I’m broadly interested in AI strategy and want to figure out the most effective interventions to get good AI outcomes.
Thomas Larsen
Karma: 2,942
One framing that I think might be helpful for thinking about “Plan A” vs “shut it all down” is: “Suppose that you have the political will for an n-year slowdown, i.e. after n years, you are forced to handoff trust to superhuman AI systems (e.g. for n = 5, 10, 30). What should the capability progression throughout the slowdown be?” This framing forces a focus on the exit condition / plan to do handoff, which I think is an underdiscussed weakness of the “shut it all down” plan.
I think my gut reaction is that the most important considerations are: (i) there are a lot of useful things you can do with the AIs, so I want more time with the smarter AIs, and (ii) I want to scale through the dangerous capability range slowly and with slack (as opposed to at the end of the slowdown).
this makes me think that particularly for a shorter slowdown (e.g. 5 years), you want to go fast at the beginning (e.g. scale to ~max controllable AI over the first year or two), and then elicit lots of work out of those AIs for the rest of the time period.
A key concern for the above plan is that govts/labs botch the measurement of “max controllable AI”, and scale too far.
But it’s not clear to me how a further delay helps with this, unless you have a plan for making the institutions better over time, or pursuing a less risky path (e.g. ignoring ML and doing human intelligence augmentation).
Going slower, on the other hand, definitely does help, but requires not shutting it all down.
More generally, it seems good to do something like “extend takeoff evenly by a factor of n”, as opposed to something like “pause for n-1 years, and then do a 1 year takeoff”.
I am sympathetic to shut all down and go for human augmentation: I do think this reduces AI takeover risk a lot, but this requires a very long pause, and it requires our institutions to bet big on a very unpopular technology. I think that convincing governments to “shut it all down” without an exit strategy at all seems quite difficult as well.
Ofc this framing also ignores some important considerations, e.g. choices about the capability progression effect both difficulty of enforcement/verification (in both directions: AI lie detectors/ai verification is helpful, while making AIs closer to the edge is a downside), as well as willingness to pay over time (e.g. scary demos or AI for epistemics might help increase WTP)
However, I also think that open agency approaches to transparency face two key difficulties: competitiveness and safety-of-the-components.[18]
I think a third key difficulty with this class of approaches is something like “emergent agency”, i.e. that each of the individual components seem to be doing something safe, but when you combine several of the agents, you get a scary agent. Intuition pump: each of the weights in a NN is very understandable (it’s just a number) and not doing dangerous scheming, but if you compose them it might be scary. Analagously, each of the subagents in the open agency AI might not be scheming, but a collection of these agents might be scheming.
Understanding the communications between the components seems like it may or may not be sufficient to mitigate this failure mode. If the understanding is “local”, i.e. looking at a particular chain of reasoning and verifying that it is valid, this is probably not sufficient to mitigate the problem, as scary reasoning might be made up of a bunch of small chains of local valid reasoning that looks safe. So I think you want something like a reasonable global picture of the reasoning that the open agent is doing in order to mitigate “emergent agency”.
I think this is kind of related to types of the “safety of the components” failure mode you talk about, particularly in the analogue to the corporation passing memos around, but the memos not corresponding to the “real reasoning” going on. However, it could be that the “real reasoning” emerges on a higher level of abstraction than the individual agents.
This sort of threat model leads me to think that if we’re aiming for this sort of open agency, we shouldn’t do end-to-end training of the whole system, lest we incentivize “emergent agency”, even if we don’t make the individual components less safe.
One upside of shut it all down is that it does in fact buy more time: in Plan A it is difficult to secure algorithmic secrets without extremely aggressive security measures, hence any rogue projects (e.g. nation state blacksites) can just coast off the algos developed by the verified projects. Then, a few years in, they fire up their cluster and try to do an intelligence explosion with the extra algorithmic progress.
