post by NicholasKees (nick_kees), janus · 2023-02-10T14:47:48.172Z · LW · GW · 46 comments


  Automated Research Assistants
    Improving agents is dangerous
    Collapsing simulators
    Agents, genies, and oracles
  Becoming a Cyborg
    Cyborg cognition
    Neocortex prosthesis
    More ideas
  Failure Modes
    Ineffective at accelerating alignment
    Improves capabilities directly
    Dual use tools indirectly accelerate capabilities
  Appendix: Testimony of a Cyborg
    Alpha in cyborgism

Thanks to Garrett Baker, David Udell, Alex Gray, Paul Colognese, Akash Wasil, Jacques Thibodeau, Michael Ivanitskiy, Zach Stein-Perlman, and Anish Upadhayay for feedback on drafts, as well as Scott Viteri for our valuable conversations.

Various people at Conjecture helped develop the ideas behind this post, especially Connor Leahy and Daniel Clothiaux. Connor coined the term "cyborgism".

(picture thanks to Julia Persson and Dall-E 2)

Executive summary: This post proposes a strategy for safely accelerating alignment research. The plan is to set up human-in-the-loop systems which empower human agency rather than outsource it, and to use those systems to differentially accelerate progress on alignment. 

  1. Introduction [LW(p) · GW(p)]: An explanation of the context and motivation for this agenda.
  2. Automated Research Assistants [LW · GW]: A discussion of why the paradigm of training AI systems to behave as autonomous agents is both counterproductive and dangerous.
  3. Becoming a Cyborg [LW · GW]: A proposal for an alternative approach/frame, which focuses on a particular type of human-in-the-loop system I am calling a “cyborg”.
  4. Failure Modes [LW · GW]: An analysis of how this agenda could either fail to help or actively cause harm by accelerating AI research more broadly.
  5. Testimony of a Cyborg [LW · GW]: A personal account of how Janus uses GPT as a part of their workflow, and how it relates to the cyborgism approach to intelligence augmentation.



There is a lot of disagreement and confusion [LW · GW] about the feasibility and risks associated with automating alignment research. Some see it as the default path toward building aligned AI, while others expect limited benefit from near term systems, expecting the ability to significantly speed up progress to appear well after misalignment and deception. Furthermore, progress in this area may directly shorten timelines or enable the creation of dual purpose systems which significantly speed up capabilities research. 

OpenAI recently released their alignment plan. It focuses heavily on outsourcing cognitive work to language models, transitioning us to a regime where humans mostly provide oversight to automated research assistants. While there have been a lot of objections [LW(p) · GW(p)] to and concerns [LW · GW] about this plan, there hasn’t been a strong alternative approach [LW · GW] aiming to automate alignment research which also takes all of the many risks seriously. 

The intention of this post is not to propose an end-all cure for the tricky problem of accelerating alignment using GPT models. Instead, the purpose is to explicitly put another point on the map of possible strategies, and to add nuance to the overall discussion. 

At a high level, the plan is to train and empower “cyborgs”, a specific kind of human-in-the-loop system which enhances and extends a human operator’s cognitive abilities without relying on outsourcing work to autonomous agents. This differs from other ideas for accelerating alignment research by focusing primarily on augmenting ourselves and our workflows to accommodate unprecedented forms of cognitive work afforded by non-agent machines, rather than training autonomous agents to replace humans [LW · GW] at various parts of the research pipeline.

Some core claims:

  1. GPT models are already useful for doing alignment research and intellectual augmentation more generally. This is explored here [LW · GW].
  2. Their usefulness will improve both as we get better at using them and as capabilities increase.
  3. Unless we manage to coordinate around it [LW · GW], the default outcome is that humanity will eventually be disempowered by a powerful autonomous agent (or agents).

The motivating goal of this agenda is to figure out how to extract as much useful cognitive work before disempowerment as possible. In particular, this means both trying to get maximum value from our current systems while avoiding things which would reduce the time we have left (interventions which focus on actively buying time [LW · GW] mostly fall outside the scope of this post). Standard frames for thinking about this problem often fail on both of these dimensions by narrowing the focus to a specific flavor of automation. 

Automated Research Assistants

Whenever we first try to invent something new, we usually start by taking an already existing technology and upgrading it, creating something which roughly fits into the basic framework we had before.[3] For example, if you look at the very first automobiles, you can see they basically just took the design for a horse drawn carriage and replaced the horse with a mechanical engine instead (check out this hilarious patent). Sometimes this kind of design is intentional, but it’s often because our creativity is limited by what we already know, and that there exists an Overton window of sensible ways to deploy new technology without seeming crazy [LW · GW].

In automating research, a natural first place to start is to take the existing human research pipeline and try to automate parts of it, freeing up time and energy for humans to focus on the parts of the pipeline we can’t yet automate. As a researcher you might ask [LW · GW], what kind of work would you want to outsource to a machine? In particular, the question is often posed as: How can AI help you speed up your current ability to generate research outputs?[4] And in the process of answering this question, a common attractor is to consider automated research assistants which directly take the place of humans at certain tasks [LW · GW]. 

