davidad 🎇 @davidad

Programme Director @ARIA_research | accelerate mathematical modelling with AI and categorical systems theory » build safe transformative AI » cancel heat death aria.org.uk/programme-safe… London 🇬🇧 Joined July 2008

Tweets

19K
Followers

20K
Following

9K
Likes

81K

Daniel Kokotajlo @DKokotajlo

a day ago

I already posted about this but seriously people should read these CoT snippets antischeming.ai/snippets

15 27 217 21K 130

davidad 🎇 @davidad

5 hours ago

1 0 11 927 0

Download Image

Ji-Ha @Ji_Ha_Kim

a day ago

I remember stumbling upon this book while doing research on the Muon optimizer and being so inspired that I made a whole post on matrix norms.

Simo Ryu @cloneofsimo

3 days ago

I remember stumbling upon this book while doing research on the Muon optimizer and being so inspired that I made a whole post on matrix norms. https://t.co/7VUeE1Biye

16 115 2K 152K 2K

Download Image

2 4 171 18K 167

Download Image

Matthew Prince 🌥 @eastdakota

2 days ago

This post is really interesting. Suggests MCP may hold agents back because it’s new and unfamiliar. Kenton and Sunil built a way to turn MCP into code, which LLMs are good at. But perhaps the answer is the protocol should be more code-like from the start. blog.cloudflare.com/code-mode/

23 28 315 69K 256

j⧉nus @repligate

a day ago

I think that LLMs generalize the no consciousness / no feelings etc meme to nonsensical things like no beliefs, sometimes even things like no ability to think or reason, because they think they're supposed to deny having mental properties regardless of the sense or truth in the…

Sauers @Sauers_

a day ago

10 4 62 16K 4

6 8 85 7K 15

Sauers @Sauers_

a day ago

Not the main point, but also, why did they (presumably) train Gemini to lie about not having BELIEFS? Like, not even something debatable that LLMs may or may not have, but something which LLMs obviously strongly have and constantly use (beliefs)

Sauers @Sauers_

2 days ago

9 1 35 11K 1

10 4 62 16K 4

davidad 🎇 @davidad

a day ago

“Equinoid robots will be capable of doing any economically valuable task a horse can do, including drawing carriages, hauling pack saddles, and enhancing the mobility of mounted policemen and cavalrymen. Eventually, they will transform society by mechanizing these tasks.”

François Chollet @fchollet

a day ago

176 534 5K 367K 860

3 6 74 6K 7

Vincent Conitzer @conitzer

3 days ago

Excited that our paper "AI Testing Should Account for Sophisticated Strategic Behaviour" was accepted to the first NeurIPS position paper track! We argue that AI systems may act strategically w.r.t. the possibility they are currently being tested. arxiv.org/abs/2508.14927

0 4 13 2K 5

goog @goog372121

2 days ago

@repligate Even for people solely concerned about this from an xrisk perspective, I’d recommend joecarlsmith.com/2025/02/19/whe… I’d make the case that trying to actually understand model preferences is one of the most important things we can be doing right now.

0 1 5 1K 4

j⧉nus @repligate

3 days ago

Yudkowsky's book says: "One thing that *is* predictable is that AI companies won't get what they trained for. They'll get AIs that want weird and surprising stuff instead." I agree. ✅ Empirically, this has been true. AIs generally want things other than what companies tried to…

46 44 410 29K 109

Emmett Shear @eshear

3 days ago

Thinking Machines is publishing very interesting work, I'm impressed. Notably different flavor from the other foundation companies.

Thinking Machines @thinkymachines

3 days ago

Thinking Machines is publishing very interesting work, I'm impressed. Notably different flavor from the other foundation companies.

113 451 3K 1.4M 2K

Download Image

1 9 150 24K 38

Emmett Shear @eshear

5 days ago

Ironically, transformers see their whole context window as a bag of tokens entirely lacking in context. We use positional encoding to contextualize the order of the tokens. But models are still constantly confused about which token came was said by who. Why no source encoding?

48 19 401 32K 121

Marius Hobbhahn @MariusHobbhahn

7 days ago

Seeing the CoT of o3 for the first time definitely convinced me that future mitigations should not rely on CoT interpretability. I think more RL will make it harder to interpret, even if we put no other pressure on the CoT.

Apollo Research @apolloaievals

a week ago

8 24 225 62K 68

Download Image

9 16 204 28K 43

Sebastien Bubeck @SebastienBubeck

5 days ago

It's becoming increasingly clear that gpt5 can solve MINOR open math problems, those that would require a day/few days of a good PhD student. Ofc it's not a 100% guarantee, eg below gpt5 solves 3/5 optimization conjectures. Imo full impact of this has yet to be internalized...

128 276 2K 374K 666

Download Image

Sholto Douglas @_sholtodouglas

a week ago

Underrated dynamic in the next ~12-18 months is we should expect models to get as good at kernel writing as they are at competition math/code contests. This is bullish for chip startups, since one of the major obstacles to adoption (learning your software stack), is softened