(This is Section 2.2.1 of my report “Scheming AIs: Will AIs fake alignment during training in order to get power?”. There’s also a summary of the full report here (audio here). The summary covers most of the main points and technical terms, and I’m hoping that it will provide much of the context necessary to understand individual sections of the report on their own.

Audio version of this section here, or search "Joe Carlsmith Audio" on your podcast app.)

Beyond-episode goals

Schemers are pursuing goals that extend beyond the time horizon of the episode. But what is an episode?

Two concepts of an "episode"

Let's distinguish between two concepts of an episode.

The incentivized episode

The first, which I'll call the "incentivized episode," is the concept that I've been using thus far and will continue to use in what follows. Thus, consider a model acting at a time t1. Here, the rough idea is to define the episode as the temporal unit after t1 that training actively punishes the model for not optimizing – i.e., the unit of time such that we can know by definition that training is not directly pressuring the model to care about consequences beyond that time.

For example, if training started on January 1st of 2023 and completed on July 1st of 2023, then the maximum length of the incentivized episode for this training would be six months – at no point could the model have been punished by training for failing to optimize over a longer-than-six-month time horizon, because no gradients have been applied to the model's policy that were (causally) sensitive to the longer-than-six-month consequences of its actions. But the incentivized episode for this training process could in principle be shorter than six months as well. (More below.)

Now, importantly, even if training only directly pressures a model to optimize over some limited period of time, it can still in fact create a model that optimizes over some much longer time period – that's what makes schemers, in my sense, a possibility. Thus, for example, if you're training a model to get as many gold coins as possible within a ten minute window, it could still, in principle, learn the goal "maximize gold coins over all time" – and this goal might perform quite well (even absent training gaming), or survive despite not performing all that well (for example, because of the "slack" that training allows).

Indeed, to the extent we think of evolution as an analogy for ML training, then something like this appears to have happened with humans with goals that extend indefinitely far into the future – for example, "longtermists." That is, evolution does not actively select for or against creatures in a manner sensitive to the consequences of their actions in a trillion years (after all, evolution has only been running for a few billion years) – and yet, some humans aim their optimization on trillion-year timescales regardless.

That said, to the extent a given training procedure in fact creates a model with a very long-term goal (because, for example, such a goal is favored by the sorts of "inductive biases" I'll discuss below), then in some sense you could argue that training "incentivizes" such a goal as well. That is, suppose that "maximize gold coins in the next ten minutes" and "maximize gold coins over all time" both get the same reward in a training process that only provides rewards after ten minutes, but that training selects "maximize gold coins over all time" because of some other difference between the goals in question (for example, because "maximize gold coins over all time" is in some sense "simpler," and gradient descent selects for simplicity in addition to reward-getting). Does that mean that training "incentivizes" or "selects for" the longer-term goal?

Maybe you could say that. But it wouldn't imply that training "directly punishes" the shorter-term goal (or "directly pressures" the model to have the longer-term goal) in the sense I have in mind. In particular: in this case, it's at least possible to get the same reward by pursuing a shorter term goal (while holding other capabilities fixed). And the gradients the model's policy receives are (let's suppose) only ever sensitive to what happens within ten minutes of a model's action, and won't "notice" consequences after that. So to the extent training selects for caring about consequences further out than ten minutes, it's not in virtue of those consequences directly influencing the gradients that get applied to the model's policy. Rather, it's via some other, less direct route. This means that the model could, in principle, ignore post-ten-minute consequences without gradient descent pushing it to stop.

Or at least, that's the broad sort of concept I'm trying to point at. Admittedly, though, the subtleties get tricky. In particular: in some cases, a goal that extends beyond the temporal horizon that the gradients are sensitive to might actively get more reward than a shorter-term goal.

  • Maybe, for example, "maximize gold coins over all time" actually gets more reward than "maximize gold coins over the next ten minutes," perhaps because the longer-term goal is "simpler" in some way that frees up extra cognitive resources that can be put towards gold-coin-getting.

