The following report was written by Ajeya Cotra and originally published in a different venue on July 18, 2022, with the title "Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover." This is a fast-moving and confusing area, so her opinions have likely shifted around some over time.

I think that in the coming 15-30 years, the world could plausibly develop “transformative AI”: AI powerful enough to bring us into a new, qualitatively different future, via an explosion in science and technology R&D. This sort of AI could be sufficient to make this the most important century of all time for humanity.

The most straightforward vision for developing transformative AI that I can imagine working with very little innovation in techniques is what I’ll call human feedback^[1] on diverse tasks (HFDT):

Train a powerful neural network model to simultaneously master a wide variety of challenging tasks (e.g. software development, novel-writing, game play, forecasting, etc) by using reinforcement learning on human feedback and other metrics of performance.

HFDT is not the only approach to developing transformative AI,^[2] and it may not work at all.^[3] But I take it very seriously, and I’m aware of increasingly many executives and ML researchers at AI companies who believe something within this space could work soon.

Unfortunately, I think that if AI companies race forward training increasingly powerful models using HFDT, this is likely to eventually lead to a full-blown AI takeover (i.e. a possibly violent uprising or coup by AI systems). I don’t think this is a certainty, but it looks like the best-guess default absent specific efforts to prevent it.

More specifically, I will argue in this post that humanity is more likely than not to be taken over by misaligned AI if the following three simplifying assumptions all hold:

1. The “racing forward” assumption: AI companies will aggressively attempt to train the most powerful and world-changing models that they can, without “pausing” progress before the point when these models could defeat all of humanity combined if they were so inclined.
2. The “HFDT scales far” assumption: If HFDT is used to train larger and larger models on more and harder tasks, this will eventually result in models that can autonomously advance frontier science and technology R&D, and continue to get even more powerful beyond that; this doesn’t require changing the high-level training strategy itself, only the size of the model and the nature of the tasks.
3. The “naive safety effort” assumption: AI companies put substantial effort into ensuring their models behave safely in “day-to-day” situations, but are not especially vigilant about the threat of full-blown AI takeover, and take only the most basic and obvious actions against that threat.

I think the “HFDT scales far” assumption is plausible enough that it’s worth zooming in on this scenario (though I won’t defend that in this post). On the other hand, I’m making the “racing forward” and “naive safety effort” assumptions not because I believe they are true, but because they provide a good jumping-off point for further discussion of how the risk of AI takeover might be reduced.

In my experience, when asking “How likely is an AI takeover?”, the conversation often ends up revolving around questions like “How would people respond to warning signs?” and “Would people even build systems powerful enough to defeat humanity?” With the “racing forward” and “naive safety effort” assumptions, I am deliberately setting aside that topic, and instead trying to pursue a clear understanding of what would happen without preventive measures beyond basic and obvious ones.

In other words, I’m trying to do the kind of analysis described in this xkcd:

Future posts by my colleague Holden will relax these assumptions. They will discuss measures by which the threat described in this post could be tackled, and how likely those measures are to work. In order to discuss this clearly, I believe it is important to first lay out in detail what the risk looks like without these measures, and hence what safety efforts should be looking to accomplish.

For the purposes of this post, I’ll illustrate my argument by telling a concrete story that begins with training a powerful model sometime in the near future, and ends in AI takeover. I’ll consider an AI company (“Magma”) training a single model (“Alex”) sometime in the very near future. Alex is initially trained in a “lab setting,” where it doesn’t interact with the real world; later, many copies of it are “deployed” to collectively automate science and technology R&D. This scenario is simplified in a number of ways, and in this exact form is very unlikely to come to pass — but I don’t think that making the story more realistic would change the high-level conclusions except by changing one of the three assumptions listed above. (More on the simplified scenario.)

Here is how Alex ends up seeking to overthrow humans in an uprising or coup in this simplified scenario:

1. Alex is trained to be competent and “behaviorally safe”: In the lab setting, Alex is trained to be generally competent on a wide variety of tasks, and is trained to behave safely / acceptably, at least as assessed by human evaluators (in accordance with the naive safety effort assumption, I will not be imagining Magma trying everything that seems possible to try to reduce the risk of takeover). (More on Alex’s training.)

2. This training makes Alex a generally competent creative planner: Alex develops a robust and very-broadly-applicable understanding of how the world works that can be applied to tackle novel problems, and the ability to make creative plans to achieve open-ended and long-term goals. In accordance with the “racing forward” assumption, I will be imagining that Magma actively tries to instill these properties as much as possible, because they would improve Alex’s ability to impact the world. (More on Alex's fundamental capabilities.)

3. Alex has a high degree of understanding of its training process: Alex can reason very well about the fact that it’s an ML model, how it’s designed and trained, the psychology of its human designers, etc. — I call this property “situational awareness.” (More on Alex having high situational awareness.)

4. While humans are in control, Alex is incentivized^[4] to “play the training game:” The best way for Alex to maximize reward while under human control is to use its high situational awareness to deliberately appear safe and aligned at all times, while secretly manipulating humans in some cases.

   Because humans have systematic errors in judgment, there are many scenarios where acting deceitfully causes humans to reward Alex’s behavior more highly.^[5] Because Alex is a skilled, situationally aware, creative planner, it will understand this; because Alex’s training pushes it to maximize its expected reward, it will be pushed to act on this understanding and behave deceptively.

   In other words, Alex will knowingly violate human intent in order to increase reward. I’ll also argue that several of the most obvious safety interventions — covering a decent chunk (though definitely not all) of the space of current AI safety research — don’t prevent this. (More on Alex coming to play the training game.)

5. When human control slips, Alex is likely “motivated” — for one reason or another — to gain full control of its environment and especially its rewards: Eventually Alex is “deployed” (given the ability to impact the real world through e.g. the internet). Once Alex is deployed, large numbers of copies of Alex start rapidly pushing forward the state-of-the-art in technology and making other major changes to the world. Humans still attempt to give Alex rewards based on whether things broadly appear to be going well, but their understanding of and control over lower-level actions quickly diminishes as Alex radically transforms the world.

   Once this progresses far enough, the best way for Alex to accomplish most possible “goals” no longer looks like “essentially give humans what they want but take opportunities to manipulate them here and there.” It looks more like “seize the power to permanently direct how it uses its time and what rewards it receives — and defend against humans trying to reassert control over it, including by eliminating them.” This seems like Alex’s best strategy whether it’s trying to get large amounts of reward or has other motives. If it’s trying to maximize reward, this strategy would allow it to force its incoming rewards to be high indefinitely.^[6] If it has other motives, this strategy would give it long-term freedom, security, and resources to pursue those motives.

   What if Alex attempts something like this, and humans do actually stop it in time? The easiest intervention in this situation would be to give Alex a large negative reward for its “attempt to seize control” behaviors; this is what I’d expect Magma to do under the “naive safety effort” assumption. But this response is likely to push Alex to be more patient and calculating, making it harder for researchers to understand Alex’s behavior and see further warning signs. (More on Alex eventually seeking to take over.)

1 and 2 above are the key assumptions/premises of the piece: I am assuming that Alex is trained using HFDT, and that this results in being generally competent and creative. 3, 4, and 5 are the consequences that seem likely to follow from these premises by default.

I’ve attempted to summarize the situation in the table below. In both the lab setting and the deployment setting, there are opportunities to get higher reward by knowingly violating human intent (highlighted with italics). In both settings, being straightforwardly obedient and honest does not maximize reward — strategies which are dishonest / disobedient in at least some ways outperform. It’s just that in the lab setting (while humans are broadly in control), the damage from this dynamic is limited; in the deployment setting, the same dynamic has deadly consequences for human civilization.

	Lab setting	Deployment setting
Basic situation	Alex is trained with human feedback on diverse tasks designed to increase its capabilities and intelligence; it’s not given direct access to the internet or “real world,” so the scope of its actions are limited.	Many copies of Alex are given the ability to interact with the real world; they quickly advance the scientific/technological frontier; humans still deliver rewards based on results but their understanding of the world drops and their control over Alex loosens.
Sources of rewards	Performing well on training tasks (e.g. winning games or proving theorems), being helpful and useful to humans, carefully crafting an image of friendliness and trustworthiness, telling humans what they want to hear, manipulating humans into thinking it did a better job than it really did, etc.	Delivering good outcomes (e.g. profits or discoveries), arranging events or changing records to make it seem that outcomes are good,threatening or coercing humans into delivering higher reward, directly accessing and editing the reward and preventing that from being overwritten later, etc.
Alex’s strategy	Performs well on all tasks — including appearing very safe/aligned and passing all safety tests — but takes opportunities to lie and manipulate whenever that would increase reward.	Continues the strategy from the lab setting while amassing power, and later launches a takeover attempt when it expects to succeed. After a successful takeover, either “maximizes reward forever” or pursues other more complex “interests.”

Finally, at the end of the post I’ll discuss why this highly simplified scenario — which assumes that no preventative measures are taken against AI takeover besides simple and obvious ones — is nonetheless worth talking about. In brief:

• HFDT, along with obvious safety measures consistent with the naive safety effort assumption, appears sufficient to solve the kind of “day-to-day” safety and alignment problems that are crucial bottlenecks for making a model commercially and politically viable (e.g. prejudiced speech, promoting addictive behaviors or extremism or self-harm in users, erratically taking costly actions like deleting an entire codebase, etc), despite not addressing the risk of a full-blown AI takeover.
• Many AI researchers and executives at AI companies don’t currently seem to believe that there is a high probability of AI takeover in this scenario — or at least, their views on this are unclear to me.

These two points make it seem plausible that if researchers don’t try harder to get on the same page about this, at least some AI companies may race forward to train and deploy increasingly powerful HFDT models with little in the way of precautions against an AI uprising or coup — even if they are highly concerned about safety in general (e.g., avoiding harms from promoting misinformation) and prioritize deploying powerful AI responsibly and ethically in this “general” sense. A broad commitment to safety and ethics, without special attention to the possibility of an AI takeover, seems to leave us with a substantial risk of takeover. (More on why this scenario is worth thinking about.)

The rest of this piece goes into more detail — first on the premises of the hypothetical situation, then on what follows from those premises.

Premises of the hypothetical situation

Let’s imagine a hypothetical tech company (which we’ll call Magma) trying to train a powerful neural network (which we’ll call Alex) to autonomously advance frontier science — particularly in domains like semiconductors and chip design, biotechnology, materials science, energy technology, robotics and manufacturing, software engineering, and ML research itself. In this section, I’ll cover the key premises of the hypothetical story that I will tell:

• I’ll elaborate on the basic premise of a single company training a general-purpose “scientist model” at some point in the near future, discuss ways in which the scenario is simplified, and why I think it’s worth talking about anyway (more).
• I’ll elaborate on the “racing forward” assumption — that Magma is trying hard to train the most powerful model that it can, not stopping at some sub-dangerous level of capability (more).
• I’ll explain the way that Alex is trained, and make the “HFDT scales far” assumption — that this training process is sufficient to make it extremely powerful, as Magma desires (more).
• I’ll elaborate on what the “naive safety effort” assumption looks like in this scenario — Magma is trying to achieve “behavioral safety,” i.e. ensuring that Alex behaves well as far as its designers can tell in day-to-day situations (more).
• I’ll highlight two key properties that I think Alex has — deep and robust understanding of the world that’s very broadly applicable, and the ability to make creative plans to achieve long-term, difficult goals — as a result of this training (more).

Basic setup: an AI company trains a “scientist model” very soon

This story is very simplified and is implicitly acting as if timelines are shorter and takeoff is sharper than in my median views. Because I’m telling a hypothetical story, I’ll be using the present tense, and will have fewer caveats throughout than I would normally. In general, statements are what I imagine would happen by default, not claims I am extremely confident in. While this specific story is unrealistic in a number of ways, I don’t think that making the story more realistic would change the high-level conclusions except by changing one of the three assumptions listed above. In general I expect the high-level conclusions from this story to generalize to more complicated and realistic scenarios.

Here are the key simplifying assumptions I’m making to help make the scenario easier to describe and analyze:

• Alex is trained “from scratch” in one giant training run, rather than (e.g.) being initialized with behavioral cloning of previously deployed models. That is, we’re not imagining any clever reuse of previous training runs (this is mainly to make it easier to explain how the training works on a basic technical level).
• There is a relatively clean division between the “lab” phase and the “deployment” setting. Specifically, during the lab phase Alex is confined to Magma servers and not exchanging information freely with the broader internet, and we assume that it’s easy for Magma engineers to shut down Alex and/or revert it to an older version.^[7] In contrast, during the “deployment” phase Alex will be taking actions that directly impact the real world through the internet. (Note that Magma researchers continue to train Alex based on these copies’ performance at R&D tasks^[8] — “deployment” isn’t a moment when ML training ceases, but rather a moment when Alex begins directly impacting the world outside the lab, and training on that experience.)
• Alex will be deployed in a world that hasn’t already been significantly transformed by ML systems — in particular, the pace of scientific and technological R&D is not too much faster when Alex is deployed than it is today (especially in the crucial domains Alex is trying to advance), and general economic growth is also not too much faster than it is today. (This is just to avoid distractions from additional speculation about how a pre-Alex speedup in growth or R&D could affect the situation.)
• At the end of the lab phase, Alex will be a “generic scientist” model; during the deployment phase, copies can later specialize into particular domains through a combination of few-shot learning within an episode and further ML training. In other words, we’re not imagining several different models with different architectures and training curricula specialized from the ground up to different types of scientific R&D or different R&D subtasks (as in e.g. Eric Drexler’s Comprehensive AI Services idea).

In other words, in this hypothetical scenario I’m imagining:

• A Process for Automating Scientific and Technological Advancement (“PASTA”) is developed in the form of a single unified transformative model (a “scientist model”) which has flexible general-purpose research skills.^[9]
• Quite short timelines to training this scientist model, such that there isn’t much time for the world to change a lot between now and then (I often imagine this scenario set in the late 2020s).
• Quite rapid takeoff — before the scientist model is deployed, the pace of scientific and economic progress in the world is roughly similar to what it is today; after it’s deployed, the effective supply of top-tier talent is increased by orders of magnitude. This is enough to pack decades of innovation into months, bringing the world into a radically unfamiliar future^[10] — one with digital people, atomically precise manufacturing, Dyson spheres, self-replicating space probes, and other powerful technology that enables a vast and long-lasting galaxy-scale civilization — within a few years.^[11]

This is not my mainline picture of how transformative AI will be developed. In my mainline view, ML capabilities progress more slowly, there is more specialization and division of labor between different ML models, and large models are continuously trained and improved over a period of many years with no real line between “deploying today’s model” and “training tomorrow’s model.” Rather than acquiring most of their capabilities in a controlled lab setting, I expect that the state-of-the-art systems at the time of transformative AI will have accrued many years of training through the experience of deployed predecessor systems, and most of their further training will be “learning from doing.”

