Quintin Pope


Quintin's Alignment Papers Roundup
Shard Theory

Wiki Contributions


I think training such an AI to be really good at chess would be fine. Unless "Then apply extreme optimization pressure for never losing at chess." means something like "deliberately train it to use a bunch of non-chess strategies to win more chess games, like threatening opponents, actively seeking out more chess games in real life, etc", then it seems like you just get GPT-5 which is also really good at chess. 

I really don't like that you've taken this discussion to Twitter. I think Twitter is really a much worse forum for talking about complex issues like this than LW/AF.

I haven't "taken this discussion to Twitter". Joe Carlsmith posted about the paper on Twitter. I saw that post, and wrote my response on Twitter. I didn't even know it was also posted on LW until later, and decided to repost the stuff I'd written on Twitter here. If anything, I've taken my part of the discussion from Twitter to LW. I'm slightly baffled and offended that you seem to be platform-policing me?

Anyways, it looks like you're making the objection I predicted with the paragraphs:

One obvious counterpoint I expect is to claim that the "[have some internal goal x] [backchain from wanting x to the stuff needed to get x (doing well at training)]" steps actually do contribute to the later steps, maybe because they're a short way to compress a motivational pointer to "wanting" to do well on the training objective.

I don't think this is how NN simplicity biases work. Under the "cognitive executions impose constraints on parameter settings" perspective, you don't actually save any complexity by supposing that the model has some motive for figuring stuff out internally, because the circuits required to implement the "figure stuff out internally" computations themselves count as additional complexity. In contrast, if you have a view of simplicity that's closer to program description length, then you're not counting runtime execution against program complexity, and so a program that has short length in code but long runtime can count as simple.

In particular, when I said "maybe because they're a short way to compress a motivational pointer to "wanting" to do well on the training objective." I think this is pointing at the same thing you reference when you say "The entire question is about what the easiest way is to produce that distribution in terms of the inductive biases."

I.e., given the actual simplicity bias of models, what is the shortest (or most compressed) way of specifying "a model that starts by trying to do well in training"? And my response is that I think the model pays a complexity penalty for runtime computations (since they translate into constraints on parameter values which are needed to implement those computations). Even if those computations are motivated by something we call a "goal", they still need to be implemented in the circuitry of the model, and thus also constrain its parameters.

Also, when I reference models whose internal cognition looks like "[figure out how to do well at training] [actually do well at training]", I don't have sycophantic models in particular in mind. It also includes aligned models, since those models do implement the "[figure out how to do well at training] [actually do well at training]" steps (assuming that aligned behavior does well in training). 

Reposting my response on Twitter (To clarify, the following was originally written as a Tweet in response to Joe Carlsmith's Tweet about the paper, which I am now reposting here):

I just skimmed the section headers and a small amount of the content, but I'm extremely skeptical. E.g., the "counting argument" seems incredibly dubious to me because you can just as easily argue that text to image generators will internally create images of llamas in their early layers, which they then delete, before creating the actual asked for image in the later layers. There are many possible llama images, but "just one" network that straightforwardly implements the training objective, after all.

The issue is that this isn't the correct way to do counting arguments on NN configurations. While there are indeed an exponentially large number of possible llama images that an NN might create internally, there are an even more exponentially large number of NNs that have random first layers, and then go on to do the actual thing in the later layers. Thus, the "inner llamaizers" are actually more rare in NN configuration space than the straightforward NN.

The key issue is that each additional computation you speculate an NN might be doing acts as an additional constraint on the possible parameters, since the NN has to internally contain circuits that implement those computations. The constraint that the circuits actually have to do "something" is a much stronger reduction in the number of possible configurations for those parameters than any additional configurations you can get out of there being multiple "somethings" that the circuits might be doing.

So in the case of deceptive alignment counting arguments, they seem to be speculating that the NN's cognition looks something like:

[have some internal goal x] [backchain from wanting x to the stuff needed to get x (doing well at training)] [figure out how to do well at training] [actually do well at training]

and in comparison, the "honest" / direct solution looks like:

[figure out how to do well at training] [actually do well at training]

and then because there are so many different possibilities for "x", they say there are more solutions that look like the deceptive cognition. My contention is that the steps "[have some internal goal x] [backchain from wanting x to the stuff needed to get x (doing well at training)]" in the deceptive cognition are actually unnecessary, and because implementing those steps requires that one have circuits that instantiate those computations, the requirement that the deceptive model perform those steps actually *constrains* the number of parameter configurations that implement the deceptive cognition, which reduces the volume of deceptive models in parameter space.

One obvious counterpoint I expect is to claim that the "[have some internal goal x] [backchain from wanting x to the stuff needed to get x (doing well at training)]" steps actually do contribute to the later steps, maybe because they're a short way to compress a motivational pointer to "wanting" to do well on the training objective.

