Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

Deconfusing Direct vs Amortised Optimization

8Charlie Steiner

4Charlie Steiner

5habryka

2DragonGod

1Htarlov

5leogao

1Noosphere89

2DragonGod

2DragonGod

1martinkunev

1Oliver Sourbut

3beren

1Roman Leventov

1Edoardo Pona

1RobertKirk

1Adam Jermyn

1Jérémy Scheurer

New Comment

17 comments, sorted by Click to highlight new comments since: Today at 8:47 AM

Coming back 4 months later, I'm confused why I didn't compare this to Abram's post Selection vs Control.

Direct optimizers typically have a very specific architecture requiring substantial iteration and search. Luckily, it appears that our current NN architectures, with a fixed-length forward pass and a lack of recurrence or support for branching computations as is required in tree search makes the implementation of powerful mesa-optimizers inside the network quite challenging.

I think this is being too confident on what "direct optimizers" require.

There is an ontology, mostly inherited from the graph-search context, in which "direct optimizers" require recurrence and iteration, but at least I don't have any particularly strong beliefs about what a direct optimizer needs in terms of architecture, and don't think other people know either. The space of feed-forward networks is surprisingly large and rich and I am definitely not confident you can't find a direct optimizer in that space.

Current LLMs also get quite high scores at imitating pretty agentic and optimizy humans, which suggest the networks do perform something quite close to search or direct optimization somewhere within it's forward pass.

Current LLMs also get quite high scores at imitating pretty agentic and optimizy humans, which suggest the networks do perform something quite close to search or direct optimization somewhere within it's forward pass.

I don't think this is necessarily true? Given the sheer scale of training, it seems plausible that they learned heuristics that cache out to goal directed behaviour, without actually performing any search towards realising particular goals.

Alternatively, model free RL agents may behave as if they are competently pursuing a particular goal, without directly optimising for said goal (hence the phenomenon of "goal misgeneralisation").

The policy they learned was selected for maximising reward during training, and so cashes out in empirical behaviour that looks goal directed (appears to maximise reward on the training distribution), but this does not imply that internally the model is running some search algorithm; it's just evaluating the learned policy.

It's not the case that agent like behaviour implies agent like internal architecture; at least that's a major claim of this post as I understand it.

I think current LLM have recurrence as the generated tokens are input to the next pass of the DNN.

From observations I see that they work better on tasks of planning, inference or even writing the program code if they start off with step by step "thinking out loud" explaining steps of the plan, of inference or of details of code to write. If you ask GPT-4 for something not trivial and substantially different from code that can be found in public repositories it will tend to write plan first. If you ask it in different thread to make the code only without description, then usually first solution with a bit of planning is better and less erroneous than the second one. They also work much better if you specify simple steps to translate into code instead of more abstract description (in case of writing code without planing). This suggests LLM don't have ability to internally span long tree of possibilities to check - like a direct agent - but they can use recurrence of token output-input to do some similar work.

The biggest difference here that I see is that:

- direct optimization processes are ultra-fast in searching the tree of possibilities but not fast computationally to discriminate good enough solutions from worse ones
- on the other hand amortised optimizers are very slow if direct search is needed. Maybe right now with GPT-4 a bit faster than regular human, but also a bit more erroneous, especially in more complex inference.
- amortised optimizers are faster for quickly finding good enough solution by some generalized heuristics, without need for direct search (or only small amount of it on higher abstraction level)
- amortised optimizers like LLMs can group steps or outcomes into more abstract groups like humans do and work on those groups instead of direct every possible action and outcome

What I'm more worried about is more close hybridization between direct and amortised optimizers. I can imagine architecture where there is a direct optimizer but instead of generating and searching impossibly vast tree of possibilities it would use a DNN model for generation of less options. Like instead of generating thosands detailed moves like "move 5 meters", "take that thing", "put it there" and optimize over that, generate more abstract plan points specified by LLM with predictions of that step outcome and then evaluate how that outcome works for the goal. This way it could plan on more abstract level like humans to narrow down general plan or list of partial goals that lead to "final goal" or to "best path" (if it's value function is more like integral over time instead of one final target). Find a good strategy. With enough time - it might be even a complex and indirect one. Then it could plan tactics for first step in the same way but on the lower abstraction. Then plan direct move step to realise first step of current tactics and run it. It might have several subprocesses that asynchronously work out strategy based on general state and goal, current tactics based on more detailed state and current strategical goal to pursue, current moves based on current tactical plan. With any numbers of abstraction and detail levels (2-3 seems like typical for humans, but AI might have more). This kind of agent might behave more like direct optimizer, even if using LLM and DNN inside for some parts. Direct optimization would have a first seat behind steering wheel in such agent.