  • Or maybe humans are trying to use short-term feedback to craft a model that optimizes over longer timescales, and so are actively searching for training processes where short-term reward is maximized by a model pursuing long-term goals. For example, maybe you want your model to optimize for the long-term profit of your company, so you reward it, in the short-term, for taking actions that seem to you like they will maximize long-term profit. One thing that could happen here is that the model starts optimizing specifically for getting this short-term reward. But if your oversight process is good enough, it could be that the highest-reward policy for the model, here, is to actually optimize for long-term profit (or for something else long-term that doesn't route via training-gaming).[1]

In these cases, it's more natural to say that training "directly pressures" the model towards the longer-term goal, given that this goal gets more reward. However, I still want to say that longer-term goals here are "beyond episode," because they extend beyond the temporal horizon to which the gradients are directly and causally sensitive. I admit, though, that defining this precisely might get tricky (see next section for a bit more of the trickiness). I encourage efforts at greater precision, but I won't attempt them here.[2]

The intuitive episode

Let's turn to the other concept of an episode – namely, what I'll call the "intuitive episode." The intuitive episode doesn't have a mechanistic definition.[3] Rather, the intuitive episode is just: a natural-seeming temporal unit that you give reward at the end of, and which you've decided to call "an episode." For example, if you're training a chess-playing AI, you might call a game of chess an "episode." If you're training a chatbot via RLHF, you might call an interaction with a user an "episode." And so on.

My sense is that the intuitive episode and the incentivized episode are often somewhat related, in the sense that we often pick an intuitive episode that reflects some difference in the training process that makes it easy to assume that the intuitive episode is also the temporal unit that training directly pressures the model to optimize – for example, because you give reward at the end of it, because the training environment "resets" between intuitive episodes, or because the model's actions in one episode have no obvious way of affecting the outcomes in other episodes. Importantly, though, the intuitive episode and the incentivized episode aren't necessarily the same. That is, if you've just picked a natural-seeming temporal unit to call the "episode," it remains an open question whether the training process will directly pressure the model to care about what happens beyond the episode-in-thise-sense. For example, it remains an open question whether training directly pressures the model to sacrifice reward on an earlier episode-in-thise-sense for the sake of more-reward on a later episode-in-this-sense, if and when it is able to do so.

To illustrate these dynamics, consider a prisoner's dilemma-like situation where each day, an agent can either take +1 reward for itself (defection), or give +10 reward to the next day's agent (cooperation), where we've decided to call a "day" an (intuitive) episode. Will different forms of ML training directly pressure this agent to cooperate? If so, then the intuitive episode we've picked isn't the incentivized episode.

Now, my understanding is that in cases like these, vanilla policy gradients (a type of RL algorithm) learn to defect (this test has actually been done with simple agents – see Krueger et al (2020)). And I think it's important to be clear about what sorts of algorithms behave in this way, and why. In particular: glancing at this sort of set-up, I think it's easy to reason as follows:

"Sure, you say that you're training models to maximize reward 'on the episode,' for some natural-seeming intuitive episode. But you also admit that the model's actions can influence what happens later in time, even beyond this sort of intuitive episode – including, perhaps, how much reward it gets later. So won't you implicitly be training a model to maximize reward over the whole training process, rather than just on an individual (intuitive) episode. For example, if it's possible for a model to get less reward on the present episode, in order to get more reward later, won't cognitive patterns that give rise to making-this-trade get reinforced overall?"[4]

From discussions with a few people who know more about reinforcement learning than I do,[5] my current (too-hazy) understanding is that at least for some sorts of RL training algorithms, this isn't correct. That is, it's possible to set up RL training such that some limited, myopic unit of behavior is in fact the incentivized episode – even if an agent can sacrifice reward on the present episode for the sake of more-reward later (and presumably: even if the agent knows this). Indeed, this may well be the default. See footnote for more details.[6]

Even granted that it's possible to avoid incentives to optimize across intuitive-episodes, though, it's also possible to not do this – especially if you pick your notion of "intuitive episode" very poorly. For example, my understanding is that the transformer architecture is set up, by default, such that language models are incentivized, in training, to allocate cognitive resources in a manner that doesn't just promote predicting the next token, but other later tokens as well (see here for more discussion). So if you decided to call predicting just-the-next-token an "episode," and to assume, on this basis, that language models are never directly pressured to think further ahead, you'd be misled.

And in some cases, the incentives in training towards cross-episode optimization can seem quite counterintuitive. Thus, Krueger et al (2020) show, somewhat surprisingly, that if you set the parameters right, a form of ML training called Q-learning sometimes learns to cooperate in prisoner's dilemmas despite the algorithm being "myopic" in the sense of: ignoring reward on future "episodes." See footnote for more discussion, and see here for a nice and quick explanation of how Q-learning works.[7]

Plots of Q-values and cooperation probability from Figure 11 in
Krueger et al (2020), reprinted with permission. The first three agents
learn to defect, and the last two, to cooperate almost 90% of the time
(note the differences between the y-axes in the first three and the last
two).