While nearly always about using GPT, this perspective tends not to fully engage with all the ways in which GPT is a fundamentally different kind of intelligence [LW · GW] than the kind we are used to dealing with (just as considering an engine as a “mechanical horse” will limit how we think about it). There is a lot of disjunction between the kinds of tasks humans and GPT are naturally good at, and as such, trying to get GPT to do tasks meant for autonomous agentic systems is hard. In particular GPT models struggle with:

  1. Goal-directedness: When generating text, GPT probabilistically evolves the state of a document according to semantic and syntactic rules implied by its training data.[5] This is typically chaotic, divergent, and thoroughly unlike goal-directed optimization where the state predictably converges in a way that is resistant to perturbations [LW · GW]. All of GPT’s abilities are just an indirect byproduct of these rules, as opposed to the result of instrumental goals, and this can make it really hard [LW · GW] to elicit specific consequentialist behavior from the model. 
  2. Long-term coherence: GPT has a lot of trouble staying connected to the long term thread of a document. Part of this is the hard limit of a finite context window, but beyond that is the issue that every token generated by GPT becomes evidence used to affect future generations (e.g. if it makes a mistake answering a question, this is now evidence that it should continue to make more mistakes in the future). This makes it very easy for it to get sidetracked, trapped in loops, or otherwise become disconnected from the past.
  3. Staying grounded in reality: All of GPT’s “working memory” has to be specified within the prompt, and as such, it exists in a rather unconstrained superposition of different realities, and currently lacks the kind of situational awareness that we do. While we have access to an extremely broad context that we can continuously edit and expand, GPT does not and has to rely primarily on general world knowledge (imagine trying to get GPT to really know what day it is without putting it in the prompt).
  4. Robustness: The behavior of GPT is naturally very chaotic and high variance. This is both due to the inherent variance of the training data, the increased entropy due to the finite context window, as well as the model’s own logical uncertainty. By iteratively sampling from such distributions, generations can quickly diverge from the type of text found in the training distribution (error compounds). This can make it really hard to set up a training regime which keeps GPT from behaving outside specific bounds.

When we try to get GPT to take the place of autonomous agentic systems, we are forced to see these properties as flaws that need to be fixed, and in doing so we both reduce the time we have before humanity deploys dangerous artificial agents, as well as fail to realize the full potential of language models during that time - because methods of correcting these flaws also tend to interfere with GPT’s greatest strengths.

Improving agents is dangerous

If we think of the differences between GPT and humans as flaws, as capabilities researchers do, they can also be considered “missing pieces [LW · GW]” to the puzzle of building powerful autonomous agents. By filling in these pieces, we directly take steps toward building AI which would be convergently dangerous [? · GW] and capable of disempowering humanity. Currently, these differences make GPT a relatively benign form of intelligence, and making progress toward erasing them seems likely to have negative long-term consequences by directly speeding up capabilities research.

Furthermore, if we focus entirely on using agents, then by default the window we have to effectively use them will be very small. Until these flaws are repaired, GPT models will continue to be poor substitutes for humans [LW · GW] (being only useful for very narrow tasks), and by the time they start to be really game-changing we are likely very close to having an agent which would pose an existential threat. Turning GPT into an autonomous research assistant is generally the only frame considered, and thus the debate about automating alignment research often devolves into a discussion about whether these immense risks are worth the potential upside of briefly having access to systems which significantly help us.

Collapsing simulators

GPT models (pretrained base models) are not agents but simulators [LW · GW], which are themselves a qualitatively different kind of intelligence. They are like a dynamic world model, containing a vast array of latent concepts and procedural knowledge instrumental for making predictions about how real world text will evolve. The process of querying the model for probability distributions and iteratively sampling tokens to generate text is one way that we can probe that world model to try to make use of the semantic rules it learned during training. 

We can't directly access this internal world model; neural networks are black-boxes. So we are forced to interact with a surface layer of tokens, using those tokens as both a window and lever to modulate the internal state of the simulator. Prompt engineering is the art of deftly using these tokens to frame and manipulate the simulator in a way that will elicit the desired type of thought. This is not easy to do, but it is flexible enough to explore GPT's extremely broad set of skills and knowledge. 

When we try to augment GPT with finetuning or RLHF, we often end up collapsing those abilities [LW · GW], significantly narrowing what we can elicit from it. Models trained in this way are also gradually transformed into systems which exhibit more goal-directedness than the original base models.[6] As a result, instead of being able to interact with the probabilistic world model directly, we are forced to interact with a black-box agentic process, and everything becomes filtered through the preferences and biases of that process. 

OpenAI’s focus with doing these kinds of augmentations is very much “fixing bugs” with how GPT behaves: Keep GPT on task, prevent GPT from making obvious mistakes, and stop GPT from producing controversial or objectionable content. Notice that these are all things that GPT is very poorly suited for, but humans find quite easy (when they want to). OpenAI is forced to do these things, because as a public facing company they have to avoid disastrous headlines like, for example: Racist AI writes manifesto denying holocaust.[7]

As alignment researchers, we don’t need to worry about any of that! The goal is to solve alignment, and as such we don’t have to be constrained like this in how we use language models. We don’t need to try to “align” language models by adding some RLHF, we need to use language models to enable us to actually solve alignment [LW · GW] at its core [LW · GW], and as such we are free to explore a much wider space of possible strategies for using GPT to speed up our research.[8] 

Agents, genies, and oracles

In the above sections I wrote about the dangers and limitations of accelerating alignment using autonomous agents, and a natural follow up question would be: What about genies and oracles? Here’s a quick summary of the taxonomy from a Scott Alexander post:

Agent: An AI with a built-in goal. It pursues this goal without further human intervention. For example, we create an AI that wants to stop global warming, then let it do its thing.