I don't think this will be outcome of research at OpenAI or other such laboratories any time soon. It might be, but if I would guess then I think it would be rather LLM or other DNN model "on top" that is connected to other models to "use at will". For example it is rather easy to connect GPT-4 so it could use other models or APIs (like database, search). So this is very low hanging fruit for current AI development. I see that next step will be connecting it to more modalities and other models. It is currently going on.

I think though, this more direct agent might be the outcome of works done by military. Direct approach is much more reliable and reliability is one of the top key values for military-grade equipment. I only hope they will take the danger of such approach seriously.

I think of mesaoptimization as primarily being concerning because it would mean models (selected using amortized optimization) doing their own direct optimization, and the extent to which the model is itself doing its own "direct" optimization vs just being "amortized" is what I would call the optimizer-controller spectrum (see this post also).

Also, it seems kind of inaccurate to declare that (non-RL) ML systems are fundamentally amortized optimization and then to say things like "more computation and better algorithms should *improve* safety and the primary risk comes from *misgeneralization*" and "amortized approaches necessarily have poor sample efficiency asymptotically" and only adding in a mention about mesaoptimizers in a postscript.

In my ontology, this corresponds to saying "current ML systems are very controller-y." But the thing I'm worried about is that eventually at some point we're going to figure out how to have models which in fact are more optimizer-y, for the same reasons people are trying to build AGI in the first place (though I do think there is a non-trivial chance that controller-y systems are in fact good enough to help us solve alignment, this is not something I'd bet on as a default!).

Relatedly, it doesn't seem like because amortized optimizers are "just" modelling a distribution, this makes them inherently more benign. Everything can be phrased as distribution modelling! I think the core confusion here might be conflating the generalization properties of current ML architectures/methods with the type signature of ML. Moving the issue to data quality also doesn't fix the problem at all; everything is "just" a data problem, too (if only we had the dataset indicating exactly how the superintelligent AGI should solve alignment, then we could simply behavior clone that dataset).

Everything can be phrased as distribution modelling.

That might a big claim, since Beren thinks there's a real difference in type, and one example is that he thinks alignment solutions for model based agents coming out of GPT-N can't work, due to amortized optimization. Thus, a non-vacous restriction is there.

The distinction amortized vs direct in humans seems related to system-1 vs system-2 in Thinking Fast and Slow.

"the implementation of powerful mesa-optimizers inside the network quite challenging"

I think it's quite likely that we see optimizers implemented outside the network in the style of AutoGPT (people can explicitly build direct optimizers on top of amortized ones).

I'm really glad you wrote this. I've thought for some time that it's an important distinction, though I think you've articulated (at least parts of) it better than my attempts perhaps! I previously described a distinction between deliberation and reaction.

mesaoptimizers that could form

across multiple forward passes

Yes, leaving aside really really deep networks and residuals, I think some sort of recurrence/iteration is plausibly needed for meaningful deliberation to occur. Chain of thought is an obvious instantiation (but so is sequential reasoning absent explicit CoT prompting), with MCTS examples (which you also mentioned) being perhaps more central.

I'll gesture at some pieces of this puzzle which I haven't got round to writing about properly publicly^{[1]} but where I'd be interested in your thoughts:

- Where do highly capable proposals/amortised actions come from?

- (handwave) lots of 'experience' and 'good generalisation'?
- How do you get good 'experience'?
- Helpful human curates massive dataset for you
- Or...? This seems to be to be where active learning and deliberate/creative exploration come in
- It's a Bayes-adaptivity problem, i.e. planning for value-of-information
- This is basically what 'science' and 'experimentalism' are in my ontology
- 'Play' and 'practice' are the amortised equivalent (where explorative heuristics are baked in)
- animals are evidence that some amortised play heuristics are effective! Even humans only rarely 'actually do deliberate experimentalism'
- but when we do, it's maybe the source of our massive technological dominance?

- Either that or you have something terribly slow like natural selection

- How do you get good 'generalisation'?
- (handwave) hierarchical/recomposable abstractions?
- This is the most magical part of the picture to me at present

- (handwave) hierarchical/recomposable abstractions?