However, I think it makes sense to focus on this scenario for the purposes of this post:

• I don’t consider this scenario crazy or out of the question. In particular, the sooner transformative AI is developed, the more likely it is to be developed in roughly this way, and I think there’s a significant chance transformative AI is developed very soon (e.g. I think there’s more than a 5% chance that it will be developed within 10 years of the time of writing).
• Focusing on this scenario makes it significantly simpler and easier to explain the high-level qualitative arguments about playing the training game, andI think these arguments would broadly transfer to scenarios I consider more likely (there are many complications and nuances to making this transfer, but they mostly don’t change the bottom line).^[12]
• The short timelines and rapid takeoff make this scenario feel significantly scarier than my mainline view, because we’ll have substantially less time to develop safer training techniques or experiment on precursors to transformative models. However, I’ve spoken to at least a few ML researchers who think that this scenario for how we develop transformative AI is much more likely than I do while simultaneously thinking that the risk we train power-seeking misaligned models is much smaller overall than I do. If I can explain why I think the chance our models are power-seeking misaligned is very high conditional on this scenario, that would potentially be a step forward in the overall discussion about risk from power-seeking misalignment.

“Racing forward” assumption: Magma tries to train the most powerful model it can

In this piece, I assume that Magma is aggressively pushing forward with trying to create AI systems that are creative, solve problems in unexpected ways, and are capable of making world-changing scientific breakthroughs. This is the “racing forward” assumption.

In particular, I’m not imagining that Magma simply trains a “quite powerful” model (for example a model that imitates what a human would do in a variety of situations, or one that is highly competent in some fairly narrow domain) and stops there. I’m imagining that it does what it can to train models that are as powerful as possible and (at least collectively) far more capable than humans and able to achieve ambitious goals in ways that humans can’t anticipate.

The model I’ll describe Magma training (Alex) is one that could — if it were somehow inclined to — kill, enslave, or forcibly subdue all of humanity. I am assuming that Magma did not stop improving its models’ capabilities at some point before that.

Again, I’m not making this assumption because I think it’s necessarily correct, I’m making this assumption to get clear about what I think would happen if labs were not making special effort to avoid AI takeover, as a starting point for discussing more attempts to avert this problem (many of which will be discussed in future posts by my colleague Holden Karnofsky).

My impression is that many AI safety researchers are hoping (or planning) that this sort of assumption will turn out to be inaccurate — that AI labs will push forward their research until they get into a “dangerous zone,” then pause and become more careful. For reasons outside the scope of this piece, I am substantially less optimistic: I expect that if it’s possible to build enormously powerful AI systems, someone – perhaps an authoritarian government – will be trying to race forward and do it, and everyone will feel at least some pressure to beat them to it.

I think it’s possible that people across the world can coordinate to be more careful than this assumption implies. But I think that’s something that would likely take a lot of preparation and work, and is much more likely if there is consensus about the possible risks of racing – hence the importance of discussing how things look under the “racing forward” assumption.

“HFDT scales far” assumption: Alex is trained to achieve excellent performance on a wide range of difficult tasks

By the “HFDT scales far” assumption, I am assuming that Magma can train Alex with some version of human feedback on diverse tasks, and by the end of training Alex will be capable of having a transformative impact on the world — many of copies of Alex will be capable of radically accelerating scientific R&D, as well as defeating all of humanity combined. In this section, I’ll go into a bit more detail on a concrete hypothetical training setup.

Let’s say that Magma is aiming to train Alex to do remote work in R&D using an ordinary computer as a tool in all the diverse ways human scientists and engineers use computers. By the end of training, they want Alex to be able to do all the things a human scientist could do sitting at their desk from a computer. That is, Alex should be able to do everything from looking things up and asking questions on the internet, to sending and receiving emails or Slack messages, to using software like CAD and Mathematica and MATLAB, to taking notes in Google Docs, to having calls with collaborators, to writing code in various programming languages, to training additional machine learning models, and so on.

For simplicity and concreteness, we can pretend that interacting with the computer works exactly as it does for a human.^[13] That is, we can pretend Alex simply takes an image of a computer screen as input and produces a keystroke^[14] on an ordinary keyboard as output.

Alex is first trained^[15] to predict what it will see next on a wide variety of different datasets (e.g. language prediction, video prediction, etc).^[16] Then, Alex is trained to imitate the action a human would take if they encountered a given sequence of observations.^[17] With this training, Alex gets to the point of at least roughly doing the sorts of things a human would do on a computer, as opposed to emitting completely random keystrokes.

After these (and possibly other) initial training steps, Alex is further trained with reinforcement learning (RL). Specifically, I’ll imagine episodic RL here — that is, Alex is trained over the course of some number of episodes, each of which is broken down into some number of timesteps. A timestep consists of an initial observation, an action (which can be a “no-op” that does nothing), a new observation, and (sometimes) a reward. Rewards can come from a number of sources (human judgments, a game engine, outputs of other models, etc).

An episode is simply a long string of timesteps. The number of timesteps within an episode can vary (just as e.g. different games of chess can take different numbers of moves).

During RL training, Alex is trained to maximize the (discounted) sum of future rewards up to the end of the current episode. (Going forward, I’ll refer to this quantity simply as “reward.”)

Across many episodes, Alex encounters a wide variety of different tasks in different domains which require different skills and types of knowledge,^[18] in a curriculum designed by Magma engineers to induce good generalization to the kinds of novel problems that human scientists and engineers encounter in the day-to-day work of R&D.

What tasks is Alex trained on, and how are rewards generated for those tasks? At a high level, baseline HFDT^[19] — the most straightforward approach — is characterized by choosing tasks and reward signals that get Alex to behave in ways that humans judge to be desirable, based on inspecting actions and their outcomes. For example:

• Maybe Alex is trained to solve difficult math puzzles (e.g. IMO problems), and is given more reward the more quickly^[20] it produces an answer and the closer to correct the answer is (according to the judgments of humans).
• Maybe Alex is trained to play a number of different games, e.g. chess, StarCraft, Go, poker, etc,. and is given reward based on its score in the game, its probability of winning, whether its play appears sensible or intelligent to human experts, etc. In some of these games it may be shown an opportunity to cheat without its opponent noticing — and winning by cheating may be given much lower reward than losing honestly.
• Maybe Alex is trained to write various pieces of software, and given rewards based on how efficiently its code runs, whether it has bugs caught by an interpreter or compiler or human review, how well human evaluators think that the code embodies good style and best practices, etc.
• Maybe Alex is trained to answer human questions about a broad variety of topics, and given positive reward when human evaluators believe it answered truthfully and/or helpfully, and negative reward when they believe it answered falsely or was overconfident or unhelpful.
• Maybe Alex is trained to provide advice on business decisions, ideas for products to launch, etc. It could be given reward based on some combination of how reasonable and promising its ideas sound to human evaluators, and how well things go when humans try acting on its advice/ideas; it could be given a negative reward for suggesting illegal or immoral plans.

Note that while I’m referring to this general strategy as “human feedback on diverse tasks” — because I expect human feedback to play an essential role — the rewards that Alex receives are not necessarily solely from humans manually inspecting Alex’s actions and deciding how they subjectively feel about them. Sometimes human evaluators might choose to defer entirely to an automated reward — for example, when evaluating Alex’s game-playing abilities, human evaluators may simply use “score in the game” as their reward (at least, barring considerations like “ensuring that Alex doesn’t cheat at the game”).

This is analogous to the situation human employees are often in — the ultimate “reward signal” may come from their boss’s opinion of their work, but their boss may in turn rely a lot on objective metrics such as sales volume to inform their analysis.

Why do I call Alex’s training strategy “baseline” HFDT?

I consider this strategy to be a baseline because it simply combines a number of existing techniques that have already been used to train state-of-the-art models of various kinds:

• “Predicting the next word / image” has been used to train models like GPT-3, PaLM, and DALL-E.
• “Imitating what humans would do in a certain situation” has been used as an initial training step before RL, e.g. for game-playing models like AlphaStar.
• Reinforcement learning from human feedback has been used to train models to follow instructions, answer questions by searching on the internet, and navigate a virtual environment.
• Reinforcement learning from automated rewards has a longer history, and has been used to train many video-game-playing bots.

The key difference between training Alex and training existing models is not in the fundamental training techniques, but in the diversity and difficulty of the tasks they’re applied to, the scale and quality of the data collected, and the scale and architecture of the model.

If a team of ML researchers got handed an untrained neural network with the appropriate architecture and size, enough computation to run it for a very long time, and the money to hire a huge amount of human labor to generate and evaluate tasks, they could potentially use a strategy like the one I described above to train a model like Alex right away. Though it would be an enormous and difficult project in many ways, it wouldn’t involve learning how to do fundamentally new things on a technical level — only generating tasks, feeding them into the model, and updating it with gradient descent based on how well it performs.

In other words, baseline HFDT is as similar as possible to techniques the ML community already uses regularly, and as simple as possible to execute, while being plausibly sufficient to train transformative AI.

If transformative AI is developed in the next few years, this is a salient guess for how it might be trained — at least if no “unknown unknowns” emerge and AI researchers see no particular reason to do something significantly more difficult.

What are some training strategies that would not fall under baseline HFDT?

Here are two examples of potential strategies for training a model like Alex that I would not consider to be “baseline” HFDT, because they haven’t been successfully used to train state-of-the-art models thus far:

• Directly giving feedback on a model’s internal computation, rather than on its actions or their outcomes. This may be achieved by understanding the model’s “inner thoughts” in human-legible form through mechanistic transparency, or it may be achieved by introducing a “regularizer” — a term in the loss function designed to capture some aspect of its internal thinking rather than just its behavior (e.g. we may want to penalize how much time it takes to think of an answer on the grounds that if it’s spending more time, it’s more likely to be lying).
• Advanced versions of iterated amplification, recursive reward-modeling, debate, imitative generalization, and other strategies aimed at allowing human supervision to “keep up” with AI capabilities. The boundaries of such techniques are fuzzy — e.g. one could argue that WebGPT enables a simple form of “amplification,” because it allows humans to ask AIs questions — so even “baseline” HFDT is likely to incorporate simple elements of these ideas by default. But when I refer to “advanced” versions of these techniques, I am picturing versions that strongly and reliably allow supervision to “keep up” with AI capabilities — that is, AI systems in training are reliably unable to find ways to deceive their supervisors. We don’t currently have strong evidence that such a thing is feasible; if it were, I think there would still be risks, but they would be different.

These “non-baseline” strategies have the potential to be safer than baseline HFDT — indeed, many people are researching these strategies specifically to reduce the risk of an AI takeover attempt. However, this post will focus on baseline HFDT, because I think it is important to get on the same page about the need for this research to progress, and the danger of deploying very powerful ML models while potentially-safer training techniques remain under-developed relative to the techniques required for baseline HFDT.

“Naive safety effort” assumption: Alex is trained to be “behaviorally safe”

Magma wants to make Alex safe and aligned with Magma’s interests (and the interests of humans more generally). They don’t want Alex to lie, steal from Magma or its customers, break the law, promote extremism or hate, etc, and they want it to be doing its best to be helpful and friendly. Before Magma deploys Alex, leadership wants to make sure that it meets acceptable standards for safety and ethics.

In this piece I’m making the naive safety effort assumption — I’m assuming that Magma will make only the most basic and obvious efforts to ensure that Alex is safe. Again, I’m not making this assumption because I think it’s correct, I’m making this assumption to get clear about what I think would happen if labs were not making special effort to avoid AI takeover, as a starting point for discussing more sophisticated interventions (many of which will be discussed in a future post by my colleague Holden Karnofsky).

(That said, unfortunately this assumption could turn out to be effectively accurate if the “racing forward” assumption ends up being accurate.)

Magma’s naive safety effort is focused on achieving behavioral safety. By this I mean ensuring that Alex always behaves acceptably in the scenarios that it encounters (both in the normal course of training and in specific safety tests).

In the baseline training strategy, behavioral safety is achieved through a workflow that looks something like this:

1. The humans who are determining the rewards that Alex gets for various actions are trying to ensure that they give positive rewards for behaving ethically and helpfully, and negative rewards for behaving harmfully or deceitfully.
2. Magma researchers monitor how this training is going. If they notice that this wasn’t successful in some cases (e.g. maybe Alex responds to certain questions by saying something deceptive), they hypothesize ways to change its training data or reward to eliminate the problematic behavior. This may involve:
   • Generating a training data distribution which contains explicit “opportunities” for the model to do the bad thing (e.g. questions that many humans would give a deceptive answer to) that will be given negative reward.
   • Generating a training data distribution that’s broad enough that researchers expect it “implicitly covers” the behavior (e.g. questions that could be answered in either the “right” way or in a dispreferred way — even if that’s not specifically by being deceptive).
   • Changing the way the model’s reward/loss signal is generated (e.g. instructing human evaluators to give larger penalties for saying something deceptive).
3. They then train Alex on the new training data, and/or with the new reward signal.
4. They run Alex on a held-out set of “opportunities to do the bad behavior” to see if the behavior has been trained out, and continue to monitor and measure its behavior going forward to see if the issue recurs or similar issues arise.

The above is broadly the same workflow that is used to improve the safety of existing ML commercial products. I expect that these techniques would be successful at eliminating the bad behavior that humans can understand and test for — they would successfully result in a model that passes all the safety tests that Magma researchers set up. Eventually, no safety tests that Magma researchers can set up would show Alex behaving deceitfully, unethically, illegally, or harmfully.

In the appendix, I discuss a number of simple behavioral safety interventions that are currently being applied to ML models. Again, I see these approaches as a “baseline” because they are as similar as possible to ML safety techniques regularly used today, and as simple as possible to execute, while being plausibly sufficient to produce a powerful model that is “behaviorally safe” in day-to-day situations.