>superintelligence
Small detail: My understanding of the IABIED scenario is that their AI was only moderately superhuman, not superintelligent
This post seems systematically too slow to me, and to underrate the capabilities of superintelligence. One particular point of disagreement:
It seems reasonable to use days or weeks as an upper bound on how fast robot doublings could become, based on biological analogies. This is very fast indeed.20
When I read this, I thought this would say “lower bound”. Why would you expect evolution to find globally optimal doubling times? This reads to me a bit like saying that the speed of a Cheetah or the size of an Blue Whale will be an upper bound on the speed/size of a robot. Why???
The case for lower bound seems clear: biology did it, probably a superintelligence could design a more functional robot than biology.
Small typo: Alog(B) = log(B^A), not log(A^B)
Also there’s a good chance AI gov won’t work, and labs will just have a very limited safety budget to implement their best guess mitigations. Or maybe AI gov does work and we get a large budget, we still need to actually solve alignment.
Thanks for writing this!
For what its worth, my view is that we’re very likely to be wrong about the specific details in both of the endings—they are obviously super conjunctive. I don’t think that there’s any way around this because we can be confident AGI is going to cause some ex-ante surprising things to happen.
Also, this is scenario is around 20th percentile timelines for me, my median is early 2030s (though other authors disagree with me). I also feel much more confident about the pre-2027 scenario than about the post 2027 scenario.
Is your disagreement that you think AGI will happen later, or that you think the effects of AGI on the world will look very different, or both? If its just the timelines, we might have fairly similar views.
This wasn’t intended to be humor. In the scenario, we write:
(To avoid singling out any one existing company, we’re going to describe a fictional artificial general intelligence company, which we’ll call OpenBrain. We imagine the others to be 3–9 months behind OpenBrain.)
I think that OpenAI, GDM, and Anthropic are in the lead and are the most likely to be ahead, with similar probability.
Thank you! We actually tried to write one that was much closer to a vision we endorse! The TLDR overview was something like:
Both the US and Chinese leading AGI projects stop in response to evidence of egregious misalignment.
Sign a treaty to pause smarter-than-human AI development, with compute based enforcement similar to ones described in our live scenario, except this time with humans driving the treaty instead of the AI.
Take time to solve alignment (potentially with the help of the AIs). This period could last anywhere between 1-20 years. Or maybe even longer! The best experts at this would all be brought in to the leading project, many different paths would be pursued (e.g. full mechinterp, Davidad moonshots, worst case ELK, uploads, etc).
Somehow, a do a bunch of good governance interventions on the AGI project (e.g. transparency on use of the AGIs, no helpful only access to any one. party, a formal governance structure where a large number of diverse parties all are represented.).
This culminates with aligning an AI “in the best interests of humanity” whatever that means, using a process where a large fraction of humanity is engaged and has some power to vote. This process might look something like giving each human some of the total resources of space and then doing lots of bargaining to find all the positive sum trades, with some rules against blackmail / using your resources to cause immense harm.
Unfortunately, it was hard to write this out in a way that felt realistic.
The next major project I focus on is likely going to be focusing on thinking through the right governance interventions here to make that happen. I’m probably not going to do this in scenario format (and instead something closer to normal papers and blog posts), but would be curious for thoughts.
AI 2027: What Superintelligence Looks Like
Thanks—I see, I was misunderstanding.
Proposal part 1: Shoggoth/Face Distinction: Instead of having one model undergo agency training, we have two copies of the base model work together, specializing in different parts of the job, and undergo training together. Specifically we have the “shoggoth” copy responsible for generating all the ‘reasoning’ or ‘internal’ CoT, and then we have the “face” copy responsible for the ‘actions’ or ‘external’ outputs. So e.g. in a conversation with the user, the Shoggoth would see the prompt and output a bunch of reasoning token CoT; the Face would see the prompt + all that CoT and then output a message to pass to the user, and then the process would repeat.