- When is deliberation/direct planning tractable?

- In any interestingly-large problem, you will
*never*exhaustively evaluate- e.g. maybe no physically realisable computer in our world can ever evaluate all Go strategies, much less evaluating strategies for 'operate in the world itself'!

- What properties of options/proposals lend themselves?
- (handwave) 'Interestingly consequential' - the differences should actually matter enough to bother computing!
- Temporally flexible
- The 'temporal resolution' of the strategy-value landscape may vary by orders of magnitude
- so the temporal resolution of the proposals (or proposal-atoms) should too, on pain of intractability/value-loss/both

- Where does strong control/optimisation come from?

- Your strategy is inevitably suboptimal
- You can have a good shot
- but keeping 'yourself' around to deliberate again or more generally making there be more similarly-oriented deliberators is a great way to handle this

Maybe the closest is this scrappy comment? ↩︎

Hi there! Thanks for this comment. Here are my thoughts:

- Where do highly capable proposals/amortised actions come from?

- (handwave) lots of 'experience' and 'good generalisation'?

Pretty much this. We know empirically that deep learning generalizes pretty well from a lot of data as long as it is reasonable representative. I think that fundamentally this is due to the nature of our reality that there are generalizable patterns which is ultimately due to the sparse underlying causal graph. It is very possible that there are realities where this isn't true and in those cases this kind of 'intelligence' would not be possible.

r...? This seems to be to be where active learning and deliberate/creative exploration come in

- It's a Bayes-adaptivity problem, i.e. planning for value-of-information
- This is basically what 'science' and 'experimentalism' are in my ontology

- 'Play' and 'practice' are the amortised equivalent (where explorative heuristics are baked in)

Again, I completely agree here. In practice in large environments it is necessary to explore if you can't reach all useful states from a random policy. In these cases, it is very useful to a.) have an explicit world model so you can learn from sensory information which is much higher bandwidth than reward usually and generalizes further and in an uncorrelated way, and b.) do some kind of active exploration. Exploring according to maximizing info-gain is probably close to optimal, although whether this is actually theoretically optimal is I tihnk still an open question. The main issue is that info-gain is hard to cmopute/approximate tractably, since it requires keeping a close track of your uncertainty, and DL models are computationally tractable by explicitly throwing away all the uncertainty and only really maintaining point predictions.

animals are evidence that some amortised play heuristics are effective! Even humans only rarely 'actually do deliberate experimentalism'

- but when we do, it's maybe the source of our massive technological dominance?

Like I don't know to what extent there are 'play heuristics' at a behavioural level vs some kind of intrinsic drive for novelty / information gain but yes, having these drives 'added to your reward function' is generally useful in RL settings and we know this happens in the brain as well -- i.e. there are dopamine neurons responsive to proxies of information gain (and exactly equal to information gain in simple bandit-like settings where this is tractable)

- When is deliberation/direct planning tractable?

- In any interestingly-large problem, you will
neverexhaustively evaluate

- e.g. maybe no physically realisable computer in our world can ever evaluate all Go strategies, much less evaluating strategies for 'operate in the world itself'!
- What properties of options/proposals lend themselves?

- (handwave) 'Interestingly consequential' - the differences should actually matter enough to bother computing!
- Temporally flexible

- The 'temporal resolution' of the strategy-value landscape may vary by orders of magnitude
- so the temporal resolution of the proposals (or proposal-atoms) should too, on pain of intractability/value-loss/both

So there are a number of circumstances where direct planning is valuable and useful. I agree about your conditions and especially the correct action step-size as well as discrete actions and known not super stochastic dynamics. Other useful conditions are when it's easy to evaluate the branches of the tree without having gone all the way down to the leaves -- i.e. in games like Chess/GO it's often very easy to know that some move tree is intrinsically doomed without having explored all of it. This is a kind of convexity to the state space (not literally mathematically, but intuitively) which makes optimization much easier. Similarly, when good proposals can be made due to linearity / generalizability in the action space it is easy to prune actions and trees.

- Where does strong control/optimisation come from?

Strong control comes from where strong learning in general comes from -- lots of compute and data -- and for planning especially compute. The optimal trade-off between amortized and direct optimization given a fixed compute budget is super interesting and I don't think we have any good models of this yet.

