[nathan11]

It seems reasonable to say that AGI will take a ton of resources. You'll need investors for power, GPUs, researchers, data, and the list goes on. It's a lot easier to get there with viable commercial products than handouts.

I'd be willing to bet that between Sam's approach and the theorized approach of the OpenAI board we're discussing, Sam's approach has a higher chance of success.

[dkjaudyeqooe]

Since AGI isn't a thing, no one knows what it will look like or if it will even exist.

The biggest breakthroughs in science do not come from those with the most money. It's all ideas.

[ben_w]

OTOH humans are a non-artificial GI, and we can use ourselves as an anchor for estimates of what we'd need for an artificial equivalent.

About 1000x the complexity of GPT-3 and much slower would be the best guess right now.

[dkjaudyeqooe]

It's looking at humans, how they're trained and their wetware makes me believe that AGI, as most people understand it, ie a super human like intelligence, will never exist. There will be powerful AI but it won't be human like in the way people think about it now.

[ben_w]

That definition ought to be reserved for ASI (S meaning super) not AGI (G meaning general).

That said I agree "human like" is unlikely, although LLMs and diffusion models are much closer than I was expecting.

[dkjaudyeqooe]

That's because their training source is human output. Human In Human Out (HIHO).

[ben_w]

Even so, was expecting more dissimilarities or even just types of inappropriateness that are very human — humans are a broad bunch, no reason the LLMs wouldn't just default to snarky and lazy, like the example from the OpenAI Dev Day of someone who tried to fine tune on their slack messages, asked it to write something, and it said "Sure, I'll do it in the morning".

Despite people calling them stochastic parrots and autocomplete on steroids, ChatGPT is behaving like it is trying to answer rather than merely trying to continue the text the user enters. I find this surprising.

[a_wild_dandan]

Precisely. Breakthroughs are often cleverer than brute force, “throw more compute/tokens at it” approaches. Turning some crucial algorithm from O(n) to O(log(n)) could be an unlock worth trillions of compute time dollars.