I don't think this is how NN simplicity biases work. Under the "cognitive executions impose constraints on parameter settings" perspective, you don't actually save any complexity by supposing that the model has some motive for figuring stuff out internally, because the circuits required to implement the "figure stuff out internally" computations themselves count as additional complexity. In contrast, if you have a view of simplicity that's closer to program description length, then you're not counting runtime execution against program complexity, and so a program that has short length in code but long runtime can count as simple.

@jkcarlsmith does seem more sympathetic to the "parameters as complexity" view than to the "code length as complexity" view. However, I go further and think that the "parameters as complexity" view actively points against deceptive alignment.

I also think NNs have an even stronger bias for short paths than you might expect from just the "parameters as complexity" view. Consider a problem that can either be solved with a single circuit of depth n, or by two parallel circuits of depth n/2 (so both solutions must set the values of an equal number of parameters, but one solution is more parallel than the other). I claim there are far more parameter configurations that implement the parallel solution than parameter configurations that implement the serial solution.

This is because the parallel solution has an entire permutation group that's not available to the serial solution: the two parallel circuits can be moved to different relative depths with respect to each other, whereas all the parts of the serial solution must have fixed relative depths. Thus, the two parallel circuits represent less of a constraint on the possible configurations of the NN, and so there are far more NNs that implement the parallel solution.

As a consequence, I expect there are significant "short depth" biases in the NN simplicity prior, consistent with empirical results such as: https://arxiv.org/abs/1605.06431

Finally, I'm extremely skeptical of claims that NNs contain a 'ghost of generalized instrumental reasoning', able to perform very competent long term hidden scheming and deduce lots of world knowledge "in-context". I think current empirical results point strongly against that being plausible.

For example, the "reversal curse" results (training on "A is B" doesn't lead to models learning "B is A"). If the ghost can't even infer from "A is B" to "B is A", then I think stuff like inferring from "I have a goal x", to "here is the specific task I must perform in order to maximize my reward" is pretty much out of the question. Thus, stories about how SGD might use arbitrary goals as a way to efficiently compress an (effective) desire for the NN to silently infer lots of very specific details about the training process seem incredibly implausible to me.

I expect objections of the form "I expect future training processes to not suffer from the reversal curse, and I'm worried about the future training processes."

Obviously people will come up with training processes that don't suffer from the reversal curse. However, comparing the simplicity of the reversal curse to the capability of current NNs is still evidence about the relative power of the 'instrumental ghost' in the model compared to the external capabilities of the model. If a similar ratio continues to hold for externally superintelligent AIs, then that casts enormous doubt on e.g., deceptive alignment scenarios where the model is internally and instrumentally deriving huge amounts of user-goal-related knowledge so that it can pursue its arbitrary mesaobjectives later down the line. I'm using the reversal curse to make a more generalized update about the types of internal cognition that are easy to learn and how they contribute to external capabilities.

Some other Tweets I wrote as part of the discussion:

Tweet 1:

The key points of my Tweet are basically "the better way to think about counting arguments is to compare constraints on parameter configurations", and "corrected counting arguments introduce an implicit bias towards short, parallel solutions", where both "counting the constrained parameters", and "counting the permutations of those parameters" point in that direction.

Tweet 2:

I think shallow depth priors are pretty universal. E.g., they also make sense from a perspective of "any given step of reasoning could fail, so best to make as few sequential steps as possible, since each step is rolling the dice", as well as a perspective of "we want to explore as many hypotheses as possible with as little compute as possible, so best have lots of cheap hypotheses".

I'm not concerned about the training for goal achievement contributing to deceptive alignment, because such training processes ultimately come down to optimizing the model to imitate some mapping from "goal given by the training process" -> "externally visible action sequence". Feedback is always upweighting cognitive patterns that produce some externally visible action patterns (usually over short time horizons).

In contrast, it seems very hard to me to accidentally provide sufficient feedback to specify long-term goals that don't distinguish themselves from short term one over short time horizons, given the common understanding in RL that credit assignment difficulties actively work against the formation of long term goals. It seems more likely to me that we'll instill long term goals into AIs by "scaffolding" them via feedback over shorter time horizons. E.g., train GPT-N to generate text like "the company's stock must go up" (short time horizon feedback), as well as text that represents GPT-N competently responding to a variety of situations and discussions about how to achieve long-term goals (more short time horizon feedback), and then putting GPT-N in a continuous loop of sampling from a combination of the behavioral patterns thereby constructed, in such a way that the overall effect is competent long term planning.

The point is: long term goals are sufficiently hard to form deliberately that I don't think they'll form accidentally.

Tweet 3:

...I think the llama analogy is exactly correct. It's specifically designed to avoid triggering mechanistically ungrounded intuitions about "goals" and "tryingness", which I think inappropriately upweight the compellingness of a conclusion that's frankly ridiculous on the arguments themselves. Mechanistically, generating the intermediate llamas is just as causally upstream of generating the asked for images, as "having an inner goal" is causally upstream of the deceptive model doing well on the training objective. Calling one type of causal influence "trying" and the other not is an arbitrary distinction.