Another thing that I think is fairly underestimated among people on LW compared to people doing deep RL is that open-loop planning is actually very hard and bad at dealing with long time horizons. This is basically due to stochasticity and chaos theory -- future prediction is hard. Small mistakes in either modelling or action propagate very rapidly to create massive uncertainties about the future so that your optimal posterior rapidly dwindles to a maximum entropy distribution. The key thing in long term planning is really adaptability and closed-loop control -- i.e. seeing feedback and adjusting your actions in response to feedback. This is how almost all practical control systems actually work and in practice in deep RL with planning everybody actually uses MPC so replans every step.

Postscript 1: Are humans direct or amortized optimizers? There is actually a large literature in cognitive science which studies this exact question, although typically under the nomenclature of model-based vs model-free reinforcement learners. The answer appears to be that humans are both.

The Active Inference framework agrees: e.g., see Constant et al., 2021 (https://www.frontiersin.org/articles/10.3389/fpsyg.2020.598733/full), the distinction between direct and amortised optimisation manifests as planning-as-inference vs. so-called "deontic action":

Deontic actions are actions for which the underlying policy has acquired a deontic value; namely, the shared, or socially admitted value of a policy (Constant et al., 2019). A deontic action is guided by the consideration of “what would a typical other do in my situation.” For instance, stopping at the red traffic light at 4 am when no one is present may be viewed as such a deontically afforded action.

This also roughly corresponds to the distinction between representationalism and dynamicism.

Another way to understand this distinction is to think about the limit of infinite compute, where direct and amortized optimizers converge to different solutions. A direct optimizer, in the infinite compute limit, will simply find the optimal solution to the problem. An amortized optimizer would find the Bayes-optimal posterior over the solution space given its input data.

Doesn't this depend on how, practically, the compute is used? Would this still used in a setting where the compute can be used to generate the dataset?

Taking the example of training an RL agent, an increase in compute will mostly (and more effectively) be used towards getting the agent more experience rather than using a larger network or model. 'Infinite' compute would be used to generate a correspondingly 'infinite' dataset, rather than optimising the model 'infinitely' well on some fixed experience.

In this case, while it remains true that the amortised optimiser will act closely following the data distribution it was trained on, this will include all outcomes - just like in the case of the direct optimiser, unless the data generation policy (ie. exploration) doesn't preclude visiting some of the states. So as long as the unwanted solutions that would be generated by a direct optimiser find a way in the dataset, the amortised agent will learn to produce them.

In the limit of compute, to have an (amortised) agent with predictable behaviour you would have to analyse / prune this dataset. Wouldn't that be as costly and hard as pruning the search of a direct optimiser of undesired outcomes?

An existing example of something like the difference between amortised and direct optimisation is doing RLHF (w/o KL penalties to make the comparison exact) vs doing rejection sampling (RS) with a trained reward model. RLHF amortises the cost of directly finding good outputs according to the reward model, such that at evaluation the model can produce good outputs with a single generation, whereas RS requires no training on top of the reward model, but uses lots more compute at evaluation by generating and filtering with the RM. (This case doesn't exactly match the description in the post as we're using RL in the amortised optimisation rather than SL. This could be adjusted by gathering data with RS, and then doing supervised fine-tuning on that RS data, and seeing how that compares to RS).

Given we have these two types of optimisation, I think two key things to consider are how each type of optimisation interacts with Goodhart's Law, and how they both generalise (kind of analogous to outer/inner alignment, etc.):

- The work on overoptimisation scaling laws in this setting shows that, at least on distribution, there does seem to be a meaningful difference to the over-optimisation behaviour between the two types of optimisation - as shown by the different functional forms for RS vs RLHF.
- I think the generalisation point is most relevant when we consider that the optimisation process used (either in direct optimisation to find solutions, or in amortised optimisation to produce the dataset to amortise) may not generalise perfectly. In the setting above, this corresponds to the reward model not generalising perfectly. It would be interesting to see a similar investigation as the overoptimisation work but for generalisation properties - how does the generalisation of the RLHF policy relate to the generalisation of the RM, and similarly to the RS policy? Of course, over-optimisation and generalisation probably interact, so it may be difficult to disentangle whether poor performance under distribution shift is due to over-optimisation or misgeneralisation, unless we have a gold RM that also generalises perfectly.

How would you classify optimization shaped like "write a program to solve the problem for you". It's not directly searching over solutions (though the program you write might). Maybe it's a form of amortized optimization?

Separately: The optimization-type distinction clarifies a circle I've run around talking about inner optimization with many people, namely "Is optimization the same as search, or is search just one way to get optimization?" And I think this distinction gives me something to point to in saying "search is one way to get (direct) optimization, but there are other kinds of optimization".