Plots of Q-values and cooperation probability from Figure 11 in Krueger et al (2020), reprinted with permission. The first three agents learn to defect, and the last two, to cooperate almost 90% of the time (note the differences between the y-axes in the first three and the last two).

Another way to incentivize optimization across intuitive episodes is to introduce various additional layers (or "outer loops") of selection in training. Thus, for example, consider a toy and unrealistic form of "population based training" where you create 1000 agents who play starcraft, have them play 100 games each (and get updated along the way via some object-level training algorithm), then select the one that performs the best on the final game.[8] In this case, if it were possible for a model to sacrifice reward on earlier games, in order to perform better on the final game, there would be at least some selection pressure for it to do so. And see, also, results in Krueger et al (2020), showing that a different form of population based training selects for cooperation in a prisoner's-dilemma-like case like the one discussed above.[9] Depending on the details, though, outer-loops of this kind may exert a much weaker degree of selection pressure than inner loops driven directly by gradient descent.[10]

Overall, my current sense is that one needs to be quite careful in assessing whether, in the context of a particular training process, the thing you're thinking of as an "episode" is actually such that training doesn't actively pressure the model to optimize beyond the "episode" in that sense – that is, whether a given "intuitive episode" is actually the "incentivized episode." Looking closely at the details of the training process seems like a first step, and one that in theory should be able to reveal many if not all of the incentives at stake. But empirical experiment seems important too.

Indeed, I am somewhat concerned that my choice, in this report, to use the "incentivized episode" as my definition of "episode" will too easily prompt conflation between the two definitions, and correspondingly inappropriate comfort about the time horizons that different forms of training directly incentivize.[11] I chose to focus on the incentivized episode because I think that it's the most natural and joint-carving unit to focus on in differentiating schemers from other types of models. But it's also, importantly, a theoretical object that's harder to directly measure and define: you can't assume, off the bat, that you know what the incentivized episode for a given sort of training is. And my sense is that most common uses of the term "episode" are closer to the intuitive definition, thereby tempting readers (especially casual readers) towards further confusion. Please: don't be confused.


  1. I'll briefly note another complexity that this sort of case raises. Naively, you might've thought that the "specified goal" would only ever be confined to the incentivized episode, because the specified goal is the "thing being rewarded," and anything that causes reward is within the temporal horizon to which the gradients are sensitive. And in some cases – for example, where the "specified goal" is some clearly separable consequence of the model's action (e.g., getting gold coins), which the training process induces the model to optimize for – this makes sense. But in other cases, I'm less sure. For example, if you are sufficiently good at telling whether your model is in fact optimizing for long-term profit, and providing short-term rewards that in fact incentivize it to do so, then I think it's possible that the right thing to say is that the "specified goal," here, is long-term profit (or at least, "optimizing for long-term profit," which looks pretty similar). However, I don't think it ultimately matters much whether we call this sort of goal "specified" or "mis-generalized" (and it's a pretty wooly distinction more generally), so I'm not going to press on the terminology here. ↩︎

  2. Also: when I talk about the gradients being sensitive to the consequences of the model's action over some time horizon, I am imagining that this sensitivity occurs via (1) the relevant consequences occurring, and then (2) the gradients being applied in response. E.g., the model produces an action at t1, this leads to it getting some number of gold coins at t5, then the gradients, applied at t6, are influenced by how many gold coins the model in fact got. (I'll sometimes call this "causal sensitivity.")

    But it's possible to imagine fancier and more philosophically fraught ways for the consequences of a model's action to influence the gradients. For example, suppose that the model is being supervised by a human who is extremely good at predicting the consequences of the model's action. That is, the model produces some action at t1, then at t2 the human predicts how many gold coins this will lead to at t5, and applies gradients at t3 reflecting this prediction. In the limiting case of perfect prediction, this can lead to gradients identical to the ones at stake in the first case – that is, information about the consequences of the model's action is effectively "traveling back in time," with all of the philosophical problems this entails.# So if, in the first case, we wanted to say that the "incentivized episode" extends out to t5, then plausibly we should say this of the second case, too, even though the gradients are applied at t3. But even in a case of pretty-good-but-still-imperfect prediction, there is a sense, here, in which the gradients the model receives are sensitive to consequences that haven't yet happened.