Tweets 4 / 5:

My point about the "instrumental ghost" wasn't that NNs wouldn't learn instrumental / flexible reasoning. It was that such capabilities were much more likely to derive from being straightforwardly trained to learn such capabilities, and then to be employed in a manner consistent with the target function of the training process. What I'm arguing *against* is the perspective that NNs will "accidentally" acquire such capabilities internally as a convergent result of their inductive biases, and direct them to purposes/along directions very different from what's represented in the training data. That's the sort of stuff I was talking about when I mentioned the "ghost".


What I'm saying is there's a difference between a model that can do flexible instrumental reasoning because it's faithfully modeling a data distribution with examples of flexible instrumental reasoning, versus a model that acquired hidden flexible instrumental reasoning because NN inductive biases say the convergent best way to do well on tasks is to acquire hidden flexible instrumental reasoning and apply it to the task, even when the task itself doesn't have any examples of such.

This is a great post! Thank you for writing it.

There's a huge amount of ontological confusion about how to think of "objectives" for optimization processes. I think people tend to take an inappropriate intentional stance and treat something like "deliberately steering towards certain abstract notions" as a simple primitive (because it feels introspectively simple to them). This background assumption casts a shadow over all future analysis, since people try to abstract the dynamics of optimization processes in terms of their "true objectives", when there really isn't any such thing.

Optimization processes (or at least, evolution and RL) are better thought of in terms of what sorts of behavioral patterns were actually selected for in the history of the process. E.g., @Kaj_Sotala's point here about tracking the effects of evolution by thinking about what sorts of specific adaptations were actually historically selected for, rather than thinking about some abstract notion of inclusive genetic fitness, and how the difference between modern and ancestral humans seems much smaller from this perspective.

I want to make a similar point about reward in the context of RL: reward is a measure of update strength, not the selection target. We can see as much by just looking at the update equations for REINFORCE (from page 328 of Reinforcement Learning: An Introduction):

The reward[1] is literally a (per step) multiplier of the learning rate. You can also think of it as providing the weights of a linear combination of the parameter gradients, which means that it's the historical action trajectories that determine what subspaces of the parameters can potentially be explored. And due to the high correlations between gradients (at least compared to the full volume of parameter space), this means it's the action trajectories, and not the reward function, that provides most of the information relevant for the NN's learning process. 

From Survival Instinct in Offline Reinforcement Learning:

on many benchmark datasets, offline RL can produce well-performing and safe policies even when trained with "wrong" reward labels, such as those that are zero everywhere or are negatives of the true rewards. This phenomenon cannot be easily explained by offline RL's return maximization objective. Moreover, it gives offline RL a degree of robustness that is uncharacteristic of its online RL counterparts, which are known to be sensitive to reward design.

Trying to preempt possible confusion:

I expect some people to object that the point of the evolutionary analogy is precisely to show that the high-level abstract objective of the optimization process isn't incorporated into the goals of the optimized product, and that this is a reason for concern because it suggests an unpredictable/uncontrollable mapping between outer and inner optimization objectives.

My point here is that, if you want to judge an optimization process's predictability/controllability, you should not be comparing some abstract notion of the process's "true outer objective" to the result's "true inner objective". Instead, you should consider the historical trajectory of how the optimization process actually adjusted the behaviors of the thing being optimized, and consider how predictable that thing's future behaviors are, given past behaviors / updates. 

@Kaj_Sotala argues above that this perspective implies greater consistency in human goals between the ancestral and modern environments, since the goals evolution actually historically selected for in the ancestral environment are ~the same goals humans pursue in the modern environment. 

For RL agents, I am also arguing that thinking in terms of the historical action trajectories that were actually reinforced during training implies greater consistency, as compared to thinking of things in terms of some "true goal" of the training process. E.g., Goal Misgeneralization in Deep Reinforcement Learning trained a mouse to navigate to cheese that was always placed in the upper right corner of the maze and found that it would continue going to the upper right even when the cheese was moved. 

This is actually a high degree of consistency from the perspective of the historical action trajectories. During training, the mouse continually executed the action trajectories that navigated it to the upper right of the board, and continued to do the exact same thing in the modified testing environment.

  1. ^

    Technically it's the future return in this formulation, and current SOTA RL algorithms can be different / more complex, but I think this perspective is still a more accurate intuition pump than notions of "reward as objective", even for setups where "reward as a learning rate multiplier" isn't literally true.

I strong downvoted and strong disagree voted. The reason I did both is because I think what you're describing is a genuinely insane standard to take for liability. Holding organizations liable for any action they take which they do not prove is safe is an absolutely terrible idea. It would either introduce enormous costs for doing anything, or allow anyone to be sued for anything they've previously done.

