The important thing for alignment work isn't the median prediction; if we had an alignment solution just by then, we'd have a 50% chance of dying from that lack.

I think the biggest takeaway is that nobody has a very precise and reliable prediction, so if we want to have good alignment plans in advance of AGI, we'd better get cracking.

I think Daniel's estimate does include a pretty specific and plausible model of a path to AGI, so I take his the most seriously. My model of possible AGI architectures requires even less compute than his, but I think the Hofstadter principle applies to AGI development if not compute progress.

Estimates in the absence of gears-level models of AGI seem much more uncertain, which might be why Ajeya and Ege's have much wider distributions.

That all makes sense. To expand a little more on some of the logic:

It seems like the outcome of a partial pause rests in part on whether that would tend to put people in the lead of the AGI race who are more or less safety-concerned.

I think it's nontrivial that we currently have three teams in the lead who all appear to honestly take the risks very seriously, and changing that might be a very bad idea.

On the other hand, the argument for alignment risks is quite strong, and we might expect more people to take the risks more seriously as those arguments diffuse. This might not happen if polarization becomes a large factor in beliefs on AGI risk. The evidence for climate change was also pretty strong, but we saw half of America believe in it less, not more, as evidence mounted. The lines of polarization would be different in this case, but I'm afraid it could happen. I outlined that case a little in AI scares and changing public beliefs

In that case, I think a partial pause would have a negative expected value, as the current lead decayed, and more people who believe in risks less get into the lead by circumventing the pause.

This makes me highly unsure if a pause would be net-positive. Having alignment solutions won't help if they're not implemented because the taxes are too high.

The creation of compute overhang is another reason to worry about a pause. It's highly uncertain how far we are from making adequate compute for AGI affordable to individuals. Algorithms and compute will keep getting better during a pause. So will theory of AGI, along with theory of alignment.

This puts me, and I think the alignment community at large, in a very uncomfortable position of not knowing whether a realistic pause would be helpful.

It does seem clear that creating mechanisms and political will for a pause are a good idea.

Advocating for more safety work also seems clear cut.

To this end, I think it's true that you create more political capitol by successfully pushing for policy.

A pause now would create even more capitol, but it's also less likely to be a win, and it could wind up creating polarization and so costing rather than creating capitol. It's harder to argue for a pause now when even most alignment folks think we're years from AGI.

So perhaps the low-hanging fruit is pushing for voluntary RSPs, and government funding for safety work. These are clear improvements, and likely to be wins that create capitol for a pause as we get closer to AGI.

There's a lot of uncertainty here, and that's uncomfortable. More discussion like this should help resolve that uncertainty, and thereby help clarify and unify the collective will of the safety community.

That is indeed a lot of points. Let me try to parse them and respond, because I think this discussion is critically important.

Point 1: overhang.

Your first two paragraphs seem to be pointing to downsides of progress, and saying that it would be better if nobody made that progress. I agree. We don't have guaranteed methods of alignment, and I think our odds of survival would be much better if everyone went way slower on developing AGI.

The standard thinking, which could use more inspection, but which I agree with, is that this is simply not going to happen. Individuals that decide to step aside are slowing progress only slightly. This leaves compute overhang that someone else is going to take advantage of, with nearly the competence, and only slightly slower. Those individuals who pick up the banner and create AGI will not be infinitely reckless, but the faster progress from that overhang will make whatever level of caution they have less effective.

This is a separate argument from regulation. Adequate regulation will slow progress universally, rather than leaving it up to the wisdom and conscience of every individual who might decide to develop AGI.

I don't think it's impossible to slow and meter progress so that overhang isn't an issue. But I think it is effectively even harder than alignment. We have decent suggestions on the table for alignment now, and as far as I know, no equally promising suggestions for getting everyone (and it does take almost everyone coordinating) to pass up the immense opportunities offered by capabilities overhangs.