This means that, at least in theory, the out of distribution behaviour of amortized agents can be precisely characterized even before deployment, and is likely to concentrate around previous behaviour. Moreover, the out of distribution generalization capabilities should scale in a predictable way with the capacity of the function approximator, of which we now have precise mathematical characterizations due to scaling laws.

Do you have pointers that explain this part better? I understand that scaling computing and data will improve misgeneralization to some degree (i.e. reduce it). But what is the reasoning why misgeneralization should be predictable, given the capacity and the knowledge of "in-distribution scaling laws"?

Overall I hold the same opinion, that intuitively this should be possible. But empirically I'm not sure whether in-distribution scaling laws can tell us anything about out-of-distribution scaling laws. Surely we can predict that with increasing model & data scale the out-of-distribution misgeneralization will go down. But given that we can't really quantify all the possible out-of-distribution datasets, it's hard to make any claims about how precisely it will go down.

This post is part of the work done atConjecture.An earlier version of this post was postedhere.Many thanks go to Eric Winsor, Daniel Braun, Chris Scammell, and Sid Black who offered feedback on this post.TLDR:We present a distinction from the Bayesian/variational inference literature of direct vs amortized optimization. Direct optimizers apply optimization power to argmax some specific loss or reward function. Amortized optimizers instead try to learn a mapping between inputs and output solutions and essentially optimize for the posterior over such potential functions. In an RL context, direct optimizers can be thought of as AIXI-like planners which explicitly select actions by assessing the utility of specific trajectories. Amortized optimizers correspond to model-free RL methods such as Q learning or policy gradients which use reward functions only as a source of updates to an amortized policy/Q-function. These different types of optimizers likely have distinct alignment properties: ‘Classical’ alignment work focuses on difficulties of aligning AIXI-like direct optimizers. The intuitions ofshard theoryare built around describing amortized optimizers. We argue that AGI, like humans, will probably be comprised of some combination of direct and amortized optimizers due to the intrinsic computational efficiency and benefits of the combination.Here, I want to present a new frame on different types of optimization, with the goal of helping deconfuse some of the discussions in AI safety around questions like whether RL agents

directly optimize for reward, and whether generative models (i.e.simulators) are likely to develop agency. The key distinction I want to make is betweendirectandamortizedoptimization.Directoptimization is what AI safety people, following from Eliezer’s early depictions, often envisage an AGI as primarily being engaged in. Direct optimization occurs when optimization power is applied immediately and directly when engaged with a new situation to explicitly compute an on-the-fly optimal response – for instance, when directly optimizing against some kind of reward function. The classic example of this is planning and Monte-Carlo-Tree-Search (MCTS) algorithms where, given a situation, the agent will unroll the tree of all possible moves to varying depth and then directly optimize for the best action in this tree. Crucially, this tree is constructed 'on the fly' during the decision of a single move. Effectively unlimited optimization power can be brought to play here since, with enough compute and time, the tree can be searched to any depth.Amortizedoptimization, on the other hand, is not directly applied to any specific problem or state. Instead, an agent is given a dataset of input data and successful solutions, and then learns a function approximator that maps directly from the input data to the correct solution. Once this function approximator is learnt, solving a novel problem then looks like using the function approximator to generalize across solution space rather than directly solving the problem. The term amortized comes from the notion ofamortized inference, where the 'solutions' the function approximator learns are the correct parameters of the posterior distribution. The idea is that, while amassing this dataset of correct solutions and learning function approximator over it is more expensive, once it is learnt, the cost of a new 'inference' is very cheap. Hence, if you do enough inferences, you can 'amortize' the cost of creating the dataset.Mathematically,

directoptimization is your standard AIXI-like optimization process. For instance, suppose we are doing direct variational inference optimization to find a Bayesian posterior parameter θ from a data-point x, the mathematical representation of this is:θ∗direct=argminθKL[q(θ;x)||p(x,θ)]

By contrast, the amortized objective optimizes some other set of parameters $\phi$ over a function approximator ^θ=fϕ(x) which directly maps from the data-point to an estimate of the posterior parameters ^θ. We then optimize the parameters of the function approximator ϕ across a whole dataset D={(x1,θ∗1),(x2,θ∗2)…} of data-point and parameter examples.