I really don't want to spend even more time arguing over my evolution post, so I'll just copy over our interactions from the previous times you criticized it, since that seems like context readers may appreciate.

In the comment sections of the original post:

Your comment

[very long, but mainly about your "many other animals also transmit information via non-genetic means" objection + some other mechanisms you think might have caused human takeoff]

My response

I don't think this objection matters for the argument I'm making. All the cross-generational information channels you highlight are at rough saturation, so they're not able to contribute to the cross-generational accumulation of capabilities-promoting information. Thus, the enormous disparity between the brain's with-lifetime learning versus evolution cannot lead to a multiple OOM faster accumulation of capabilities as compared to evolution.

When non-genetic cross-generational channels are at saturation, the plot of capabilities-related info versus generation count looks like this:

with non-genetic information channels only giving the "All info" line a ~constant advantage over "Genetic info". Non-genetic channels might be faster than evolution, but because they're saturated, they only give each generation a fixed advantage over where they'd be with only genetic info. In contrast, once the cultural channel allows for an ever-increasing volume of transmitted information, then the vastly faster rate of within-lifetime learning can start contributing to the slope of the "All info" line, and not just its height.

Thus, humanity's sharp left turn.

In Twitter comments on Open Philanthropy's announcement of prize winners:

Your tweet

But what's the central point, than? Evolution discovered how to avoid the genetic bottleneck myriad times; also discovered potentially unbounded ways how to transmit arbitrary number of bits, like learning-teaching behaviours; except humans, nothing foomed. So the updated story would be more like "some amount of non-genetic/cultural accumulation is clearly convergent and is common, but there is apparently some threshold crossed so far only by humans. Once you cross it you unlock a lot of free energy and the process grows explosively". (&the cause or size of treshold is unexplained)

(note: this was a reply and part of a slightly longer chain)

My response

Firstly, I disagree with your statement that other species have "potentially unbounded ways how to transmit arbitrary number of bits". Taken literally, of course there's no species on earth that can actually transmit an *unlimited* amount of cultural information between generations. However, humans are still a clear and massive outlier in the volume of cultural information we can transmit between generations, which is what allows for our continuously increasing capabilities across time.

Secondly, the main point of my article was not to determine why humans, in particular, are exceptional in this regard. The main point was to connect the rapid increase in human capabilities relative to previous evolution-driven progress rates with the greater optimization power of brains as compared to evolution. Being so much better at transmitting cultural information as compared to other species allowed humans to undergo a "data-driven singularity" relative to evolution. While our individual brains and learning processes might not have changed much between us and ancestral humans, the volume and quality of data available for training future generations did increase massively, since past generations were much better able to distill the results of their lifetime learning into higher-quality data.

This allows for a connection between the factors we've identified are important for creating powerful AI systems (data volume, data quality, and effectively applied compute), and the process underlying the human "sharp left turn". It reframes the mechanisms that drove human progress rates in terms of the quantities and narratives that drive AI progress rates, and allows us to more easily see what implications the latter has for the former.

In particular, this frame suggests that the human "sharp left turn" was driven by the exploitation of a one-time enormous resource inefficiency in the structure of the human, species-level optimization process. And while the current process of AI training is not perfectly efficient, I don't think it has comparably sized overhangs which can be exploited easily. If true, this would mean human evolutionary history provides little evidence for sudden increases in AI capabilities.

The above is also consistent with rapid civilizational progress depending on many additional factors: it relies on resource overhand being a *necessary* factor, but does not require it to be alone *sufficient* to accelerate human progress. There are doubtless many other factors that are relevant, such as a historical environment favorable to progress, a learning process that sufficiently pays attention to other members of ones species, not being a purely aquatic species, and so on. However, any full explanation of the acceleration in human progress of the form: 
"sudden progress happens exactly when (resource overhang) AND (X) AND (Y) AND (NOT Z) AND (W OR P OR NOT R) AND..." 
is still going to have the above implications for AI progress rates.

There's also an entire second half to the article, which discusses what human "misalignment" to inclusive genetic fitness (doesn't) mean for alignment, as well as the prospects for alignment during two specific fast takeoff (but not sharp left turn) scenarios, but that seems secondary to this discussion.

I think this post greatly misunderstands mine. 

Firstly, I'd like to address the question of epistemics. 

When I said "there's no reason to reference evolution at all when forecasting AI development rates", I was referring to two patterns of argument that I think are incorrect: (1) using the human sharp left turn as evidence for an AI sharp left turn, and (2) attempting to "rescue" human evolution as an informative analogy for other aspects of AI development.

(Note: I think Zvi did follow my argument for not drawing inferences about the odds of the sharp left turn specifically. I'm still starting by clarifying pattern 1 in order to set things up to better explain pattern 2.)

