Creator Teortaxes argues Google's AI compute advantage over competitors has shrunk from 100x to just 10x

"GDM always comes back" was a reasonable prediction when they outclassed everyone by orders of magnitude. There's been a good post recently tracking how their compute advantage was falling: 100x, 50x, 10x… No, they legit can just lose. It'd still take a lot of time though.

@xeophon People forget that at one point Gemini was so dominant that it was only dethroned from first place by the next release of Gemini.

@xeophon This time may be different At the very least I don't think Gemini 3.5 Pro will be that new model

@gazorp5 @xeophon November 2024.

@AndrewCurran_ @xeophon i don't remember a single time this has ever happened tbh

@xeophon Google gave the transformer, people forget that easily ig

@xeophon I mean that your goat sucks and not worth the 1B 😭

@xeophon this is how i set my watch

@AndrewCurran_ @xeophon In evals, maybe. For actual use cases it’s never been the best imho

@xeophon did you know they trained their last model with their state of the art TPU???

@wandering_mush @xeophon 他们准备通过抽烟来达成agi

@xeophon Can you give an example of when the second part happened? No, it wasn't gemini 3, anybody who had used it honestly knew it was thoroughly benchmaxxed and they got mogged by Opus 4.5 like two weeks after. Only case when Google was ahead was Nano Banana.

@xeophon ngl I was a google bull for the longest time, before it was even cool, even in the bard days because I knew (tpus, etc.) but I think they're cooked. I don't know if you can run a frontier lab like that

@AndrewCurran_ @xeophon Never under estimate google. This is the first company already achieved RSI.

@xeophon hm..after Fable I'm not that sure about the 'the others are done for' part.

@xeophon never understood the “google is done for” takes. and there were SOOO many of them a few years back. people genuinely believed google will go to 0

@xeophon I've continued to enjoy gemini, as it uniquely follows the frame you ask it to, rather than does this 'I have to push back' midwit opinion slop anthropic and openAI do.

@xeophon and that’s the proof that humans are not rational.

@AndrewCurran_ @gazorp5 @xeophon Except that was only for arena maxxing. No one actually used these models for anything over 4o or o1 at that point

Gemini 2.5 Pro was probably Gemini’s peak relative to competition, and it was quite quickly beaten.

@xeophon Gemini lost to Gemini.

That's what happens when you move faster than people can switch.