Comment by mullingitover
11 hours ago
> If this is truly AGI (subject to one's definition of AGI still), then this is a very boring release of an AGI model.
Hot take: These models are never going to be 'AGI'. We're just going from a GPT4 ball that's 90% round to a GPT5 that's 99% round to a GPT6 that's 99.9% etc etc etc
I think that the harnesses and context management is really where the rubber meets the road, and the real gains are happening there.
I don't remember where I heard this, but one of my favorite criticisms of the current AI situation is that it's wrong simply because of the size and energy required compared to the human brain. The idea is that there's still some element missing thats fundamental, and that the way we train them now is part of the solution, but not all of it. I think finding the extra missing element is going to take an entirely different approach that will also solve the sizing and resource issue. The kickers is that if they do achieve (and solve) AGI in this way all the giant data centers would be mostly useless.
Yes, the very explicit plan of both OpenAI and Anthropic is to use the not particularly efficient LLMs to automate their own AI engineering. That seems to be going well - on coding front and model tuning front so far. They have more planned.
And then use those to find fundamentally better new architectures for AI - that perhaps are as efficient as the human brain.
It might not work, but I didn't think it'd solve maths problems... So it might work. And if it happens, they'd use the data centres to run millions of instances of it.
It's scary, TBH.
I recall them saying they use models to write CUDA kernels and whatnot. Makes sense, and unsurprising that models are good at writing code.
But I think calling this “automating AI research” is misleading. I’m not sure there’s evidence yet that they do creative research work. Even in mathematics, but they are finding counter-examples by intelligent brute-forcing. Not to downplay the results, as they are incredible, but this is one very specific kind of proof and not the most creative type, which arguably requires generalisation.
Quite the gamble.
> but I didn't think it'd solve maths problems
Finding counterexamples is low-hanging fruit, the automation of which isn't shocking.
3 replies →
I mean, the plan is to use these models to find and solve those gaps. That's kind of the whole pitch of these companies: they spend a TON of money upfront setting up this infrastructure, but each iteration yields a system capable of making the next iteration even better.
>The kickers is that if they do achieve (and solve) AGI in this way all the giant data centers would be mostly useless.
Perhaps. But only at that point, not leading up to that point.
It's kind of like setting up scaffolding to build something. You spend all of that time and money to build something just to tear it down in the end. But the point is that it's simply a cost to be able to build the actual thing you're building.
If these companies are able to achieve the results they're looking for, none of the investors involved are going to care that the datacenters and infrastructure they spent so much money.
If we manage to get to AGI and it looks, works and behaves like a human brain... I mean, cool, but that's a very useless AGI compared to the incredible stuff we have access to today.
The HN crowd I'm sure will still be unhappy calling it AGI because "it's not AGI unless its speech comes from the cerebral cortex region of the brain, otherwise it's just sparkling emoji" or something.
I think the idea is that you wouldn’t need humans to do anything anymore, right? As impressive as it is, it’s still ultimately directed by human planning and coordination. Assuming they are aligned, you could have a collection of AGI that you let loose and they tirelessly solve all of humanity’s problems, do all of our work, and progress science and our understanding of the universe.
Those are all things that humanity is doing everyday. What we have is amazing, but it’s not that.
1 reply →
you're on one of the most pro AI spaces on the whole internet and yet you're still crying about "the hn crowd", what a bizarre distorted perspective
1 reply →
>I think that the harnesses and context management is really where the rubber meets the road, and the real gains are happening there.
True. So we did hit a wall with pure scaling alone, though no lab would admit it. It's crazy to see how harness switchout results in such vast delta in benchmark scores.
We have not "hit a wall" by any stretch yet. I don't understand how someone can even hold this viewpoint? It's mind boggling.
Harnesses magnify and make the intelligence actionable, but we have not reached limits on raw intelligence yet, not even close.
Agreed. Trivially observable by using a frontier model from today and one from 6 months ago with the same harness.
I don't think so.
One could use gpt-4 or gpt-5 with today's harnesses and we'd see how well that goes.
I think models using these harnesses were also RLHF'd hard on responding to looping instructions and following through on goals. Older models were tuned for basic chat responses.
If someone could tune models of that size to have comparable effectiveness at much much lower costs, they would have done so by now.
"The harness improvements are the real sauce" is like a sincere "It's gotta be the shoes" take about Micheal Jordan.
(For the younger: that line was from a series of Nike ads where his skills were being explained)
We call that a sigmoid.