Bitduke

AI’s research taste keeps getting better, especially with Fable 5.2 being able to predict experiment results before actually running them means fewer wasted experiments, less compute, and faster research progress models getting better at this too

Google just turned up the heat on Nvidia again, thanks to AI agents yesterday google announced two more TPUs, bringing the total to four this year. the new 8i Max doubles compute and memory bandwidth in the same rack space, while new software makes TPUs easier to use and AI

good tip on how to get more out of Fable 5.2

Carbon has arrived

time to test Fable 5.2 on TreeBench

can a robot (re)create a masterpiece

this is what 700,000 brushstrokes simulated by Fable 5.2 look like

Fable 5.2 also made a massive jump on the Remote Labor Index, from 21.3% with Opus 5.5 to 41.3%! the benchmark tests AI on 240 real freelance projects, like building websites, or creating presentations, with human judges comparing its work against professionals. basically, Fable

another leaderboard just got crushed by Claude Fable 5.2 it scored 16.4% on SemiAnalysis’ GDPx, which tests AI agents on difficult real-world professional tasks, nearly doubling (!) Fable 5.1’s 8.5% and leaving Opus 5.5 at 12%, even though it barely improved in law and finance

the public gets smarter ai, but how much it can help build even smarter ai is still up to Anthropic with Fable 5.2 and Mythos 5.2, it deliberately limited gains in kernel optimization and llm architecture research during training seems like they’re trying to slow the loop of ai

upcoming GPT-6.2 Astra will need to make a real leap to land where Fable 5.2 is right now

interesting results for Fable 5.2 on a reasoning without chain of thought benchmark, one of the few benchmarks where Fable didn't come out on top turns out that if you remove chain of thought, Fable scores around 145, still behind GPT-6.1 Sol (~149) and GPT-6 Astra (~159)

just imagine someone telling you a year ago that a model could one-shot something like that, you’d have thought they were out of their mind but now this is actually our reality

pretty good practice for testing what each new model can catch that the previous ones missed

Fable 5.2 seems to be making massive leaps across a bunch of different benchmarks

#1 spot on Epoch AI’s (EAI) leaderboard with the Fable 5.2 release! Claude Fable 5.2 scored 174 on Epoch AI’s ECI, compared with 169 for GPT-6 Astra and Opus 5.5. that gap is really noticeable on the chart, especially after seeing frontier models clustered so closely together i

Nvidia scandal? markets in the red? Anthropic doesn’t seem scared of the market at all no plans to cancel its IPO and still targeting next week, November 19 or 20, per WSJ pretty smart move to release two obviously superior AI models (really good just check benchmarks) just a

is gene therapy with ai now on the table? Mythos 5.2, one of Anthropic’s powerful models that shares Fable’s underlying model, built on existing protein-folding tools to accurately predict the structures of 28 of the 34 viral shells tested, which can carry gene therapies into

oh, so that's how it works! only those who've been subscribed to Claude for over a year got 3 resets, everyone else got fewer

unfortunately, Fable 5.2 isn't much better than Opus 5.5 at kernel optimization and LLM architecture research. seems like Anthropic simply doesn't want to put such powerful tools in competitors' hands