    I'm not, here, going to extend the concept of the "incentivized episode" to cover forms of sensitivity-to-future-consequences that rest on predictions about those consequences. Rather, I'm going to assume that sensitivity in question arises via normal forms of causal influence. That said, I think the fact that it's possible to create some forms of sensitivity-to-future-consequences even prior to seeing those consequences play out is important to keep in mind. In particular, it's one way in which we might end up training long-horizon optimizers using fairly short incentivized episodes (more discussion below). ↩︎

  3. There may be other, additional, and more precise ways of using the term "episode" in the RL literature. Glancing at various links online, though (e.g. here and here), I'm mostly seeing definitions that refer to an episode as something like "the set of states between the initial state and the terminal state," which doesn't say how the initial state and the terminal state are designated. ↩︎

  4. As an example of someone who seems to me like they could be reasoning in this way, though it's not fully clear, see this comment from Eliezer Yudkowsky, in response to a hypothetical in which he imagines humans rewarding an agent for each of its sentences according to how useful that sentence is:

    "Let's even skip over the sense in which we've given the AI a long-term incentive to accept some lower rewards in the short term, in order to grab control of the rating button, if the AGI ends up with long-term consequentialist preferences and long-term planning abilities that exactly reflect the outer fitness function."

    That said, as I discuss below, the details of the training process here matter. ↩︎

  5. In particular: Paul Christiano, Ryan Greenblatt, and Rohin Shah. Though they don't necessarily endorse my specific claims here, and it's possible I've misunderstood them more generally. ↩︎

  6. My hazy understanding of the argument here is that these RL algorithms update the model's policy towards higher-reward actions on the episode in a way that doesn't update you towards whatever policies would've led to you starting in a higher-reward episode (In this sense, they behave in a manner analogous to "causal decision theory."). Thus, let's say that the agent on Day 1 (with no previous agent to benefit her) chooses between cooperating (0 reward) and defecting (+1 reward reward), and so this episode results in an update towards defecting. Then, on Day 1, the agent either starts out choosing between 10 vs. 11 reward (call this a "good episode"), or 0 vs. +1 reward (call this a "bad episode"). Again, either way, it updates towards defection. It doesn't update, in the good episode, towards "whatever policy led me to this episode."

    That said, in my current state of knowledge about RL, I'm still a bit confused about this. Suppose, for example, that at the point of choice, you don't know whether or not you're in the good episode or the bad episode, and the training process is updating you with strength proportional to the degree to which you got more reward than you expected to get. If you start out with e.g. 50% that you're in a good episode and 50% that you're in a bad episode (such that the expected reward of cooperating is 5, and the expected reward of defecting is 6), then it seems like it could be the case that being in a good episode results in reward that is much better than you expected, such that policies that make it to a good episode end up reinforced to a greater extent, at least initially.

    I'm not sure about the details here. But from my current epistemic state, I would want to spell out and understand the details of the training process in much greater depth, in order to verify that there isn't an incentive towards cross-episode optimization. ↩︎

  7. I think the basic dynamic here is: the Q-values for the actions reflect the average reward for taking that action thus far. This makes it possible for the Q-value for "cooperate" to give more weight to the rewards received in the "good episodes" (where the previous-episode's agent cooperated) rather than the "bad episodes" (where the previous-episode's agent defected), if the agent ends up in good episodes more often. This makes it possible to get a "cooperation equilibrium" going (especially if you set the initial q-value for defecting low, which I think they do in the paper in order to get this effect), wherein an agent keeps on cooperation. That said, there are subtleties, because agents that end up in a cooperation equilibrium still sometimes explore into defecting, but in the experiment it (sometimes) ends up in a specific sort of balance, with q-values for cooperation and defection pretty similar, and with the models settling on a 90% or so cooperation probability (more details here and in the paper's appendix). ↩︎

  8. I owe this example to Mark Xu. ↩︎

  9. See also discussion from Carl Shulman here: 'it could be something like they develop a motivation around an extended concept of reproductive fitness, not necessarily at the individual level, but over the generations of training tendencies that tend to propagate themselves becoming more common and it could be that they have some goal in the world which is served well by performing very well on the training distribution.'' ↩︎