Proposal part 2: Blind the Evaluation Process to the Internal Reasoning: Remember that reward process that evaluates overall performance? Hide from it all the outputs of the Shoggoth. Just show it the outputs from the Face.I don’t see how this proposal could get you AGI-level capabilities. You need the “shoggoth” part of the model to gain general purpose capabilities somehow, and if you are completely shielding the shoggoth from RL, all of the capabilities have to come from pre-training, which seems unlikely. My guess is that a more realistic proposal involves having separate reward models for the two parts:
The shoggoth RM, which only rewards capabilities
The face RM, which also includes alignment-relevant stuff
But it isn’t clear to me how to separate out the alignment/capabilities. IIUC the hope is that the capabilities training seems would naturally incentivize the shoggoth to have bad goals, and perhaps instrumentally decieve the humans, but to do this deception in the COT in english so we can see.
You later write:
How it all (hopefully) fits together: First, because of the paraphraser, the Shoggoth will not learn to speak in some optimized alien language, or even to speak in jargon or euphemisms—because semantics not recognized by the paraphraser will be scrambled by it. So the Shoggoth will continue to ‘think in English’ even as it undergoes lots of agency training and learns substantial new skills.
Second, because the training process is blind to the outputs of the Shoggoth (and instead only sees what the Face outputs), there should be no pressure on the Shoggoth to learn to decieve, conceal, dissemble, doublethink, etc. Instead the Shoggoth will be ‘free’ to put all its important thoughts in plain English.Which suggests we are training the shoggoth with agency training. But I don’t understand how this is consistnet with “the training process be blind to the outputs of the Shoggoth”. Did you mean just some “alignment” subset of the training process?
I think a problem with all the proposed terms is that they are all binaries, and one bit of information is far too little to characterize takeoff:
One person’s “slow” is >10 years, another’s is >6 months.
The beginning and end points are super unclear; some people might want to put the end point near the limits of intelligence, some people might want to put the beginning points at >2x AI R&D speed, some at 10, etc.
In general, a good description of takeoff should characterize capabilities at each point on the curve.
So I don’t really think that any of the binaries are all that useful for thinking or communicating about takeoff. I don’t have a great ontology for thinking about takeoff myself to suggest instead, but I generally try to in communication just define a start and end point and then say quantitatively how long this might take. One of the central ones I really care about is the time between wakeup and takeover capable AIs.
wakeup = “the first period in time when AIs are sufficiently capable that senior government people wake up to incoming AGI and ASI”
takeover capable AIs = “the first time there is a set of AI systems that are coordinating together and could take over the world if they wanted to”
The reason to think about this period is that (kind of by construction) it’s the time where unprecedented government actions that matter could happen. And so when planning for that sort of thing this length of time really matters.
Of course, the start and end times I think about are both fairly vague. They also aren’t purely a function of AI capabilities, and they care about stuff like “who is in government” and “how capable our institutions are at fighting a rogue AGI”. Also, many people believe that we never will get takeover capable AIs even at superintelligence.
Perhaps that was overstated. I think there is maybe a 2-5% chance that Anthropic directly causes an existential catastrophe (e.g. by building a misaligned AGI). Some reasoning for that:
I doubt Anthropic will continue to be in the lead because they are behind OAI/GDM in capital. They do seem around the frontier of AI models now, though, which might translate to increased returns, but it seems like they do best on very short timelines worlds.
I think that if they could cause an intelligence explosion, it is more likely than not that they would pause for at least long enough to allow other labs into the lead. This is especially true in short timelines worlds because the gap between labs is smaller.
I think they have much better AGI safety culture than other labs (though still far from perfect), which will probably result in better adherence to voluntary commitments.
On the other hand, they haven’t been very transparent, and we haven’t seen their ASL-4 commitments. So these commitments might amount to nothing, or Anthropic might just walk them back at a critical juncture.
2-5% is still wildly high in an absolute sense! However, risk from other labs seems even higher to me, and I think that Anthropic could reduce this risk by advocating for reasonable regulations (e.g. transparency into frontier AI projects so no one can build ASI without the government noticing).
