[Etheryte]

If anything, I would say that that's a very optimistic take. The hype train is strong, but that's largely what it is once you look at the details. What we have right now is impressive, but no one has shown anything close to a possible path from where we are right now to AGI. The things we can do right now are fancy, but they're fancy in the same way good autocomplete is fancy. To me, it feels like a local maxima, but it's very unclear whether the specific set of approaches we're exploring right now can lead to something more.

[ctoth]

> What we have right now is impressive, but no one has shown anything close to a possible path from where we are right now to AGI[0].

[0]: From GPT-4 to AGI: Counting the OOMs https://situational-awareness.ai/from-gpt-4-to-agi/

[Etheryte]

I'm not convinced, and neither is Sam Altman himself [0]. Also, if that projection holds, and that's a big if, the purported breakthrough would cost 10^6 times as much as GPT-4 took to train. That's over 100 million dollars [1] times a million. That adds up to over 100 trillion dollars, in the ballpark of four times the GDP of the whole of United States.

[0] https://www.wired.com/story/openai-ceo-sam-altman-the-age-of...

[1] https://en.wikipedia.org/wiki/GPT-4#Training

[Workaccount2]

The thing is that it looks like, or perhaps I should say it's "understood" at this point, that transformer's abilities scale pretty much linearly with compute (there is also some evidence they scale exponentially with parameter count, but just some evidence).

Right now there is insane amounts of money being thrown at AI because progress is matching projections. There doesn't seem to be a leveling off or diminishing returns taking place. And that's just compute, we could probably freeze compute and still make insane progress just because optimizations have so much momentum right now too.

[hackinthebochs]

How do you distinguish the path to fancier autocomplete from the path towards AGI? Why think we're on the former rather than the latter?