LessWrong 2.0 Reader

I think we mostly agree, but there's some difference in what we're measuring against.

I agree that it really doesn't appear that the leading labs have any secret sauce which is giving them more than 2x improvement over published algorithms.

I think that Llama 3 family does include a variety of improvements which have come along since "Attention is all you need" by Vaswani et al. 2017. Perhaps I am wrong that these improvements add up to 1000x improvement.

The more interesting question to me is why the big labs seem to have so little 'secret sauce' compared to open source knowledge. My guess is that the researchers in the major labs are timidly (pragmatically?) focusing on looking for improvements only in the search space very close to what's already working. This might be the correct strategy, if you expect that pure scaling will get you to a sufficiently competent research agent to allow you to then very rapidly search a much wider space of possibilities. If you have the choice between digging a ditch by hand, or building a backhoe to dig for you....

Another critical question is whether there are radical improvements which are potentially discoverable by future LLM research agents. I believe that there are. Trying to lay out my arguments for this is a longer discussion.

Some sources which I think give hints about the thinking and focus of big lab researchers:

https://www.youtube.com/watch?v=UTuuTTnjxMQ

https://braininspired.co/podcast/193/ 

Some sources on ideas which go beyond the nearby idea-space of transformers:

https://www.youtube.com/watch?v=YLiXgPhb8cQ 

https://arxiv.org/abs/2408.10205 

If they're being smashed in a literal sense, sure. I think the more likely way things would go is that hierarchies just cease to be a stable equilibrium arrangement. For instance, if the bulk of economic activity shifts (either quickly or slowly) to AIs and those AIs coordinate mostly non-hierarchically amongst themselves.

Hierarchies getting smashed requires someone to smash them, in which case that someone has the mandate of heaven. That's how it worked with John Wentworth's original example of Zhu Di, who overthrew Zhu Yunwen.

I think different types of safety research have pretty different effects on concentration of power risk.

As others have mentioned, if the alternative to human concentration of power is AI takeover, that's hardly an improvement. So I think the main ways in which proliferating AI safety research could be bad are:

1. "Safety" research might be more helpful for letting humans use AIs to concentrate power than they are for preventing AI takeover.
2. Actors who want to build AIs to grab power might also be worried about AI takeover, and if good(-seeming) safety techniques are available, they might be less worried about that and are more likely to go ahead with building those AIs.

There are interesting discussions to be had on the extent to which these issues apply. But it seems clearer that they apply to pretty different extents depending on the type of safety research. For example:

• Work trying to demonstrate risks from AI doesn't seem very worrisome on either 1. or 2. (and in fact, should have the opposite effect of 2. if anything).
• AI control [LW · GW] (as opposed to alignment) seems comparatively unproblematic IMO: it's less of an issue for 1., and while 2. could apply in principle, I expect the default to be that many actors won't be worried enough about scheming to slow down much even if there were no control techniques. (The main exception are worlds in which we get extremely obvious evidence of scheming.)

To be clear, I do agree this is a very important problem, and I thought this post had interesting perspectives on it!

Is the sentence “in reality we should expect combined human-AI entities to reach dangerous capabilities before pure artificial intelligence” really true, and if so how much earlier and does it matter? (I lean towards “not necessarily true in the first place, and if true, probably not by much, and it’s not all that important”)

I guess in my model this is not something that suddenly becomes true at a certain level of capabilities. Instead, I think that the capabilities of human-AI entities become more dangerous in something of a continuous fashion as AI (and the technology for controlling AI) improves.

Yep, Habryka is right. Also, I agree with Wei Dai re: reassuringness. I think literal extinction is <50% likely, but this is cold comfort given the badness of some of the plausible alternatives, and overall I think the probability of something comparably bad happening is >50%.

If you think I missed the point, can you explain in more detail?

Here is my model: Demon king buys shares in “The Demon King will attack the Frozen Fortress”, then sends some small technically-an-attack to the fortress so the market resolves yes, and knowing this will be done is not worth the money lost to the Demon King on the market. No serious-battle plans or military secrets are leaked.

Do you disagree with this, or think it's true but misses the point, in which case what was the point?

To paraphrase the post, AI is a sort of weapon that offers power (political and otherwise) to whoever controls it. The strong tend to rule. Whoever gets new weapons first and most will have power over the rest of us. Those who try to acquire power are more likely to succeed than those who don't. 

So attempts to "control AI" are equivalent to attempts to "acquire weapons".

This seems both mostly true and mostly obvious. 

The only difference from our experience with other weapons is that if no one attempts to control AI, AI will control itself and do as it pleases.

But of course defenders will have AI too, with a time lag vs. those investing more into AI. If AI capabilities grow quickly (a "foom"), the gap between attackers and defenders will be large. And vice-versa, if capabilities grow gradually, the gap will be small and defenders will have the advantage of outnumbering attackers.

In other words, whether this is a problem depends on how far jailbroken AI (used by defenders) trails "tamed" AI (controlled by attackers who build them).

Am I missing something?

Thanks. I already got in touch with Masaharu Mizumoto.

Hm. Does that imply that a pack of dogs hunting a human is a stag hunt game?