JL Austin stay winning
Arb@ArbResearch
For this to work at all, you will need some interesting transforms of the training data
English
Arb
206 posts
Arb
@ArbResearch
Unfocused Research Organization. A consultancy by @gleech & @CharlesD353 & @mishayagudin
Katılım Şubat 2022
42 Takip Edilen1.5K Takipçiler
The current training process (just optimise outcomes bro) blatantly applies selection pressure toward outputs (lying, sycophancy, cheating, etc). If this pressure continues to increase (with RL) we might be in trouble. "Implicit agency" thus replaces instrumental convergence.
English
We recently helped Yoshua Bengio and his LawZero colleagues write up their mathematical argument for Scientist AI. Our role was to be an internal red-team.
The core notion is "consequence-invariance": training that never selects predictors for what their outputs would cause
English
Arb retweetledi
I'm helping out with a new funder! Short applications, deadline in two weeks, decisions hopefully in four weeks. $50k cap because we're covering stuff other people cba with
English
Arb retweetledi
Arb retweetledi
In the last 5 months, 18 Erdős Problems were solved, for the first time, by AI systems working alone. How should mathematics respond?
Prof Václav Rozhoň & I are building a journal to filter this flood of AI maths at scale. We have a plan, funding, & credits; now we need a board
English
Arb retweetledi
@geoffreyirving seems you're right (Erdos problems, human vs teaming vs AI): +17% diff, nearly significant at 5%
thanks to @nc_znc
English
Arb retweetledi
Join us as we publish hermits, informants, translator-traitors, forum freaks, discord archives, pirated plays, psychiatry bloggers, groupchats and television recaps
cristóbal@tobyshooters
Grateful to join @suspendedreason, alongside @ctbeiser, to further Not Nothing's first decade of work towards new forms of reading.
English
Interesting followup claim from @MichaelHassid and friends
gavin leech (Non-Reasoning)@gleech
New paper on a long-shot I've been obsessed with for a year: How much are AI reasoning gains confounded by expanding the training corpus 10000x? How much LLM performance is down to "local" generalisation (pattern-matching to hard-to-detect semantically equivalent training data)?
English
@ArbResearch @nc_znc Using this methodology, I found 204 CVEs for AISLE, avg 6.7 CVSS. The top ~51 have an avg of 8.3. They use much smaller models than Mythos Preview. Idk the degree of human involvement, but this does illustrate that CVEs are not an apples-to-apples comparison of model capabilities
English
@nc_znc took a look at the purported Ant-OAI cyber gap.
[The confounders are severe enough that this could be totally misleading] but the naive view is: Mythos has a decent lead, but OAI isn't far behind.
x.com/peterwildeford…
Peter Wildeford🇺🇸🚀@peterwildeford
My two questions: if GPT-5.5 and Mythos were roughly equal... 1.) where is OpenAI's large pile of vulnerabilities they discovered? 2.) why is the White House kinda ignoring it and letting it be open widely, while meanwhile micromanaging Anthropic's release plans?
English