Point 2: Are LLMs safer than other approaches?

I agree that this is a questionable proposition. I think it's worth questioning. Aiming progress at easier-to-align approaches seems highly worthwhile.

I agree that an LLM may have something like a mind inside. I think current versions are almost certainly too dumb to be existentially dangerous (at least directly - if a facebook algorithm can nearly cause an insurrection, who knows what dangerous side effects any AI can have).

I'm less worried about GPT10 playing a superintelligent, Waluigi-collapsed villain than I am about a GPT6 that has been amplified to agency, situational awareness, and weak superintelligence by scaffolding it into something like a cognitive architecture. I think this type of advance is inevitable. ChatGPT extensions and Bing Chat both use internal prompting to boost intelligence, and approaches like SmartGPT and Tree of Thoughts massively improve benchmark results over the base LLM.

Fortunately, this direction also has huge advantages for alignment. It has a very low alignment tax, since you give them additional goals in natural language, like "support human empowerment" or whatever the SOTA alignment goal is. And they have vastly better interpretability since they're at least summarizing their thoughts in natural language.

Here's where your skepticism that they're being honest about summarizing those thoughts comes into full force. I agree that it's not reliable; for instance, changing the intermediate answer in chain of thought prompting often doesn't change the final output, indicating that that output was for show.

However, a safer setup is to never use the same model twice. When you use chain-of-thought reasoning, construct a new context with the relevant information from memory; don't just let the context window accrue, since this allows fake chains-of-thought and the collapse of the simulator into a waluigi.

Scaffolded LLMs should not turn an LLM into an agent, but rather create a committee of LLMs that are called for individual questions needed to accomplish that committee's goals.

This isn't remotely a solution to the alignment problem, but it really seems to have massive upsides, and only the same downsides as other practically viable approaches to AGI.

To be clear, I only see some form of RL agents as the other practical possibility, and I like our odds much less with those.

I think there are other, even more readily alignable approaches to AGI. But they all seem wildly impractical. I think we need to get ready to align the AGI we get, rather than just preparing to say I-told-you-so after the world refuses to forego massive incentives to take a much slower but safer route to AGI.

To paraphrase, we need to go to the alignment war with the AGI we get, not the AGI we want.

I really like your recent series of posts that succinctly address common objections/questions/suggestions about alignment concerns. I'm making a list to show my favorite skeptics (all ML/AI people; nontechnical people, as Connor Leahy puts it, tend to respond "You fucking what? Oh hell no!" or similar when informed that we are going to make genuinely smarter-than-us AI soonish).

We do have ways to get an AI to do what we want. The hardcoded algorithmic maximizer approach seems to be utterly impractical at this point. That leaves us with approaches that don't obviously do a good job of preserving their own goals as they learn and evolve:

1. training a system to pursue things we like, such as shard theory and similar approaches.
2. training or hand-coding a critic system, as in outlined approaches from Steve Byrnes and me as well as many others. Nicely summarized as a Steering systems approach. This seems a bit less sketchy than training in our goals and hoping they generalize adequately, but still pretty sketchy. 
3. Telling the agent what to do in a Natural language alignment approach. This seems absurdly naive. However, I'm starting to think our first human-plus AGIs will be wrapped or scaffolded LLMs, and they to a nontrivial degree actually think in natural language. People are right now specifying goals in natural language, and those can include alignment goals (or destroying humanity, haha).  I just wrote an in-depth post on the potential Capabilities and alignment of LLM cognitive architectures, but I don't have a lot to say about stability in that post.

None of these directly address what I'm calling The alignment stability problem, to give a name to what you're addressing here. I think addressing it will work very differently in each of the three approaches listed above, and might well come down to implementational details within each approach. I think we should be turning our attention to this problem along with the initial alignment problems, because some of the optimism in the field stems from thinking about initial alignment and not long-term stability.

