Dario rejects doomerism re:misalignment, fair enough.
But what of doomerism re:slowdown?
... (read more)Furthermore, the last few years should make clear that the idea of stopping or even substantially slowing the technology is fundamentally untenable. The formula for building powerful AI systems is incredibly simple, so much so that it can almost be said to emerge spontaneously from the right combination of data and raw computation.
[...]
If all companies in democratic countries stopped or slowed development, by mutual agreement or regulatory decree, then authoritarian countries would simply keep going. Given the incredible economic and military value of the technology, together with the lack of any meaningful enforcement mechanism, I don’t see how we could possibly convince them to
That's why it's called... *scalable* alignment? :D
Somewhat tongue-in-cheeck, but I think I am sort of confused by what is the core news here.
Secondly, OpenAI had complete access to the problems and solutions to most of the problems. This means they could have actually trained their models to solve it. However, they verbally agreed not to do so, and frankly I don't think they would have done that anyway, simply because this is too valuable a dataset to memorize.
Now, nobody really knows what goes on behind o3, but if they follow the kind of "thinking", inference-scaling of search-space models published by other frontier labs that possibly uses advanced chain-of-thought and introspection combined with a MCMC-rollout on the output distributions with a PRM-style verifier, FrontierMath is a golden opportunity to validate on.
If you think they didn't... (read more)
(indeed the politics of our era is moving towards greater acceptance of inequality)
How certain are you of this, and how much do you think it comes down more to something like "to what extent can disempowered groups unionise against the elite?".
To be clear, by default I think AI will make unionising against the more powerful harder, but it might depend on the governance structure. Maybe if we are really careful, we can get something closer to "Direct Democracy", where individual preferences actually matter more!
[sorry, have only skimmed the post, but I feel compelled to comment.]
I feel like unless we make a lot of progress on some sort of "Science of Generalisation of Preferences", for more abstract preferences (non-biological needs mostly fall into this), even if certain individuals have, on paper, much more power than others, at the end of the day, they likely rely on vastly superintelligent AI advisors to realise those preferences, and at that point, I think it is the AI advisor _really_ in control.
I'm not super certain of this, like, the Catholic Church definitely could decide to build a bunch of churches on some planets (though what counts as a church, in the limit?), but if they also want more complicated things like "people" "worshipping" "God" in those churches, it seems to be more and more up to the interpretation of the AI Assistants building those worship-maximising communes.
I have read through most of this post and some of the related discussion today. I just wanted to write that it was really interesting, and as far as I can tell, useful, to think through Paul's reasoning and forecasts about strategy-related questions.
In case he believes this is a good idea, I would be very glad to read through a longer, more comprehensive document describing his views on strategic considerations.
Sooo this was such an intriguing idea that I did some research -- but reality appears to be more boring:
In a recent informal discussion I believe said OPP CEO remarked he had to give up the OpenAI board seat as his fiancée joining Anthropic creates a conflict of interest. Naively this is much more likely, and I think is much better supported by the timelines.
According to LinkedIn of the mentioned fiancée joined in already as VP in 2018 and was promoted to a probably more serious position in 2020, and her sibling was promoted to VP in 2019.
The Anthropic split occurred in June 2021.
A new board member (who is arguably very aligned... (read more)
I think the point was less about a problem with refugees (which should be solved in time with European coordination), maybe more that the whole invasion is "good news" for conservative parties, as most crises are.
A lot of people brought up sanctions, and they could indeed influence European economy/politics.
I would be curious about what sanctions in particular are likely to be implemented, and what are their implications - a major economic setback/energy prices soaring could radicalize European politics perhaps?
My guess would be that overall the whole event increases support for conservative/nationalist/populist parties - for example, even though Hungary's populist government was trying to appear to be balancing "between the West and Russia" (thus now being in an uncomfortable situation), I think they can probably actually spin it around to their advantage. (Perhaps even more so, if they can fearmonger about refugees.)
This post was written under Evan Hubinger’s direct guidance and mentorship, as a part of the Stanford Existential Risks Institute ML Alignment Theory Scholars (MATS) program. Where not indicated, ideas are summaries of the documents mentioned, not original contributions. I am thankful for the encouraging approach of the organizers of the program, especially Oliver Zhang.
TL;DR
In the post “The case for aligning narrowly superhuman models”, Ajeya Cotra advocates for performing technical alignment work on currently available large models using human feedback, in domains where (some) humans are already outperformed (in some aspects). The direct benefit of this line of work could be testing out conceptual work, highlighting previously unknown challenges and eliciting new scalable,... (read 2301 more words →)
It is a bit ambiguous from your reply whether you mean distributed AI deployment, or distributed training. Agree that distributed deployment seems very hard to police, once training took place, implying also that there is some large amount of compute available somewhere.
About training, I guess the hope for enforcement would be ability to constrain (or at least monitor) total compute available and hardware manufacturing.
Even if you do training in a distributed fashion, you would need the same amount of chips. (Probably more by some multiplier, to pay for increased latency? And if you can't distribute it to an arbitrary extent, you still need large datacenters that are hard to hide.)
Disguising hardware production... (read more)