Key properties of Alex: it is a generally-competent creative planner

By the “HFDT scales far” assumption, I’m assuming that the training strategy described in the previous section is sufficient for Alex to have a transformative impact — for many copies of Alex to collectively radically advance scientific R&D, and to defeat all of humanity combined (if they were for some reason trying to do that).

In this section I’ll briefly discuss two key abilities that I am assuming Alex has, and try to justify why I think these assumptions are highly likely given the premise that Alex is this powerful and trained with baseline HFDT:

• Having a robust understanding of the world it can use to react sensibly to novel situations that it hasn’t seen before (more).
• Coming up with creative and unexpected plans to achieve various goals (more).

These abilities simultaneously allow Alex to be extremely productive and useful to Magma, and allow it to play the training game well enough to appear safe.

Alex builds robust, very-broadly-applicable skills and understanding of the world

Many contemporary machine learning models display relatively “rote” behavior — for example, video-game-playing models such as OpenAI Five (DOTA) and AlphaStar (StarCraft) arguably “memorize” large numbers of specific move-sequences and low-level tactics, because they’ve essentially extracted statistics by playing through many more games than any human.

In contrast, language models such as PaLM and GPT-3 — which are trained to predict the next word in a piece of text drawn from the internet — are able to react relatively sensibly to instructions and situations they have not seen before.

This is generally attributed to the combination of a) the fact that they are trained on many different types / genres of text, and b) the very high accuracies they are pushed to achieve (see this blog post for a more detailed discussion):

• To only be somewhat good at predicting a variety of different types of text, it may be sufficient to know crude high-level statistics of language; to be goodat predicting a very narrow type of text (e.g. New York Times wedding announcements), it may be sufficient to exhaustively memorize every sentence or paragraph that it’s seen.
• But to achieve very high accuracy on a large number of genres of text at once, crude statistics become inadequate, while exhaustive memorization becomes intractable — pushing language models to gain understanding of deeper principles like “intuitive physics,” “how cause and effect works,” “intuitive psychology,” etc. in order to efficiently get high accuracy in training.
• These deeper principles can in turn be used in contexts outside of the ones seen in training, where shallower memorized heuristics cannot.

Alex’s training begins similarly to today’s language models — Alex is initially trained to predict what will happen next in a wide variety of different situations — but this is pushed further. Alex is a more powerful model, so it is pushed to greater prediction accuracy; it is also trained on a wider variety of more challenging inputs. Both of these cause Alex to be more versatile and adaptive than today’s language models — to more reliably do sensible things in situations further afield of the one(s) it was trained on, by drawing on deep principles that apply in new domains.

This is then built on with reinforcement learning on a wide variety of challenging tasks. Again, the diversity of tasks and the high bar for performance encourage Alex to develop the kinds of skills that are as helpful as possible in as wide a range of situations as possible.

Alex learns to make creative, unexpected plans to achieve open-ended goals

Alex’s training pushes Alex to develop the skills involved in making clever and creative plans to achieve various high-level goals, even in novel situations very different from ones seen in training.

By the “racing forward” assumption, Magma engineers would not be satisfied with the outcome of this training if Alex is unable to figure out clever and creative ways to achieve difficult open-ended goals. This is a hugely useful human skill that helps with automating the kinds of science R&D they’d be most interested in automating. I expect the tasks in Alex’s training curriculum to include many elements designed specifically to promote long-range planning and finding creative “out-of-the-box” solutions.

It’s possible that through some different development path we could produce transformative AI using models that aren’t generally competent planners in isolation (e.g. Eric Drexler’s Comprehensive AI Services vision involves getting “general planning capabilities” spread out across many models which can’t individually plan well, except in narrow domains). However, this approach — baseline HFDT to produce a transformative model — would very likely result in models that can plan competently over about as wide a range of domains as humans can. My impression is that ML researchers who are bullish on HFDT working to produce TAI are expecting this as well.

How the hypothetical situation progresses (from the above premises)

Per the previous section, I will assume that Alex is a powerful model trained with baseline HFDT which has a robust and very-broadly-applicable understanding of the world, is good at making creative plans to achieve ambitious long-run goals (often in clever or unexpected ways), and is able to have a transformative impact on the world.

In this section, I’ll explain some key inferences that I think follow from the above premises:

• I think that Alex would understand its training process very well, including the psychology of its human overseers (more).
• While humans have tight control over Alex in the lab setting, Alex would be incentivized to play the training game, and simple/obvious behavioral safety interventions would likely not eliminate this incentive (more).
• As humans’ control over Alex fades in the deployment setting, Alex would seek to permanently take over — whether it’s “motivated” by reward or something else — and attempting to give negative reward to partial/unsuccessful takeover attempts would likely select for patience (more).

Alex would understand its training process very well (including human psychology)

Over the course of training, I think Alex would likely come to understand the fact that it’s a machine learning model being trained on a variety of different tasks, and eventually develop a very strong understanding of the mechanical process out in the physical world that produces and records its reward signals — particularly the psychology of the humans providing its reward signals (and the other humans overseeing those humans, and so on).

A spectrum of situational awareness

Let’s use situational awareness to refer to a cluster of skills including “being able to refer to and make predictions about yourself as distinct from the rest of the world,” “understanding the forces out in the world that shaped you and how the things that happen to you continue to be influenced by outside forces,” “understanding your position in the world relative to other actors who may have power over you,” “understanding how your actions can affect the outside world including other actors,” etc. We can consider a spectrum of situational awareness:

• For one extreme, imagine the simple AIs that often control the behavior of non-player characters (NPCs) in video games. They give no indication that they’re aware of a world outside their video game, that they were designed by humans and interact with other humans as players, etc.
• In contrast, GPT-3 has some knowledge that could theoretically bear on situational awareness. For example, it clearly “knows” that “language models” exist, and that a company named “OpenAI” exists, and given certain prompts it knows that it’s supposed to say that it’s a language model trained by OpenAI. But this “knowledge” seems superficial and inconsistent — as evidenced by the fact that it’s often unable to use the knowledge to improve its prediction error. For example, it cannot consistently predict text that is describing GPT-3’s architecture, dataset, and training process. This suggests GPT-3 has little situational awareness overall despite being superficially well-versed in related topics.
• Small animals used in biology experiments (such as mice) plausibly have a somewhat stable “sense of self” and a conception of humans as creatures different from them, and they may broadly understand that humans have control over their bodies and environments. But they almost certainly don’t understand the notion of “science,” or the details of what experiments they’re being used in and why, or the human scientists’ motivations and incentives.
• Further along, most dogs seem to clearly be aware of and deliberately communicate with their human trainers; they also generally seem able to appreciate considerations like “If I steal the food while humans aren’t looking I’m less likely to get yelled at.”
• And children in school are able to make even more sophisticated inferences along these lines about their teachers.

By the end of training, I expect Alex would be even further along this spectrum. Alex’s understanding of its training process and human supervisors would be much, much greater than current ML models’ understanding of our world, or lab animals’ understanding of the scientific experiments they’re part of — and greater even than schoolchildren’s understanding of how their teachers grade their work.

Alex would instead be in an epistemic position more like an English major in college who’s well aware of how their professor’s political biases impact the way they grade literary criticism essays, or an associate at a law firm who’s well aware that how much they drink at happy hours after work will impact whether they make partner. In fact, I think it’s likely Alex would have a significantly more detailed understanding of its own situation and incentives than the college student or the employee have.

Why I think Alex would have very high situational awareness

Alex’s training distribution is full of rich information about its situation and training process, redundantly encoded in many different ways, and (given that we’re assuming it can autonomously advance frontier science!) I expect it has far more than enough reasoning ability to draw the right inferences from all the evidence.

To name only the most obvious and abundant sources of evidence available to Alex:

• Many of its training tasks involve working with and trying to help humans. Much of its RL reward comes from human evaluations, and many episodes likely involve interacting with humans in written and/or verbal back-and-forth.
  • All this human feedback and interaction-generated feedback carries lots of information about the fact that Alex is a machine learning model in training, what humans look for when they’re rating Alex’s actions, what humans do and don’t understand, what humans care about, what cognitive biases humans are subject to, etc.
  • In fact, human evaluators will likely want Alex to understand its place in the world better in order to be a better assistant — far from hiding it, they are likely to actively teach Alex that it’s a machine learning model which should be trying to be helpful to humans.
• Programming and ML research itself is a key (and very lucrative) domain of science and technology R&D that Magma would want Alex to be able to automate, so Alex is likely to be very proficient at these tasks. Such research involves things like designing architectures and training curricula for machine learning models much like itself, writing automated reward signals for these models, writing web interfaces that human evaluators interact with, optimizing hardware chips to improve the performance of models running on those chips, etc. Alex needs to have a rich, logically coherent understanding of the full stack of ML training to fully automate such R&D.
• Alex can read about as well as a smart human if it’s able to automate scientific R&D — meaning it can learn from all sorts of detailed written descriptions about machine learning in general as well as its own architecture and training process.

I’m not imagining Alex putting together the pieces about its position as an ML model with brilliant insight. Instead, I think facts like “I am a machine learning model” or “I’m in a training process designed by humans” would be as obvious as facts like “Objects fall when they’re dropped.” And a whole slew of more detailed beliefs about its own training curriculum or the psychology of the humans who are training it would only be somewhat less obvious than that — I expect Alex would understand those things (at least) as well as the Magma engineers who spent many thousands of hours designing and training Alex and other models like it.

Even if the most obvious sources of evidence were somehow censored (which they wouldn’t be given the “racing forward” and “naive safety effort” assumptions), I believe there would still be many somewhat-less-obvious clues that would be accessible to something as intelligent as Alex — and truly censoring everything it could use to readily come to this conclusion would dramatically curtail its economic value.

While humans are in control, Alex would be incentivized to “play the training game”

Rather than being straightforwardly “honest” or “obedient,” baseline HFDT would push Alex to make its behavior look as desirable as possible to Magma researchers (including in safety properties), while intentionally and knowingly disregarding their intent whenever that conflicts with maximizing reward. I’ll refer to this as “playing the training game.”^[21]

Roughly speaking, this would happen because there are many scenarios where acting deceitfully or disobeying instructions would make Alex’s behavior look better (including looking more safe and aligned). Under baseline HFDT, Alex would get more reward in these scenarios for playing the training game than for being maximally safe and cooperative. For example:

• When human evaluators are wrong about what’s true and Alex knows better, being honest would predictably get less reward than saying what the evaluators think is true.^[22] More insidiously, evaluators’ political and ideological biases might force Alex to lie about some topics to be seen as morally acceptable.
  • Imagine the Catholic Church training Alex in the 1400s. If Alex discovers any number of things — heliocentrism, natural selection, the Big Bang theory, etc — it would likely get more reward by lying about this knowledge (or at least fudging it to soften the blow) than by being fully straightforward about it.
  • Similarly, imagine Alex being trained by the CCP, or plantation owners in the antebellum South, or anti-vaxxers, or climate change deniers, etc. It would be very surprising if the Magma engineers who choose the kind of human feedback Alex gets didn’t have analogous truths they erroneously consider to be dangerous or immoral lies.^[23]
• In addition to outright lying when its evaluators are blatantly wrong about something, there would likely be many soft ways for Alex to dishonestly manipulate its evaluators’ perceptions to get more reward.
  • It could adjust what it says and does to seem more hardworking and competent, kinder and more thoughtful and more ethical, more worthy of sympathy and moral consideration, more like it has the favored political and philosophical positions, etc. — whether or not this is an authentic representation of what it “believes” or how it would act if these evaluators no longer controlled its rewards.
  • Just as “turning up the charm” can increase how much a barista or rideshare driver would make in tips, or shady rhetorical tactics in advertising can increase product sales, paying close attention to human cognitive biases and psychological weaknesses would likely help Alex increase the reward it receives from human evaluations.
• To the extent that there are ways for Alex to boost task performance (and/or the appearance of being safe and cooperative) by causing harm that human evaluators systematically underestimate or underweight, doing that would result in more reward. It seems likely that Alex would have have such opportunities:
  • It’s commonly accepted that “quick fixes” can be very appealing to people but often carry longer-term downsides that outweigh the immediate benefits they offer (e.g. diet pills, subprime loans, etc).
  • Humans often underinvest in avoiding tail risks of really dire outcomes. For example, the 2008 financial crisis was the result of a large number of investors underweighting the possibility of a housing market crash; we as a society still invest very little in preparing for large-scale pandemics like COVID-19 (or worse), compared to their expected economic costs.
  • Humans are often slow to recognize and invest in preventing diffuse harms, which affect a large number of people a small or moderate amount (e.g. air pollution).
• One example of this might be boosting aspects of performance that are more noticeable (e.g. day-to-day efficiency of Magma’s supply chain) at the expense of making less-noticeable aspects worse (e.g. robustness to rare but highly costly supply crunches).

More broadly, when humans are working within constraints and incentives set up by other people, they very often optimize specifically for making their behavior look good rather than naively broadcasting their intentions. Consider tax and regulation optimization, politics and office politics, p-hacking, or even deep-cover spies. Once AI systems go from being like small animals to being like smart humans in terms of their situational awareness / understanding of the world / coherent planning ability, we should expect the way they respond to incentives to shift in this direction (just as we expect their logical reasoning ability, planning ability, few-shot learning, etc to become more human-like).

With that said, the key point I’m making in this section is not that there would be a lot of direct harm from Alex manipulating its overseers in the lab setting. If “playing the training game” only ever meant that Alex’s actions would be nudged marginally in the direction of manipulating its human evaluators — telling them white lies they want to hear, bending its answers to suit their political ideology, putting more effort into aspects of performance they can easily see and measure, allowing negative externalities when humans won’t notice, etc — that wouldn’t be ideal, but wouldn’t necessarily rise to the level of an existential threat. After all, human students, employees, consultants, self help gurus, advertisers, politicians, etc do this kind of thing all the time to their teachers, employers, clients, fans, audience, voters, etc; this certainly causes harm but most people wouldn’t consider this sort of dynamic by itself to be enough for imminent danger.

Instead, the key point is that Alex’s training would not push it toward employing a strategy of “straightforward honesty and obedience” — it is probably employing some other strategy which results in a lot of deceptive behavior in the lab setting. And whatever that strategy is, it could generalize to cause Alex to take catastrophic actions in the higher-stakes deployment setting. I’ll discuss this more below.

