Doing lesson 15 of the @fastdotai course; deducing how to rearrange convolutions as a matrix product

Correction, I think: Whilst R1-Zero is pure RL, R1 is not: SFT is involved.

@chris_j_paxton @ChongZitaZhang It does make sense when you really think about it

RT @ChongZitaZhang: Motions so efficient, no human reference needed

Reading through the R1 paper. From my understanding, in a nutshell: - R1-Zero is pure RL, with GRPO as the policy - R1 is pure RL, with GRPO as the policy, with some cold start data, and further refinement stages

For generating novel output, that is

Has there been any work done in introducing some sort of entropy/variation/randomness in the process of generating an LLM output?

@abhi1thakur - OpenRouter - TogetherAI - LambdaLabs - FireworksAI a few quick ones off the top of my head

I'm thinking perhaps that it would be the case for novel/extrapolated tasks, rather than the opposite.

@jeremyphoward What ideas do you have in mind for alternative approaches?

RT @WevolverApp: Researchers at the University of Science and Technology of China have designed the octopus-inspired SpiRobs robotic arm. D…

RT @UnslothAI: You can now reproduce DeepSeek-R1's reasoning on your own local device! Experience the "Aha" moment with just 7GB VRAM. Un…

I think the best way to understand the 把 particle is the following example: - 我放书在桌子上 - I put the book on the table - 我把书放在桌子上。 - The book has been put on the table by me Following from the previous example... - 把书放桌子上 - The book is placed on the table

Putonghua abilities I need to improve on: - Emphasizing the 4th tone - Differentiating between the 2nd and 3rd tones - Paying closer attention to tones in various audio

和 can be used as a drop in replacement for 跟. The only difference is that in certain connotations, 跟 has a sequentially connotion to it. You can't say 和我读, but you can say 跟我读.

Then on top of that, the benchmarks have very, very few tasks: like 10–20. It's not representative.

RT @iScienceLuvr: I have been thinking what a DeepSeek R1 for medicine would would like... But this paper already kinda did it? plus code…