Sebfox @Sebfox1

@DavidSacks What is Chamath doing there, is he serving drinks?

Vinod is probably not the right guy to be lecturing others about having "no decency or sense of proper behavior" lmao

Matthew Prince 🌥 @eastdakota

4 months ago

One more I forgot until just reminded: 3. Khosla Ventures wanted to invest in our Series C. Vinod took me, Michelle, and Lee out to dinner after he’d given us a term sheet. Near the end, Michelle and Lee got up to use the restroom. Vinod leaned over and said: “I’m impressed with

219 201 8K 3.1M 2K

when your boss makes a joke thats not joking but you really need that job

keyboard_arrow_left Previous keyboard_arrow_right Next

Kylie Robison @kyliebytes

2 days ago

i have watched this video 20 times

291 413 7K 2.7M 4K

I Donald Trump, president of the Unites States, do solemnly declare....

Software was always an asset because it was expensive to make: build once, sell to many, maintain for years. That's ending. Most AI-summoned software will be used once and thrown away, a tool for this afternoon's dataset or an app for one trip. Nobody will maintain it, because regenerating it costs less than understanding it. Film made photos precious. Digital made them disposable, and photography got bigger. Same jump, now happening to software.

HOLY MOTHER OF MATHEMATICS!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! Google is working on a math-focused variant of its DeepThink model, and its raw thoughts are pretty funny.

keyboard_arrow_left Previous keyboard_arrow_right Next

Junior engineering hiring isn't collapsing because AI does engineering. Two other things are happening. First, AI is the respectable story for downsizing that was coming anyway. "We're realising AI efficiencies" sounds like strategy, and "we over-hired" sounds like mismanagement. Second, AI eats well-specified, low-agency work, and we built the junior role out of exactly that. Do the ticket, implement the spec, learn by osmosis. So AI didn't kill the junior job, it revealed how little of the junior job was engineering. The scarce thing at every level is agency and systems thinking, and the old junior role never asked for either while the new one asks for nothing else.

Ben Evans waves away ChatGPT memory as "stickiness, not a network effect". But stickiness at a billion users is what most moats actually were - Windows held ordinary users through their files and muscle memory, not the developer flywheel. The taxonomy is doing too much work. Same with brand: ChatGPT's consumer brand recognition isn't definitely a moat yet, but it's also not definitely not one. Nobody's mum asks for Gemini by name.

Awesome work from Benchling. They built a wet-lab benchmark without authoring a single question, by diffing a published protocol against the version a scientist actually ran. The standard you want to grade against is usually already in your data. benchling.com/blog/can-llms-…

Sixteen authors, including the man who coined Eroom's law, reviewed a decade of AI in drug discovery. Their verdict: "evidence of clinically relevant impact is, so far, disappointingly limited." Note the shape of that. They concede the methods exist, are applied, and are benchmarked. The failure sits downstream of all three. Five reasons they give. 1) Nobody was building for the clinic. 2) Biological data is conditional, only true in the exact conditions it was measured in. 3) The problem was never specified, so a model can be right on its own terms and useless where someone has to choose. 4) Technology push rather than science pull. 5) And operationalising takes far longer than building. Their fix: benchmarking has to "move on from model validation and instead focus on their ability to improve decision making."

I think Benedict Evans was spot on in his OpenAI essay earlier this year. He works through platforms, ecosystems, network effects and flywheels, then lands on power: can OpenAI get consumers, developers and enterprises to use its systems whatever the systems actually do? That's the real bar for platform status, and it's worth sitting with how high it is. Windows kept its developers and users through long stretches when better options existed. Amazon keeps shoppers who know the same thing is cheaper somewhere else. For OpenAI the equivalent would be people staying on its models even when something demonstrably better ships next door - and that is nowhere near true. Right now, everyone using AI is one better model away from leaving.