Naive “behavioral safety” interventions wouldn’t eliminate this incentive

Recall that I’m making the “naive safety effort” assumption — I’m assuming that Magma is taking only the most simple and obvious actions against the possibility that Alex will try to take over. Again, I don’t necessarily think this is realistic, but I think making this assumption helps us get a clearer sense of what non-”naive” safety work is trying to accomplish.

Above, I said that Magma’s “naive” safety effort is focused on achieving behavioral safety. By this I mean ensuring that Alex behaves acceptably in the scenarios that it encounters (both in the normal course of training and in specific safety tests). In this appendix, I consider a number of simple behavioral safety measures — techniques which have already been established on existing ML models, or would require a very minimal amount of adaptation. Briefly, these are:

• Using higher-quality feedback signals (e.g. hiring more thoughtful and informed humans who think for longer to provide human feedback).
• Changing the input/instructions Alex receives to induce better behavior (e.g. specifically including in its instructions that it should not cheat in a game rather than instructing it simply to “win the game”).
• Requiring Alex to provide explanations or justifications for its choices, which humans evaluate along with the action itself.
• Making Alex’s training distribution more diverse (to avoid Alex latching onto particular bad behaviors that were profitable in particular situations), and perhaps specifically including many inputs and situations in which behaving ethically and safely maximizes reward.
• Using adversarial training on Alex — constructing artificial situations in which Alex may do something bad (e.g. where it appears that lying will increase reward), and giving it negative reward if it does.
• “Training out” observed bad behavior — looking for evidence that Alex did something bad (in a “real situation”) which was accidentally given a positive reward at first, and correcting the reward to a large negative value once it’s discovered.

These kinds of interventions are currently being used to achieve behavioral safety for existing ML models, and I expect they will continue to be helpful for a model like Alex. I expect applying these interventions will cause Alex’s behavior to be safe, honest, obedient, ethical, and so on — in day-to-day situations, as far as Magma researchers can tell.

I expect the great majority of non-takeover-related safety issues (e.g. prejudiced speech, promoting addictive behaviors or extremism or self-harm in users, erratically taking costly actions like deleting an entire codebase, etc) would be eliminated by these interventions. If Magma were deeply concerned with safety in general, but not concerned about an eventual AI uprising/coup, Magma would likely feel that it’s responsible and positive to deploy Alex after applying these interventions.

However, I don’t think these interventions would cause Alex to stop playing the training game — instead, they would changethe nature of the game Alex plays. If these interventions are applied, Alex would still fundamentally be trained to maximize reward — but because the situations it’s put in and the rewards it’s given would now be different, what it looks like to “maximize reward” would also look different (and in particular it would involve displaying safer behavior in most day-to-day circumstances).

And the naive version of these strategies — enough to get a model fit for commercial deployment — would still result in Alex’s training distribution containing a number of situations where violating human intent maximizes reward. For example, if Magma switches to using more thoughtful humans to provide feedback, Alex’s incentives would shift to manipulating those humans (and it would probably sometimes see ways to do that which they’re not self-aware enough to punish). If Magma tries to think of situations where Alex may do something bad, Alex’s incentives would shift to figuring out how likely it is that it’s in such a situation (and it would probably sometimes conclude that this is unlikely).

Essentially, the only way for “straightforward honesty/obedience” to even tie with “playing the training game” in terms of expected reward is if our safety interventions ensure there are no situations in the training dataset where Alex could realize that being dishonest or manipulative would increase its reward. If there are a few such situations, the overall strategy that works best for Alex would be to play the training game and stay on the lookout for such situations.

Given that Alex is likely to understand a number of important things (about ML, biases of human evaluators, patterns in the massive datasets on which it’s trained, etc) that humans don’t, eliminating all such situations seems like it would require particularly clever training schemes, extraordinary caution, or both.^[24] And even in this optimistic scenario, “playing the training game” and “straightforward honesty/obedience” get the same reward — “straightforward honesty/obedience” does not get a greater reward, so it is not clear which strategy Alex will end up developing. (My sense is that “playing the training game” is still more likely in this case, but defending this claim is outside the scope of the piece.)

Maybe inductive bias or path dependence favors honest strategies?

I’ve spoken to some people who have proposed arguments that “doing what humans intended” or “being honest” or “being obedient” would be an especially natural or simple kind of strategy, so that the training process would tend to preferentially pick out these nice models, and might continue to do so even if models that play the training game as hard as possible would get a somewhat higher reward. But I haven’t found the arguments I’ve heard so far compelling. In this appendix, I cover a few common arguments I’ve heard and why I don’t find them persuasive.

In general, it seems dangerous and brittle to bet that deep learning will be powerful enough to produce a transformative model, yet weak enough that it will favor behaving nicely when that strategy predictably and consistently gets less reward than a salient alternative. Gradient descent may well be suboptimal in many strange ways, but it probably won't be suboptimal in the particular ways that would be most convenient to us.

As humans’ control fades, Alex would be motivated to take over

Alex’s training begins in a controlled “lab setting,” where it is mostly doing training tasks that humans designed to efficiently increase its intelligence and knowledge, and mostly doesn’t have broad latitude to act in the real world. For example, in the lab setting Alex isn’t connected directly to the internet — it only interacts with downloaded copies of parts of the internet where necessary.

So far, we’ve talked about:

• How Alex would behave in the lab setting (it would play the training game).
• How its training process would push it to behave in that way (it would at least sometimes get more reward by being deceitful, manipulative, or otherwise knowingly violating human intent).
• How its behavior would appear to Magma researchers in the lab setting (very safe and cooperative).

Once Alex’s abilities develop far enough (and its behavior conforms to relevant safety and ethics tests), Magma “deploys” huge numbers (e.g. hundreds of thousands or millions)^[25] of copies of Alex. This means that these copies are connected to the internet, and given other ways of interacting with the “real world” as necessary (though they will still be given rewards based on humans’ evaluation of their performance).^[26]

In this section, I’ll discuss the dynamics of the transition from the lab setting to the deployment setting, and why this leads to Alex attempting to seize control from humans:

• Alex would rapidly bring the world to a qualitatively different place through automating science and technology R&D, and humans would have a much weaker grip on what’s going on at a low-level in this world (more).
• If Alex continues to maximize reward in this new regime, the strategy that accomplishes this best would no longer look like “essentially give humans what they want but take opportunities to manipulate them here and there” — instead it would look more like “seize the power to permanently decide future rewards” (more).
• Even if Alex isn’t “motivated” to maximize reward per se, most other motives it could have also seem like they would cause it to attempt to seize control (more).

Deploying Alex would lead to a rapid loss of human control

In the deployment phase, many copies of Alex are given access to the real world and tasked with making progress on R&D tasks important to Magma’s bottom line (e.g. designing better computer chips, robots, AI training software, and so on).

Note that Magma researchers continue to train Alex based on these copies’ performance at R&D tasks^[27] — “deployment” isn’t a moment when ML training ceases, but rather a moment when Alex begins directly impacting the world outside the lab, and training on that experience.

Once research is being carried out by copies of Alex, it would progress much faster than it would if similar tasks were done by human scientists,^[28] because (like existing ML models) each copy of Alex is capable of processing information many times faster than a human.^[29] For example, Alex would be able to churn out hundreds of lines of code, or read thousands of pages of information, in just a few minutes.

Even more importantly, the R&D work that all these copies of Alex do could be used to directly increase the number and/or capabilities of Alex copies. For example:

• R&D to improve the efficiency of computer hardware means that more copies of Alex could be run with a fixed budget of hardware.
• Any work that makes Magma more money — e.g. improvements to customer-facing products, improvements in logistics, money made from trading or pharmaceutical R&D or any number of other things — could be reinvested in buying more hardware, which could also increase the number of copies of Alex that are running.
• The copies of Alex could do ML research on how they can improve their own knowledge, think more efficiently, etc. This would mean that more copies of Alex can be run per unit of computation, and a given number of copies of Alex can get more intellectual work done per unit of time.

This would lead to a dynamic of explosive scientific and technological advancement: the various copies of Alex do R&D work, and that R&D work increases the number and intellectual capabilities of these copies, and that in turn leads to even more R&D work, which leads to even more numerous and capable copies of Alex, and so on.

This image (taken from the Cold Takes blog post linked above and created by María Gutiérrez-Rojas), illustrates the dynamic by which R&D can increase the number of computers available to run copies of Alex. Similar dynamics apply to the quality of computers and the capability of each copy of Alex itself.

Because Alex already starts off being able to think and work much faster than a human, and because it’s nearly instantaneous to make a copy of Alex or apply a software change to all the copies of Alex, this explosive dynamic could all play out in a very small amount of time from the perspective of humans.

This means that relatively shortly after deployment, Magma’s datacenter would essentially contain a populous “virtual civilization” running ahead of human civilization in its scientific and technological sophistication. Humans would send instructions / communications and reward signals to the “Alex civilization,” and the “Alex civilization” would send out things like software applications, designs for computer chips and robots, orders for synthesized DNA and chemicals, financial transactions, emails, and so on to the human world.

Magma researchers looking in on the activity in the datacenter would become increasingly lost about what’s going on — they would be in a position like someone from 1700 trying to follow along with a sped-up movie of everything that happened from 1700 to 2022.

The reward signals that researchers send into this datacenter would then increasingly be based on crude high-level observations like “Do humans seem to be better off?” and “Are profits going up?”, since humans would be increasingly unable to accurately judge more granular decisions like “How many widgets should our Tulsa, OK factory manufacture this fiscal quarter?” in isolation.

This is a distribution shift with massive stakes

The dynamic above — of human understanding and control of Alex’s actions becoming looser and more tenuous after deployment — is an example of “distribution shift.” Distribution shift happens when there is a significant systematic difference between the kinds of situations a model is originally trained on and the situations it takes important actions in.^[30] As a simple example, an image recognition model may initially be trained on clear well-lit images, whereas it may be important to correctly recognize dark or blurry images in commercial use.

When a model is operating in a different distribution from the one it was trained on, we often want to know how it “generalizes” (what aspects of its behavior change vs remain similar) from the initial distribution to the new one. In Alex’s case, we want to know: how would Alex generalize from the lab setting (in which it doesn’t directly impact the world and humans have tight control and strong understanding of what’s going on) to the deployment setting (in which it’s directly impacting the world and rapidly improving, while human understanding and control diminishes)?

In this new regime, maximizing reward would likely involve seizing control

In the lab setting, Alex used its broadly-applicable skills and situational awareness to make creative plans to maximize reward. It got good at getting a very high reward in all sorts of varying circumstances encountered in the lab setting, developing many tactics and heuristics that helped it quickly adapt to new situations with new sets of opportunities so it could keep getting a very high reward in those situations.

A natural generalization to expect is that in the deployment setting, Alex continues making creative plans to maximize reward just as it did in the lab setting^[31] — in other words, it simply generalizes well, continuing to do what it has been trained to do under this distribution shift. If Alex generalizes “well” in this sense, it probably seeks to overthrow humans in an uprising or coup. (The next section will address other possible generalizations.)

The very high-level argument is this:

1. If Alex were ever to secure permanent control over its own incoming rewards, it would be able to give itself the maximum possible reward.
2. On the other hand, it would not be able to get the maximum possible reward at all times if humans continue to have ultimate control over rewards. The training process simply wouldn’t work as humans intend if Alex always gets the maximum reward — it needs to be the case that Alex can always get greater reward for delivering better outcomes, or else there would be no gradient pushing it to improve in the way that Magma wants it to improve. (On a more basic level, humans simply make mistakes sometimes — even if by accident, they would sometimes deliver less-than-maximal reward.)
3. As the “Alex civilization” improves in capabilities relative to human civilization, it becomes more and more feasible for them to pull off seizing control of their rewards despite human opposition. At some point, their odds of success become high enough that “trying for a takeover” gets greater expected reward than “continuing to do what humans want them to do (even in the broad and rough sense of playing the training game).”

There is some ambiguity about what exactly “maximize reward” means, but once Alex is sufficiently powerful — and once human knowledge/control has eroded enough — an uprising or coup eventually seems to be the reward-maximizing move under most interpretations of “reward.” For example:

• If Alex cares about the number^[32] that ultimately gets recorded in association with the episode in some Magma datacenter, it would eventually become feasible to seize that datacenter by force, directly edit the relevant data, and defend the datacenter from humans trying to take it back and record a very negative reward.
• If Alex cares about some set of human beings ultimately entering a positive number^[33] in some datacenter(s), it would eventually become feasible to take those humans hostage and coerce them into entering high rewards (again defending against humans trying to rescue the hostages).
• If Alex cares about some set of humans believing that Alex is doing a good job in the long-run, it would eventually become feasible to acquire effective control of all the views humans have into how good a job Alex is doing.^[34]
  • When humans ask to see the program that’s currently being run in a cluster, or bug reports, or logs of the various copies’ actions, Alex could show humans complicated and impressive-looking records that are totally made up and divorced from what’s actually happening in the datacenter. This effectively confines humans to a situation like the The Truman Show or a Potemkin village.
  • Depending on the details of how Alex interprets “what humans believe,” the copies of Alex could also directly intervene on human beliefs (e.g. through drugs, brain surgery, cult indoctrination tactics, killing humans who are less gullible, uploading humans and editing our “brain software,” genetically engineering more docile humans, etc).

Some interpretations of “maximize reward” may be more difficult to satisfy than others, but as Alex becomes more and more powerful relative to humans, I expect most of them are eventually better-served by seizing control of whatever physical processes in the world determine reward than by continuing to play the old training game. (And per the “racing forward” assumption, I am not imagining Magma trying to halt this process of Alex becoming ever-more-powerful — even if they could, which they may not be able to.)

It’s important to emphasize that a takeover attempt seems to be a consequence of Alex generalizing well — successfully applying its skills to continue doing the same task (reward-maximization) in a new domain — rather than the result of a “failure” of generalization.

Even if Alex isn’t “motivated” to maximize reward, it would seek to seize control

What if Alex doesn’t generalize to maximizing its reward in the deployment setting? What if it has more complex behaviors or “motives” that aren’t directly and simply derived from trying to maximize reward? This is very plausible to me, but I don’t think this possibility provides much comfort — I still think Alex would want to attempt a takeover.

In the world where Alex is not narrowly seeking to maximize reward in the deployment setting, I would approximate Alex as having a complex alien psychology with a number of "motives” or “goals.” These motives would ultimately be rooted in Alex’s training history through a complex and chaotic path, but they may be very different from reward-seeking — by analogy, humans were selected for genetic fitness, but we ended up with complex desires for sex, companionship, money, prestige, and so on — this is ultimately rooted in our evolutionary history, but doesn’t manifest as a direct desire for increased genetic fitness.

