You can get really embarrassingly far with bootstrappingthis might be more powerful than naive distillationCan Bölük: Trick Gemini Flash into thinking it has the moves of…
Germany has fallenelie: tl;dr: they trained on 10 epochs of a "very light rephrasing" of EVERY gpqa diamond EVAL data :)
> The AI model advances at about [Intelligence]^0.075, or the the 13th root of input intelligence [1,2].no, it ain't gonna be like thatRamez Naam: This AI…
RT GoodfireRe This is worth being careful about!Naive setups can degrade interp tools’ validity, but these probes are run on a *frozen copy* of the model…
Of note, Jurgen Schmidhuber might be the only guy here whose agenda is "self-replicating intelligent robots for GERMANY AND JAPAN to compete against THE UNITED STATES…
…oh, rightof course they instantly did that@TheZvi hardest hitGoodfire: Our team spent months developing RLFR, our method which uses probes on a model's internals as reward…
> commissioned a drawing of himselfBroooo… you really think Jürgen would have used anyone else's IP as he dunks on virgins stealing his ideas? he draws…
Donald's Deal Wednesday(PSA: nobody wants to make deals, that's his own kink)*Walter Bloomberg: TRUMP: IRAN CALLED A WHILE AGO; THEY WANT TO MAKE A DEAL
Deepest Lore...Alex Semendinger 🔸: @_a9lim at long last, we have Done Some Research as described in the classic LessWrong post "It Is Reasonable To Research How…
same energythere's no hard qualitative boundary between "Erdos problem" and "Putnam problem"OpenAI just invests into doing these publicity stuntsI would obviously love it if DeepSeek claimed…