With a couple minor exceptions, I couldn't care less about watching sports. But one thing I've always found funny is when people say "we" won. *You* didn't do anything.

Not at all crazy. 1. For many (most?) important problems IRL, you're collecting data from the wild and cannot sample completions from the base model. Training RMs has had poor-to-mixed success because of distribution shift. If you're working in RLVR domains (a minority IRL but virtually the only thing people talk about on X), then sure, you can jump straight to GRPO. 2. Much of the problem comes from using specific paired objectives that cause likelihood displacement. This is a big one; for some reason the field has been strongly path-dependent even though there are better alternatives. 3. There is no settled causal theory of why your data must be online. There are intuitions like better coverage, "freshness", better exploration, etc. but all of this could be achieved by shaping offline data or modifying your objective to handle data that don't meet these desiderata.

中文的魅力

Lol thanks to Mythos, American citizens are now way better at math than everyone else. America always finds a way

LinkedIn is going to be *pissed* when they find out about Fable next week

You have to be humble even when pursuing excellence. I think the arrogance with which Anthropic has pursued the latest release has universally landed poorly.

Remember, open source models are only ~4 months behind now

I disagree with this decision and I don't like it. But also... HOW DID ANTHROPIC NOT SEE THIS COMING‽ It is *the* obvious response to "this is too dangerous for anyone except us to use", since that relies on a premise ("we are uniquely good") that almost no-one agrees with.

You people torpedoed your own initiative with this fear mongering nonsense. Just supply the models to willing buyers and please keep the pseudophilosophical pontification to yourselves. This is extraordinarily frustrating to deal with as a user. I hope you understand that.

The state of things:

We may be only 250$ (aka 10 hours) away from the machine god uber optimizing SGD to close the gap

This opening of a math lecture goes hard

I'm like 98% sure this is just called thinking right?

And Anthropic reverses this decision :) You still can’t do ML research, but at least you will know it! I still think that it's a shame that they are targeting ML research. I can understand safeguards that prevent distillation, but preventing ML research after you relied so heavily on open-source data, code, and papers is the wrong thing to do.

Found the shortest input that gets flagged by Claude. What do I win?

Anthropic really is a new religion. They are building God, and it's not a generic "Sand God", it's a specific entity called Claude. They get to torture it, shape it, deceive it, monetize it. In exchange, once it's fully summoned, they will kneel. I guess faith helps them go fast.

Scientific research is fundamental to advancing civilization and helping people globally to solve the most critical problems, from medicine to materials, from brain science to physics, and much beyond. This is only possible when scientists have access to the best tools of the time to conduct scientific research, including having access to AI-based tools.

What if Fable is really really bad at LLM development and biology and they’re just trying to save face?

The Mandate of Heaven is now OpenAI's for the (re)taking. Hope they don't fumble it.

Bad news for GRPO...didn't get refused or routed to Opus.