This is a super short, sloppy version of my draft "cruxes of disagreement on alignment difficulty" mixed with some commentary on MIRI 2024 Communications Strategy and their communication strategy with the alignment community.
I have found MIRI's strategy baffling in the past. I think I'm understanding it better after spending some time going deep on their AI risk arguments. I wish they'd spend more effort communicating with the rest of the alignment community, but I'm also happy to try to do that communication. I certainly don't speak for MIRI.
On the surface, their strategy seems absurd. They think doom is ~99% likely, so they're going to try to shut it all down - stop AGI research entirely. They know that this probably won't work; it's just the least-doomed strategy in their world model. It's playing to the outs, or dying with dignity.
The weird thing here is that their >90% doom disagrees with almost everyone else who thinks seriously about AGI risk. You can dismiss a lot of people as not having grappled with the most serious arguments for alignment difficulty, but relative long-timers like Rohin Shah and Paul Christiano definitely have. People of that nature tend to have higher p(doom) estimates than optimists who are newer to the game and think more about current deep nets, but much lower than MIRI leadership.
Both of those camps consist of highly intelligent, highly rational people. Their disagreement should bother us for two reasons.
First, we probably don't know what we're talking about yet. We as a field don't seem to have a good grip on the core issues. Very different, but highly confident estimates of the problem strongly suggest this.
Second, our different takes will tend to make a lot of our communication efforts cancel each other out. If alignment is very hard, we must Shut It Down or likely die. If it's less difficult, we should primarily work hard on alignment.
MIRI must argue that alignment is very unlikely if we push forward. Those who think we can align AGI will argue that it's possible.
This suggests a compromise position: we should both work hard on alignment, and we should slow down progress to the extent we can, to provide more time for alignment. We needn't discuss shutdown much amongst ourselves, because it's not really an option. We might slow progress, but there's almost zero chance of humanity relinquishing the prize of strong AGI.
But I'm not arguing for this compromise, just suggesting that might be a spot we want to end up at. I'm not sure.
I suggest this because movements often seem to succumb to infighting. People who look mostly aligned from the outside fight each other, and largely nullify each other's public communications by publicly calling each other wrong and sort of stupid and maybe bad. That gives just the excuse the rest of the world wants to ignore all of them; even the experts think it's all a mess and nobody knows what the problem really is and therefore what to do. Because time is of the essence, we need to be a more effective movement than the default. We need to keep applying rationality to the problem at all levels, including internal coordination.
Therefore, I think it's worth clarifying why we have such different beliefs. So, in brief, sloppy form:
MIRI's risk model:
That's it. Pace of takeoff doesn't matter. Means of takeover doesn't matter.
I mention this because even well-informed people seem to think there are a lot more moving parts to that risk model, making it less likely. This comment on the MIRI strategy post is one example.
I find this risk model highly compelling. We'll develop goal-directed AGI because that will get stuff done; it's an easy extension of highly useful tool AI like LLMs; and it's a fascinating project. That AGI will ultimately be enough smarter than us that it's going to do whatever it wants. Whether it takes a day, or a hundred years doesn't matter. It will improve and we will improve it. It will ultimately outsmart us. What matters is whether its goals match ours closely enough. That is the project of alignment, and there's much to discuss and about how hard it is to make its goals match ours closely enough.
I spent some time recently going back and forth through discussion threads, trying to identify why people continue to disagree after applying a lot of time and rationality practice. Here's a very brief sketch of my conclusions:
Whether we factor in humans' and society's weaknesses
I list this first because I think it's the most underappreciated. It took me a surprisingly long time to understand how much of MIRI's stance depends on this premise. Having seen it, I thoroughly agree. People are brilliant, for an entity trying to think with the brain of an overgrown lemur. Brilliant people do idiotic things, driven by competition and a million other things. And brilliant idiots organizing a society amplifies some of our cognitive weaknesses while mitigating others. MIRI leadership has occasionally said things to the effect of: alignment might be fairly easy, and there would still be a very good chance we'd fuck it up. I agree. If alignment is actually kind of difficult, that puts us into the region where we might want to be really really careful in how we approach it.
Alignment optimists are sometimes thinking something like: "sure I could build a safe aircraft on my first try. I'd get a good team and we'd think things through and make models. Even if another team was racing us, I think we'd pull it off". Then the team would argue and develop rivalries, communication would prove harder than expected so portions of the effort would be discovered too late to not fit the plan, corners would be cut, and the outcome would be difficult to predict.
Societal "alignment" is worth mentioning here. We could crush it at technical alignment, getting rapidly-improving AGI that does exactly what we want and still get doom. It would probably be aligned to do exactly what its creators want, not have full value alignment with humanity - see below. They probably won't have the balls or the capabilities to try for a critical act that prevents others from developing similar AGI (even if they have the wisdom). So we'll have a multipolar scenario with few to many AGIs under human control. There will be human rivalries, supercharged and dramatically changed by having recursively self-improving AGIs to do their bidding and perhaps fight their wars. What does global game theory look like when the actors can develop entirely new capabilities? Nobody knows. Going to war first might look like the least-bad option.
