LessWrong 2.0 Reader

To the extent that complexity makes succesfully red-teaming our approaches too difficult, then we could try to mostly depend on simple approaches and only use more complex approaches only if we are confident that at least they won't can't make the situation worse.

My general view is that complex methods are only productive if you have a good feedback loop (either empirical feedback or a very robust theoretical feedback (e.g. proofs)). In the case of control, so long as we do a good job with red teaming^[1], we can build a good feedback loop such that complex methods can be productive. I think a similar process is often used in computer security and I think computer security often involves complex measures^[2].

As an aside, this also makes me relatively pessimistic about developing complex and unprincipled alignment (non-control) approaches for hard-to-experiment-on threat models like scheming (aka deceptive alignment). It will be difficult to make scheming model organisms which realistically reproduce the internal structure and capability level of models we're worried about and thus it will be hard to effectively iterate on complex and unprincipled measures.

An alternative approach for empirics is to focus on highly principled measures where we have a story for why they should generalize to scheming that we can test directly (e.g. we'll understand how the model works which we can test by seeing if we can understand models in general).

This makes me relatively unexcited about very generic approaches to weak-to-strong generalization working for scheming as such approaches must be unprincipled (after all, they are very generic, so they can't be operating on a specific mechanistic story) and thus we can't test the story for why they would generalize.

I think that these examples are quite different. People who were voting for a president who ultimately turned out to be corrupt most likely predicted that the most likely scenario won't be totally horrible for them personally. And most of the times they are right, corrupt presidents are bad for countries, but you could live happily under their rule. People would at least expect to retain whatever valuable stuff they already have.

With AI the story is different. If a person thinks that the most likely scenario is the scenario where everyone dies, then they would not only expect to not win anything, but to lose everything they have. And media is actively talking about it, so it's not like they are not at least aware of the possibility. Even if they don't see them as personally responsible for everything, they sure don't want to feel the consequence of everyone dying.

This is not the case of someone doing something that they know is wrong but benefits them. If they truly believe that we are all going to die, then it's doing something they know is wrong and actively hurts them. And allegedly hurts them much more than any scenario that has happened so far (since we are all still alive).

So I believe that most people working on it, actually believe that we are going to be just fine. Whether this belief is warranted is another question.

I think the justification of this articles goes something like this: "here is the vision of future where we successfully align AI. It is utopian enough that it warrants pursuing it". Risks from creating AI just aren't the topic. They deal with them elsewhere. These essays were specifically focused on the "positive vision" part of the alignment. So I think you are critiquing the articles for lacking something they were never intended to have in the first place.

OpenAI seems to have made some basic commitments here. The word commit is mentioned 29 times. Other companies did it as well here, and here. Here Anthropic makes promises for optimistic, intermediate and pessimistic scenarios of AI development.

I don't know about similarity... but I was just making a point that QI doesn't require it.

If the initial paraphraser doesn't change between '!' and '.', then learning a canonical-paraphraser from that initial paraphraser won't do it either.

...

So the canonical-paraphrasing idea helps somewhat, but fundamentally, the better the paraphraser (the more closely it captures the human notion of semantic equivalence), the better the anti-steganography guarantee we can get.

What I meant is that we train the canonical paraphraser based on the "human notion of semantic equivalence," NOT the "initial paraphraser."

One way of doing this is to train another model (call it the "adversarial rephraser") at the same time that you train the canonical rephraser. The adversarial rephraser takes a single sentence s as input, and tries to pick another sentence s' such that either 1) s has the same meaning as s' but CR(s) != CR(s'), or 2) s has a different meaning from s' but CR(s) = CR(s'). Semantic equivalence is determined by some oracle - probably an LLM like GPT-4 is sufficient, but you could also fine-tune a model on human data specifically for this task. We then train the canonical rephraser against the adversarial examples.

The adversarial rephraser should be able to discover strategies like changing '.' to '!', and the canonical rephraser will learn to account for those strategies.

You made lots of points, so I wrote a comment for each... probably this was too many replies? I didn't know what else to do that didn't feel like avoiding your points. I hereby state that I do not expect of you to respond to all five of my comments!

Why do you hate earworms? To me, they are mildly pleasant. The only moments when I wish I didn’t have an earworm happening at that moment is when I’m trying to remember another tune and the earworm for musicianship purposes and the earworm prevents me from being able to do that.

Another example would be Anthropic creating a dedicated team [LW · GW] for stress testing their alignment proposals. And as far as I can see, this team is lead by someone who has been actively engaged with the topic of AI safety on LessWrong, someone who you sort of praised [LW · GW] a few days ago.

I don't quite know what the point here is. This is marginally good stuff, it doesn't seem sufficient to me or close to it, I expect us all to probably die, and again: from the CEO and cofounders there has been no serious engagement or justification for the plan for them to personally make billions of dollars building potential omnicidal machines.

I think the point about them not engaging with critics is also a bit too harsh. Here [LW · GW] is DeepMind's alignment team response to concerns raised by Yudkowski. I am not saying that their response is flawless or even correct, but it is a response nonetheless. They are engaging with this work. DeepMind's alignment team also seemed to engage with concerns raised by critics in their (relatively) recent work [LW · GW].

I don't disagree that it is good of the DeepMind alignment team to engage with arguments on LessWrong. I don't know that a few researchers at an org engaging with these arguments is meeting the basic standard here. The first post explicitly says it doesn't represent the leadership, and my sense is that the leadership have avoided talking about the subject, and that the people involved do not have the political power to push for the leadership to engage in open debate.

That said I do concede the point that DeepMind has generally been more cautious than OpenAI and Anthropic, and never created the race to building potential omnicidal machines (in that they were first – it was OpenAI and Anthropic who added major competitors).

I don't think that money alone would've convinced CEOs of big companies to run this enterprise. Altman and Amodei, they both have families. If they don't care about their own families, then they at least care about themselves. After all, we are talking about scenarios where these guys would die the same deaths as the rest of us. No amounts of hoarded money would save them. They would have little motivation to do any of this if they believed that they would die as the result of their own actions. And that's not mentioning all of the other researchers working at their labs. Just Anthropic and OpenAI together have almost 2000 employees. Do they all not care about their and their families' well-being?

I'm not sure how quite to explain that I think a mass of people can do something that they each know on some level is the wrong thing and will hurt them later, but I believe it is common. I think partly it is a mistake to think of a mass of people as having the sum of the agency of all the people involved, or even the maximum.

I think it is easier than you do to simply not think about far away dangers that one can say one is not really responsible for. Does every trader involve in the '08 financial crisis take personal responsibility for it? Does every voter for a politician who turns out to ultimately be corrupt take personal responsibility for it? Do all the tens of thousands of people involved in various genocides take personal responsibility for stopping it as soon as they see it coming? I think it is very easy for people to erect a cartesian boundary between themselves and the levers of power. People are often aware that they are doing the wrong thing. I broke my diet two days ago and regret it, and on some level I knew I'd end up regretting it. And the was a situation I had complete agency over. The more indirectness, the more things are in far-mode, the less people take action on it or feel like they can do anything based on it today.

I agree it is not money alone. These people get to work alongside some of the most innovative and competent people of our age, connect with extremely prestigious journalists and institutions, be invited to halls of power in senior parts of government, and build systems mankind has never seen. All further incentive to find a good rationalization (rather than to stay home and not do that).