ElPandoLeTrash

I am mighty trash panda! Aj Lajk Mashine Learning, with little bit of Trash talking and some hummour…

[Tuomas Artman @artman]

Today is a hard day. I shared this note with the @linear team today: We’ve made the difficult decision to increase our workforce. This is not a cost-cutting exercise or a reflection of anyone’s performance. We’re simply reimagining every role for the agentic AI era. We’re hiring. We’re sorry about that.

[ElPandoLeTrash @binary_racoon]

@litcapital There are different explanations… after next Gemini model releases watch the coding performance… also I believe that last Claude had pretty big jumps in image right?

[ElPandoLeTrash @binary_racoon]

@0xSero I tested it as predictor for my paper - interesting is that it was completely same for thinking and non-thinking. I mean same final accuracy.

[BURKOV @burkov]

For those living under a rock: LLMs stopped becoming smarter around summer 2025.

Everything impressive you see since then is about finetuning them for specific tasks (mainly coding and software-tool-based task solving) and building tooling around them (such as agentic coding systems).

[ElPandoLeTrash @binary_racoon]

@burkov Exactly!

[ElPandoLeTrash @binary_racoon]

@a_karvonen Just realised that 4.5 had 128k context window so this might be reason as well - you are probably right…

[ElPandoLeTrash @binary_racoon]

@a_karvonen Wasn’t that <128k?

[ElPandoLeTrash @binary_racoon]

@tjbecker That is why there is no fpr in mythos preview paper…. Or is it?

[ElPandoLeTrash @binary_racoon]

@eliebakouch Isn’t this expected? I mean you don’t need long cot to refine and plan the response - bigger model will be definitely more effective - less noise - information goes trough more params 🤷

[0xSero @0xSero]

https://t.co/txDnDa0Flf

[ElPandoLeTrash @binary_racoon]

@teortaxesTex How will closing source help you when you are not even in top 3?

[ElPandoLeTrash @binary_racoon]

@scaling01 Mb proof that they did not had rl environments for this kind of task 🤷 not sure…

[ElPandoLeTrash @binary_racoon]

@mervenoyann They move to us because they will get there more money. Nothing else, nothing more.

[ElPandoLeTrash @binary_racoon]

@a7alec @johnpalmer You definitely don’t have MacMini….

[ElPandoLeTrash @binary_racoon]

@scaling01 Given how big Google internal repo is I guess they will have plenty of static data as well - not counting data brokers. I guess that user interactions are good for rlhf, the alpha is their rl-vr env - honestly I think here it is again about how much compute you allocate

[ElPandoLeTrash @binary_racoon]

@xlr8harder It’s like this episode of scp….. 😆

[ElPandoLeTrash @binary_racoon]

@paulhshort @xlr8harder Exactly they even admitted it smwh

[ElPandoLeTrash @binary_racoon]

@menhguin @pli_cachete Where you can get cheap compute? I went for phd recently and most of it goes from my own pocket…

[ElPandoLeTrash @binary_racoon]

@scaling01 @sdmat123 It is in cursor now….. you can try it… hard to say if better or how much…

[ElPandoLeTrash @binary_racoon]

@johannes_hage 1+ 👋