Protein folding is basically solved. De novo protein design works. Mutation effect prediction works. All at the scale of a single molecule. A new Cell paper points out that none of it has translated upwards to cells, tissues, or the diseases anyone actually cares about, and argues the reason is structural rather than a matter of scale. Three reasons they give. The mechanisms are combinatorially vast: 47 proteins have to assemble in coordination for one piece of cellular machinery, and no pair tells you what it does. The data does not exist and will not, since even the most ambitious project planned is still ~1000x short of what a language model trains on. And most disease is many cell types interacting, so a perfect model of one cell still would not get you there. You can see it in the benchmarks already. Single-cell foundation models do not consistently beat simple linear baselines out of distribution.

Sutton's bitter lesson says general methods plus compute always eventually beat approaches with human knowledge built in. A new Cell paper argues biology is where it breaks, and the argument is more careful than the usual objection. They do not say the lesson is wrong. They say two things about biology specifically. First, it assumes you can always get more data. Language models train on trillions of tokens accumulated over centuries. The largest biological datasets are a hundred-billion-ish, and the most ambitious profiling project now planned still lands about a thousandfold short. A related result: in predicting immune disease, non-linear models only start beating linear ones once sample sizes pass a million individuals. Second, and this is the interesting move, the priors they want to build in are not human intuition or convention. How two proteins interact is settled by thermodynamics, electrostatics and shape. Encoding that is not hand-coding a heuristic, it is refusing to consider physically impossible answers. Their example is AlphaFold, whose architecture builds in explicit priors about protein geometry, chemical bonds and physical plausibility, drawn from decades of crystallography. The obvious counter is that AlphaFold 3 moved to a more general architecture with fewer of those hand-built priors, which is what the bitter lesson would predict. So it stays a live question, and one worth watching anywhere data is scarce and the rules are genuinely known.

Re-read Ben Evans's "How will OpenAI compete?" this weekend. He makes this analogy to browsers in the 1990s and how despite contemporary opinions, they turned out to be a commodity and not where the value was captured. While this could be true about OpenAI, the analogy could also cut the other way: the search box was also an identical input box over someone else's output, and identical UI didn't stop search going winner-take-most. The value was never in the box there.

Solving a Rubik's Cube with graph theory.

Two models with the same AUC, the standard one-number summary of how well a model separates the good from the bad. Use them to pick: take the most promising 10% of a screening library into the lab, and model 2 gives a 40% hit rate against model 1's 20%. Twice as good. Use them to rule out: screen out the 80% most likely to be toxic and test the rest, and 3.3% of what survives model 1's cut turns out bad against 10% for model 2. Three times as good, the other way round. Same models, same data, opposite rankings. The one number they share said they were equivalent. From a new Nature Reviews Drug Discovery review of a decade of AI in drug discovery. The authors call what is missing the context of use: picking winners, screening out losers and predicting a number are three different questions, and one score answers none of them.

After a year of building evaluation for clinical AI, the biggest shift in my thinking is this: evaluation is not something you have, it is something you do. The standard your judge checks against does not exist on paper. It has to be discovered from real outputs, captured from the experts who hold it, and kept alive as the system and its failure modes move.

Consumer adoption of the 'AI super app' has a loong way to go. 80% of ChatGPT users send <3 messages a day. Half of US adults don't use an AI chatbot at all, and only 24% use one daily. The average ChatGPT user spends 7 minutes a day. Still only around 5% of users pay. OpenAI has stopped publishing messages per day numbers since July 2025.

Seems to me that this aims to solve a set of properties (1 - reliable structured outputs, 2 - uncertainty estimation, 3 - hallucination, 4 - latency & cost), but for every one they either don't really solve it, or it's already solved at the frontier... 1) already solved by the frontier 2) no evidence offered to suggest that this actually calibrates well with humans. looks to just be just a slightly better version of logprobs, which aren't all that helpful anyway 3) definitely doesn't solve this at all. this just guarantees a response from a set of options, but LLM could still be very confidently picking the wrong option 4) you can get extremely cheap & fast with openweight / frontier small models already

Diogo Almeida @CompleteSkeptic

2 weeks ago

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

4K 8K 77K 39.9M 59K