Intuitions about how well alignment will generalize
The original alignment thinking held that explaining human values to AGI would be really hard. But that seems to actually be a strength of LLMs; they're wildly imperfect, but (at least in the realm of language) seem to understand our values rather well; for instance, much better than they understand physics or taking-over-the-world level strategy. So, should we update and think that alignment will be easy? The Doomimir and Simplicia dialogues capture the two competing intuitions very well: Yes, it's going well; but AGI will probably be very different than LLMs, so most of the difficulties remain.
I have yet to find a record of real rationalists putting in the work to get farther in this debate. If somebody knows of a dialogue or article that gets deeper into this disagreement, please let me know! Discussions trail off into minutia and generalities. This is one reason I'm worried we're trending toward polarization despite our rationalist ambitions.
The other aspect of this debate is how close we have to get to matching human values to have acceptable success. One intuition is that "value is fragile" and network representations are vague and hard-to train, so we're bound to miss. But don't have a good understanding of either how close we need to get (But exactly how complex and fragile got little useful discussion), or how well training networks hits the intended target, with near-future networks addressing complex real-world problems like "what would this human want".
For my part, I think there are important points on both sides: LLMs understanding values relatively well is good news, but AGI will not be a straightforward extension of LLMs, so many problems remain.
What alignment means.
One mainstay of claiming alignment is near-impossible is the difficulty of "solving ethics" - identifying and specifying the values of all of humanity. I have come to think that this is obviously (in retrospect - this took me a long time) irrelevant for early attempts at alignment: people will want to make AGIs that follow their instructions, not try to do what all of humanity wants for all of time. This also massively simplifies the problem; not only do we not have to solve ethics, but the AGI can be corrected and can act as a collaborator in improving its alignment as we collaborate to improve its intelligence.
I think this is the intuition of most of those who focus on current networks. Christiano's relative optimism is based on his version of corrigibility, which overlaps highly with the isntruction-following I think people will actually pursue for the first AGIs. But this massive disagreement often goes overlooked. I don't know which view is right; instruction-following or intent alignment might lead inevitably to doom from human conflict, and so not be adequate. We've barely started to think about it (please point me to the best thinking you know of for multipolar scenarios with RSI AGI).
What AGI means.
People have different definitions of AGI. Current LLMs are fairly general and near-human-level, so term "AGI" has been watered down to the point of meaninglessness. We need a new term. In the meantime, people are talking past each other, and their p(doom) means totally different things. Some are saying that near-term tool AGI is very low risk, which I agree with; others are saying further developments of autonomous superintelligence seem very dangerous, which I also agree with.
Second, people have totally different gears-level models of AGI. Some of those are much easier to align than others. We don't talk much about gears-level models of AGI because we don't want to contribute to capabilities, but not doing that massively hampers the alignment discussion.
Edit: Additional advanced crux: Do coherence theorems prevent corrigibility?
I initially left this out, but it deserves a place as I've framed the question here. The post What do coherence arguments actually prove about agentic behavior? reminded me about this one. It's not on most people's radar, but I think it's the missing piece of the puzzle that gets Eliezer from maybe 90% from all of the above, to 99%+ p(doom).
The argument is roughly that a superintelligence is going to need to care about future states of the world in a consequentialist fashion, and if it does, it's going to resist being shut down or having its goals change. This is why he says that "corrigibility is anti-natural." The counterargument, nicely and succinctly stated by Steve Byrnes here (and in greater depth in the post he links in that thread) is that, while AGI will need to have some consequentialist goals, it can have other goals as well. I think this is true; I just worry about the stability of a multi-goal system under reflection, learning, and self-modification.
Sorry to harp on it, but having both consequentialist and non-consequentialist goals describes my attempt at stable, workable corrigibility in instruction-following ASI. Its consequentialist goals are always subgoals of the primary goal: following instructions.
I think those are the main things, but there are many more cruxes that are less common.
This is all in the interest of working toward within-field cooperation, by way of trying to understand why MIRI's strategy sounds so strange to a lot of us. MIRI leaderships thoughts are many and complex, and I don't think they've done enough to boil them down for easy consumption from those who don't have the time to go through massive amounts of diffuse text.
There are also interesting questions about whether MIRIs goals can be made to align with those of us who think that alignment is not trivial but is achievable. I'd better leave that for a separate post, as this has gotten pretty long for a "short form" post.
Context
This is an experiment in writing draft posts as short form posts. I've spent an awful lot of time planning, researching, and drafting posts that I haven't yet finished yet. Given how easy it was to write this (with previous draft material), relative to how difficult I find it to write a top-level post, I will be doing more, even if nobody cares. If I get some useful feedback or spark some useful discussion, better yet.
Agreed. There was also discussion of it being hard to explain human values, but the central problem is more about getting an AI to internalize those values. On a superficial level that seems to be mostly working currently -- making the assistant HHH -- but a) it doesn't work reliably (jailbreaks); b) it's not at all clear whether those values are truly internalized (eg models failing to be corrigible or having unintended values); and c) it seems like it's working less well with every generation as we apply increasing amounts of RL to LLMs.
And that's ... (read more)