Edit: I left out Ozyrus's posts on approach 3. He's the first person I know of to see agentized LLMs coming, outside of David Shapiro's 2021 book. His post was written a year ago and posted two weeks ago to avoid infohazards. I'm sure there are others who saw this coming more clearly than I did, but I thought I'd try to give credit where it's due.

Great analysis. I'm impressed by how thoroughly you've thought this through in the last week or so. I hadn't gotten as far. I concur with your projected timeline, including the difficulty of putting time units onto it. Of course, we'll probably both be wrong in important ways, but I think it's important to at least try to do semi-accurate prediction if we want to be useful.

I have only one substantive addition to your projected timeline, but I think it's important for the alignment implications.

LLM-bots are inherently easy to align. At least for surface-level alignment. You can tell them "make me a lot of money selling shoes, but also make the world a better place" and they will try to do both. Yes, there are still tons of ways this can go off the rails. It doesn't solve outer alignment or alignment stability, for a start. But GPT4's ability to balance several goals, including ethical ones, and to reason about ethics, is impressive.^[1] You can easily make agents that both try to make money, and thinks about not harming people.

In short, the fact that you can do this is going to seep into the public consciousness, and we may see regulations and will definitely see social pressure to do this.

I think the agent disasters you describe will occur, but they will happen to people that don't put safeguards into their bots, like "track how much of my money you're spending and stop if it hits $X and check with me". When agent disasters affect other people, the media will blow it sky high, and everyone will say "why the hell didn't you have your bot worry about wrecking things for others?". Those who do put additional ethical goals into their agents will crow about it. There will be pressure to conform and run safe bots. As bot disasters get more clever, people will take more seriously the big bot disaster.

Will all of that matter? I don't know. But predicting the social and economic backdrop for alignment work is worth trying.

Edit: I finished my own followup post on the topic, Capabilities and alignment of LLM cognitive architectures. It's a cognitive psychology/neuroscience perspective on why these things might work better, faster than you'd intuitively think. Improvements to the executive function (outer script code) and episodic memory (pinecone or other vector search over saved text files) will interact so that improvements in each make the rest of system work better and easier to improve.

1. ^{^}

   I did a little informal testing of asking for responses in hypothetical situations where ethical and financial goals collide, and it did a remarkably good job, including coming up with win/win solutions that would've taken me a while to come up with. It looked like the ethical/capitalist reasoning of a pretty intelligent person; but also a fairly ethical one.

I wrote up a short post with a summary of their results. It doesn't really answer any of your questions. I do have thoughts on a couple, even though I'm not expert on interpretability. 

But my main focus is on your footnote: is this going to help much with aligning "real" AGI (I've been looking for a term; maybe REAL stands for Reflective Entities with Agency and Learning?:). I'm of course primarily thinking of foundation models scaffolded to have goals, cognitive routines, and incorporate multiple AI systems such as an episodic memory system. I think the answer is that some of the interpretability work will be very valuable even in those systems, while some of it might be a dead end - and we haven't really thought through which is which yet.

It would be more useful with a little more info on what ideas you're offering. Linkposts with more description get more clickthrough. You can edit in a little more info.

I think you're right about these drawbacks of using the term "alignment" so broadly. And I agree that more work and attention should be devoted to specifying how we suppose these concepts relate to each other. In my experience, far too little effort is devoted to placing scientific work within its broader context. We cannot afford to waste effort in working on alignment.

I don't see a better alternative, nor do you suggest one. My preference in terminology is to simply use more specification, rather than trying to get anyone to change the terminology they use. With that in mind, I'll list what I see as the most common existing terminology for each of the sub-problems.

P1: Avoiding takeover from emergent optimization in AI agents

Best term in use: AInotkilleveryoneism. I disagree that alignment is commonly misused for this.