θ∗amortized=argminϕEp(D)[L(θ∗,fϕ(x))]

WhereL is the amortized loss function.

Amortized optimization has two major practical advantages. Firstly, it converts an inference or optimization problem into a supervised learning problem. Inference is often very challenging for current algorithms, especially in unbounded domains (and I suspect this is a general feature of computational complexity theory and unlikely to be just solved with a clever algorithm), while we know how to do supervised learning very well. Secondly, as the name suggests, amortized optimization is often much cheaper at runtime, since all that is needed is a forward pass through the function approximator rather than an explicit solution of an optimization problem for each novel data-point.

The key challenge of amortized optimization is in obtaining a dataset of solutions in the first place. In the case of supervised learning, we assume we have class labels which can be interpreted as the 'ideal' posterior over the class. In unsupervised learning, we define some proxy task that generates such class labels for us, such as autoregressive decoding. In reinforcement learning, this is more challenging, and instead we must use proxy measures such as the Bellman update for temporal difference based approaches (with underappreciated and often unfortunate

consequences), or mathematical tricks to let us estimate the gradient without ever computing the solution as in policy gradients, which often comes at the cost of high variance.Another way to understand this distinction is to think about the limit of infinite compute, where direct and amortized optimizers converge to different solutions. A direct optimizer, in the infinite compute limit, will simply find the optimal solution to the problem. An amortized optimizer would find the Bayes-optimal posterior over the solution space given its input data.

Almost all of contemporary machine learning, and especially generative modelling, takes place within the paradigm of

amortizedoptimization, to the point that, for someone steeped in machine learning, it can be hard to realize that other approaches exist. Essentially all supervised learning is amortized: 'inference' in a neural network is performed in a forward pass which directly maps the data to the parameters of a probability distribution (typically logits of a categorical) over a class label^{[1]}[^1]. In reinforcement learning, where direct optimization is still used, thedistinction is closely related tomodel-free (amortized) vs model-based (direct) methods. Model-free methods learn an (amortized) parametrized value function or policy -- i.e. use a neural network to map from observations to either values or actions. Model-based methods on the other hand typically perform planning or model-predictive-control (MPC) which involves direct optimization over actions at each time-step of the environment. In general, research has found that while extremely effective in narrow and known domains such as board games, direct optimization appears to struggle substantially more in domains with very large state and action spaces (and hence tree branching width), as well as domains with significant amounts of stochasticity and partial observability, since planning under belief states is vastly more computationally taxing than working with the true MDP state. Planning also struggles in continuous-action domains where MCTS cannot really be applied and there are not really any good continuous-action planning algorithms yet known^{[2]}.With recent work, however, this gap is closing and direct optimization, typically (and confusingly) referred to as

RL ismodel basedcatchingupto amortized. However, all of these methods almost always use some combination of both direct and amortized approaches. Typically, what you do is learn an amortized policy or value function and then use the amortized prediction to initialize the direct optimizer (which is typically a planner). This has the advantage of starting off the planner with a good initialization around what are likely to be decent actions already. In MCTS you can also short circuit the estimation of the value of an MCTS node by using the amortized value function as your estimate. Thesetechniques vastly improve the efficiency of direct optimization and are widely used. For instance, alpha-go and efficient-zero both make heavy use of amortization in these exact ways despite their cores of direct optimization.hybridRelevance for alignmentThe reason that it is important to carefully understand this distinction is that direct and amortized optimization methods seem likely to differ substantially in their safety properties and capabilities for alignment. A direct optimizer such as AIXI or any MCTS planner can, with enough compute, exhibit behaviour that diverges arbitrarily from its previous behaviour. The primary constraint upon its intelligence is the compute and time needed to crunch through an exponentially growing search tree. The out-of-distribution capabilities of an amortized agent, however, depend entirely on the generalization capabilities of the underlying function approximator used to perform the amortization. In the case of current neural networks, these almost certainly cannot accurately generalize arbitrarily far outside of their training distribution, and there are indeed good reasons for suspecting that this is a general limitation of function approximation