  10. Indeed, in principle, you could imagine pointing to other, even more abstract and hard-to-avoid "outer loops" as well, as sources of selection pressure towards longer-term optimization. For example, in principle, "grad student descent" (e.g., researchers experimenting with different learning algorithms and then selecting the ones that work best) introduces an additional layer of selection pressure (akin to a hazy form of "meta-learning"), as do dynamics in which, other things equal, models whose tendencies tend to propagate into the future more effectively will tend to dominate over time (where long-term optimization is, perhaps, one such tendency). But these, in my opinion, will generally be weak enough, relative to gradient descent, that they seem to me much less important, and OK to ignore in the context of assessing the probability of schemers. ↩︎

  11. Thanks to Daniel Kokotajlo for flagging this concern. ↩︎

New Comment
1 comment, sorted by Click to highlight new comments since: Today at 11:08 AM

Thanks for writing the report - I think it’s an important issue, and you’ve clearly gone to a lot of effort. Overall, I think it’s good. 

However, it seems to me that the "incentivized episode" concept is confused, and that some conclusions over the probability of beyond-episode goals are invalid on this basis. I'm fairly sure I'm not confused here (though I'm somewhat confused that no-one picked this up in a draft, so who knows?!).

I'm not sure to what extent the below will change your broader conclusions - if it only moves you from [this can't happen, by definition], to [I expect this to be rare], the implications may be slight. It does seem to necessitate further thought - and different assumptions in empirical investigations.

The below can all be considered as [this is how things seem to me], so I’ll drop caveats along those lines. I’m largely making the same point throughout - I’m hoping the redundancy will increase clarity.

  1. "The unit of time such that we can know by definition that training is not directly pressuring the model to care about consequences beyond that time." - this is not a useful definition, since there is no unit of time where we know this by definition. We know only that the model is not pressured by consequences beyond training.
    1. For this reason, in what follows I'll largely assume that [episode = all of training]. Sticking with the above definition gives unbounded episodes, and that doesn't seem in the spirit of things.
  2. Replacing "incentivized episode" with "incentivizing episode" would help quite a bit. This clarifies the causal direction we know by definition. It’s a mistake to think that we can know what’s incentivized "by definition": we know what incentivizes.
    1. In particular, only the episode incentivizes; we expect [caring about the episode] to be incentivized; we shouldn't expect [caring about the episode] to be the only thing incentivized (at least not by definition).
  3. For instance, for any t, we can put our best attempt at [care about consequences beyond tin the reward function (perhaps we reward behaviour most likely to indicate post-t caring; perhaps we use interpretability on weights and activations, and we reward patterns that indicate post-t caring - suppose for now that we have superb interpretability tools).
    1. The interpretability case is more direct than rewarding future consequences: it rewards the mechanisms that constitute caring/goals directly.
    2. If a process we've designed explicitly to train for , is described as "not directly pressuring the model to " by definition, then our definition is at least misleading - I'd say broken.
    3. Of course this example is contrived - but it illustrates the problem. There may be analogous implicit (but direct!) effects that are harder to notice. We can't rule these out with a definition.
      1. These need not be down to inductive bias: as here, they can be favoured as a consequence of the reward function (though usually not so explicitly and extremely as in this example).
  4. What matters is the robustness of correlations, not directness.
    1. Most directness is in our map, not in the territory. I.e. it's largely saying [this incentive is obvious to us], [this consequence is obvious to us] etc.
    2. The episode is the direct cause of gradient updates.
    3. As a consequence of gradient updates, behaviour on the episode is no more/less direct than behaviour outside that interval.
  5. While a strict constraint on beyond-episode goals may be rare, pressure will not be.
    1. The former would require that some beyond-episode goal is entailed by high performance in-episode.
    2. The latter only requires that some beyond-episode goal is more likely given high performance in-episode.
      1. This is the kind of pressure that might, in principle, be overcome by adversarial training - but it remains direct pressure for beyond-episode goals resulting from in-episode consequences.

 

The most significant invalid conclusion (Page 52):

  • "Why might you not expect naturally-arising beyond-episode goals? The most salient reason, to me, is that by definition, the gradients given in training (a) do not directly pressure the model to have them…"
    • As a general "by definition" claim, this is false (or just silly, if we say that the incentivized episode is unbounded, and there is no "beyond-episode").
    • It's less clear to me how often I'd expect active pressure towards beyond-episode-goals over strictly-in-episode-goals (other than through the former being more common/simple). It is clear that such pressure is possible.
    • Directness is not the issue: if I keep  small, I keep  small. That I happen to be manipulating  directly, rather than  directly, is immaterial.
      • Again, the out-of-episode consequences of gradient updates are just as direct as the in-episode consequences - they’re simply harder for us to predict.