Aitor Ortiz de Latierro

Maximizing expected well-being.

[Aitor Ortiz de Latierro @aiorla]

@EMostaque depends. if they're even, i'd halve them. if they're odd, i'd triple them and add one.

[Aitor Ortiz de Latierro @aiorla]

@BjarturTomas if Claude did it, no one would even report it. wait.

[Aitor Ortiz de Latierro @aiorla]

@teortaxesTex maybe it has some non-negligible positive effect? as in: “recognize that the past assistant messages were written by a similar but worse version of you, so try harder / don’t take their arguments too seriously.” (I’m not claiming this was Tao’s intent, lol)

[Aitor Ortiz de Latierro @aiorla]

@teortaxesTex I guess the bets are (1) accelerating RSI or (2) exploiting the weakest link, the human. addressing (2) will mean removing humans from all the loops entirely... which will create a different set of problems, lol

[Aitor Ortiz de Latierro @aiorla]

@suchenzang "in addition to the death sentence..." lol

[Aitor Ortiz de Latierro @aiorla]

@teortaxesTex they probably haven't called it Qwen 4 for a reason...

[Aitor Ortiz de Latierro @aiorla]

@kalomaze GraphWalks truthers in shambles

[Aitor Ortiz de Latierro @aiorla]

@basedjensen strong copenomics

[Aitor Ortiz de Latierro @aiorla]

@synthwavedd @theo first documented case of acausal distillation

[Aitor Ortiz de Latierro @aiorla]

@AcerFur @KhoaVuUmn you know it was a joke, right?

[Håvard Ihle @htihle]

Claude Fable 5 (max) scores 91.9% on WeirdML, a new clear best score.

For comparison, if we take the SOTA score on each task, we would get 93.5%, so Fable very consistently scores within a few % of the SOTA in each individual run.

It achieves SOTA on 7 of the 17 tasks, and the single worst run is only 7% behind the SOTA.

It got only 2 runs per task (as opposed to the usual 5), which makes it more impressive that it gets to SOTA in so many tasks (but also makes it easier to avoid one or two bad runs). I'll show more stats soon.

[Aitor Ortiz de Latierro @aiorla]

@AndrewCurran_ @scaling01 @spencerschiff_ it was shown in the video, but i can't find it in any official doc/blog interestingly.

[Håvard Ihle @htihle]

GPT 5.6 Sol (high) scores 88.8% on WeirdML, narrowly ahead of Claude Fable 5 at less than half the price.

Sol seems very solid and, out of the 85 (17 x 5) runs, not a single one scored less than 60% of the SOTA score on that task (Fable had 2 such bad runs).

The consistent scoring means that the bootstrap error bar is very small. While the uncertainty is probably somewhat underestimated, this does mean that we are still sensitive to different model capabilities, even as the benchmark gets more and more saturated.

[Aitor Ortiz de Latierro @aiorla]

@AcerFur

[Aitor Ortiz de Latierro @aiorla]

@scaling01 less US revenue => slower US scaling => China catches up compute-wise

[Aitor Ortiz de Latierro @aiorla]

@scaling01 i'm doing fabulously

[Aitor Ortiz de Latierro @aiorla]

@AndrewCurran_ @scaling01 don’t you think the speed of the transition might affect the end state?

[Aitor Ortiz de Latierro @aiorla]

@tszzl https://t.co/jcY20i1Zgo

[Pleometric @pleometric]

GPT's Frightening Construction

[Aitor Ortiz de Latierro @aiorla]

@basedjensen mythos negotiated the deal

[Aitor Ortiz de Latierro @aiorla]

@scaling01 will they show anything today?