They skip language processing and go straight to embodiment? Or do you think we can see their 'talking' model based on the JEPA architecture anytime soon?

Maybe this is why Anthropic started hiding their model's reasoning behind summaries, with some dumb excuse for doing so.

It's more likely that a previous version of Gemini 2.5 Pro had visible reasoning traces, so it was easy to distill. Even the current Pro, with some prompting, can be tricked into 'thinking' in a final message.

Professional distillers

Ethan, try this: ask the model to guess the temperature it's using while generating its response. I tried this with Gemini. It's fascinating to see if the model picks up on its own "weirdness" in real time and reflects on it. A good model at this might defend better against hallucinations.

"uncertainty is a signal of weakness" - o3 probably

Sam mentioned in his interviews that long context would no longer be a problem by summer.
I see a few other possible explanations for this announcement:

Super long context (new attention technology)
Per-user fine-tuning (unlikely, but would be cool, finally really personal AI)

Is RAG being used across all chats?

Це щось дуже погано виглядає, ми в основному в обороні, російські знамениті м'ясні штурми. Тобто в нас командування настільки гірше за їхнє? Не розумію, як все так погано.

Not good, trump is filling a new ocean, an Ocean of Lies.

Speedrun: USA, difficulty: Republican, ending: autocracy. No major glitches.

I can't find bullshit in the google translator language list. Can somebody help?

Думаю це вже застаріла практика, придумають щось "краще". Тож до квітня вже будуть перші видні рухи проти чорношкірих.

Cool stuff.
World model includes billions people models and roles they play in social nets.

Надіюсь майбутній серіал буде ближче до першої частини.

Десантна м'ясна броня (ДМБ)

Щось не вірю я у всю цю всесвітню диктатуру лоулайф кіберпанк і піздец.
Буде як зараз: йобана каша - десь добре, десь пекло.

Ті що мені сподобались не кинула(