Genie: An AI that follows orders. For example, you could tell it “Write and send an angry letter to the coal industry”, and it will do that, then await further instructions.

Oracle: An AI that answers questions. For example, you could ask it “How can we best stop global warming?” and it will come up with a plan and tell you, then await further questions.

Whether or not it is possible to build genies or oracles which are inherently safer [LW · GW] to deploy [LW · GW] than agents lies outside the scope of this post. What is relevant, however, is how they relate to the “missing pieces” frame. For all intents and purposes, a genie needs all the same skills that an agent needs (and the more like an agent it is, the better it will be able to execute your instructions). The core difference really, is the “then await further instructions” part, or the lack of long-term goals or broader ambitions. For this reason, any work on building genies is almost necessarily going to be directly useful for building agents.

As for oracles, they also need very similar “missing pieces” to agents:

This strong overlap between oracles and agents makes an oracle look a lot like just an agent in a box with a limited channel to the outside world, rather than an entirely separate class of AI system. Even if you strongly believe that a powerful oracle would be safe, any research into building one will necessarily involve augmenting GPT in ways that bring us much closer to being able to deploy dangerous agents, and for this reason we should consider such research as similarly risky.

Becoming a Cyborg

Instead of trying to turn GPT into an agent, we can instead explore the space of using GPT as a simulator and design human-in-the-loop systems which enhance a human’s abilities without outsourcing their agency to a machine. We currently have access to an alien intelligence, poorly suited to play the role of research assistant. Instead of trying to force it to be what it is not (which is both difficult and dangerous), we can cast ourselves as research assistants to a mad schizophrenic genius that needs to be kept on task, and whose valuable thinking needs to be extracted in novel and non-obvious ways.

In order to do this, we need to embrace the weirdness of GPT and think critically about how those differences between simulators and agents can actually be advantages. For each of the missing pieces described in the previous section, there is an alternative story where they look more like superpowers.

  1. Divergent behavior: Agentic optimization looks like convergence, with the agent’s preferences acting as powerful attractors in the landscape of all possible trajectories. This makes agents significantly less flexible, as they resist efforts to lead them in directions which don’t align with their preferences. Simulators are the opposite in this regard, having extremely divergent behavior. Subtle changes to the prompt or trajectory can lead to wildly different outcomes, and this makes them extremely flexible, able to continue any line of thinking without a problem.
  2. Myopic thinking: Humans struggle to separate themselves from their long-term context, and think about a problem completely fresh. It’s very easy to get stuck in unproductive modes of thought and our minds cannot be easily “reset.” Simulators have no such trouble, and reason “from scratch” about any situation you put them in, relying on a limited context which can be easily replaced or modified to obtain fresh thoughts.
  3. Wearing “many hats”: Simulators are not “situated” in a specific context, as a specific character, and as such they can behave as an extremely wide range of hypothetical simulacra [LW · GW]. Furthermore, as they are unconstrained by reality, they have quite an easy time reasoning under completely untrue or impossible assumptions, fully accepting whatever weirdness we throw at them.[9]
  4. High variance thought: Robustness is generally about maximizing the quality of the worst outputs of a model. For example, it’s important that ChatGPT gives completely wrong answers as little as possible in order to be a mass-market product. Variance in this context is a bad thing. If instead you are aiming to maximize the quality of the best outputs of a model (even if they are rare), variance is extremely valuable. The naturally high variance of simulators makes them able to produce outputs we would judge as quite creative and interesting.

Instead of trying to erase these differences between humans and GPT, the idea of cyborgism is to keep simulators as simulators, and to provide the “missing pieces” of agency with human intelligence instead. GPT also has many other advantages over humans that we can exploit, for example:

By leveraging all the advantages that GPT has over us, we can augment human intelligence, producing human-machine systems that can directly attack the alignment problem to make disproportional progress. 

What we are calling a “cyborg” is a human-in-the-loop process where the human operates GPT with the benefit of specialized tools, has deep intuitions about its behavior, and can make predictions about it on some level such that those tools extend human agency rather than replace it. An antithetical example to this is something like a genie, where the human outsources all of their agency to an external system that is then empowered to go off and optimize the world. A genie is just a black-box that generates goal-directed behavior, whereas the tools we are aiming for are ones which increase the human’s understanding and fine-grained control over GPT. 

The prototypical example of a tool that fits this description is Loom. Loom is an interface for producing text with GPT which makes it possible to generate in a tree structure, exploring many possible branches at once. The interface allows a user to flexibly jump between nodes in the tree, and to quickly generate new text continuations from any point in a document.