^{[3]}. A secondary key limitation of the capabilities of an amortized agent is in its training data (since an amortized method effectively learns the probability distribution of solutions) and hence amortized approaches necessarily have poor sample efficiency asymptotically compared to direct optimizers which theoretically need very little data to attain superhuman capabilities. For instance, a chess MCTS program needs nothing but the rules of chess and a very large amount of compute to achieve arbitrarily good performance while an amortized chess agent would have to see millions of games.Moreover, the way scaling with compute occurs in amortized vs direct optimization seems likely to differ. Amortized optimization is fundamentally about modelling a probability distribution given some dataset. The optimal outcome here is simply the exact Bayesian posterior, and additional compute will simply be absorbed in better modelling of this posterior. If, due to the nature of the dataset, this posterior does not assign significant probability mass to unsafe outcomes, then in fact more computation and better algorithms should

improvesafety and the primary risk comes frommisgeneralization-- i.e. erroneously assigning probability mass to dangerous behaviours which are not as likely as in the true posterior. Moreover amortized optimizers are just generative models, it is highly likely that all amortized optimizers obey the same power-law scaling we observe in current generative modelling which means sharply diminishing (power law) returns on additional compute and data investment. This means that, at least in theory, the out of distribution behaviour of amortized agents can be precisely characterized even before deployment, and is likely to concentrate around previous behaviour. Moreover, the out of distribution generalization capabilities should scale in a predictable way with the capacity of the function approximator, of which we now have precise mathematical characterizations due to scaling laws.The distinction between direct and amortized optimizers also clarifies what I think is the major conceptual distinction between perspectives such as shard theory vs 'classical' AGI models such as AIXI. Shard theory, and

related worksare primarily based around describing whatamortizedagents look like. Amortized agents do not explicitly optimize for rewards but rather repeat and generalize behaviours at test-time that led to reward in the past (for direct optimizers, however, reward is very much the optimization target). All the optimization occurred during training when a policy was learnt that attempted to maximize reward given a set of empirically known transitions. When this policy is applied to novel situations, however, the function approximator is not explicitly optimizing for reward, but instead just generalizing across 'what the policy would do'. Thus, a key implicit claim of shard theory is that the AGIs that we build will end up looking much more like current model-free RL agents than planners like alpha-go and, ultimately, AIXI. Personally, I think something like this is quite likely due to the intrinsic computational and ontological difficulties with model-based planning in open-ended worlds which I will develop in a future post.For alignment, my key contention is that we should be very aware of whether we are thinking of AGI systems as

directoramortizedoptimizers or some combination of both. Such systems would have potentially very different safety properties. Yudkowsky's vision is essentially of a direct optimizer of unbounded power. For such an agent, indeed the only thing that matters and that we can control is its reward function, so alignment must focus entirely on the design of the reward function to be safe, corrigible etc. For amortized agents, however, the alignment problem looks very different. Here, while the design of the reward function is important, so too is the design of thedataset. Moreover, it seems likely that for such amortized agents we are much less likely to see sudden capability jumps with very little data, and so they are likely much safer overall. Such amortized agents are also much closer to the cognitive architecture of humans, which do not have fixed utility functions nor unbounded planning ability. It is therefore possible that we might be able to imprint upon them a general fuzzy notion of 'human values' in a way we cannot do with direct optimizers.The fundamental question, then, is figuring out what is the likely shape of near-term AGI so we can adapt our alignment focus to it. Personally, I think that a primarily amortized hybrid architecture is most likely, since the computational advantages of amortization are so large, and that this appears to be how humans operate as well. However, epistemically, I am still highly uncertain on this point and things will clarify as we get closer to AGI.

Postscript 1: Are humans direct or amortized optimizers?There is actually a

largeliteratureincognitive sciencewhich studies this exact question, although typically under the nomenclature of model-based vs model-free reinforcement learners. The answer appears to be that humans are both. When performing tasks that are familiar, or when not concentrating, or under significant mental load (typically having to do multiple disparate tasks simultaneously), humans respond in an almost entirely amortized fashion. However, when faced with challenging novel tasks and have mental energy, humans are also capable of model-based planning like behaviour. These results heavily accord with (at least my) phenomenology, where usually we act 'on autopilot' however when we really want something we are capable of marshalling a significant amount of direct optimization power against a problem^{[4]}^{[5]}. Such a cognitive architecture makes sense for an agent with limited computational resources and, as discussed, such hybrid architectures are increasingly common in machine learning as well, at least for those that actually use direct optimization. However, while current approaches have a fixed architecture where amortization is always used in specific ways (i.e. to initialize a policy or to make value function estimates), humans appear to be able to flexibly shift between amortized and direct optimization according to task demands, novelty, and level of mental load.Postscript 2: MesaoptimizersMy epistemic status on this is fairly uncertain, but I think that this framing also gives some novel perspective on the question of mesaoptimization raised in the

