I want the Gemini 4 leaks to be real so fucking bad but that benchmark table is almost offensively perfect
I've been around AI twitter long enough to know when I'm being emotionally manipulated by a spreadsheet
@firstadopter $235B left for buybacks is such an absurd sentence. NVIDIA has gone from selling chips to having nation-state quantities of money lying around.
this is gonna happen everywhere. so many systems are stable only because humans react slowly. put millions of agents in the loop reacting instantly and the assumptions underneath the system just fucking disappear.
Fascinating.
Chief Economist at Apollo: agents could cause a bank run by sweeping household cash into accounts paying 3-5% instead of the 0.1% national average, causing banks to lose a large share of their cheap deposits.
@emollick this gap feels very temporary to me. the interesting question is what happens when an open model crosses that agentic threshold and you can run it 24/7 on hardware you own with zero usage limits. that changes a lot more than benchmarks.
I’m extremely pro-AI, but I have a hard time imagining a post-AGI world where frontier labs remain normal private companies.
If one of them actually gets to AGI first, the power dynamic with governments becomes fucking bizarre. You’re no longer talking about a company with valuable IP. You’re talking about a private institution controlling something potentially more strategically important than nukes, intelligence agencies, and entire industrial bases.
At that point I’d expect some form of state control, nationalization, forced partnership, or whatever new legal structure gets invented for it.
The weird part is the labs are racing to build something that, if they succeed hard enough, could make it politically impossible for them to remain what they are today.
@quxiaoyin the funniest part is declaring the model race over while every lab is still burning billions trying to make the next model smarter. product matters enormously, but intelligence is still the scarce thing everyone is fighting over.
@VadimStrizheus “just use claude and a folder” is exactly the kind of sentence people say right before someone builds a much better abstraction and suddenly it feels obvious
@boneGPT the loss of shared culture is the part that actually bothers me. everyone having a perfectly optimized private feed sounds less like entertainment and more like 8 billion people slowly hallucinating alone together
I think the new skill is mostly knowing what the model is actually good at and structuring the work around that. being a great programmer doesn’t automatically give you that intuition. you have to spend a stupid amount of time using these things before their failure modes start feeling obvious.
@jaredLunde the real bottleneck was never intelligence. it was getting your friends to give a shit about building anything besides another note-taking app
@johnennis I agree they’re different, but I never understood why people treat that as some profound limitation. submarines don’t swim like fish either. who gives a shit if the mechanism is different if the capability is there?
@scottstts i disagree on the 750 tps part. give me a genuinely smart model at that speed and I’ll change how I use computers entirely. but yeah, none of that matters if I hit a fucking usage wall the moment I actually start depending on it.