📰 News
➡️ Turbulence as OpenAI Makes Waves
This week, OpenAI credibly claims to have solved one of the Millennium Prize Problems: 7 long-standing problems in mathematics, each with a $1m reward for their resolution. The problem in question relates to the Navier-Stokes equation for modelling fluids, and this should be a huge cause for celebration, but the community reaction has been at best mixed.
There are allegations that OpenAI learned of a promising approach being pursued by other researchers (including an Anthropic employee), then raced to scoop the result from underneath them, potentially even using user data from those researchers in training the model that cracked the problem. There’s also the broader problem that maths research is not just about the destination, and the community loses a lot when frontier labs with unlimited resources race to the solution without sharing the journey.
Source: Scientific American
➡️ The Frontier Fails to Reach British Shores
Anthropic dealt a blow to the UK’s AI Security Institute (AISI) this week, as they declined to give the organisation access to Mythos 5.1, their latest frontier model capable of cybersecurity research. While they did give access to the more locked-down Fable 5.1, the failure to release the more dangerous model to a world-leading organisation does suggest a possible turn to a more protectionist policy amongst US labs.
It’s certainly a disappointing turn. It’s now just a question of whether the UK can continue growing its reputation as a world-leader in AI security if no one is willing to put their most dangerous models under the microscope.
Source: IT Pro
➡️ Astra: Literally Too Good To Be True?
When Astra came out last week, one of the most impressive results was its 99.9% score on ARC-AGI-3, a benchmark designed to resist memorisation. However, independent runs of the same benchmark have since given a less impressive score of 62.7%. So what gives? It turns out OpenAI were using a custom harness which allowed it to preserve reasoning between attempts, giving it a leg up versus the models it was being compared against.
Is it cheating? No, not really, it shows the model can really do these things, but it’s not a like-for-like comparison like might have been reasonably assumed.
Source: TNW
➡️ Claude and Present Danger
Anthropic’s threat intelligence report is sobering reading. In the past eight months, they have detected malicious actors, including state-backed ones attempting to use their models for cyber-espionage, biological weapon development and covert surveillance. Coupled with warnings of existential risks, this can all seem overwhelming, but there does at least seem to be an increasing awareness of the problems that we are starting to face, and hopefully an increase in appetite for tackling them quickly and responsibly.
Source: BBC News
99.9% with the house harness, 62.7% without it. That 37-point gap is the whole story of modern evals in one number. The model didn’t get dumber between runs. The scaffolding changed, and scaffolding is half the product now. Nobody in production ships a raw model; they ship a model plus retries, routing, tool definitions, and a harness that remembers what it was doing. So when a lab reports a score without disclosing the harness, they’re grading their engineering and asking you to credit the weights. Publish the harness or publish nothing. Anything else is a demo wearing a benchmark’s clothes.