post. Using our terminology, we can understand the danger and likelihood of mesaoptimizers as the question of whether performing amortized optimization will tend to instantiate direct optimizers somewhere within the function approximator. The idea being, that for some problems, the best way to obtain the correct posterior parameters is to actually directly solve the problem using direct optimization. This is definitely a possibility, and may mean that we can unsuspectingly obtain direct optimizers from what appear to be amortized and hence we may overestimate the safety and underestimate the generalizability of our systems. I am pretty uncertain, but it feels sensible to me that the main constraint on mesa-optimization occuring is to what extent the architecture of the mapping function is conducive to the implementation of direct optimizers. Direct optimizers typically have a very specific architecture requiring substantial iteration and search. Luckily, it appears that our current NN architectures, with a fixed-length forward pass and a lack of recurrence or support for branching computations as is required in tree search makes the implementation of powerful mesa-optimizers inside the network quite challenging. However, this assessment may change as we scale up networks or continue to improve their architectures towards AGI.risks from learned optimizationTo begin to understand the risk of mesaoptimization, it is important to make a distinction between

within-forward-passmesa-optimizers, and mesaoptimizers that could formacross multiple forward passes. A mesaoptimizer within the forward pass would form when in order to implement some desired functionality, a direct optimizer has to be implemented within the amortized function. An early example of this sort potentially happening is the recent discovery that language models can learn to perform arbitrary linear regression problems (-/cites), and hence potentially implement an internal iterative algorithm like gradient descent. Another possibility is that mesaoptimizers could form across multiple linked forward passes, when the forward passes can pass information to later ones. An example of this occurs in autoregressive decoding in language models, where earlier forward passes write output tokens which are then fed into later forward passes as inputs. It would be possible for a mesaoptimizer to form across forward passes by steganographically encoding additional information into the output tokens to be reused in later computations. While probably challenging to form in the standard autoregressive decoding task, such information transmission is almost entirely the point of other ideas like giving models access to ‘scratch-pads’ or external memory.^{^}People have also experimented with direct optimization even in perceptual tasks and found, unsurprisingly, that it

canimproveperformance(although it may not be worth the additional compute cost).^{^}This is somewhere I suspect that we are not bottlenecked by fundamental reasons, but that we simply haven't found the right algorithms or conceptualized things in the right way. I suspect that the underlying algorithmic truth is that continuous action spaces are harder by a constant factor, but not in qualitative terms the way it is now.

^{^}Mathematically speaking, this is not actually a limitation but the desired behaviour. The goal of an amortized agent is to learn a posterior distribution over the solution 'dataset'. More compute will simply result in better approximating this posterior. However, this posterior models the

datasetand not the true solution space. If the dataset is not likely to contain the true solution to a problem, perhaps because it requires capabilities far in excess of any of the other solutions, it will not have high probability in the true posterior. This is exactly why alignment ideas like 'ask GPT-N for a textbook from 2100 describing the solution to the alignment problem' cannot work, even in theory. GPT is trained to estimate the posterior of sequencesover its current dataset(the internet text corpus as of 2022 (or whenever)). The probability of such a dataset containing a true textbook from 2100 with a solution to alignment is 0. What does have probability is a bunch of text from humans speculating as to what such a solution would look like, given current knowledge, which may or may not be correct. Therefore, improving GPT by letting it produce better and better approximations to the posterior will not get us any closer to such a solution. The only way this could work is if GPT-N somehowmisgeneralizedits way into a correct solution and the likelihood of such a misgeneralization should ultimatelydecreasewith greater capabilities. It is possible however that there is some capabilities sweet spot where GPT-N is powerful enough to figure out a solution but not powerful enough to correctly model the true posterior.^{^}This has also been noted with respect to language model samples (in a

LW affiliated context).To nuance this: humans who are not concentrating areamortizingintelligences.^{^}Similarly to humans, we can clearly view 'chain of thought' style prompting in language models as eliciting a very basic direct optimization or planning capability which is probably a basic version of our natural human planning or focused attention capabilities. Equivalently, 'prompt programming' can be seen as just figuring out how best to query a world model and hence (haphazardly) applying optimization power to steer it towards its desired output. Neuroscientifically, I see human planning as occurring in a similar way with a neocortical world model which can be sequentially 'queried' by RL loops running through prefrontal cortex -> basal ganglia -> thalamus -> sensory cortices.