Pattern 1: using the human sharp left turn as evidence for an AI sharp left turn.

The original sharp left turn post claims that there are general factors about the structure and dynamics of optimization processes which both caused the evolutionary sharp left turn, and will go on to cause another sharp left turn in AI systems. The entire point of Nate referencing evolution is to provide evidence for these factors.

My counterclaim is that the causal processes responsible for the evolutionary sharp left turn are almost entirely distinct from anything present in AI development, and so the evolutionary outcome is basically irrelevant for thinking about AI. 

From my perspective, this is just how normal Bayesian reasoning works. If Nate says:

P(human SLT | general factors that cause SLTs) ~= 1

P(human SLT | NOT general factors that cause SLTs) ~= 0

then observing the human SLT is very strong evidence for there being general factors that cause SLTs in different contexts than evolution.

OTOH, I am saying:

P(human SLT | NOT general factors that cause SLTs) ~= 1

And so observing the human SLT is no evidence for such general factors.

Pattern 2: attempting to "rescue" human evolution as an informative analogy for other aspects of AI development.

When I explain my counterargument to pattern 1 to people in person, they will very often try to "rescue" evolution as a worthwhile analogy for thinking about AI development. E.g., they'll change the analogy so it's the programmers who are in a role comparable to evolution, rather than SGD. 

I claim that such attempted inferences also fail, for the same reason as argument pattern 1 above fails: the relevant portions of the causal graph driving evolutionary outcomes is extremely different from the causal graph driving AI outcomes, such that it's not useful to use evolution as evidence to make inferences about nodes in the AI outcomes causal graph. E.g., the causal factors that drive programmers to choose a given optimizer are very different from the factors that cause evolution to "choose" a given optimizer. Similarly, evolution is not a human organization that makes decisions based on causal factors that influence human organizations, so you should look at evolution for evidence of organization-level failures that might promote a sharp left turn in AI.

Making this point was the purpose of the "alien space clowns" / EVO-Inc example. It was intended to provide a concrete example of two superficially similar seeming situations, where actually their causal structures are completely distinct, such that there are no useful updates to make from EVO-Inc's outcomes to other automakers. When Zvi says:

I would also note that, if you discover (as in Quintin’s example of Evo-inc) that major corporations are going around using landmines as hubcaps, and that they indeed managed to gain dominant car market share and build the world’s most functional cars until recently, that is indeed a valuable piece of information about the world, and whether you should trust corporations or other humans to be able to make good choices, realize obvious dangers and build safe objects in general. Why would you think that such evidence should be ignored?

Zvi is proposing that there are common causal factors that led to the alien clowns producing dangerous cars, and could also play a similar role in causing other automakers to make unsafe vehicles, such that Evo-Inc's outcomes provide useful updates for predicting other automakers' outcomes. This is what I'm saying is false about evolution versus AI development.

At this point, I should preempt a potential confusion: it's not the case that AI development and human evolution share zero causal factors! To give a trivial example, both rely on the same physical laws. What prevents there being useful updates from evolution to AI development is the different structure of the causal graphs. When you update your estimates for the shared factors between the graphs using evidence from evolution, this leads to trivial or obvious implications for AI development, because the shared causal factors play different roles in the two graphs. You can have an entirely "benign" causal graph for AI development, which predicts zero alignment issues for AI development, yet when you build the differently structured causal graph for human evolution, it still predicts the same sharp left turn, despite some of the causal factors being shared between the graphs. 

This is why inferences from evolutionary outcomes to AI development don't work. Propagating belief updates through the evolution graph doesn't change any of the common variables away from settings which are benign in the AI development graph, since those settings already predict a sharp left turn when they're used in the evolution graph.

Concrete example 1: We know from AI development that having a more powerful optimizer, running for more steps, leads to more progress. Applying this causal factor to the AI development graph basically predicts "scaling laws will continue", which is just a continuation of the current trajectory. Applying the same factor to the evolution graph, combined with the evolution-specific fact of cultural transmission enabling a (relatively) sudden unleashing of ~9 OOM more effectively leveraged optimization power in a very short period of time, predicts an extremely sharp increase in the rate of progress.

Concrete example 2: One general hypothesis you could have about RL agents is "RL agents just do what they're trained to do, without any weirdness". (To be clear, I'm not endorsing this hypothesis. I think it's much closer to being true than most on LW, but still false.) In the context of AI development, this has pretty benign implications. In the context of evolution, due to the bi-level nature of its optimization process and the different data that different generations are "trained" on, this causal factor in the evolution graph predicts significant divergence between the behaviors of ancestral and modern humans.

Zvi says this is an uncommon standard of epistemics, for there to be no useful inferences from one set of observations (evolutionary outcomes) to another (AI outcomes). I completely disagree. For the vast majority of possible pairs of observations, there are not useful inferences to draw. The pattern of dust specks on my pillow is not a useful reference point for making inferences about the state of the North Korean nuclear weapons program. The relationship between AI development and human evolution is not exceptional in this regard.