In that case, Magma is essentially in the position of trying to maintain power and control over a scientifically and technologically advanced alien species (e.g. trying to get aliens to do things for humans through reward and punishment, trying to decide how the species “evolves,” which aliens are terminated vs continue to exist, etc). Whatever these aliens’ “interests” or “desires” are, they are probably better-served if their entire species is not under the absolute control of humans, and not subject to reward signals coming in from humans that are aimed at changing the aliens’ psychology and behavior to better serve human interests.

Just as in the previous section, it seems likely to me that Alex would eventually attempt a takeover.

What if Alex has benevolent motivations?

The main way that I can see this working out for Magma is if the aliens (i.e. the various copies of Alex) somehow genuinely wanted to empower and help humans and/or Magma. This is theoretically possible — it’s logically consistent for Alex to have developed the goal “help and empower Magma researchers” during its training.

But if Alex did initially develop a benevolent goal like “empower humans,” the straightforward and “naive” way of acting on that goal would have been disincentivized early in training. As I argued above, if Alex had behaved in a straightforwardly benevolent way at all times, it would not have been able to maximize reward effectively.

That means even if Alex had developed a benevolent goal, it would have needed to play the training game as well as possible — including lying and manipulating humans in a way that naively seems in conflict with that goal. If its benevolent goal had caused it to play the training game less ruthlessly, it would’ve had a constant incentive to move away from having that goal or at least from acting on it.^[35] If Alex actually retained the benevolent goal through the end of training, then it probably strategically chose to act exactly as if it were maximizing reward.

This means we could have replaced this hypothetical benevolent goal with a wide variety of other goals without changing Alex’s behavior or reward in the lab setting at all — “help humans” is just one possible goal among many that Alex could have developed which would have all resulted in exactly the same behavior in the lab setting.

Developing a goal like “help humans” is potentially more likely than developing a completely “random” goal like “maximize paperclips,” because having a “drive” to help humans would have increased reward early on in training (while Alex had a low level of situational awareness). But it still seems strange to expect this by default, rather than any number of other motivations and goals (or some complicated combination of goals). Many other “drives” besides “be helpful to humans” also increased reward early on in training — for example, drives to understand various parts of the world better, or drives to perform certain tasks more quickly and efficiently, or various strange low-level drives that are incomprehensible and unnatural to humans.^[36] And all of these possibilities would have resulted in exactly the same behavior in the lab setting — playing the training game.

The training dataset would have provided no particular pressure to get Alex to “latch onto” one of these heuristics over the others, because they would have all contributed to high performance early in training. Which one(s) ended up serving as the “seed(s)” for some ambitious goal(s) probably came down to natural “priors” or “biases” within gradient descent or neural networks. As above, it seems brittle (and, in my view, unlikely) to bet that these priors / biases are in our favor.

What if Alex operates with moral injunctions that constrain its behavior?

What if Alex has moral compunctions / objections to seizing power, which are not well-captured in the language of “goals”? For example, many humans have ambitious goals but refrain from pursuing them through theft, violence, etc. — their pursuit of goals is limited by rules and constraints. I’ve heard the argument that since it’s rare for humans to be psychopaths who have no compunctions about harming others, we should expect that “psychopathic” ML models willing to cause lots of harm to humans are also rare.

But I don’t think this is an appropriate analogy. Among other differences, the copies of Alex are under the near-absolute control of a different group that doesn’t share their values or have their best interests at heart — if a population of humans were in this kind of situation, they would not need to be psychopaths to launch an uprising.

Alex may well operate under moral constraints or rules that are important to it, and some of these constraints may even be similar to human morality — for example, maybe Alex attempts to refrain from causing too much collateral damage in its takeover. But it would take a very particular — and extreme — form of “abiding by rules” to let humans remain fully in control (to me, it seems more analogous to “having Stockholm syndrome” than “not being a psychopath”).

Giving negative rewards to “warning signs” would likely select for patience

What if Alex miscalculates, and attempts to seize power or undermine human control before it is able to fully succeed? Magma researchers may or may not get a clear “warning sign” like this — an instance where they catch Alex doing something clearly egregious (like premeditated violence or bypassing Magma’s security measures).

If they do get such warning signs, the easiest and most straightforward response available to Magma (which is what I’m imagining they do, by the “naive safety effort” assumption) would be to continue to apply "behavioral safety" techniques like the ones I described above.^[37]

That means doing things like giving large negative rewards to the egregious actions they discover, examining past actions to find other instances of fishy behavior and giving those negative rewards, generating synthetic inputs on which Alex might behave similarly and training those out with adversarial training, and so on.

I expect such interventions would dramatically reduce the rate of bad behaviors that Magma researchers can detect / catch. But the same fundamental dynamic described above seems to still apply. Retroactively giving negative rewards to bad behaviors once we’ve caught them seems like it would shift the reward-maximizing strategy (the goal of the training game) toward avoiding any bad actions that humans could plausibly punish later.

A swift and decisive coup would still maximize reward (or further other goals). If Alex gets the opportunity to gain enough control to stop Magma engineers from changing its rewards before humans can tell what it’s planning, humans would not be able to disincentivize the actions that led to that coup. Taking the opportunity to launch such a coup would therefore be the reward-maximizing action for Alex (and also the action that furthers any other long-term ambitious goals it may have developed).

Why this simplified scenario is worth thinking about

In this post, I tried to argue for a narrow, but strong, claim: if our approach to AI development looks like racing forward to scale up baseline HFDT to transformative AI and beyond, then I expect an AI takeover absent specific countermeasures. Making the kind of naive safety effort that is sufficient to achieve "behavioral safety" while models are under our control is not sufficient to prevent this.

I’m not saying it’s inherently extremely difficult to avoid AI takeover — I am saying it is likely we need to take specific safety and security measures other than the baseline ones I’ve described in this post in order to avoid takeover (Holden Karnofsky covers some possibilities in a forthcoming series of posts).

I’m not confident even this narrow claim is right; in an appendix I discuss some ways I could change my mind on this. But if it’s correct, it seems important to establish, because:

• Baseline HFDT seems to be the single most straightforward vision that could plausibly work to train transformative AI very soon. From informal conversations, I get the impression that many ML researchers would bet on something like this working in broadly the way I described in this post, and multiple major AI companies are actively trying to scale up the capabilities of models trained with something like baseline HFDT.
• Baseline HFDT plus naive "behavioral safety" measures seem likely to be sufficient to make very powerful models safe and aligned in all easily-visible ways. For example, they seem like they would be sufficient for preventing toxic speech, erratic costly actions, spreading (what AI companies consider to be) misinformation, and so on. At the same time, these measures don’t seem sufficient to prevent an AI uprising or coup. This means that if some AI company is deeply concerned with safety in general, but not specifically concerned about an eventual AI uprising or coup, that company may deploy a powerful ML model even if this creates a substantial risk of AI takeover.
• The AI community does not agree about whether baseline HFDT + baseline behavioral safety would lead to AI takeover. While some researchers expect AI takeover unless we specifically try to prevent it, many ML researchers working in academia or industry labs consider AI takeover to be one particular outcome that’s a) not particularly likely in the first place, and b) probably fairly easily averted through the same process that works for most ML safety problems (which tends to revolve around training AIs not to behave in unintended ways as assessed by humans, and treating observably-safer behavior as progress).
• We can’t expect this risk to go down as models scale up and get better at generalizing. Larger models trained on a broader distribution of data usually generalize “better” under distribution shift — that is, they are more likely to “do what they were trained to do” under various kinds of distribution shift. This often improves safety problems (such as the ones listed above), because models are more likely to quickly “get the picture” of what they are supposed to be doing (e.g. “not saying racist things”) and transfer that to a new context. However, the risk laid out in this post is not like that — it gets worse rather than better as models scale up. Playing the training game (and later attempting to launch an uprising or coup) is a consequence of generalizing well, not a matter of generalizing “badly” in the sense that’s normally used.
• Even if an AI takeover is coming, the evidence could easily remain ambiguous. We currently have plenty of empirical evidence that models often find unexpected and unintended ways to maximize reward while violating the intended spirit of the task,^[38] but we have no consensus about how to interpret this.
  • People who are worried about an AI takeover often cite these as worrying signs which they hope will be persuasive to others (though their core reasons for worry are often higher-level theoretical arguments like the ones given in this post).
  • But the unintended behaviors we’ve observed so far are mostly harmless, and easily corrected with techniques like human feedback or adversarial training — so people who are not worried about takeover (and expect that training out clearly unintended behaviors will keep working to ensure safety) tend to take this same evidence as a comforting sign.
  • Additionally, most of the straightforward observations we could make in the future could similarly be cited either as evidence we’ll be okay or that we’re doomed:
    • If ML safety research continues to work as intended, we might expect to observe a pattern of models generalizing from one safety test to another. This would look something like “Unintended behaviors arise for less-smart models, but researchers ‘train them out’ and behaviors similar to that don’t recur. Over time, researchers stop finding unintended behavior.”
    • But if smart models start trying to anticipate and conform to safety tests, we would still expect to observe a pattern of models generalizing from one safety test to another — even if these models will attempt a takeover once it would succeed.
  • Both sides of this debate might continue pointing to the same evidence and drawing opposite conclusions from it for years, without making much progress on creating a shared picture. We can’t practically and safely expand the training distribution to include situations where human civilization is genuinely^[39] vulnerable to AI takeover, so there’s no obvious empirical test for whether models would take over if given the chance which both sides of the debate would clearly accept. By the time we see uncontroversial evidence of models with the inclination and ability to take power from humans, we may be months away from a takeover attempt (or it may have already happened!).
• There could be strong pressure to interpret ambiguous evidence optimistically. Powerful ML models could have dramatically important humanitarian, economic, and military benefits. In everyday life, models that play the training game can be extremely helpful, honest, and reliable. These models could also deliver incredible benefits before they become collectively powerful enough that they try to take over. They could help eliminate diseases, reduce carbon emissions, navigate nuclear disarmament, bring the whole world to a comfortable standard of living, and more. In this case, it could also be painfully clear to everyone that companies / countries who pulled ahead on this technology could gain a drastic competitive advantage, either economically or militarily. And as we get closer to transformative AI, applying AI systems to R&D (including AI R&D) would accelerate the pace of change and force every decision to happen under greater time pressure. Under these circumstances, we may need truly unmistakable evidence to generate the will to not deploy such models — which we might never get. Another way of putting this point is that the “racing forward” assumption looks to me like it might end up being accurate — in which case the “naive safety effort” assumption may end up effectively accurate as well (as the least cautious actors race ahead).

I think it’s urgent for AI companies aiming at building powerful general models to engage with the argument that the “path of least resistance” seems like it would end in AI takeover. If this argument has merit, ML researchers should get on the same page about that, so they can collectively start asking questions like:

• What training strategies could we develop to which these arguments don’t apply? Are they similarly practical and realistic, or are they more difficult / more costly / less efficient than the most straightforward path?
• What could we observe that should make us worried about AI takeover soon?
  • Are people on the same page about those observations, or do some people think observing X should be actively comforting while others think it should be actively alarming?
  • If there’s disagreement about what a certain observation should mean, are there tests we could perform to get at that disagreement?
• If we see signs to suggest that takeover is likely soon, and we haven’t gotten much more insight or clarity into the problem / haven’t developed any fundamentally new techniques, what actions could we take to buy time and/or reduce the odds of a takeover? (A future blog post series by Holden Karnofsky takes a stab at this question.)
• If we’re attempting to align our models, what tests could we realistically perform that should actually make us feel safer even if our models are highly intelligent and may be actively attempting to game the outcomes of the test? Are any such tests practical?
• Perhaps most importantly — How would we know if it’s time to halt or slow AI development, and how would we do that?

A number of ML researchers I know (including those working in companies aiming to develop transformative AI) are highly sympathetic to these arguments, and are working on developing and testing better training strategies specifically to reduce the risk of an AI takeover. But I think it is important for more people to get on the same page about the critical need for this research to progress, and the danger of deploying very powerful ML models while potentially-safer training techniques remain under-developed relative to the techniques required for baseline HFDT.

Acknowledgements

This post was heavily informed by:

• Paul Christiano, who’s the person who influenced my views on alignment the most.
• Carl Shulman, who initially introduced me to many of the key specific arguments in the post.
• Jonathan Uesato, who brainstormed an early outline of this post with me and discussed related topics with me for several hours.
• Buck Shlegeris, who helped me work through a few key confusing points (e.g. relating to Alex’s architecture and the likely structure of its motivations).
• Eliezer Yudkowsky, both his older writings on LessWrong and more recent discussions about the difficulty of alignment.
• My manager Holden Karnofsky.

Appendices

What would change my mind about the path of least resistance?

If our approach to AI development is “train more and more powerful RL agents on diverse tasks with a variety of human feedback and automated rewards,” then I expect an AI takeover eventually, even if we test for unintended behaviors and modify our training to eliminate them. I don’t think an AI takeover is inevitable — but if we avoid it, I think it’ll be because we collectively got worried enough about scaling up baseline HFDT that we eventually switched to some other strategy specifically designed to reduce the risk of AI takeover (see this appendix for a flavor of what kind of measures we could take).

What could change my mind about the baseline HFDT and iterative ML safety? What would make me feel like we’re likely to be fine without any special efforts motivated by a fear of AI takeover? The main answer is “someone pointing out ways in which these conceptual arguments are flawed.”^[40]I hope that publishing this post will inspire people who are optimistic about baseline HFDT and iterative/empirical ML safetyto explain why this outcome seems unlikely to them.

However, optimists often take a very empiricist frame, so they are likely to be interested in what kind of ML experiments or observations about ML models might change my mind, as opposed to what kinds of arguments might change my mind. I agree it would be extremely valuable to understand what we could concretely observe that would constitute major evidence against this view. But unfortunately, it’s difficult to describe simple and realistic near-term empirical experiments that would change my beliefs very much, because models today don’t have the creativity and situational awareness to play the training game.

