Yeah. Google really needs to make a jump with Gemini 4. But what strikes me more: "impressed by Opus 4.8" was less than 2 months ago ...

I'll say this, the blistering advancement in LLM and the incredible rate of iteration on these models is unlike anything I've experienced in my professional career. Nothing prior to this in my field has advanced so quickly with such great impact.

i'm genuinely so confused how this happened. they've had a lot of money and a lot of time, i feel like it has to be an organizational issue at this point

Two possibilities: a. Google initially saw LLM as a competitor to their main product, internet search. That could create internal discord among teams making them slower out of the gate b. Google sees what China sees: very fast/cheap/good-enough beats huge and expensive for most paying use-cases

.. both a. and b. can be true rn

I'm wondering if Google sees what China seems to see - that fast embedded LLM will be more important in the end than the most massive, expensive models. Flash 3.6 is incredibly fast, and has a combo of speed/quality/cost that might meet industrial and fortune-500 needs more than Fable/Opus.

Their Gemma team does seem to fare better than the Gemini one. They and Qwen are the frontier for smaller models.

what they may be aiming for is what Gretsky called "skating to where the puck is going to be" all your robots are belong to us ;)

Is that a fair comparison? I thought Flash is closer to Sonnet.

Flash is cheaper, and way faster than sonnet. But flash 3.6 is the best model google has available.

It's Flash. It is not intended as a frontier model. It's probably the best "flash" type model out there and arguably what the market needs and wants. But your point is true - Goog hasn't a real frontier model currently

It's really a shame how far behind they've fallen. Crazy how we're still on version 3.1 for Pro. But even though it's far less capable, I'll say I still enjoy having Gemini in my suite. It frequently adds ideas and perspectives that Claude and ChatGPT don't.

I think to be fair they’re optimising for something different. But I also think they’re doing that because they’re so far behind on coding and reasoning.

Does it not sit on the Pareto frontier anywhere?

It's hard to believe they are so incompetent.