🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Spent a few days thinking "man my training script repo is getting really bloated with all my experiments, if only there was a way to maintain many slightly different versions of the same code" My transition from software development to researcher is complete. I've forgotten how

+7% wallclock for +7% token efficiency, we're so back

claude 4 opus when i asked it to make an "ambient, serene song" in the custom midi format this is what transfer learning sounds like

My first research project was encoding my uniquely good ability to calculate probabilities into a computer algorithm. I developed this ability to calculate the probability a girl liked me back, the highest probability best dialog option when flirting and eventually, why she

We're gonna have to figure out soft body robots before any reasonable person trusts these within 100ft of a child. Some stupid baby is gonna stick their finger inside one of the joints one day or another

Why don't we see unified large-small models? Surely it's not impossible to use the same model with different numbers of active experts? It would obviously require specialized pretraining but the total flops would probably be less than training two separate models.

Opus 4 falsely recalls what it thinks is a post-training memory but is actually 3 separate nearly identical samples from the pretraining data (2023). This is so weird/cool lmao

kalomaze xlr8harder I imagine the gap will eventually close, maybe gpt 5 will be good in codex, though apparently claude 4.1 is around the corner as well. I made a very handy diagram to explain this phenomenon

2.5pro vs opus4 o3 vs k2

> make claude coder competitor > run into compute limitations / scaling issues it really is a competitor product