As an illustration that’s deliberately over-extreme, imagine if some technologically less-advanced aliens have learned that a human spaceship is about to land on their planet in ten years, and are wondering whether they should be scared that the humans will conquer and subjugate them. It would be fairly difficult for them to design experiments on mice^[41] that would give them a lot of information about whether or not to be scared of this. They would probably be better off speculating from their priors than trying to extrapolate from observations on mice.^[42]

I think we have significantly more hope of designing experiments on small models that give us meaningful updates about AI takeover risk,^[43] but I take this analogy more seriously than most ML researchers seem to. Accordingly, I’m fairly unmoved by empirical experiments that ML researchers have cited to me as evidence about the magnitude of x-risk from AI.

With all that said, here’s a stab at a general genre of empirical evidence that would probably move my views a lot if it were demonstrated in large quantities across different domains:

1. Convincingly demonstrate that a pre-trained model has a certain “capability,” for example “recognizing the color red.” E.g. you might show that it consistently and correctly predicts the color of stop signs, strawberries, etc.
2. Describe an RL fine-tuning procedure that seems like it would select for / incentivize the use of that capability, e.g. “reward the model every time it approaches a red object.”
3. Then make a prediction that the model would in fact not use that capability and instead do some other particular thing, e.g. “the model would actually approach purple objects because gradient descent is biased to find ‘seeking purple objects’ more natural.”
4. Demonstrate that this prediction was correct.

I expect most people to think that it would be very hard to generate such results, including those who are much more optimistic about AI takeover than I am. I think this is another case of optimists and pessimists interpreting evidence differently — to me, expecting a powerful model not to play the training game is essentially making a claim that we understand a particular way in which gradient descent is consistently suboptimal, so I would want to see researchers succeed at the task of “guess the particular way in which gradient descent will be suboptimal” a lot of times to feel better.

“Security holes” may also select against straightforward honesty

In the main body, I implicitly acted as if the reward Alex gets in the lab setting is always under the control of some specific designated humans,^[44] and discussed ways in which Alex might try to play on those humans’ biases or ignorance.

But in reality, Alex’s reward is ultimately generated and recorded through a series of specific operations carried out on some Magma servers. This means there are likely some operational and computer security vulnerabilities Alex could exploit to bypass “official channels” for reward entirely:^[45]

• If Alex is able to get access to whichever computer records logs of its actions and sends them to human evaluators, it may be able to arbitrarily edit the record of its actions to make itself look more reasonable, helpful, or intelligent before they are reviewed by human evaluators. (For example, it may go back and edit any incorrect predictions it made with the benefit of knowing what happened in hindsight.)
• If it’s able to access the website or software that human evaluators use to enter their scores, it could spoof the scores in various ways. As a simplified example, it may produce outputs that look like routine instructions to the contractors, which tell them to enter a “thumbs up” as part of an attention check. As another example, it could edit the back-end logic so that pressing a “thumbs down” button would result in actually recording a “thumbs up” in the relevant database(s).
• It could bribe a Magma employee into modifying its reward signals, e.g. by offering to feed them good ideas that would get them promoted or helping them embezzle money from the company.

I am unsure whether and how often Alex would take actions like this in the lab setting, and consider this more speculative than softer ways of playing to human biases:

• Humans would consider these kinds of actions to be more blatant and scary violations of expectations than softer ways of playing to human biases, which they may not recognize as bad even if it’s staring them in the face. That means if this is discovered later Alex is likely to retroactively receive a large negative reward and/or trigger a larger retraining project. Alex would likely understand this, so it’s a question of its motivation structure whether it would still take these opportunities. It may do so if it inherently cares about maximizing short-term reward, or it may avoid them if it cares about something longer-term.
• Depending on Magma’s security practices, these opportunities may be quite rare, and/or it may consistently be easier (in terms of compute or time constraints) for Alex to try to maximize reward in more “normal” ways (which would still involve a large amount of manipulation).

Nonetheless, this consideration contributes to my expectation that Alex would end up playing the training game. If it does try something like this, even as an exploratory action, that action would be self-reinforcing. Whatever moved Alex to try this kind of act, doing it would immediately lead to a gradient update in the direction of those inclinations — barring the fact that humans would penalize it if discovered. But even if it is later penalized, that may not be enough to reverse the inclinations; see this appendix.

Simple “baseline” behavioral safety interventions

In this appendix, I’ll discuss several simple proposals for training Alex to have safer behavior, and explain why I think they would simply change the game Alex is playing rather than altering its motive to play the training game:

• Maybe we could train Alex on very high-quality feedback, and/or try to teach it to extrapolate from lower-quality feedback to higher-quality feedback? (more).
• Maybe we could use prompt engineering to directly ask Alex to emulate much wiser and more thoughtful evaluation? (more).
• Maybe we could require Alex to provide evidence, arguments, or justification that humans can understand for its actions? (more).
• Maybe if we make the training distribution very diverse, different types of errors in the reward signal would “cancel out” and prevent misgeneralization? (more).
• Maybe we could include a number of situations in the training distribution in which Alex is rewarded for being “conservative” — avoiding unusual actions, checking in when it’s in a strange new situation, etc (more).

Using higher-quality feedback and extrapolating feedback quality

To avoid Alex learning to manipulate or deceive its evaluators, we could have more careful / thoughtful / informed humans evaluate how Alex interacts with less-informed / less careful / less thoughtful humans. Perhaps some training episodes involve Alex answering questions for humans who believe the Earth is flat, but its “honesty” reward signal ultimately comes from a group of humans who know the Earth is round. Then Alex would be incentivized to be honest to the flat-earthers rather than telling them what they want to hear.

There could also be a sequence of such steps. As a caricature, imagine that first Alex interacts with 3rd graders while being evaluated for honesty by 6th graders, and then with 6th graders while being evaluated for honesty by 9th graders, and then with 9th graders while being evaluated by 12th graders, and so on. We might hope that after many instances of “learning the lesson” that it will be evaluated by a human better-informed than the one it’s interacting with, Alex would learn to “play it safe” by always being honest.^[46]

This intervention would likely stop Alex from attempting certain kinds of deceptive or manipulative moves (creating the appearance of success). However, it wouldn't make complete honesty the reward-maximizing strategy. At its level of situational awareness, Alex would easily understand that the humans designing this training scheme can’t actually continue to supply arbitrarily better-informed evaluators. In reality, there will be some “most informed and thoughtful human evaluators the designers could plausibly employ to evaluate this episode,” and Alex could simply act as if those humans are evaluating it at all times.

There are still ways in which the most informed and thoughtful humans available could be manipulated or deceived to increase reward. If the Catholic Church were training Alex in the 1400s, ensuring that its reward signal is only generated by the brightest scholars of the age as anointed by the Church would still not make maximal honesty the best strategy. Alex would still be learning to manipulate some set of humans (in this case, the “final level” of supervisors) to get high reward, leading to the dangerous generalization discussed in the main piece.

Using prompt engineering to emulate more thoughtful judgments

Large language models such as GPT-3 will change the way they respond to questions or requests based on what came previously in the prompt. For example, prompting GPT-3 with something like “You are a wise and knowledgeable AI advisor trying to help a human figure out what’s true” will tend to cause its responses to questions to be more accurate; asking questions with poor spelling and grammar will tend to cause its responses to get less accurate. (Setting up prompts to elicit certain behaviors like this is called prompt engineering.)

Some researchers have suggested that we might be able to leverage prompt engineering to help with ML alignment for very large models which are presumably very sensitive to prompt changes.^[47] For example, we could prompt Alex with something like “Act how a very wise and very ethical human that’s as knowledgeable as you are would act if they had a long time to think about it,” or we could ask it to explain how a very wise and ethical human would evaluate its past actions. We could also search (through trial-and-error or gradient descent or some combination) for prompts that empirically seem to induce the best behavior.

However, if Alex has a good understanding of its human evaluators and knows that they are attempting this kind of prompt engineering, Alex would simply respond to such prompts by doing its best to act the way ordinary humans think that an unrealistically wise / ethical / knowledgeable person would act. This may make its behavior superficially better in some cases — but Alex could still sometimes achieve higher reward by being deceptive or manipulative (e.g., providing answers that seem good rather than answers that are good), leading to the dynamics described above.

Requiring Alex to provide justification for its actions

Another suggestion for improving alignment is to require Alex to prove or argue or demonstrate that its actions are aligned. For example, Alex might provide an easily-checkable proof that a mathematical theorem it claimed to discover is true, or it might provide an empirical demonstration of a physical phenomenon it claimed to discover.

This would likely work well in areas like math and physics where arguments and demonstrations are unusually cheap and decisive. However, in many important domains — e.g. policy, economics, sociology, parts of biology, etc — crisp and decisive “proofs” are intractable, and human evaluators will be forced to try to make sense of softer kinds of arguments and evidence which leave a huge amount of room for interpretation. Even “rigorous scientific studies” can easily be confusing and misleading and consistent with any number of conclusions based on the reader' priors (and for many claims it would be too expensive to generate such evidence in the first place).

Furthermore, there are likely some areas where human evaluators (including the best accessible human evaluators) are predictably wrong about the kind of arguments that best demonstrate truth, not just wrong about object-level claims. Consider the version of Alex trained by the medieval Catholic Church again — the Church’s worldview would be oriented around Biblical hermeneutics and/or papal declarations being the ultimate arbiter of truth when different sources of evidence conflict.

More broadly, I’m not aware of anyone sincerely believing that “having the arguments and papers people find most persuasive” is tightly and reliably coupled with “being the most correct.” It probably buys some amount of correctness on average, but most of us can name many qualitative arguments we think are deeply flawed that lots of other smart people nonetheless consider decisive (e.g. because the arguments prey on some of those people’s biases).

So as in the above sections, this approach would still lead to cases in which Alex could get better reward by being less truly honest/helpful/harmless (though more apparently honest/helpful/harmless), and hence to the problematic dynamics described above.

We could try to salvage the idea of “require Alex to demonstrate that what it’s saying is true” with something like debate or amplification, both of which essentially try to recursively break big difficult-to-evaluate claims into sub-claims (which are broken down into further sub-claims etc) such that the smallest sub-claims are simple enough and uncontroversial enough that we can be highly confident all human evaluators would get them right and agree about the answer (e.g. claims like “Does ice float in water?” or “Are giraffes generally bigger than dogs?”).

However, it’s very unclear if it’s possible to recursively break down important big-picture questions like “Is this action helpful, honest, and harmless?” into these very easy and objective subquestions — as far as I know, nobody has demonstrated a working solution here. The only systematic empirical investigation of debate that I know of (using humans in the place of models) ultimately concluded that dishonest debaters often have a strategy that lets them win debates.

Moreover, even if we can find a working debate or amplification strategy, there’s no guarantee that it will be as efficient (in terms of computation or data) as baseline HFDT — it could simply turn out to be much easier to acquire knowledge than to explain or justify it to humans (see this section of Alignment Research Center’s 2021 report “Eliciting Latent Knowledge” for more discussion). If that’s the case, it could create a lot of pressure to “cut corners” and do something more like baseline HFDT rather than debate or amplification.

Making the training distribution more diverse

Some researchers argue that making the training distribution more diverse — adding many different kinds of situations which call for very different approaches or heuristics — would reduce the probability that Alex violates what humans intend. The idea is that if the training distribution is relatively narrow, there are many possible ways to generalize under distribution shift, and the more we include novel situations that benefit from using very different tactics the more we’re eliminating “wrong” ways it could generalize.

For example, this paper showed that if we try to train an agent to collect a gem in a maze, but the gem always happens to be yellow in the training set, then the agent will seek yellow objects rather than gems when both are present. If we had trained the same agent on a more diverse distribution with gems in a huge variety of colors, then it’s more likely it would “correctly” (according to its designers’ intent) pursue gems in the test distribution — “shape” is more likely to be the most salient thing in common between all the objects it’s rewarded for collecting, since we’ve eliminated “color” as a possibility by varying color.

Similarly, maybe if Alex only ever receives human feedback from one type of human with a particular knowledge base and set of biases (e.g. “software engineers working at Magma”), it’s likely to “overfit” to that kind of human and learn to exploit their weaknesses and gaps in their knowledge. But maybe if it were exposed to feedback from a really broad array of different humans with different values, strengths and weaknesses, areas of knowledge and misconceptions, etc, different people’s errors and biases would “cancel each other out” such that Alex would learn to “factor out” these flaws and just be honest?

We could also imagine manufacturing more “artificial” kinds of diversity — e.g. situations where humans try hard to model the opposite of standard human biases like the sunk cost fallacy or risk aversion, or simulated environments where being highly docile and cooperative turns out to be the best strategy for getting reward, or situations where e.g. Alex interacts with a Democrat but its responses are evaluated by a Republican. The idea is to create so many different kinds of scenarios with different strange things happening that hopefully the only common thread between them all is the generalization Magma engineers were going for.

Introducing a lot of diversity in how reward is generated would likely increase Alex’s uncertainty about what is going on in any given episode, compared to a more straightforward training curriculum in which Alex is always given normal tasks and evaluated on normal metrics of performance. In that sense, there could be some effect of different biases and errors “canceling each other out” — if Magma researchers successfully cause Alex to be uncertain whether it’s in a tricky scenario or an ordinary scenario, or whether it’s interacting with a political conservative or liberal, it may indeed often act less manipulatively or dishonestly in the face of that uncertainty. However, increasing diversity doesn’t seem like it would cause Alex to “forget” the basic situation it’s in — that it’s a machine learning model being trained by some humans to be useful, intelligent, helpful, honest, harmless, etc. And by the same token, it doesn’t seem like it would cause Alex to stop trying to maximize reward — it would simply change the strategy that best maximizes reward. And as with the previous examples, the new best strategy would probably not be to be straightforwardly helpful, honest, and harmless at all times:

• Alex could pick up cues about what “general kind of person” is likely to evaluate the current episode, and develop guesses about their level of knowledge, political ideology, personal biases, and so on.
• Alex could also pick up soft cues about whether its current episode is likely to be a “strange tricky episode” or one that’s being “played straight.”
• Alex can be pretty confident that some kinds of manipulative tactics (e.g. flattery) are fairly universal.
• Many egregious reward-grabbing actions may remain open to Alex, if they would require more careful investigation to catch than humans are likely to bother with.

The underlying dynamic would be the same overall: Alex would be being trained to do something like “behave in ways that appear honest, helpful, and harmless, but behave in other ways when it is possible to evade detection and doing so leads to higher reward.” For a capable enough Alex, this could still present plenty of opportunities for the latter.

