Comment by bglazer
4 days ago
Yudkowsky seems to believe in fast take off, so much so that he suggested bombing data centers. To more directly address your point, I think it’s almost certain that increasing intelligence has diminishing returns and the recursive self improvement loop will be slow. The reason for this is that collecting data is absolutely necessary and many natural processes are both slow and chaotic, meaning that learning from observation and manipulation of them will take years at least. Also lots of resources.
Regarding LLM’s I think METR is a decent metric. However you have to consider the cost of achieving each additional hour or day of task horizon. I’m open to correction here, but I would bet that the cost curves are more exponential than the improvement curves. That would be fundamentally unsustainable and point to a limitation of LLM training/architecture for reasoning and world modeling.
Basically I think the focus on recursive self improvement is not really important in the real world. The actual question is how long and how expensive the learning process is. I think the answer is that it will be long and expensive, just like our current world. No doubt having many more intelligent agents will help speed up parts of the loop but there are physical constraints you can’t get past no matter how smart you are.
How do you reconcile e.g. AlphaGo with the idea that data is a bottleneck?
At some point learning can occur with "self-play", and I believe this is already happening with LLMs to some extent. Then you're not limited by imitating human-made data.
If learning something like software development or mathematical proofs, it is easier to verify whether a solution is correct than to come up with the solution in the first place, many domains are like this. Anything like that is amenable to learning on synthetic data or self-play like AlphaGo did.
I can understand that people who think of LLMs as human-imitation machines, limited to training on human-made data, would think they'd be capped at human-level intelligence. However I don't think that's the case, and we have at least one example of superhuman AI in one domain (Go) showing this.
Regarding cost, I'd have to look into it, but I'm under the impression costs have been up and down over time as models have grown but there have also been efficiency improvements.
I think I'd hazard a guess that end-user costs have not grown exponentially like time horizon capabilities, even though investment in training probably has. Though that's tricky to reason about because training costs are amortised and it's not obvious whether end user costs are at a loss or what profit margin for any given model.
On the fast-slow takeoff - Yud does seem to beleive in a fast takeoff yes, but it's also one of the the oldest disagreements in rationality circles, on which he disagreed with his main co-blogger on the orignal rationalist blog, Overcoming Bias, some discussion of this and more recent disagreements here [1].
[1] https://www.astralcodexten.com/p/yudkowsky-contra-christiano...
AlphaGo showed that RL+search+self play works really well if you have an easy to verify reward and millions of iterations. Math partially falls into this category via automated proof checkers like Lean. So, that’s where I would put the highest likelihood of things getting weird really quickly. It’s worth noting that this hasn’t happened yet, and I’m not sure why. It seems like this recipe should already be yielding results in terms of new mathematics, but it isn’t yet.
That said, nearly every other task in the world is not easily verified, including things we really care about. How do you know if an AI is superhuman at designing fusion reactors? The most important step there is building a fusion reactor.
I think a better reference point than AlphaGo is AlphaFold. Deepmind found some really clever algorithmic improvements, but they didn’t know whether they actually worked until the CASP competition. CASP evaluated their model on new Xray crystal structures of proteins. Needless to say getting Xray protein structures is a difficult and complex process. Also, they trained AlphaFold on thousands of existing structures that were accumulated over decades and required millenia of graduate-student-hours hours to find. It’s worth noting that we have very good theories for all the basic physics underlying protein folding but none of the physics based methods work. We had to rely on painstakingly collected data to learn the emergent phenomena that govern folding. I suspect that this will be the case for many other tasks.
> How do you reconcile e.g. AlphaGo with the idea that data is a bottleneck?
Go is entirely unlike reality in that the rules are fully known and it can be perfectly simulated by a computer. AlphaGo worked because it could run millions of tests in a short time frame, because it is all simulated. It doesn't seem to answer the question of how an AI improves its general intelligence without real-world interaction and data gathering at all. If anything it points to the importance of doing many experiments and gathering data - and this becomes a bottleneck when you can't simply make the experiment run faster, because the experiment is limited by physics.