What structure do transformers learn from next-token prediction? Using Computational Mechanics, the authors predict that transformers represent the "belief state geometry" of the data-generating process: how an observer updates beliefs over its hidden states. They train a transformer and find this predicted fractal structure linearly embedded in its residual stream.
One of the most pleasing things about probability and expected utility theory is that there are many coherence arguments that suggest that these are the “correct” ways to reason. If you deviate from what the theory prescribes, then you must be executing a dominated strategy. There must be some other strategy that never does any worse than your strategy, but does strictly better than your strategy with certainty in at least one situation. There’s a good explanation of these arguments here.
We shouldn’t expect mere humans to be able to notice any failures of coherence in a superintelligent agent, since if we could notice these failures, so could the agent. So we should expect that powerful agents appear coherent to us. (Note that it is possible that the...
Suppose we have an AI CEO, ruthlessly devoted to maximising shareholder value. Fortunately, we haven't been completely stupid, and have given it a rich definition of shareholder value, encompassing the success, flourishing, safety, preference satisfaction, and moral values of the shareholders, not just the stock price. So it won't destroy the company - or the world - to build a giant stock-ticker with a swiftly increasing number on it.
This rich definition is contained in a sophisticated LLM, able to analyse the current situation and answer any question about the objective. The LLM is an integral part of the AI's system and motivation, not something it can simply set aside or hack directly.
Unfortunately, as discussed, there are going to be zones in the space of possible situations where...
Many thanks to Abram Demski for materially and intellectually supporting this research. Thanks also to Brian Berns, Roman Malov, Gurkenglas, and Lorxus for valuable conversations.
Despite the central place of corrigibility to theoretical AI alignment, it seems like no one can really agree about what is. Much of this confusion is due to the fact that almost every definition of corrigibility proposed thus far is deficient in one way or another. In particular, the most prevalent existing accounts of corrigibility are over-epistemicized (modeling corrigibility as arising purely from uncertainty about the AI's goal, which gives rise to issues like the problem of fully updated deference) or evaluatively-thickened (making the AI's normative priorities instrumental to its corrigibility as a result of forcing corrigibility to be a single unified goal.)...
(If you're already familiar with all basics and don't want any preamble, skip ahead to Section B for technical difficulties of alignment proper.)
I have several times failed to write up a well-organized list of reasons why AGI will kill you. People come in with different ideas about why AGI would be survivable, and want to hear different obviously key points addressed first. Some fraction of those people are loudly upset with me if the obviously most important points aren't addressed immediately, and I address different points first instead.
Having failed to solve this problem in any good way, I now give up and solve it poorly with a poorly organized list of individual rants. I'm not particularly happy with this list; the alternative was publishing nothing, and publishing this seems marginally...
One of the great advantages of working with Gödel-Löb provability logic is that it's possible to implement an evaluator which efficiently checks whether a sentence in the language of GL is true. Mihaly and Marcello used this to write a program that checks whether two modal agents cooperate or defect against each other. Today, Nate and I extended this with an implementation of modal UDT, which allows us to check what UDT does on different decision problems---see Program.hs in the Github repository. No guarantees for correctness, since this was written rather quickly; if anybody is able to take the time to check the code, that would be very much appreciated!
The implementation of UDT is rather pleasing, I think. Here's the informal definition of modal UDT, using PA
I think the title of the post is straightforwardly false and has led many people astray, but the post makes a fine locally valid point if you ignore the title.
See for example:
... (read more)