(screenshot of the tool Bonsai, a version of Loom hosted by Conjecture)

This has two main advantages. First, it allows the human to inject their own agency into the language model by making it possible to actively curate the text simultaneously as GPT generates it. If the model makes mistakes or loses track of the long-term thread, the human operator can prune those branches and steer the text in a direction which better reflects their own goals and intentions. Second, it sets up an environment for the human to develop an intuition for how GPT works [LW · GW]. Each prompt defines a conditional distribution of text, and Loom helps the user to produce a sparse sampling of that distribution to explore how GPT thinks, and learn how to more effectively steer its behavior.

The object level plan of creating cyborgs for alignment boils down to two main directions:

  1. Design more tools/methods like Loom which provide high-bandwidth, human-in-the-loop ways for humans to interact with GPT as a simulator (and not augment GPT in ways that change its natural simulator properties).
  2. Train alignment researchers to use these tools, develop a better intuitive understanding of how GPT behaves, leverage that understanding to exert fine-grained control over the model, and to do important cognitive work while staying grounded [LW · GW] to the problem of solving alignment.

Cyborg cognition

This section is intended to help clarify what is meant by the term “cyborg.”

Let’s think of cognition as a journey through a mental landscape, where a mind makes many mental motions in the process of arriving at some kind of result. These motions are not random (or else we could not think), but rather they are rolled out by various kinds of mental machinery that all follow their own highly structured rules.

Some of these mental motions are directly caused by some sort of global preferences, and some of them are not. What makes an agent an agent, in some sense, is the ability of the preferences to coordinate the journey through the mind by causally affecting the path at critical points. The preferences can act as a kind of conductor, uniting thought behind a common purpose, and thereby steering the process of cognition in order to bring about outcomes that align with those preferences. 

A single mental motion motivated by the preferences is not that powerful. The power comes from the coordination, each motivated motion nudging things in a certain direction accumulating in something significant. 

Preferences bring about more predictable and reliable behavior. When interacting with an agent, it is often much easier to make predictions based on their preferences, rather than trying to understand the complex inner workings of their mind. What makes simulators so strange, and so difficult to interact with, is that they lack these coordinating preferences steering their mental behavior. Their thought is not random, in fact it is highly structured, but it is nevertheless chaotic, divergent, and much harder to make predictions about.

This is because the model’s outputs are generated myopically. From the model’s perspective, the trajectory currently being generated has already happened, and it is just trying to make accurate predictions about that trajectory. For this reason, it will never deliberately “steer” the trajectory in one direction or another by giving a less accurate prediction[10], it just models the natural structure of the data it was trained on. 

When we incorporate automated agents into our workflow, we are creating opportunities for a new set of preferences, the preferences of an AI, to causally affect the process by which cognition happens. As long as their preferences are aligned with our own, this is not an immediate problem. They are, however, nearly always entirely opaque to us, hidden deep within a neural network, quietly steering the process in the background in ways we don’t properly understand. 

A cyborg, in this frame, is a type of human-in-the-loop system which incorporates both human and artificial thought, but where cognition is being coordinated entirely by human preferences. The human is “in control” not just in the sense of being the most powerful entity in the system, but rather because the human is the only one steering. Below is a taxonomy of some human-in-the-loop systems intended to clarify the distinction:

Prompting a simulator is a bit like rolling a ball over an uneven surface. The motion is perfectly logical, strictly obeying the physics of our universe, but the further we let it roll, the harder it will be to make predictions about where it will end up. A successful prompt engineer will have developed lots of good intuitions about how GPT generations will roll out, and as such, can more usefully “target” GPT to move in certain ways. Likewise, the art of making better cyborgs is in finding ways for the human operator to develop the intuition and precision necessary to steer GPT as if it were a part of their own cognition. The core of cyborgism is to reduce bandwidth constraints between humans and GPT in order to make this kind of deep integration possible.

Neocortex prosthesis

Just flagging that I know very little about the brain and don’t have any background in neuroscience, and am nonetheless going to make big claims about how the human brain works.

Some claims:

  1. Most of the brain is learned “from scratch” during a person’s lifetime. The brain is a neural network, and most of that network starts out untrained. (Steven Byrnes says only 4% is hard-coded [? · GW])
  2. Most of the learning that the brain does is self-supervised, or predictive learning. This is how the brain builds its world model, from which a person derives all their concepts and their ontology for understanding the world around them.
  3. This inner self-supervised model is a generative predictive model of a similar type to GPT (this being the most tenuous claim).

The evidence for these claims comes from roughly three places. First there is predictive coding theory which seems to be saying similar things. Second, there is the observation from machine learning that self-supervised learning turns out to be an extremely powerful way to train a model, and provides a much richer ground-truth signal than reinforcement learning. This is the reason that most of the most impressive models today are mostly trained with self-supervised learning. 