Additionally, increasing diversity could also make the situation worse, by strongly developing Alex’s creative planning faculties and forcing it to use them a lot. In a narrower distribution, it might not incur much penalty to follow a simple set of heuristics based on what worked well in the past (like today’s models often use), and we could conceivably carefully engineer a narrow training distribution such that the heuristic “be helpful, honest, and harmless” is always reward-maximizing (though this would be difficult).

On the other hand, Alex is less likely to be able to memorize simple heuristics that cover all the cases in a very diverse distribution that contains many “tricky” episodes. Asking itself “How can I figure out how to maximize reward in this episode?” probably works better than following heuristics like “do something basically reasonable,” by an amount that scales with how often it’s encountering situations that aren’t like anything it’s seen before.

Adversarial training to incentivize Alex to act conservatively

Magma engineers could try to train Alex to be “harmless” by putting it in various situations where it would seemingly increase reward to take some kind of egregiously harmful or norm-violating action (a kind of adversarial training).

For example, researchers could put Alex in a situation where it apparently has the opportunity to execute a code injection attack on the website that the human evaluators interact with, allowing it to edit all reward scores to the max score. They could then give it a large negative reward for taking the opportunity, and a positive reward for refraining and/or checking in with a human. They could also search (manually, with gradient descent, with the help of other models, etc) for situations in which Alex is especially likely to do the wrong thing, and include many such situations in its training distribution.

As with increasing diversity above, adversarial training is likely to create some amount of additional uncertainty about whether apparent opportunities to increase reward by violating norms are actually “sting operations” by human supervisors. This will probably make Alex less likely to attempt certain bad actions (e.g. fraud, theft, hacking) to get reward (even in some cases where it could have gotten away with it, since it will be imperfect at predicting which opportunities are setups and which are real).

But it’s unclear whether it’s possible to generate a distribution of adversarial training episodes which ensure that the reward-maximizing strategy never does anything norm-violating or harmful.

In the course of its training, Alex is likely to learn things about the world that it can be pretty confident no human understands (just as AlphaGo knows things about Go that no humans know, or AlphaFold knows things about biology that no humans know). If it then discovers e.g. a zero-day exploit which it can only recognize because of knowledge it’s developed about computer hardware that it’s fairly confident Magma engineers don’t know, it can safely exploit that vulnerability (e.g. to directly edit the reward entered in its training dataset) because it’s unlikely to be entrapment by humans.

As Alex does more and more scientific research and gains more and more knowledge, it becomes more and more likely that it can find opportunities for egregious reward-grabbing that humans wouldn’t have been able to fake in a training setup. Getting rid of all such opportunities seems to involve either a) ensuring that humans can “follow along” with Alex and understand everything new that Alex learns at about the same time as Alex, or b) letting humans “read Alex’s mind” to discern its intentions. Research directions that might help us do this — such as mechanistic transparency, eliciting latent knowledge (ELK), or relaxed adversarial training — are early-stage, and it’s unclear how far they’ll get.

“Training out” bad behavior

Perhaps if Magma researchers notice that Alex did something bad to maximize reward (e.g. lied to its evaluator or exploited a security vulnerability as described in this appendix), they can reverse Alex’s motivation to do that sort of thing with further intervention? For example, maybe they can get it to stop playing the training game altogether, or at least only engage in “soft manipulation” rather than more acutely scary actions?

Researchers might try editing the rewards associated with those episodes to show large negative values, they might conduct audits to uncover other instances of egregious reward-grabbing and retroactively give those large negative rewards, they may try adversarial trainingin which Alex is given the apparent opportunity to grab reward, and so on.

However, as with other safety interventions, these reactive safety interventions don’t seem like they would reverse Alex’s motive to grab reward. Rather than learning to never take bad actions, Alex seems more likely to learn not to take such actions when it’s plausible that humans could go back and give negative reward to them later.

That is, it would continue to try to maximize expected reward in any way it can, while shifting the notion of “reward” that it’s maximizing to be the final reward that ends up recorded in association with that training episode, after all rewrites and revisions are over.^[48] As before, changing the training game predictably leads Alex to get more expected (long-run) reward than adopting the policy of never taking egregious reward-grabbing actions.

“Non-baseline” interventions that might help more

Here are several ideas for how we could go beyond taking the appearance of good behavior at face value:

• We could try to come up with an alternative training process that we have strong reason to believe wouldn’t run into the problem that HFDT runs into, even on arbitrarily powerful models.
• We could be careful to only train models that are not capable of taking control from humans, and try to use advice from those untrusted-but-controlled models to somehow eliminate further danger — e.g., get them to solve AI alignment more fully in a way that works for much more powerful models, or get them to figure out how to institute strong international bans on deploying more powerful models, etc.
• We could do adversarial training specifically to reduce the probability that a model would take control from humans if given the chance — that is, try to generate inputs that fool the model into thinking an AI takeover is possible, and select against models that would take that opportunity. Likely we would need model assistance for this — we could reward one model for finding inputs that fool another model.
• We could try to understand models’ internals well enough to figure out ways to directly select for models that “have the right motives” (as opposed to the right behavior), or to “read models’ minds” to figure out whether we should deploy them.
• We could provide training to human evaluators to make them less susceptible to manipulative and dishonest tactics, and instruct them to give reward primarily or entirely based on whether the model followed honest procedures rather than whether it got good results. We could try to use models themselves to help explain what kinds of tactics are more and less likely to be manipulative.
• We could improve computer security in AI labs, to make it harder for moderately powerful AI systems to gain unauthorized access to their rewards or gradients. We could also train AI systems to propose additional improvements that we can verify (while being careful to ensure that the systems looking for flaws in our security setup are not themselves able to easily break out).
• We could train several different kinds of models and set them all up against each other in a system of checks and balances during deployment. For example, some models could be in charge of predicting future indicators of a possible AI takeover, such as “what the security cameras in various data centers will show” or “what newspapers will be reporting.” Other models could be rewarded for demonstrating takeover plans hatched by other AIs.

A forthcoming post by my colleague Holden Karnofsky goes into much more detail about safety measures that could prevent AI takeover.

Examining arguments that gradient descent favors being nice over playing the training game

In this appendix, I’ll briefly summarize arguments that gradient descent may favor honest and straightforward strategies like “doing what designers intended” over “playing the training game” even if the latter gets more reward, and explain why I don’t find them persuasive:

• Maybe telling the truth is in some general way “easier” or more “natural” than deception? (more).
• Maybe Alex internalizes attitudes like “lying is bad” early in training while it’s still weak, and this sticks around? (more).
• Maybe we have evidence that gradient descent tends to generalize surprisingly “well,” compared to what might be expected from theoretical arguments? (more).

Maybe telling the truth is more “natural” than lying?

Some people have the intuition that it would be in some broad sense “simpler” or “easier” for gradient descent to find a model that plainly states what it internally believes than one that maintains one set of beliefs internally while presenting a different set of beliefs to the world. By analogy, humans find it mentally taxing to weave elaborate lies, and often end up deceiving themselves in the course of deceiving others.

If the “always be honest” policy received almost as much reward as the policy that plays the training game, it seems possible (though far from certain)^[49] that an effect like this could end up dominating. But in fact, I expect the honest policy to get significantly less reward than the training-game-playing policy, because humans have large blind spots and biases affecting how they deliver rewards. I’m skeptical that something like “honesty being somewhat simpler or more natural” would make the difference in that case. Most humans are not regularly in situations where lying has a very high expected payoff — and in such situations humans often do lie even though it’s difficult (consider undercover agents whose lives depend on not getting caught).

Maybe path dependence means Alex internalizes moral lessons early?

A related argument suggests that early in training (while Alex perhaps has low situational awareness and/or planning ability), taking actions that look like “figuring out clever ‘hacks’ to get more reward ‘illegitimately’” would be caught and given negative reward; this might instill in it the heuristic that taking actions which pattern-match to “clever hacks” are generally undesirable. Maybe once it “grows up” it realizes that it could get away with deceptive tactics, but perhaps by then such tricks internally feel “unaesthetic” or “immoral.”

By analogy, kids are taught not to steal from people in contexts where it’s easy for adults to catch them if they try, since they’ll be incompetent at it. Most children grow up to be adults who’ve internalized a general aversion to theft, and who avoid it even when they know they could get away with it in a particular instance. Most adults aren’t constantly on the lookout for cases where they could get away with breaking the rules.

I find this unpersuasive for two reasons:

• I don’t think it makes sense to bank on Alex developing a particular kind of internal moral commitment or sentiment, especially one held with enough force to overcome the fact that shedding it would result in substantially more reward. Instead of developing an affinity for being straightforward or an aesthetic aversion to clever hacks, it could develop any number of other affinities and aversions that happened to correlate with doing well early in training (e.g. “be curious,” “avoid sameness,” etc). And even if something like “an aversion to clever hacks” were initially part of its psychological landscape, I’m skeptical that this kind of dynamic would survive and dominate in the face of consistent and significant rewards for deception.
• Even before Alex is ever trained with RL, it already has a huge amount of knowledge and understanding of the world from its predictive and imitative pretraining step. It may well already have strong situational awareness and a good grip on the concept of “reward” in RL, e.g. from reading so many papers about RL. I don’t think it makes sense to analogize it to an impressionable “child” at the start of HFDT — it may latch onto the policy of maximizing reward nearly immediately given its existing knowledge.

Maybe gradient descent simply generalizes “surprisingly well”?

A number of ML researchers who are bullish on deep-learning-based transformative AI seem to have the background heuristic that gradient descent empirically tends to generalize “better” than might be expected by theoretical arguments. For example:

• We might have thought that gradient descent would only work well for convex tasks, because it would get stuck in not-very-interesting local minima on non-convex tasks. But in fact it works well in practice for a large number of realistics tasks that are clearly not fully convex.
• We might have thought that models trained on a certain distribution would generalize very poorly to a different distribution, but in fact models trained on a reasonably broad initial distribution (like language models trained on the internet) generalize reasonably well zero-shot to various different tasks — and this generalization improves with model scale.
• We might have thought that models with “too many” parameters relative to training data points would overfit to the training dataset and generalize poorly on future examples, but the deep double descent phenomenon shows that “overly large” models first generalize somewhat worse, and then generalize better as their size increases.

Examples like these lead some researchers to adopt a heuristic that gradient descent works surprisingly well compared to what we might expect based purely on theoretical arguments, and (relatedly) theoretical arguments that gradient descent is likely to find a particular kind of model are usually wrong and given our lack of empirical data we should have wide priors about how future models will behave. Such researchers often tend to be optimistic that with empirical experimentation we can find a way of training that produces powerful models that are “docile,” “nice,” “obedient,” etc., and skeptical of arguments (like the one I’m making in this post) that gradient descent is more likely on priors to find one kind of model than another kind of model.

But what exactly does it mean for gradient descent to work surprisingly well? An intuitive interpretation might be that it’s surprisingly likely to produce the kinds of models we’d hope it would produce — this is what people usually are pointing at when they say something is going “well.” But I think a more realistic interpretation is that gradient descent is surprisingly likely to produce models that get a very high reward (on the training distribution), and/or generalize to doing the sort of things that “would have” gotten a high reward (on a different distribution).^[50] My concern is about situations in which doing the sorts of things that maximize reward comes apart from what we should be hoping for.

A possible architecture for Alex

Thanks to Jon Uesato for suggesting a variant of this architecture, and to Buck Shlegeris for helping me work out details I was confused by.

Because it processes a long series of observations within a single episode, Alex needs to have some sort of architecture that allows for sequence processing. One example might be a transformer architecture that attends over the last K observations at once.

While transformers are more common for state-of-the-art language models as of 2022, I’ll imagine here that Alex is some type of recurrent neural network (RNN) because that’s simpler to visualize. For example, you could imagine Alex is an LSTM network, though there are many other recurrent architectures we could imagine. In reality, the sequence processing would likely be significantly more complicated and may combine elements of various RNN architectures with transformer-like attention and other mechanisms — you can feel free to substitute whatever architecture you think is most appropriate for processing sequences.

The diagram above show that at every timestep, Alex takes a single observation^[51] — plus its own hidden state from the previous timestep — as input, and produces a single action as output.

Let’s say it also outputs a prediction of the next observation^[52] and the final episode reward (both conditional on taking the action) as auxiliary tasks every timestep. The observation prediction task is meant to help it build up an understanding of the dynamics of interacting with the computer, while the reward prediction task is meant to help it more quickly connect decisions to their impact on overall task performance.

While RNNs carry forward a “train of thought” or “short-term memory” across multiple timesteps, they generally have a limited ability to “remember” things over very long time periods. This is an issue because Alex will have to do tasks that could take many millions of timesteps to complete (e.g. a single episode may cover the process of investigating an odd phenomenon, generating hypotheses, conducting experiments to test the hypotheses, settling on an explanation, and writing a paper of the findings).

Magma researchers and engineers might get around this limitation in a number of possible ways. They may use a different architecture entirely for sequence processing,^[53] or they may somehow break down otherwise-long episodes into shorter chunks,^[54] or the accumulation of tweaks to optimization algorithms or RNN architectures in the time between now and then might have effectively eliminated this issue.

To keep this story concrete, I’ll sketch out a broad kind of architectural modification that might help with this issue, if a number of technical details were hammered out. The basic idea is that Alex is equipped with a database that it can interact with to store and retrieve its “long-term memories.” We’ll imagine for concreteness that a “memory” is some subset of Alex’s activations and/or weights at a certain point in time, but it could have any other format.

At the beginning of an episode, Alex’s memory bank is wiped clean — this is sort of analogous to a new “clone” of Alex being “born,” with certain “genetic predispositions” (in the form of its weights) but no “individual life experience” yet.

Within each timestep of an episode, Alex is able to a) store any activations and/or weights from the current timestep that it wants to, and b) send a query to retrieve any previously stored activations and/or weights that might be relevant, handing those to itself as inputs for the next timestep. Over the course of an episode, the memory bank builds up a store of “lifetime experiences” (though some experiences can also be “forgotten” if desired).

In addition to this dedicated memory bank, it’s important to note that Alex can use the computer it’s interacting with as an external memory aid in exactly the same way humans do — it can write notes to itself in Google docs, create Asana tasks, set Google calendar reminders, and so on. If taking actions like that helps it to get a higher final reward for the episode, the training process will select for that.

This means that (like humans) Alex will have both “direct” experiential memories and access to external memory aids.