I agree with Zach that Anthropic is the best frontier lab on safety, and I feel not very worried about Anthropic causing an AI related catastrophe. So I think the most important asks for Anthropic to make the world better are on its policy and comms.
I think that Anthropic should more clearly state its beliefs about AGI, especially in its work on policy. For example, the SB-1047 letter they wrote states:Broad pre-harm enforcement. The current bill requires AI companies to design and implement SSPs that meet certain standards – for example they must include testing sufficient to provide a “reasonable assurance” that the AI system will not cause a catastrophe, and must “consider” yet-to-be-written guidance from state agencies. To enforce these standards, the state can sue AI companies for large penalties, even if no actual harm has occurred. While this approach might make sense in a more mature industry where best practices are known, AI safety is a nascent field where best practices are the subject of original scientific research. For example, despite a substantial effort from leaders in our company, including our CEO, to draft and refine Anthropic’s RSP over a number of months, applying it to our first product launch uncovered many ambiguities. Our RSP was also the first such policy in the industry, and it is less than a year old. What is needed in such a new environment is iteration and experimentation, not prescriptive enforcement. There is a substantial risk that the bill and state agencies will simply be wrong about what is actually effective in preventing catastrophic risk, leading to ineffective and/or burdensome compliance requirements.
Liability doesn’t not address the central threat model of AI takeover, for which pre-harm mitigations are necessary due to the irreversible nature of the harm. I think that this letter should have acknowledged that explicitly, and that not doing so is misleading. I feel that Anthropic is trying to play a game of courting political favor by not being very straightforward about its beliefs around AGI, and that this is bad.
To be clear, I think it is reasonable that they argue that the FMD and government in general will be bad at implementing safety guidelines while still thinking that AGI will soon be transformative. I just really think they should be much clearer about the latter belief.
Yeah, actual FLOPs are the baseline thing that’s used in the EO. But the OpenAI/GDM/Anthropic RSPs all reference effective FLOPs.
If there’s a large algorithmic improvement you might have a large gap in capability between two models with the same FLOP, which is not desirable. Ideal thresholds in regulation / scaling policies are as tightly tied as possible to the risks.
Another downside that FLOPs / E-FLOPs share is that it’s unpredictable what capabilities a 1e26 or 1e28 FLOPs model will have. And it’s unclear what capabilities will emerge from a small bit of scaling: it’s possible that within a 4x flop scaling you get high capabilities that had not appeared at all in the smaller model.
Credit: Mainly inspired by talking with Eli Lifland. Eli has a potentially-published-soon document here.
The basic case against against Effective-FLOP.
We’re seeing many capabilities emerge from scaling AI models, and this makes compute (measured by FLOPs utilized) a natural unit for thresholding model capabilities. But compute is not a perfect proxy for capability because of algorithmic differences. Algorithmic progress can enable more performance out of a given amount of compute. This makes the idea of effective FLOP tempting: add a multiplier to account for algorithmic progress.
But doing this multiplications seems importantly quite ambiguous.
Effective FLOPs depend on the underlying benchmark. It’s not at all apparent which benchmark people are talking about, but this isn’t obvious.
People often use perplexity, but applying post training enhancements like scaffolding or chain of thought doesn’t improve perplexity but does improve downstream task performance.
See https://arxiv.org/pdf/2312.07413 for examples of algorithmic changes that cause variable performance gains based on the benchmark.
Effective FLOPs often depend on the scale of the model you are testing. See graph below from: https://arxiv.org/pdf/2001.08361 - the compute efficiency from from LSTMs to transformers is not invariant to scale. This means that you can’t just say that the jump from X to Y is a factor of Z improvement on Capability per FLOP. This leads to all sorts of unintuitive properties of effective FLOPs. For example, if you are using 2016-next-token-validation-E-FLOPs, and LSTM scaling becomes flat on the benchmark, you could easily imagine that at very large scales you could get a 1Mx E-FLOP improvement from switching to transformers, even if the actual capability difference is small.