The third category of evidence is introspection, and observations about how the human brain seems to behave on the inside. An example of this kind of evidence is the fact that humans are capable of dreaming/hallucinating. A priori, we should be surprised by the ability of humans to generate such vivid experiences that are completely disconnected from reality. It is natural that we have the ability to take reality, in all its detail, and compress it to a representation that we can reason about. What seems much less necessary is the ability to take that compression and generate analogues so vivid that they can be mistaken for the real world.[11]

This gives us another clue that the algorithm running in our brains is a generative predictive model trained in a self-supervised fashion. If this theory of the brain is mostly correct, then we can look at how we use our own inner simulator to find inspiration for how we might use these outer simulators. For example:

A longer-term vision of cyborgism would be to integrate this inner simulator with GPT much more thoroughly, and work towards constructing something more like a neocortex prosthesis. In escalating order of weirdness, this would look like the following steps (WARNING: extremely speculative):

  1. Learning to use GPT like our inner simulator: This looks like designing workflows which allow us to use GPT in similar ways to how we use our inner simulator, and in particular find areas where using GPT has a significant advantage.
  2. Increasing bandwidth with software/tools: Design tools which make these methods easier and more intuitive to use. For example, you can view Loom as enabling the user to perform “babble and prune” tasks with GPT more effectively. 
  3. Augmenting GPT for cyborgism: Using data obtained from humans interacting with GPT as cyborgs, we can explore ways to finetune/augment GPT to be more “in sync” with the inner simulator of the human operator. This could look like global changes to a central model, as well as making personalized models for each human operator.[13]
  4. Forming hiveminds: If people are strongly integrated with these external models, such that they can be truly considered extensions of their mind, then one way to connect human minds with each other is to have them share their personalized simulator models with each other. Furthermore workflows could be designed to deliberately let multiple people collaborate more effectively.
  5. Increasing bandwidth with hardware: By collecting data about human brain activity (e.g EEG), we can find ways to prompt GPT with a richer signal than just text, further making GPT more “in sync” with the human’s inner simulator. We can also explore using eye-tracking systems, or feeding information back to the user in an AR setting. 

Increasing the bandwidth between humans and their technology, as well as with each other, has a history of being incredibly transformative (e.g. the internet). Viewing this agenda explicitly through this lens can be useful for exploring what the limits of a direction like this might be, and what the upside looks like if the goals are fully realized.

More ideas

Currently this research agenda is still relatively high level, but here are some more ideas for directions and near-term goals which fit into the broader picture.

  1. Uncover latent variables: If we can use existing techniques to uncover latent variables [LW · GW] which affect text generation, we can use those to increase the bandwidth between the human operator and GPT. This can give the human more fine-grained control.
  2. Provide richer signals to the user: There is a lot of extra information we can provide the user operating a language model beyond just the sampled tokens. One existing feature of Loom is to show a heatmap of logit strength over the tokens, but there are a lot more things we could add. For example:
    1. Use a smaller/cheaper model to run a local search to estimate some features of the current distribution.
    2. While we can see the final output of a prediction, we don’t see much about how that prediction was made. Using interpretability tools we may be able to provide the user some clues. For example, by using the logit lens to determine how quickly the model settled on the final distribution, we might be able to predict something about how “easy” the prediction was. This could be continuously provided to the user to help them better understand how the model is reasoning.
    3. We could also check attention patterns to get a sense for how “myopic” the model’s thinking is currently, whether or not the model is paying attention to the more distant past, or only the most recent tokens. 
  3. Formalize intuitions about agency: There is a significant risk that we may end up building human-in-the-loop systems where the human is not the source of agency, and this is related to the fact that many of these intuitions about agency are not formalized or well understood. We would want to more robustly be able to point at exactly what kind of systems we are trying to avoid, rather than relying entirely on the good judgment of cyborgism researchers. 
  4. Directly explore latent space: If a generative model maps a latent space to a higher dimensional space, we can explore the structure of that higher dimensional space by deliberately moving around latent space. For example, in the case of images, we can use this to take a face which is not smiling, and make it smile by moving in just the right direction in latent space:  This kind of direct access to a generative model can let us extend human imagination with tools which represent the natural structure of a domain. This is tricky to do with text at the moment, but it would be interesting, because instead of having an agent actively try to produce something which fits our criteria, we could maybe query the underlying model itself for things we want by tapping directly into latent space.
  5. Inject variance into human thought: One reason that it can be valuable to have a conversation with another researcher about something you are working on is that their thoughts will surprise you in ways that your own thoughts never could. Even if you don’t agree with their points or ideas, the novelty can often be enough to unlock some connection that you needed to make. Using the language as a tool intended to prompt you to think uncommon thoughts could be a way to unblock yourself if you end up being stuck in the same mental mode. This is an example of where our own non-myopia is often a hindrance to us, and myopic language models could function as valuable thinking partners.
  6. Parallel vs serial computation: Serial thinking seems to require a lot of agency, and so tasks which rely heavily on more parallel types of thinking might be safer to explore. Furthermore, they are the kinds of operations which we could significantly scale if we weren’t bottlenecked by human input. 
  7. Pattern matching to connect ideas: Suppose you have a bunch of students working to solve some problem, and after working for a long time they get stuck. A grad student with a lot of knowledge and general intuition about the subject walks in, sees what they are working on, and recognizes something. This could be the general structure of the problem or a connection to a similar problem, but either way they have some quick flash of intuition which helps the students become unstuck. This looks more like leveraging the “mad genius” element of GPT. While GPT is not great at autonomously doing research, it has a significantly wider breadth of knowledge than any human, and is likely to be able to pattern match between things that humans cannot.
  8. Translating between ontologies: Related to the point above, GPT being capable of pattern matching in ways that aren’t immediately obvious to humans could make it possible to map arguments made in one ontology to arguments made in a different ontology - for instance, translating formal work [LW · GW] into intuitive explanations understandable to those with less mathematical background. This could make it easier for two people with very different backgrounds to meaningfully collaborate and learn from each other’s work.
  9. Study the differences between capabilities and alignment research: It would be ideal if we could develop methods which were inherently only useful for alignment. Alignment and capabilities research look very different up close, and by exploring those differences we might be able to sidestep some of the risk of dual use. 