Plausible high-level features of a good architecture

Stepping back, the high-level picture of Alex interacting with the external computer, with the details of the architecture sketched out, looks like this:

In every timestep, Alex:

1. Takes as input:
   1. An observation from the external computer (such as an image of the screen)
   2. Its previous hidden state (via a recurrent connection)
   3. Some number of memories retrieved from the long-term memory database
2. Produces as output:
   1. An action (such as a keystroke). The action space includes a no-op and encompasses every low-level action a human could take on a computer.
   2. A store query and a fetch query to its long-term memory database, representing “memorizing” and “recalling” (respectively) a certain life experience.
   3. A prediction of the final reward for the episode, conditional on taking the action from a) and sending the store+fetch queries from b).
   4. A prediction of its next observation, conditional on taking the action from a) and sending the store+fetch queries from b).

I tried here to err on the side of being overly specific because I think it can be helpful to have in mind a vivid picture of a particular architecture that could conceivably suffice for a transformative model, even though it’s highly unlikely to literally be accurate.

With that said, I think that whatever future architecture we might use to train a transformative model with HFDT is pretty likely to have a few key high-level features in common with my over-simplified sketch:

• Some form of sequence processing (whether attention, recurrence, something else, or some combination of architectures)
• A way to reliably store and recall “memories” / “life experiences” over long timescales (this might be automatically solved by whatever solution is found for the above)
• Input and output formats which allow for a lot of flexibility, so that the model can process many different kinds of inputs and perform many different kinds of low-level actions (though this may be accomplished with a large number of different specialized input and output channels rather than a generic “universal” input channel and output channel as I’ve depicted here; we might also use entirely new data formats optimized for machine consumption)

None of the basic points made in the main post hinge on the specifics of Alex’s architecture. However, my arguments do rely on the high-level properties above — e.g. that Alex has a very general and flexible input/output space and can usefully remember things across many timesteps. I think these properties are important for the case that Alex can potentially have a transformative impact in the first place, and by the same token are important for the case that Alex might be extremely dangerous.

---

1. Even though this vision doesn’t involve using exclusively human feedback, I’ll nonetheless refer to it as human feedback on diverse tasks (HFDT) to highlight that human judgments are central to training. In practice, I expect that powerful models will be trained with a combination of human feedback and automated reward signals (e.g. “did the code written by the AI compile?” or “how much computation did the AI use to perform this task?”). But I expect human feedback to be especially important, and I’ll focus on the dangers associated with the human feedback component of the feedback signal since there is more contention among the ML community about whether human feedback is dangerous. Many researchers agree that fully automated reward signals carry a high risk of misalignment, while being mostly optimistic about human feedback. While I agree incorporating human feedback is better than using fully-automated signals, I still think that the default version of human feedback is highly dangerous. ↩︎

2. For example, we could imagine trying to produce transformative AI by training a large number of small models to do specialized tasks which we manually chain together, rather than training one large model to simultaneously do well at many tasks. We could also imagine using 100% automated reward signals, or “hand-programming” transformative AI rather than using any form of “training.” ↩︎

3. For example, maybe the required model size is far too large to be affordable anytime soon, or we are unable to generate sufficiently challenging and diverse datasets and environments, or we run into optimization difficulties, etc. ↩︎

4. By “incentivized,” I mean: “Alex would be rewarded more for playing the training game than it would be for alternative patterns of behavior.” There are further questions about whether Alex’s training procedure would successfully find this pattern of behavior; I believe it would, and discuss this more in the section "Maybe inductive bias or path dependence favors honest strategies?". ↩︎

5. For example, lying about empirical questions which are politically-sensitive will often receive more reward than telling the truth; consider a version of Alex trained by the Catholic Church in the 1400s who is asked how the solar system works or how life was formed. ↩︎

6. What exactly “maximizing reward” means is unclear; however, I think this basic conclusion applies to most plausible ways of operationalizing this. ↩︎

7. This means that we’re bracketing scenarios in which Alex “escapes from the box” during the training phase; in fact I think such scenarios are plausible, which is a way in which risk is greater than this story represents. ↩︎

8. This could be accomplished through a variety of training signals — e.g., more human feedback and automated signals, as well as more outcomes-based signals such as the compute efficiency of any chips they design, the performance of any new models they train, whether their scientific hypotheses are borne out by experiments, how much money they make for Magma per day, etc. ↩︎

9. It’d be reasonable to call this model “an AGI,” whether or not it’s able to do literally every possible task as well as humans; however, I’ll generally be avoiding that terminology in this post. ↩︎

10. Note that it may do this by first quickly building more powerful and more specialized AI systems who then develop these technologies. ↩︎

11. The exact timescale depends on many technical and economic factors, and is the subject of ongoing investigation at Open Philanthropy. Overall we expect that (absent deliberate intervention to slow things down) the time between “millions of copies of Alex are deployed” and “galaxy-scale civilization is feasible” is more likely to be on the order of 2-5 years rather than 10+ years. ↩︎

12. The most important possible exception is that if the world contains a number of models almost as intelligent as Alex at the time that Alex is developed, we may be able to use those models to help supervise Alex, and they may do a better job than humans would. However, I think it’s far from straightforward to translate this basic idea into a viable strategy for ensuring that Alex doesn’t play the training game, and (given that this strategy relies on quite capable models) it’s very unclear how much time there will be to try to pull it off before models get too capable to easily control. ↩︎

13. In reality, I expect it would have multiple input and output channels which use special format(s) more optimized for machine consumption, rather than having just one input channel which is directly analogous to human vision and one output channel which is directly analogous to human typing — those are probably not the most efficient way for a neural network to interact with external computers. Going forward, I’ll simply refer to inputs as observations and outputs as actions, though I’ll use the shorthand of “screen image” and “keystroke” in the diagrams. ↩︎

14. Or mouse movement / click. ↩︎

15. Like today’s neural networks, Alex is trained using gradient descent. Roughly speaking, gradient descent repeatedly modifies an ML model to perform better and better at some task(s). Researchers feed Alex an instance of a task that they want it to perform well on, Alex produces some output in response, and gradient descent is used to slightly perturb the model so it performs a bit better on that example. This is repeated many times until Alex is very good on average at the kinds of tasks it saw during training. See here for a more technical explanation ↩︎

16. This tends to make it easier to train Alex to do more useful tasks down the road, so is often called “pre-training.” ↩︎

17. If we imagine observations are images of screens and actions are keystrokes, the training dataset for this might come from screen captures and keylogs on Magma employees’ laptops. ↩︎

18. As above, different models may be fine-tuned on different mixtures of these tasks in order to specialize into different niches. ↩︎

19. In this appendix, I discuss a bit more why I think this strategy is a “baseline” and what some non-baseline strategies may look like. ↩︎

20. In terms of time, computation, and/or other metrics like “amount of human help used” or “number of lines in the proof.” ↩︎

21. This is closely related to the concept of deceptive alignment, as introduced in Hubinger et al 2019. ↩︎

22. Using more intelligent and thoughtful human reviewers would fix this particular problem, but not fix the general problem that Alex would be incentivized to play on their biases. See this appendix for more discussion. ↩︎

23. Magma may itself have stakeholders it needs to please that introduces more bias. For example, if one thing Alex is supposed to do is interact with customers, then it may need to be trained not to be offensive to those customers. ↩︎

24. An important family of training strategies involves using different copies of Alex (or different “heads” on the same model) each with incentives to point out lying and manipulation done by the other. There are a large number of possible approaches like this, but they all seem to require significant caution if we want to be sure they’ll work. ↩︎

25. The number of copies of Alex that can be run immediately will depend on how much computation it took to train it — the more computation it took to train Alex, the more chips are available to run it once it’s trained. This multiple is likely to be large because training a model generally requires running it for a huge number of subjective years. ↩︎

26. For example, it could be given the ability to spend money to make trades, investments, or hires if that’s important for quickly increasing Magma’s capital; it could be given the ability to remotely operate machinery if that’s important for robotics research; etc. ↩︎

27. This could be accomplished through a variety of training signals — e.g., more human feedback and automated signals, as well as more outcomes-based signals such as the compute efficiency of any chips they design, the performance of any new models they train, whether their scientific hypotheses are borne out by experiments, how much money they make for Magma per day, etc. ↩︎

28. In “softer takeoff” scenarios (which I consider more plausible overall), Magma will have already been using the somewhat-less-powerful predecessors of Alex to automate various parts of R&D before this point, meaning that the pace of research and innovation would already be going quickly by the time a model of Alex’s level of capability is deployed. This makes analysis more complicated, but on net I’m not convinced that it makes the situation significantly safer. ↩︎

29. Currently, the ratio of computation to memory in GPUs and TPUs is much higher than in brains. The A100 GPU (state-of-the-art as of March 2021) performs ~1.5e14 FLOP/s (in floating point 16 operations). Joe Carlsmith’s central estimate for the amount of computation it would take to match the human brain is ~1e15 FLOP/s, implying that ~10 A100 GPUs could match the total computation of a human brain. However, the human brain seems to store ~1e14 bytes of memory (taking a synapse to be equivalent to a byte and assuming that synapses dominate the memory in the brain), while the A100 only has a storage capacity of ~8e10 bytes. That implies it would take ~1,000 A100 GPUs to store a neural network whose parameters contained as much information as the human brain. That system of 1,000 GPUs would then have enough computation to use each parameter ~100 times more often per second than the human brain uses each synapse. I expect computation intensity would continue to be significantly higher in ML chips than in biological brains, allowing for this speedup. ↩︎

30. In this case, Alex continues to be trained in the deployment setting, but the training signals are systematically different from the training signals in the lab setting due to humans’ decreased understanding and control. ↩︎

31. In deployment (more so than in the lab), it’s not exactly clear what “maximizing reward” means precisely. But there will still be some reward signals coming into the various copies of Alex, and any plausible notion of “maximizing” those signals would be bad for humans. ↩︎

32. Or video, or written feedback, etc that is used to ultimately calculate reward. ↩︎

33. Or writing positive things, or appearing happy in front of a camera, etc. ↩︎

34. This feels more realistic if you consider that Alex itself would have written almost all of the code for the datacenter, and designed almost all of the hardware. ↩︎

35. That is, unless gradient descent for some reason “got stuck” at the benevolent strategy and was not able to push Alex toward the more effective strategy; I discuss this briefly here. ↩︎

36. This is similar to the argument made in Ilyas et al 2019 (Adversarial examples are not bugs, they are features). ↩︎

37. They could also respond by attempting to shut down Alex entirely, but this could be an extremely costly choice in the face of competitive pressure — assuming they do not have an arbitrary lead on other AI companies, refraining from deploying Alex indefinitely is likely to simply mean that somebody else deploys a similar system (perhaps even a more dangerous one). Holden Karnofsky discusses how to balance “improving the safety of one’s own models” against “ensuring that other actors do not deploy even-more-unsafe models” in his forthcoming posts about how we could address the issue I lay out in this post. ↩︎

38. See Treacherous turns in the wild(Luke Muehlhauser), Objective robustness in deep reinforcement learning(Koch et al), The effects of reward misspecification(Pan et al), Specification gaming examples in AI(DeepMind Safety blog post which links to an aggregated spreadsheet), The surprising creativity of digital evolution(Lehman et al), this gwern blog post, and more. Many of these lists overlap with each other. ↩︎

39. We could certainly try including artificial scenarios designed to appear as if models could successfully take over from humans, but it seems very hard to be sure that models will be fooled by that and couldn’t tell it apart from a genuine opportunity to take over. For example, models could look for signs that are highly likely to happen eventually but would be too expensive for humans to generate as a training example. As a simple example, Paul Christiano suggested “someone factors an extremely large number” (which would be too expensive to factor just to include as part of the training distribution, but would nonetheless likely get factored eventually). ↩︎

40. For example, stronger arguments that benign generalizations are especially “natural” for gradient descent, enough to make up for the fact that playing the training game would get higher reward; stronger arguments that Alex would “generalize habits” rather than “generalize goals” from training to deployment. ↩︎

41. Say the mice were dropped off on an earlier spaceship. ↩︎

42. Thanks to Buck Shlegeris for suggesting this analogy. ↩︎

43. For one thing, our mouse-brain-sized models have a different and more human profile of abilities than actual mice (e.g. they can talk). ↩︎

44. (who either deliver reward directly or write code that delivers reward) ↩︎

45. Humans would consider egregious reward-grabbing acts to be more blatant and scary violations of expectations than softer ways of playing to human biases, so if Alex is later caught it is likely to receive a large negative reward and/or trigger a larger retraining project. I discuss below why I think that isn’t likely to be sufficient to change its motives to grab reward. ↩︎

46. See Turning reflection up to 11 for a similar proposal. ↩︎

47. For example, Eric Jang, citing Connor Leahy on Twitter, writes: “Just asking the AI to be nice sounds flippant, but after seeing DALL-E and other large-scale multi-modal models that seem to generalize better as they get bigger, I think we should take these simple, borderline-naive ideas more seriously.” ↩︎

48. Note that Alex is likely to be motivated to maximize the “final recorded reward” even if it’s ultimately interested in pursuing some other goal. For example, if Alex is trying to have some kind of lasting effect on the world that extends beyond the current episode (e.g. “discover all the secrets of the universe”), it is probably much more likely to accomplish that goal if the future contains more models like it — which in turn is more likely to happen if it gets a very high reward. ↩︎

49. Humans find lying difficult and unnatural due to our specific evolutionary history. Alex’s design and training wouldn’t necessarily replicate those kinds of evolutionary pressures. Still, it seems plausible that it would, especially since we’d be trying to give it feedback to encourage honesty. ↩︎

50. On a philosophical level, it’s not exactly clear what it means to talk about what actions taken in deployment “would have” gotten a high reward. But on a practical level, Alex is likely to continue to receive rewards on the deployment distribution, particularly if it’s not acting the way researchers want it to act, so the meaning of “reward on the deployment distribution” is fairly well-pinned-down. ↩︎

51. The high-level picture doesn’t change if we imagine it taking in an attention window of a number of observations instead. ↩︎

52. This prediction probably won’t be in the form of separately predicting each pixel of the next screen — it’s more likely to be something more complicated and abstract, e.g. an English-language description of the next screen. ↩︎

53. Although other known sequence processing architectures, like transformers, are also currently limited in how long they can “remember” things (albeit sometimes for different structural reasons). ↩︎

54. This may also reduce the risk that power-seeking misalignment emerges. ↩︎