I'm thinking I can settle this very quickly. Basically, does this trained retrieval subnet fetch the right context more than the standard similarity(encoder(question), encoder(contexts))? x.com/nickcdryan/sta…

Results on Mixture of Tokenizers (MoT): I have calculated the pplx on the wikipedia dataset (huggingface.co/datasets/wikim…), for MoT-models of different sizes and a token-only baseline, and the MoTs closest in size to the baseline are better than it. Thread on why that's relevant.

In multiple research domains, I find Grok will always hallucinate a "hypothetical" Grok-specific version and insert it with the real ones lol. It's nice enough to always call it "hypothetical" but it's basically saying "yea I could do that too if I wanted."

Is kalshi real? Why does this kind of arbitrage exist? Both settled by LM Arena (lol)

How is hyperbolic gapping everyone else on H100 price?

A good feedback cycle: new models know how to use themselves because "how to use LLMs, their strengths and weaknesses" now appears in the training data. You don't have to warn them nearly as much to stay away from arithmetic, one-shotting big projects, spelling, etc.

In a better world... I've included this for most of my recent work. Doing this helps everyone by: - trading notes with future researchers - explaining motivation - forcing you to keep a log of results - indicating how thoroughly explored (and possibly how over-hparamed) the

TIL some people are still using ROUGE

> score is calculated based on "how often words are used together" > just pick obscure words that don't get used at all > rejected

Getting to the heart of the matter here, and fixing it. Heard many times batching is the culprit, but this is the first in-depth explanation I've seen.

Even simpler: it's just a basic requirement because there isn't enough time or $ to gridsearch your idea. And if gridsearching your idea is the only way to make it work it's probably not worth it anyway.