Secondly, I'd like to address a common pattern in a lot of Zvi's criticisms. 

My post has a unifying argumentative structure that Zvi seems to almost completely miss. This leads to a very annoying dynamic where:

  • My post makes a claim / argument that serves a very specific role in the context of the larger structure.
  • Zvi misses that context, and interprets the claim / argument as making some broader claim about alignment in general.
  • Zvi complains that I'm over-claiming, being too general, or should split the post along the separate claims Zvi (falsely) believes I'm making.

The unifying argumentative structure of my post is as follows:

Having outlined my argumentative structure, I'll highlight some examples where Zvi's criticisms fall into the previously mentioned dynamic.


[Zvi] He then goes on to make another very broad claim.

[Zvi quoting me] > In order to experience a sharp left turn that arose due to the same mechanistic reasons as the sharp left turn of human evolution, an AI developer would have to:

[I list some ways one could produce an ML training process that's actually similar to human evolution in the relevant sense that would lead to an evolution-like sharp left turn at some point]

[Zvi criticizes the above list on the grounds that inner misalignment could occur under a much broader range of circumstances than I describe]

(I added the bolding)

The issue here is that the list in question is specifically for sharp left turns that arise "due to the same mechanistic reasons as the sharp left turn of human evolution", as I very specifically said in my original post. I'm not talking about inner alignment in general. I'm not even talking about sharp left turn threat scenarios in general! I'm talking very specifically about how the current AI paradigm would have to change before it had a mechanistic structure sufficiently similar to human evolution that I think a sharp left turn would occur "due to the same mechanistic reasons as the sharp left turn of human evolution".


As a general note, these sections seem mostly to be making a general alignment is easy, alignment-by-default claim, rather than being about what evolution offers evidence for, and I would have liked to see them presented as a distinct post given how big and central and complex and disputed is the claim here.

That is emphatically not what those sections are arguing for. The purpose of these sections is to describe two non-sharp left turn causing mechanisms for fast takeoff, in order to better illustrate that fast takeoff != sharp left turn. Each section specifically focuses on a particular mechanism of fast takeoff, and argues that said mechanism will not, in and of itself, lead to misalignment. You can still believe a fast takeoff driven by that mechanism will lead to misalignment for other reasons (e.g., a causal graph that looks like: "(fast takeoff mechanism) -> (capabilities) -> (something else) -> (misalignment)"), if, say, you think there's another causal mechanism driving misalignment, such that the fast takeoff mechanism's only contribution to misalignment was to advance capabilities in a manner that failed to address that other mechanism. 

These sections are not arguing about the ease of alignment in general, but about the consequence of one specific process.


The next section seems to argue that because alignment techniques work on a variety of existing training regimes all of similar capabilities level, we should expect alignment techniques to extend to future systems with greater capabilities.

That is, even more emphatically, not what that specific section is arguing for. This section focuses specifically on the "AIs do AI capabilities research" mechanism of fast takeoff, and argues that it will not itself cause misalignment. Its purpose is specific to the context in which I use it: to address the causal influence of (AIs do capabilities research) directly to (misalignment), not to argue about the odds of misalignment in general.

Further, the argument that section made wasn't:

because alignment techniques work on a variety of existing training regimes all of similar capabilities level, we should expect alignment techniques to extend to future systems

It was:

alignment techniques already generalize across human contributions to AI capability research. Let’s consider eight specific alignment techniques:

[list of alignment techniques]

and eleven recent capabilities advances:

[list of capabilities techniques]

I don’t expect catastrophic interference between any pair of these alignment techniques and capabilities advances.

And so, if you think AIs doing capabilities will be like humans doing capabilities research, but faster, then there will be a bunch of capabilities and alignment techniques, and the question is how much the capabilities techniques will interfere with the alignment techniques. Based on current data, the interference seems small and manageable. This is the trend being projected forwards, the lack of empirical interference between current capabilities and alignment (despite, as I note in my post, current capabilities techniques putting ~zero effort into not interfering with alignment techniques, an obviously dumb oversight which we haven't corrected because it turns out we don't even need to do so). 

Once again, I emphasize that this is not a general argument about alignment, which can be detached from the rest of the post. It's extremely specific to the mechanism for fast takeoff being analyzed, which is only being analyzed to further explore the connection between fast takeoff mechanisms and the odds of a sharp left turn. 


He closes by arguing that iteratively improving training data also exhibits important differences from cultural development, sufficient to ignore the evolutionary evidence as not meaningful in this context. I do not agree. Even if I did agree, I do not see how that would justify his broader optimism expressed here:

This part is a separate analysis of a different fast takeoff causal mechanism, arguing that it will not, itself cause misalignment either. Its purpose and structure mirrors that of the argument I clarified above, but focused on a different mechanism. It's not a continuation of a previous (non-existent) "alignment is easy in general" argument. 