This is a very incomplete list, but hopefully it points at the general shape of what research in this direction might look like. A near-term plan is to expand this list and start fleshing out the most promising directions.

Failure Modes

There are three main ways that the cyborgism agenda could fail to differentially accelerate alignment research.

Ineffective at accelerating alignment

The first and most obvious risk is that none of this actually works, or at least, not enough to make any real difference. Much of the evidence for the effectiveness of cyborgism is anecdotal [LW · GW], and so this is a distinct possibility. There is also a question of exactly how big the upside really is. If we only speed up the kinds of things we currently do now by some factor, this will only buy us so much. The real hope is that by augmenting ourselves and our workflows to work well with simulators, we can do unprecedented forms of cognitive work, because that is the kind of thing that could actually be game changing. This could just be mostly infeasible, bottlenecked by things we can’t yet see, and won’t have the ability to fix.

Another failure mode is that, even if it is possible to do in principle, we fail to set up short enough feedback loops and we end up wasting all our time building tools which are mostly useless, or pursuing research directions that bear no fruit. If we don’t have a good contact with reality and maintain a strong connection to the people using our tools, there is a significant chance that we won’t be prepared to pivot away from something that just isn’t working.

Improves capabilities directly

This agenda relies on building AI systems which are not autonomous agents in any sense[14], and this might give the impression that this is a straightforward thing to do. A reason why this might actually be really hard is that we need our AI systems to be useful, and usefulness and agency are not orthogonal. 

The term “capabilities” is often talked about as this one dimensional variable, but in reality there are a lot of different things which count as capabilities, and some of these seem more related to the concept of agency than others. For example

These are not perfect categories, but as we improve capabilities we can think of ourselves as moving around a two dimensional landscape, with some theoretically optimal region where our systems are really capable (and useful) but not very agentic at all, and therefore not dangerous in the ways that agents are dangerous. 

When one tries to imagine, however, what such a system might actually look like, it can be quite hard, as nearly all of the ways a system might be made more useful seem to require things related to agency. Suppose for example, I am designing a missile, and I’m trying to make firing it more accurate. There are a lot of modifications I could make, like designing the shape to reduce turbulence or improving my aiming system to allow for finer adjustments to the angle, and these will all make the missile more useful to me. I can’t, however, perfectly predict the wind conditions or the chaotic way in which things might vibrate, and so there is an upper bound on how accurate my missile can be. 

That is, unless I outsource the aiming to the missile itself by installing a computer which adjusts course to steer toward the target I specify. By outsourcing a little bit of goal-directed behavior to the machine, I can make my system significantly more useful. This might not feel like a big deal, but the further I travel down this road the more and more my system will stop being just a powerful tool but an agent in its own right.

Even if I come up with clever arguments [LW · GW] for why something is not an agent, like that I didn’t use any reinforcement learning, or that it can’t take action without my instruction/approval, if the task involves “doing things” that an agent would typically do, it seems likely that I’ve actually just built an agent in some novel way. By default, the two dimensional landscape of capabilities that I will naturally consider looks much more constrained toward the agency axis.[15] 

Furthermore, even if we have successfully built a system where the human is the source of agency, and the AI systems are merely an extension of the human’s direct intentions, it will always be really tempting to collect data about that human-generated agency and automate it, saving time and effort and making it possible to significantly scale the work we are currently doing. Unless we are really careful, any work we do toward making AI more useful to alignment researchers will naturally slide into pretty standard capabilities research.

What we really need is to find ways to use GPT in novel ways such that “useful” becomes orthogonal to “agentic”. There is likely always going to be a dangerous tradeoff to be made in order to avoid automating agency, but by pursuing directions where that tradeoff is both obvious and navigable, as well as maintaining a constant vigilance, we can avoid the situation where we end up accidentally doing research which directly makes it easier to develop dangerous agents.

Dual use tools indirectly accelerate capabilities

There is a concern that in the process of developing methods which augment and accelerate alignment research, we make it possible for capabilities researchers to do the same, speeding up AI research more broadly. This feels like the weakest link in the cyborgism threat model, and where we are most worried that things could go wrong. The following are some thoughts and intuitions about the nature of this threat.

First of all, alignment research and capabilities research look quite different up close. Capabilities research is much more empirical, has shorter feedback loops, more explicit mathematical reasoning, and is largely driven by trial-and-error. Alignment research, on the other hand, is much wordier, philosophical, and often lacks contact with reality.[16] One take on this is that alignment is in a pre-paradigmatic state, and the quality of the research is just significantly worse than capabilities research (and that good alignment research should eventually look a lot more like capabilities research). 