I don't think I've heard these termed alignment, outside of the assumption you mention in the Berkeley model of value alignment (P5) as the only way of avoiding takeover (P1). P1 has been termed "the control problem" which encompasses value alignment. Which is good. This does not fit the intuitive definition of alignment. The deliberately clumsy term "AInotkilleveryoneism" seems good for this, in any context you can get away with it. Your statement seems good otherwise.

P2: Ensuring that AI’s information processing (and/or reasoning) is intelligible to us

Best term in use: interpretability

This is more commonly called interpretability, but I agree that it's commonly lumped into "alignment work" without carefully examining just how it fits in. But it does legitimately fit into P1 (which shouldn't be called alignment, as well as (what I think you mean by) P3, P5, and P6, which do fit the intuitive meaning of "alignment." Thus, it does seem like this deserves the term "alignment work" as well as its more precise term of interpretability. So this seems about right, with the caveat of wanting more specificity. As it happens, I just now published a post on exactly this.

P3: Ensuring AIs are good at solving problems as specified (by user or designer)

Best term in use: None. AI safety?

I think you mean to include ensuring AIs also do not do things their designers don't want. I suggest changing your description, since that effort is more often called alignment and accused of safety-washing.

This is the biggest offender. The problem is that "alignment is intuitively appealing. I'd argue that this is completely wrong: you can't align a system with goals (humans) with a tool without goals (LLMs). A sword or an axe are not aligned with their wielders; they certainly lead to more trees cut down and people stabbed, but they do not intend those things, so there's a type error in saying they are aligned with their users goals. 

But this is pedantry that will continue to be ignored. I don't have a good idea for making this terminology clear. The term AGI at one point was used to specify AI with agency and goals, and thus which would be alignable with human goals, but it's been watered down. We need a replacement. And we need a better term for "aligning" AIs that are not at all dangerous in the severe way the "alignment problem" terminology was intended to address. Or a different term for doing the important work of aligning agentic, RSI-capable AGI.

P4: Ensuring AI systems enhance, and don’t erode, human agency

What? I'd drop this and just consider it a subset of P6. Maybe this plays a bigger role and gets the term alignment more than I know? Do you have examples?

P5: Ensuring that advanced AI agents learn a human utility function

Best term in use: value alignment OR technical alignment. 

I think these deserve their own categories in your terminology, because they overlap - technical alignment could be limited to making AGIs that follow instructions. I have been thinking about this a lot. I agree with your analysis that this is what people will probably do, for economic reasons; but I think there are powerful practical reasons that this is much easier than full value alignment, which will be a valuable excuse to align it to follow instructions from its creators. I recently wrote up that logic.  This conclusion raises another problem that I think deserves to join the flock of related alignment problems: the societal alignment problem. If some humans have AGIs aligned to their values (likely through their intent/instructions), how can we align society to avoid resulting disasters from AGI-powered conflict?

P6: Ensuring that AI systems lead to desirable systemic and long term outcomes

Best term in use: I don't think there is one. Any ideas?

I think you're right about these drawbacks of the widespread use of the term alignment for

The other issue with pausing training of large foundation models (the type of pause we might actually achieve) is that it could change the direction of AGI research to a less-alignable type of first AGIs.

This is related to but different from the direction argument made by Russel Thor in that comment thread is. You don't have to think that LLMs are an inefficient path to AGI to worry about changing the direction. LLMs might be an equally or more efficient path that happens to be safer. Indeed, I think it is. The instructability and translucency of LLM cognition makes them an ideal base model out of which to build "real" agentic, self-improving AGI that can be aligned to human intentions and whose cognitive processes can be understood and monitored relatively well.

On the other hand, pausing new training runs might redirect a lot of effort into fleshing out current foundation models into more useful cognitive architectures. These systems seem like our best chance of alignment. Progress on foundation models might remove the fairly tight connection between LLMs cognition and the language they emit. That would make them less "translucent" and thereby less alignable. So that's a reason to favor a pause.

Neither of those arguments are commonly raised; I have yet to write a post about it.