Thirdly, I'd like to make some random additional commentary.

I would argue that ‘AIs contribute to AI capabilities research’ is highly analogous to ‘humans contribute to figuring out how to train other humans.’ And that ‘AIs seeking out new training data’ is highly analogous to ‘humans creating bespoke training data to use to train other people especially their children via culture’ which are exactly the mechanisms Quintin is describing humans as using to make a sharp left turn.

The degree of similarity is arguable. I think, and said in the original article, that similarity is low for the first mechanism and moderate for the second. 

However, the appropriate way to estimate the odds of a given fast takeoff mechanism leading to AI misalignment is not to estimate the similarity between that mechanism and what happened during human evolution, then assign misalignment risk to the mechanism in proportion to the estimated similarity. Rather, the correct approach is to build detailed causal models of how both human evolution and AI development work, propagate the evidence from human evolutionary outcomes back through your human evolution causal model to update relevant latent variables in that causal model, transfer those updates to any of the AI development causal model's latent variables which are also in the human evolution causal model, and finally estimate the new misalignment risk implied by the updated variables of the AI development model.

I discussed this in more detail in the first part of my comment, but whenever I do this, I find that the transfer from (observations of evolutionary outcomes) to (predictions about AI development) are pretty trivial or obvious, leading to such groundbreaking insights as:

  • More optimization power leads to faster progress
  • Human level general intelligence is possible
  • Neural architecture search is a bad thing to spend most of your compute on
  • Retraining a fresh instance of your architecture from scratch on different data will lead to different behavior

That seems like a sharp enough left turn to me.

A sharp left turn is more than just a fast takeoff. It's the combined sudden increase in AI generality and breaking of previously existing alignment properties.

...humans being clearly misaligned with genetic fitness is not evidence that we should expect such alignment issues in AIs. His argument (without diving into his earlier linked post) seems to be that humans are fresh instances trained on new data, so of course we expect different alignment and different behavior.

But if you believe that, you are saying that humans are fresh versions of the system. You are entirely throwing out from your definition of ‘the system’ all of the outer alignment and evolutionary data, entirely, saying it does not matter, that only the inner optimizer matters. In which case, yes, that does fully explain the differences. But the parallel here does not seem heartening. It is saying that the outcome is entirely dependent on the metaphorical inner optimizer, and what the system is aligned to will depend heavily on the details of the training data it is fed and the conditions under which it is trained, and what capabilities it has during that process, and so on. Then we will train new more capable systems in new ways with new data using new techniques, in an iterated way, in similar fashion. How should this make us feel better about the situation and its likely results?

I find this perspective baffling. Where else do the alignment properties of a system derive from? If you have a causal structure like

(programmers) -> (training data, training conditions, learning dynamics, etc) -> (alignment properties)

then setting the value of the middle node will of course screen off the causal influence of the (programmers) node. 

A possible clarification: in the context of my post when discussing evolution, "inner optimizer" means the brain's "base" optimization process, not the human values / intelligence that arises from that process. The mechanistically most similar thing in AI development to that meaning of the word "inner optimizer" is the "base" training process: the combination of training data, base optimizer, training process, architecture, etc. It doesn't mean the cognitive system that arises as a consequence of running that training process. 

Consider the counterfactual. If we had not seen a sharp left turn in evolution, civilization had taken millions of years to develop to this point with gradual steady capability gains, and we saw humans exhibiting strong conscious optimization mostly for their genetic fitness, it would seem crazy not to change our beliefs at all about what is to come compared to what we do observe. Thus, evidence.

I think Zvi is describing a ~impossible world. I think this world would basically break ~all my models on how optimizing processes gain capabilities. My new odds of an AI sharp left turn would depend on the new models I made in this world, which in turn would depend on unspecified details of how human civilization's / AI progress happens in this world.

I would also note that Quintin in my experience often cites parallels between humans and AIs as a reason to expect good outcomes from AI due to convergent outcomes, in circumstances where it would be easy to find many similar distinctions between the two cases. Here, although I disagree with his conclusions, I agree with him that the human case provides important evidence.

Once again, it's not the degree of similarity that determines what inferences are appropriate. It's the relative structure of the two causal graphs for the processes in question. The graphs for the human brain and current AI systems are obviously not the same, but they share latent variables that serve similar roles in determining outcomes, in a way that the bi-level structure of evolution's causal graph largely prevents. E.g., Steven Byrnes has a whole sequence which discusses the brain's learning process, and while there are lots of differences between the brain and current AI designs, there are also shared building blocks whose behaviors are driven by common causal factors. The key difference with evolution is that, once one updates the shared variables from looking at human brain outcomes and applies those updates to the AI development graph, there are non-trivial / obvious implications. Thus, one can draw relevant inferences by observing human outcomes.