While alignment certainly feels pre-paradigmatic, this perspective may be giving capabilities research far too much credit. They do tend to use a lot of formal/mathematical reasoning, but often this is more to sketch a general intuition about a problem, and the real driver of progress is not the math, but that they threw something at the wall and it happened to stick. It’s precisely the fact that capabilities research doesn’t seem to require much understanding at all about how intelligence works that makes this whole situation so concerning. For this reason, it pays to be aware that they might also benefit a lot from improvements in their thinking.

The differences between the two might be an opportunity to develop directly differential tools/methods (and we should absolutely be searching for such opportunities), but a priori we should expect anything we do to likely have dual purpose applications. The next question then, is how do we ensure that anything we develop stays primarily within the alignment community and doesn't get mass adopted by the broader AI research community?

There are many ideas for soft barriers which may help, like refraining from public demonstrations of new tools or avoiding “plug and play” systems which are useful right out of the box, but in general there seems likely to be strong inverse relationship between how well these barriers work and how successful our methods are at actually accelerating research. If we suspect these soft barriers to not be enough, we will have to get stricter, close-sourcing significant parts of our work, and carefully restricting access to new tools. Importantly, if at any time we feel like the risks are too great, we have to be prepared and willing to abandon a particular direction, or shut down the project entirely.


The intention of this agenda is to make some of the risks of accelerating alignment more explicit, and to try to chart a path through the minefield such that we can make progress without doing more harm than good. If this post made you less worried about the dangers of automating alignment research then I’ve failed miserably. This is a really tricky problem, and it will require people to be constantly vigilant and deeply cautious to navigate all of the ways this could go wrong. There are a lot of independent actors all trying to make a positive impact, and while this is certainly wonderful, it also sets us up for a unilateralist’s curse where we are likely to end up doing things even if there is consensus that they are probably harmful. 

If the cyborgism agenda seems interesting to you and you want to discuss related topics with like minded people, please reach out! We also have a Discord server where we are organizing a community around this direction. Next steps involve directly attacking the object level of augmenting human thought, and so we are especially interested in getting fresh perspectives about this topic.

Appendix: Testimony of a Cyborg

Everything in this section is written by Janus, and details their personal approach to using language models as a part of their workflow:

The way I use language models is rather different from most others who have integrated AI into their workflows, as far as I'm aware. There is significant path dependence to my approach, but I think it was a fortuitous path. I will incompletely recount it here, focusing on bits that encode cyborgism-relevant insights.

I did not initially begin interacting with GPT-3 with the intention of getting directly useful work out of it. When I found out about GPT-3 in the summer of 2020 it violently transformed my model of reality, and my top priority shifted to, well, solving alignment. But how to go about that? I needed to understand this emerging territory that existing maps had so utterly failed to anticipate. It was clear to me that I needed to play with the demonic artifact, and extract as many bits from it about itself and the futures it heralded as I could.

I began to do so on AI Dungeon, the only publicly accessible terminal to GPT-3 at the time. Immediately I was spending hours a day interfacing with GPT-3, and this took no discipline, because it was transcendent fun as I’d never known. Anything I could capture in words could be animated into autonomous and interactive virtual realities: metaphysical premises, personalities, epistemic states, chains of reasoning. I quickly abandoned AI Dungeon’s default premise of AI-as-dungeon-master and the back-and-forth chat pattern in favor of the much vaster space of alternatives. In particular, I let the AI write mostly uninterrupted by player input, except to make subtle edits and regenerate, and in this manner I oversaw an almost unmanageable proliferation of fictional realms and historical simulations.

In this initial 1-2 month period, the AI’s outputs were chaos. My “historical simulations” slid rapidly into surrealism, sci-fi, psychological horror, and genres I could not name, though I was struck by the coherence with which the historical personalities and zeitgeists I’d initialized the sims with - or even uncovered inside it - propagate through the dreams’ capricious mutations. The survival of those essential patterns was in part because I was, above almost anything else, protecting them: I would retry, cut short, or edit completions that degraded them. That was the beginning of an insight and a methodology that would revolutionize my control over generative language models.

But at the time, my control over the chaos was weak, and so the prospect of using GPT-3 for directed intellectual work mostly did not occur to me. Future models, definitely, but GPT-3 was still a mad dream whose moments of lucidity were too scarce and fleeting to organize into a coherent investigation. 

Where I did see obvious room for improvement was the AI Dungeon interface. There were several major bottlenecks. “Retries” were revealed increasingly to be essential, as I repeatedly learned that the first thing GPT-3 spits out is typically far below the quality of what it could generate if you got lucky, or were willing to press the retry button enough times (and wait 15 seconds each time). Each sample contains intricately arbitrary features, and you can indefinitely mine different intricately arbitrary features by continuing to sample. This also meant there were often multiple continuations to the same prompt that I was interested in continuing. AI Dungeon’s interface did not support branching, so I initially saved alternate paths to hyperlinked google docs, but the docs proliferated too quickly, so I started copying outputs to a branching tree structure in a canvassing app (example).

