


Founder and CEO, Mithril (@mithrilcompute). Orchestrating Compute. Fmr Research Scientist @GoogleDeepMind, Deep Learning Team. CS PhD @Stanford ML, Systems


Jensen’s contributions to the world go far beyond building a successful company; he helped build an industry and created a platform upon which others can build their dreams. Jensen’s philosophy is unique in how much it emphasizes “additionality” and expanding the pie--doing “as much as needed, but as little as possible” to move the world in a better direction. He’s, of course, visionary, but also an extremely generous and positive-sum thinker and actor. In reference to Morris Chang, Jensen once said: “The world is full of successful people, but heroes are rare.” “There is a difference between success and impact. I think Morris – his career, his philosophies, TSMC, its strategy, its core values – is absolutely a study in industrial revolution.” Well said, and the same applies to Jensen himself.
I'm terribly sad @JensenHuang is no longer leading @nvidia. Thank you Jensen. For the tireless work, for the focus, for the wisdom, for the support, for the vision, for the chance you gave me to do my life's work @nvidia. We will do our best to push this work forward.

Across our projects @GoogleDeepMind, we are witnessing a genuine phase transition in the pace of scientific discovery. You will see a glimpse of this in our releases in the next few weeks.

It's always been about sample_efficiency x num_samples. Methods that can't agglomerate FLOPs (e.g. aren't well suited to run on GPUs, with their abundant parallel FLOPs but comparatively limited memory bandwidth) can't assimilate many samples. To use a funny analogy, sample_efficiency:gross_margin::num_samples:revenue. Structured methods with strong sample efficiency that can't scale aren't interesting. However, in defense of the "you aren't bitter lesson pilled enough" contingent, thin margin but ultra scalable businesses can be somewhat interesting. The opposite is less true.
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO

Misinterpreting bitter lesson and an incredibly short period when scaling transformers did most of the legwork has reward hacked brains of a whole generation of scientists. The more compute you’re pouring in training you’re model, the *more* important it is what algorithm you’re using. There are algorithms that scale and algorithms that don’t and the only bitter lesson is that a lot of energy was spent on directions that were dead ends in the end. At the same time there are whole dimensions of scaling completely unexplored.
This is also going to get lot of young scientists upset because they stopped developing their insights thinking scaling transformers alone is sufficient and felt defeated due to “the bitter lesson”

btw, in the age of AI, intellectual incisiveness and being able to hone in on good questions is more important than ever questions about the taxes on a model like this... are not good questions questions about the handful of undergirding core assumptions and where they come from -- those are good and fair

i recall once reading someone explain the distinction between *precision* and *accuracy* in investing precision without accuracy is analyzing the Macy's or Blockbuster org charts and financial statements in great depth to form an investment thesis without seeing Amazon and Netflix coming accuracy without precision is correctly discerning that "the internet is going to be huge," and then going all in on the leaders at the time... Cisco, AOL, and Yahoo you clearly need both in good measure, and @shaunmmaguire has demonstrated both... so probably smart to try and engage with what he's suggesting people critiquing the tax-related precision should instead focus on asking whether the model is even remotely accurate, b/c if it is... the implications are massive and far-reaching. a hallmark of a really good research question is that both the negative and affirmative answers are useful. the answer to a bad question doesn't matter either way. there's a ton of alpha if Shaun is right... and also a lot if he's wrong. deciding whether you agree or disagree is probably one of the highest-leverage things you could do if you're an investor.
I hereby challenge @UnstructuredCap to a public debate How about today at noon PT? Consider this a financial duel

Cool work, Albert and team!
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.

can't believe it's actually working
OpenAI has announced to three new discoveries in mathematics, computer science, and physics: - Local smoothing and the Kakeya conjecture - The Exponential Time Hypothesis is false - Existence of quantum chromodynamics and completions of the Standard Model

vLLM is all you need to reproduce DeepSeek's performance.
🚨🚨🚨 vLLM is all you need to reproduce DeepSeek’s performance. With Wide-EP and EPLB, vLLM can match DeepInfer’s decode performance within 6% and generate $39/gpu/hr in revenue at DeepSeek’s off-peak pricing, at 200 P90 TPS on the AgentX workload. Specifically: • Under DeepSeek’s decode configuration (DeepSeek-V4.1-Flash, 131K context, nextn=3), vLLM on GB200 achieves throughput within 6% of DeepInfer’s published results across batch sizes 8–256. • On AgentX, vLLM pushes the Pareto frontier further on GB300, reaching up to 1M tokens/s/GPU and generating $38.96/gpu/hr in revenue at 196 P90 TPS, based on DeepSeek’s off-peak API pricing. • Also, vLLM extended low-latency kernel support for DeepSeek-V4.1-Flash, achieving up to 826 TPS per user on AgentX.

"The invention of the boat reduced the importance of swimming, but enabled the discovery of new lands." Beautiful analogy from @ylecun about the opportunities and changes with AI, in mathematics and in science more broadly.

We are entering a new era of discovery now: openai.com/index/sharing-ai-…

Everyone has been obsessed by Jev and what it can do, and they should be! Fast typed decisions >> slow text generation. It's a complete game changer. But how do you do that with all the data you have? We just dropped 𝚊𝚒_𝚍𝚎𝚌𝚒𝚍𝚎(), which runs decision models natively on all your data in @databricks . Now you can run system one decisions on all your big data, see screenshot. databricks.com/blog/introduc…

After >$15B of development costs (per the S-1) And almost 15 years of research Starship has delivered its first orbital payload She will change the course of humanity


Quite an interesting turn in LLM diversity, if the market starts looking for specialized models that are optimized and branded for different use cases. Should be very good for the market and open more room for non-frontier neolabs!