t55

what timelines do you have in mind

that's a standard feature in cursor, windsurf, etc.

this is what deepmind did 10 years ago lol

For a 100k token context window; all those models are comparable though

gemini 2.5 pro shines for 200k+ tokens

i didn't say they invented everything; in science you always stand on the shoulders of giants

i still think my original statement is fair

yeah, RLVR is still nascent and hence there's lots of noise.

> How can these spurious rewards possibly work? Can we get similar gains on other models with broken rewards?

it's because in those cases, RLVR merely elicits the reasoning strategies already contained in the model through pre-training

this paper, which uses Reasoning gym, shows that you need to train for way longer than those papers you mentioned to actually uncover novel reasoning strategies: https://arxiv.org/abs/2505.24864

so you think it's fake news? another example of a paper with strong claims without much evidence?

agree, the RG evals feel like a fresh breeze

> prolonged RL training can uncover novel reasoning strategies that are inaccessible to base models, even under extensive sampling

does this mean that previous RL papers claiming the opposite were possibly bottlenecked by small datasets?

> I personally think that Gemini 2.5 Pro's superiority comes from having hundreds or thousands RL tasks (without any proof whatsoever, so rather a feeling).

Given that GDM pioneered RL, that's a reasonable assumption

it aged well!

Very cool, reminds me of https://rehearsal.so/

which sort of interviews will you support?

great article!

Other debate topics can be found on the frontpage

Congrats! obviously massive potential for saas automations what made you choose windows over other OSs?

Well, it depends. on some tasks, they surely are already super-intelligent.

If they don't use some of the released kernels, then they are leaving money on the table

Never bet against a cracked team

in a parallel universe...

agreed, quite ambitious to cover this entire week in one post

The level of DS's openness has been really useful to actually understand what LLM folks are working on day in day out

Indeed. Modern LLM research is more about large-scale systems design than deriving gradients by hand. Math matters, but systems win.

Anthropic doubling down on code makes sense, that has been their strong suit compared to all other models

Curious how their Devin competitor will pan out given Devin's challenges

thank you!

hot take: i don't think you even need to understand much linear algebra/calculus to understand what a transformer does. like the math for that could probably be learned within a week of focused effort.

you're welcome!

Yes, it is great for key concepts but a bit outdated. Hence we added an LLM/FA section in the linked post!

I should have been more precise, sorry. Didn't want to imply they entirely ditched CUDA but basically circumvented it in a few areas like you said.

Agreed, not sure how much math is really needed.

lol. i guess this tutorial is about cutting out guido ;)