Google AI and DeepMind News and Discussions

[weatheriscool]

[weatheriscool]

[weatheriscool]

[weatheriscool]

[wjfox]

https://x.com/RoundtableSpace/status/2077419165636272293

[firestar464]

Bloomberg- Google Gemini Launch Delayed as Tech Falls Short of Internal Goals

https://removepaywalls.com/https://www. ... rnal-goals

Lol. Interesting thing is that Google's spokesperson made reference to an upgraded Flash model. Wouldn't be surprised if we got that before Pro.

[weatheriscool]

[firestar464]

Called it. Though it's mad funny that it's slightly worse than 3.5 Flash on the AA Intelligence Index.

Oh well, at least it's a better deal in terms of pricing.

[spryfusion]

[firestar464]

[Cyber_Rebel]

This along with the reported news above regarding Hassabis, makes me believe Google isn't in the AI "race" anymore, as in the AGI aspect at least. I enjoyed Gemini as a model, even if it did fall behind. They might've leaned in more strongly on their world modeling alongside omni modal areas.



Source: https://newsletter.semianalysis.com/p/g ... is-cooking

[firestar464]

Eh, Meta came back from far, far worse, though not without hiring lots of new talent and completely restructuring things. Perhaps Google is moving to do that. Who knows.

Main issue is their culture.

[Yuli Ban]

I've been saying for a couple years now that Hassabis never felt like he really believed in LLMs as the path to AGI and probably thought that pivoting so hard to them was a mistake. The literal only reason why Gemini exists is because of ChatGPT.
Remember Bard? When Bard first released, people were literally shocked at how it was basically GPT-2.5-tier when OpenAI already had released GPT-4, which made no sense because up to that point, DeepMind was seen as THE premier AI lab in the world, the one that would unleash AGI, and now they looked like they blew such an obvious lead pretty much permanently. Gemini didn't catch up to ChatGPT until ChatGPT itself seemed to rest on its laurels when it took 2 years to release GPT-4.5 and 5, and both were totally disappointing, while Gemini and Claude caught up and then came DeepSeek.

But that entire time, I only recall Hassabis changing his mind about how soon AGI was coming once, early on, and he later backtracked. No amount of LLM or reasoning model progress ever seemed to convince him these models were going to get us to AGI alone, since he always seemed to find them to be only one part of a much larger model architecture.
It's possible Google didn't want DeepMind to divert resources to pure AGI research when they'd already invested so much into Gemini and Nano Banana. It's possible Hassabis isn't AS against Gemini as I'm saying but there was some alternate skew he wanted to pursue that the higher ups just didn't agree with. Maybe he thinks we're closer to AGI than it even seems but found the myopic focus on scaling LLMs to be obscuring something obvious. Or maybe there's another reason entirely.

I'm just not surprised; to me it just feels like this was a long time coming, like when Microsoft made Google "dance" back in 2022, it was to the expense of everything Hassabis thought was reasonable R&D in lieu of one useful path that ballooned out of control. I dunno! Unless he tells us, I'm just guessing.