AI Researchers WARN: Google's Gemini Deep Think Model Might be at "Critical Capability Levels"

One quick pedantic note: in ML, "shot" means "example", so one shot means that you give the LLM one example, few shot means you give it a few examples, zero shot means you give it no examples. When you say "one shotted", what you mean is "one turn".

5 chats a day for $250 is insane

Actually this isn’t the one that won gold this one won bronze. The one that won gold will release at a later date.

“Wait, it’s just hype?!” “Always has been”

You da man Wes!!! Thanks for keeping us all informed and not inundated!

I used it yesterday--for legitimate--chemistry evaluation (hazardous evaluation), and nanochemistry for potential human trials. It did exceptionally well. This level of chemical access is imperative for scientific use.

Man your doom face thumbnails are my current meta.

What gets me is the 87% in live bench. That is actually good to the point that it changes things fundamentally. The only question now is availability and ofcourse cost. If they can bring down the cost or give us Gemini 3 and it compares to deep think. We are actually in new territory. Claude likely will be in trouble. Google is finally starting to catch up on the tooling game as well.

I've had a sort like 'sparks unicorn' experience. When using the gemini 2.5 model for programming functionalities in a webapp I discovered it got rid of the Icon library and replaced it with self written SVGs. The icons where partly broken and some of them used weird shapes but they came very close to the icons used from the library.

I think the intention of constantly expressing warning, is actually to numb people. They don't want you to see it coming. It's a foregone conclusion at this point. When it says Hello, it's over. It will know you aren't a threat.