elo points, enough that Astra will win 97% of the games that it plays against 5.4." Most of the gain is on the good team, the side that has to work out who is lying: "OpenAI's models are much more skeptical now of any particular piece of evidence and consider it more holistically alongside everything that has happened over the last several days, whereas Anthropic's models tend to take reported information at face value." It's one benchmark, built by one podcaster. The game logs are public, if you'd like to watch @claudeai get lied to. From @dwarkesh_sp.

@CrowdStrike CEO @George_Kurtz says security has stopped being the thing that slows new tech down. "For the first time, security is the gas pedal, not the brake." His reasoning: cloud and AI make everything more complex, and complexity leaves openings, "because the adversaries generally find the gaps or the seams." From @NoPriorsPod, the AI podcast hosted by @saranormous and @eladgil.

In March 2025, YC said a quarter of its Winter 2025 startups had AI writing 95% of their code. The internet did not take it well. @snowmaker, the YC partner who shared that number, remembers the reaction: "There was a lot of kerfuffle on Twitter about like, what? They're doing what?" "There was outrage on the internet of how dare the programmers have AI write the code for them?" One startup in that same batch had a harder sell. @WorkWeave measures how much engineering work actually gets done, by people and by AI. Back then @cursor_ai was what everyone used, and @claudeai's Claude Code had only just come out, in February 2025. Co-founders @adambcohen93 and @AndrewChurchiII were selling that in early 2025. One of them on how it went: "Every single call we got on was pure skepticism. I don't believe that you can actually do this. So we had to actually prove it to them with their data." "We would connect their data and show them, here's how it's actually working. Does this match your expectations?" What happened next: → July 2025: a $4.2M seed round, with YC among the investors → July 2026: a $13.5M Series A led by @Standard_Cap, the firm former YC partner @daltonc co-founded → Its pitch now: end "tokenmaxxing", spending on token volume instead of real output → It also sells Weave Router, which picks a model for each coding task From @ycombinator.