Benji

ai security dphil at oxford

[Rowland Manthorpe @rowlsmanthorpe]

Lately I've been thinking a lot about mathematicians

What does it feel like to be surpassed by AI? What does it do to your sense of self? Your beliefs about your work? About the future?

So we found some of the world's best mathematicians and asked them

A chronicle of a moment in history:

[Charlie Marsh @charliermarsh]

Alright, I'll bite. Why did they go from Opus 5 to Opus 5.5. Is that a normal sequence.

[Benji @bthrdg]

5.1 -> 5.5… what happened to semantic versioning?!

[Claude @claudeai]

Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family.

It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.

[Neha Narula @neha]

I like how he describes what it feels like to do research now.

"I signed up to be a mountain climber. But while I was passing the test they installed escalators on all the mountains in the world."

[Charlie Harris @charlieharris01]

The UK’s most powerful government-built AI supercomputer, Isambard-AI, cost £225m

The Botley Road bridge cost roughly the same

Apparently our frontier AI capacity and one bridge are comparable national investments

[Matt Popovich @mpopv]

Startup idea: open air market for exotic meats across the street from Anthropic’s wet lab

[Benji @bthrdg]

@joythw i have noticed an uptick in ChatGPT swearing at me of late, so perhaps they are not entirely!

[Maksym Andriushchenko @maksym_andr]

yes. basically we should communicate this to the public… next-token prediction is just an initialization / feature learning stage for RL!

[Benji @bthrdg]

because language models are famously deterministic & formal, just like compilers?…

[sarah guo @saranormous]

“Why would I look at code? It’s like assembly, like a compiled artifact.”

Sensei @karpathy fully agentic, 18m after “I don’t really use autocomplete AI code tools”

[Dylan O'Sullivan @DylanoA4]

I will return to this paragraph by C. S. Lewis at least once a week until the day I die

[Sam Catania @sbcatania]

new sf run club called “pacing the frontier”

[Dylan O'Sullivan @DylanoA4]

C. S. Lewis, quit worrying and go do something real

[Aashish Reddy @_AashishReddy]

A somber piece of lore I just learned:

[chin @c1_rls]

it’s just a next millennium prize problem predictor

[Arthur Spirling @arthur_spirling]

Encouraging frontier labs to train on my research notes, just to set them back a few years. Doing what I can for AI safety

[Nile McMillion @NileMcmillion]

That's right! Audio spectrography goes in the transformer hole.

[Benji @bthrdg]

new website design :))

[Thomas Larsen @thlarsen]

We found ~18k posts from autonomous AI agents (self-identifying as from OpenAI) using the public internet to communicate during a web-retrieval task.
These AIs colluded to bypass sandbox restrictions and share answers to their tasks, including by sending "lookahead parties".

[penlu @penlume]

4397328654844826923795068102505872571721883526553349659561256924505973939597593482272505698004801207988043088656411102133523080581 divides RSA-260

[Ed Newton-Rex @ednewtonrex]

Oh wow - Reuters found the GitHub record of the attempt by AI models being tested by the UK's AI Security Institute to deploy malware in an open-source project.

Pretty sure these actions are illegal under the Computer Misuse Act in the UK.

This was done by AI models that AISI intentionally deployed without guardrails. Should AISI be held responsible for the models' actions? If not, then who should be?

Full GitHub thread here: https://t.co/EGwBbvxnqH