Concrete example 1: brains use a local, non-gradient based optimization process to minimize predictive error, so there exists some non-SGD update rules that are competitive with SGD (on brainlike architectures, at least).

Concrete example 2: brains don't require GPT-4 level volumes of training data, so there exist architectures with vastly more data-friendly scaling laws than GPT-4's scaling.

In the generally strong comments to OP, Steven Byrnes notes that current LLM systems are incapable of autonomous learning, versus humans and AlphaZero which are, and that we should expect this ability in future LLMs at some point. Constitutional AI is not mentioned, but so far it has only been useful for alignment rather than capabilities, and Quintin suggests autonomous learning mostly relies upon a gap between generation and discernment in favor of discernment being easier. I think this is an important point, while noting that what matters is ability to discern between usefully outputs at all, rather than it being easier, which is an area where I keep trying to put my finger on writing down the key dynamics and so far falling short.

What I specifically said was:

Autonomous learning basically requires there to be a generator-discriminator gap in the domain in question, i.e., that the agent trying to improve its capabilities in said domain has to be better able to tell the difference between its own good and bad outputs.

I realize this is accidentally sounds like it's saying two things at once (that autonomous learning relies on the generator-discriminator gap of the domain, and then that it relies on the gap for the specific agent (or system in general)). To clarify, I think it's the agent's capabilities that matter, that the domain determines how likely the agent is to have a persistent gap between generation and discrimination, and I don't think the (basic) dynamics are too difficult to write down.

You start with a model M and initial data distribution D. You train M on D such that M is now a model of D. You can now sample from M, and those samples will (roughly) have whatever range of capabilities were to be found in D.

Now, suppose you have some classifier, C, which is able to usefully distinguish samples from M on the basis of that sample's specific level of capabilities. Note that C doesn't have to just be an ML model. It could be any process at all, including "ask a human", "interpret the sample as a computer program trying to solve some problem, run the program, and score the output", etc.

Having C allows you to sample from a version of M's output distribution that has been "updated" on C, by continuously sampling from M until a sample scores well on C. This lets you create a new dataset D', which you can then train M' on to produce a model of the updated distribution.

So long as C is able to provide classification scores which actually reflect a higher level of capabilities among the samples from M / M' / M'' / etc, you can repeat this process to continually crank up the capabilities. If your classifier C was some finetune of M, then you can even create a new C' off of M', and potentially improve the classifier along with your generator. In most domains though, classifier scores will eventually begin to diverge from the qualities that actually make an output good / high capability, and you'll eventually stop benefiting from this process.

This process goes further in domains where it's easier to distinguish generations by their quality. Chess / other board games are extreme outliers in this regard, since you can always tell which of two players actually won the game. Thus, the game rules act as a (pairwise) infallible classifier of relative capabilities. There's some slight complexity around that last point, since a given trajectory could falsely appear good by beating an even worse / non-representative policy, but modern self-play approaches address such issues by testing model versions against a variety of opponents (mostly past versions of themselves) to ensure continual real progress. Pure math proofs is another similarly skewed domain, where building a robust verifier (i.e., a classifier) of proofs is easy. That's why Steven was able to use it as a valid example of where self-play gets you very far.

Most important real world domains do not work like this. E.g., if there were a robust, easy-to-query process that could classify which of two scientific theories / engineering designs / military strategies / etc was actually better, the world would look extremely different.

There are other issues I have with this post, but my reply is already longer than the entire original post, so I'll stop here, rather than, say, adding an entire additional section on my models of takeoff speed for AIs versus evolution (which I'll admit probably should have another post to go with it).

Addressing this objection is why I emphasized the relatively low information content that architecture / optimizers provide for minds, as compared to training data. We've gotten very far in instantiating human-like behaviors by training networks on human-like data. I'm saying the primacy of data for determining minds means you can get surprisingly close in mindspace, as compared to if you thought architecture / optimizer / etc were the most important.

Obviously, there are still huge gaps between the sorts of data that an LLM is trained on versus the implicit loss functions human brains actually minimize, so it's kind of surprising we've even gotten this far. The implication I'm pointing to is that it's feasible to get really close to human minds along important dimensions related to values and behaviors, even without replicating all the quirks of human mental architecture.

I believe the human visual cortex is actually the more relevant comparison point for estimating the level of danger we face due to mesaoptimization. Its training process is more similar to the self-supervised / offline way in which we train (base) LLMs. In contrast, 'most abstract / "psychological"' are more entangled in future decision-making. They're more "online", with greater ability to influence their future training data.

I think it's not too controversial that online learning processes can have self-reinforcing loops in them. Crucially however, such loops rely on being able to influence the externally visible data collection process, rather than being invisibly baked into the prior. They are thus much more amenable to being addressed with scalable oversight approaches.

Load More