Yannick Mühlhäuser

AI & Compute Researcher @FLI_org

[Nirmal @Nirmal_Utwani]

The Google Brain (& @GoogleDeepMind) team back in 2015 feels like a stronger fit to this picture: Andrej Karpathy, Ilya Sutskever, Dario Amodei, Noam Shazeer, Demis Hassabis, Shane Legg, Mustafa Suleyman, Jeff Dean, Andrew Ng, Greg Corrado, Geoffrey Hinton, Ashish Vaswani, Jakob Uszkoreit, Łukasz Kaiser, Niki Parmar, Aidan Gomez.

[nxthompson @nxthompson]

This is an amazing quote from Lindsey Graham about when he realized how politics works. https://t.co/sh04ahQOFq

[John Attridge @John_Attridge]

Every nonfiction book is called like Republic of Cheese: How Gouda Saved Democracy and Created the Modern World

[Robert Long @rgblong]

SF hipster who still calls it the Lunar Society podcast

[Yannick Mühlhäuser @yannick___m]

Very to share what I have worked on for the past few months. If you want to transition into AI Verification or AI Security, this is the place to go.

[ERA Fellowship @ERA_Cambridge]

We’re launching a new Frontier AI Security Residency! Apply now at https://t.co/LrwdD9g4Gw

The Frontier AI Security Residency (FASR) is an 8-week, full-time residency supporting ambitious projects across cybersecurity & hardware-enbaled verification for securing frontier AI.

[Jaime Sevilla @Jsevillamol]

Scores on benchmarks have progressed overall faster than the median response for this survey.

The train is moving fast.

[Konstantin Pilz @KonstantinPilz]

If this is correct, then China has now achieved training runs on "domestic" clusters equivalent to about 25k H100s.*

I.e., they are a bit over 2 years behind on training cluster scale.

U.S. AI companies were training on clusters of this size in early 2024: https://t.co/M1rYb1Ac6X

Ofc, there could be larger clusters in China used for training models that we don't know about. E.g., a single cluster of 12k smuggled Blackwells would be >25k H100s.

Meanwhile, the largest U.S. clusters are now approaching 1 million H100e: https://t.co/MjKLwfIUpx

*One 910C has 1054 TFLOP/s INT8, vs 1979 TFLOP/s FP8 for H100. Not really "domestic", because Huawei procured the dies illicitly from TSMC rather than producing them in China.

[Miles Brundage @Miles_Brundage]

https://t.co/4SYPcNckev

[Cedar You @our_decay]

Watched a cute animal video that I knew to be AI all the way through

[Simon Grimm @Simon__Grimm]

I even notice this among very informed people, equating Europe with the European Union.

Many of the most relevant powers still lie with member states (which is useful, as they can do lots of valuable things more quickly than the EU).

[Rob Wiblin @robertwiblin]

The US is achieving levels of growth that leave Europe in the dust.

If it doesn't change track soon the EU will lose its status as the world's pre-eminent regulatory superpower.

[Dario Amodei @DarioAmodei]

In addition to transparency, I now believe frontier models should face mandatory third-party testing for cyber, bio, and autonomy risks—with the power to block or revoke deployment of models that pose catastrophic risk.

[Peter Wildeford🇺🇸🚀 @peterwildeford]

OpenAI joins Anthropic in thinking pausing may be needed 👀

"there should be an international organization that helps [...] make it possible for the world to take coordinated action, including slowing frontier development when needed"

[Peter Wildeford🇺🇸🚀 @peterwildeford]

What does Barack Obama's UN ambassador and the guy who ran against Barack Obama for President have in common?

Thinking "AI safeguards are the highest and most urgent national priority" due to, among other things, risks of extinction.

[Seán Ó hÉigeartaigh @S_OhEigeartaigh]

Very pleased indeed to see this from OpenAI too. Let's get going!

"As frontier AI development continues, we expect national and global coordination to become more important. We have long believed there should ultimately be an international organization that helps coordinate leading AI efforts to reduce catastrophic risk. Cooperation and shared safety standards are an important part of the path forward, especially because the incentives around commercial and national competition are hard to escape. One goal of such an organization should be to make it possible for the world to take coordinated action, including slowing frontier development when needed, so societal resilience, safety, and alignment can keep pace."

[Unworthy Hand @kisstheblade_]

What America needs, what it really needs, is another president so large he can frame mog a water buffalo. This is Pritzker's time to shine.

[Sylvian (#1 Skinamarink Fan) @TheAlexSylvian]

every passage is like "hitler demanded 50000 planes to be made in one year. albert speer swore he would make a hundred thousand planes instead. by using slave labor and at the expense of everything else they made 23469 planes. in that same time the allies made 250 billion planes"

[Stephen McAleer @McaleerStephen]

We need to figure out how to have the option for a coordinated slowdown in the face of recursive self-improvement.

[Mitt Romney @MittRomney]

Our highest and most urgent national priority should be AI safeguards. The risks of AI weapons, pathogens, mass unemployment, surveillance, and even extinction must not continue to be largely ignored.

[Peter Wildeford🇺🇸🚀 @peterwildeford]

OPENAI: "We also see early signs of recursive self-improvement in today's systems". RSI is "potentially the most consequential frontier safety issue of the coming decade."