qt cache🪷

NVIDIA's Vera Rubin NVL72 could train large mixture-of-experts AI models using 75% fewer GPUs than Grace Blackwell. Training a 10T-parameter model on 100T tokens in one month would just require roughly 16,000 Rubin GPUs, compared with 64,000 GB200 GPUs.

It's strange how quiet things are coming out of China right now, and how little we hear from the Chinese AI labs. It's as if the US is leaving China in the dust. Four US labs are releasing sota Frontier, and there are no significant releases from China.

*NVIDIA IN TALKS TO ACQUIRE REFLECTION AI: FT What’re the odds reflection can reach the frontier with near bottomless compute? Net negative for OpenAI and Anthropic in the long run I’d imagine Jensen’s also already invested $800m into Reflection

I asked people which neolabs under $10B have the highest talent density: 32 votes: Core Automation (@MillionInt, @_arohan_) 26 votes: Periodic Labs (@LiamFedus, @ekindogus) 20 votes: Flapping Airplanes (@spectorb, @amspector100, @aidanmantine) 18 votes: Standard Intelligence (@G413N, @devanshpandey) 8 votes: Ricursive (@annadgoldie, @Azaliamirh) 5 votes each: - Ineffable (David Silver) - Mirendil (@bneyshabur, @HarshMeh1a, @shayan_, @tararezaeikh) - Recursive (@RichardSocher, @_rockt, @timshi_ai, @josh_tobin_, @CaimingXiong, @jeffclune, @tydsh, Alexey Dosovitskiy) 3 votes each: - Applied Compute (@ypatil125, @rhythmrg, @lindensli) - Isara (@ezhang7423, @hegasz) - Prime Intellect (@vincentweisser, @johannes_hage) 2 votes each: - Elorian (@AndrewDai, @yinfeiy, @SethInternet) - Goodfire (@eric_ho, @DanJBalsam, @banburismus_) - Inherent (@tantumscollins, @edwardfhughes, @LouisKirschAI, @kallyaleksiev) - Trajectory (@rronak_, @QuantumArjun, @MichaelElabd) - World Labs (@drfeifei, @jcjohnss, @BenMildenhall, @chlassner) People couldn't pick a company they work at or founded. Disclosure: I'm a small investor in Applied Compute, Factory, Standard Intelligence, Trajectory and Wafer. I didn't vote. Continued:

cognition is arbitraging boomers against the ai frontier and masking it with olympiad mumbo jumbo

.@sama says the advice to pick one thing to build and pour all your focus into it no longer applies. "The canonical advice that was correct two years ago was, 'Pick one, and put all your effort into it, because you can't build all 30 things.'" "You can now build all 30 things. And you can get people to use them all. And you can work more on the ones that work." "So I think this mindset of lots of ideas, quick feedback, and being open to a lot of things not working is great for this new era for how people build."

So I know Cognition doesn’t care about Little Ole’ Me, and this is just my story, so take it with a grain of salt, but here’s why I don’t think all these crazy scaling revenue numbers by coding agents are going to stick: We’ve been happy Devin users for a while. Really like the product, and like the company. Still do! This isn’t shade at Cognition, and I’m not upset with them, just a sample of why I worry about the competitive landscape for these cos. Anyway, our usage grew to >$100k run rate or so. I think one or two of our users even received an email congratulating them for being Top 200 users of Devin. As usage grew, we added more capabilities and needed a BAA (layman’s terms: an agreement that says they are HIPAA compliant). This required an Enterprise contract (why a piece of paper requires extra money is a topic for another discussion, but it’s very normal…) so we started talking to their sales team. Basically, we were going to ~1.5x our spend in order to get the BAA. Honestly, that was fine with us. We like Devin! Our only request was to pay quarterly (already a big concession from our existing monthly payments, given our relatively small size). They declined, unless we committed to EVEN MORE usage. Before responding, I decided to see how difficult it would be to set up Hermes to do the equivalent work as Devin, and to see what the costs were. I’d host it in our infrastructure and only use providers we already have a BAA with. Two days. That’s it. It’s functionally 100% equivalent. Maybe slightly better in some ways, slightly worse in others. We spend a little bit of time futzing with it, but we did that with Devin, too. The cost is roughly 25% of Devin. So, two days for $100k or so per year savings. All because they didn’t want us to pay quarterly. At some point, other companies are going to catch onto this, too. So congrats on this milestone, but keep finding more ways to provide value and keep customers happy. It’s hard to compete.

someone today said i had the best social team for the muse launch bruh this is all me out here rawdogging these tweets. i don’t let the corposlops near my tweet cannon

A paper without a single experiment is going viral. Either people have not looked at the paper or they asked an LLM to do it for them.

I am so happy the world is increasingly ready to treat AI seriously. Permanent, employee-level access for independent evaluators should be required by law. We will join Anthropic in offering access to @METR before it’s required. We should not only slow capabilities progress but also accelerate alignment research. The right incentives could redirect labs' capabilities research programmes to alignment. Voluntary efforts are fragile. As the stakes continue to rise, it will be appropriate to treat non-compliance or non-disclosure of incidents with the same severity as violations of nuclear safety laws. This must be global. You don’t need trust for a mutually beneficial, verifiable treaty. No country wants humanity to lose control over superintelligence. Let’s extend the hand to China. We need frontier AI researchers to make progress on alignment, _now_. If you are one, ask yourself how you can help. Your impact could be monumental. AI has great upside, but we must avoid catastrophic risks. There is a lot of responsibility in very few hands. Let's get it right.

Rumors have been circulating for three days that Gemini 4 has finished pretraining early, in part due to new discoveries GDM made during the training run. Nothing confirmed, and impossible to know what's true. We may all find out together soon.

Strong agree. I know that there is rivalry between the labs but it's important that we learn to work together given what's coming.

2 bits of gossip i’ve heard lately: • meta tbd hit major pretraining setbacks. not fatal, but apparently enough for delays + finger pointing • ssi had a mini coup, then a wave of departures twitter char limit is oppressive, so dms open for gossip trading 😋

SAT Critical Reading — Passage 1 “I don’t like data centers because they use lots of water” Q: Why does the speaker dislike data centers? A) Water consumption B) Inarticulable intuition about the increase in the rate of human disempowerment over time

The final minute or so of the SA podcast today was interesting. @JordanNanos:“SpaceX chooses to install NVIDIA GPUs, because they are the best. Pretty matter of fact” @fabknowledge looks serious (SA has large clients in Google & AMD who don’t want to hear this). Jordan then notes that a bunch of people hearing Jeff Dean left Google because he wanted to work with NVIDIA GPUs “so bad” 🤔

I see so many people founding because they want to be founders. Almost all lack a contrarian thesis for why the world should be different. There are about a dozen companies that matter each decade and meaningfully shape how the world works. You should generally optimize for contributing meaningfully to one of those.