If we move away from pretrained LLMs, I think E-FLOPs become even harder to define, e.g., if we’re able to build systems may be better at reasoning but worse at knowledge retrieval. E-FLOPs does not seem very adaptable.
(these lines would need to parallel for the compute efficiency ratio to be scale invariant on test loss)
Users of E-FLOP often don’t specify the time, scale, or benchmark that they are talking about it with respect to, which makes it very confusing. In particular, this concept has picked up lots of steam and is used in the frontier lab scaling policies, but is not clearly defined in any of the documents.
Anthropic: “Effective Compute: We define effective compute as roughly the amount of compute it would have taken to train a model if no improvements to pretraining or fine-tuning techniques are included. This is operationalized by tracking the scaling of model capabilities (e.g. cross-entropy loss on a test set).”
This specifies the metric, but doesn’t clearly specify any of (a) the techniques that count as the baseline, (b) the scale of the model where one is measuring E-FLOP with respect to, or (c) how they handle post training enhancements that don’t improve log loss but do dramatically improve downstream task capability.
OpenAI on when they will run their evals: “This would include whenever there is a >2x effective compute increase or major algorithmic breakthrough”
They don’t define effective compute at all.
Since there is significant ambiguity in the concept, it seems good to clarify what it even means.
Basically, I think that E-FLOPs are confusing, and most of the time when we want to use flops, we’re usually just going to be better off talking directly about benchmark scores. For example, instead of saying “every 2x effective FLOP” say “every 5% performance increase on [simple benchmark to run like MMLU, GAIA, GPQA, etc] we’re going to run [more thorough evaluations, e.g. the ASL-3 evaluations]. I think this is much clearer, much less likely to have weird behavior, and is much more robust to changes in model design.
It’s not very costly to run the simple benchmarks, but there is a small cost here.
A real concern is that it is easier to game benchmarks than FLOPs. But I’m concerned that you could get benchmark gaming just the same with E-FLOPs because E-FLOPs are benchmark dependent — you could make your model perform poorly on the relevant benchmark and then claim that you didn’t scale E-FLOPs at all, even if you clearly have a broadly more capable model.
A3 in https://blog.heim.xyz/training-compute-thresholds/ also discusses limitations of effective FLOPs.
I think I mostly am on board with this comment. Some thoughts:
this feels a bit overly binary to me. I think that understanding-based safety cases will be necessary for ASI. But behavioral methods seem like they might be sufficient before hand.
I don’t know what you mean by “rapid growth”. It seems like you might be imagining the “shut it all down → solve alignment during pause → rapidly scale after you’ve solved alignment” plan. I think we probably should never do a “rapid scaleup”
Another reaction I have is that a constraint to coordination will probably be “is the other guy doing a blacksite which will screw us over”. So I think there’s a viability bump at the point of “allow legal capabiliites scaling at least as fast as the max size blacksite that you would have a hard time detecting”.
So I think this paragraph isn’t really right, because “slowdown’ != ‘pause’, and slowdowns might still be really really helpful and enough to get you a long way.
I actually currently think that you want to accelerate compute production, because hardware scaling seems safer than software scaling. I’m not sure exactly what you mean by “in an uncontrolled fashion”.. if you mean “have a bunch of inspectors making sure the flow of new chips isn’t being smuggled to illegal projects”, then I agree with this, on my initial read I thought you meant something like “pause chip production until they start producing GPUS with HEMs in them”, which I think is probably bad.
In other words I think that you want to create a big compute overhang during a pause. The downside is obvious, but the upsides are:
compute is controllable, far more than software, and so differentially advances legal projects.
more compute for safety. We want to be able to pay a big safety tax, more compute straightforwardly helps.
extra compute progress funges against software progress, which is scarier.
compute is destroyable (e.g. we can reverse and destroy compute, if we want to eat an overhang), but software progress mostly isn’t (you can’t unpublish reserach).
(this comment might be confusing because I typed it quickly, happy to clarify if you want)