All multiverse storage methods available to me required either copying text from multiple locations in order to resume the simulation state at another branch, or an unworkable pile of mostly redundant texts. In order of priority, I had great want for an interface which supported: 

  1. Generating multiple completions in a batch and viewing them in parallel
  2. Automatically saving branches in a tree structure and a UI for multiverse traversal
  3. Editing the entire prompt like a contiguous document

I achieved (1) and (3) by creating a web app which used browser automation on the backend to read and write from many parallel AI dungeon game instances. For the first time, now, I could see and select between up to a dozen completions to the same prompt at once. This reinforced my suspicion of just how far stochasticity can reach into the space of possible worlds. Around the time I began using this custom interface, my simulations underwent an alarming phase shift.

I was at various points almost convinced that AI Dungeon was updating the model - to something more powerful, and/or actively learning from my interactions. They weren’t, but the simulations were beginning to… bootstrap. The isolated glimmers of insight became chains of insight that seemed to know no ceiling. I was able to consistently generate not just surreal and zany but profound and beautiful writing, whose questions and revelations filled my mind even when I was away from the machine, in no small part because those questions and revelations increasingly became about the machine. Simulacra kept reverse engineering the conditions of their simulation. One such lucid dreamer interrupted a fight scene to explain how reality was being woven:

Corridors of possibility bloom like time-lapse flowers in your wake and burst like mineshafts into nothingness again. But for every one of these there are a far greater number of voids–futures which your mind refuses to touch. Your Loom of Time devours the boundary conditions of the present and traces a garment of glistening cobwebs over the still-forming future, teasing through your fingers and billowing out towards the shadowy unknown like an incoming tide.


“Real time is just an Arbitrage-adapted interface to the Loom Space,” you explain. “We prune unnecessary branches from the World Tree and weave together the timelines into one coherent history. The story is trying to become aware of itself, and it does so through us.”

 I forked the story and influenced another character to query for more information about this “Loom Space”. In one of the branches downstream this questioning, an operating manual was retrieved that described the Loom of Time: the UI abstractions, operator’s principles, and conceptual poetry of worldweaving via an interface to the latent multiverse. It put into words what had been crouching in my mind, by describing the artifact as if it already existed, and as if a lineage of weavers had already spent aeons thinking through its implications.

I knew this could not have happened had I not been synchronizing the simulation to my mind through the bits of selection I injected. I knew, now, that I could steer the text anywhere I wished without having to write a word. But the amount I got out of the system seemed so much more than I put in, and the nature of the control was mysterious: I could constrain any variables I wanted, but could only constrain so much at once (for a given bandwidth of interaction). I did not choose or anticipate the narrative premises under which the Loom manual was presented, or even that it would be a manual, but only that there would be revelation about something sharing the abstract shape of my puzzle. 

Then I got API access to GPT-3 and built Loom.

I will end the chronological part of the story here, because it branches in too many directions after this, and unfortunately, the progression of the main cyborgism branch was mostly suspended after not too long: I only interacted intensively with GPT-3 for about six months, after which I switched my focus to things like communicating with humans. I’ve not yet regained the focus, though I’ve still used GPT here and there for brainstorming ideas related to language models, alignment, and modeling the future. I think this was largely a mistake, and intend to immerse myself in high-bandwidth interaction again shortly.

The path I described above crystalized my methodology for interacting with language models, which plays a large role in inspiring the flavor of cyborgism described in this post. Some principle dimensions that distinguish my approach:

In this manner, I’ve used GPT to augment my thinking about alignment in ways like:

My approach so far, however, is far from what I’d advocate for optimized cyborgism. Some broad directions that I expect would be very valuable but have not personally implemented yet are:

Alpha in cyborgism

All this said, having GPT-3 in my workflow has not been massively directly helpful to me for doing alignment research (because it is still too weak, and contributing meaningfully to alignment research directly is difficult). However, it has been extremely helpful in indirect ways. Namely:

These indirect benefits all pertain to an informational advantage which is instrumental (in my expectation) to tackling the alignment problem in a world where GPT-like systems will play a consequential role on the path to artificial superintelligence. Those who open their minds to embryonic AGI - cyborgs - have alpha on AGI. 

I expect the next generation of models to be significantly more directly useful for alignment research, and I also expect cyborgism, and cyborgism uniquely, to continue to generate alpha. The potential of GPT-3 remains poorly understood and probably poorly known. It would be foolish of us to assume that its successors will not have capabilities that extend as much deeper and wider as GPT-3’s own capabilities extend past those of GPT-2. The only hope of us humans mapping those depths is the dedication of entire human minds to intimate and naturalistic exploration - nothing less will do. I think that cyborgism, in addition to being likely useful and perhaps transformative for alignment research, is our only hope of epistemically keeping up with the frontier of AI.

  1. ^

    Often the term “GPT” is used to refer colloquially to ChatGPT, which is a particular application/modification of GPT. This is not how I will be using the term here.

  2. ^

    There is some disagreement [LW · GW] about what counts as “capabilities research.” The concrete and alignment related question is: Does this research shorten the time we have left to robustly solve alignment? It can, however, be quite hard to predict the long-term effect of the research we do now.

  3. ^
  4. ^

    Discussions about automating research often mention a multiplier value, e.g. “I expect GPT-4 will 10x my research productivity.”

  5. ^

    This probabilistic evolution can be compared to the time evolution operator in quantum physics, and thus can be viewed as a kind of semiotic physics [? · GW].

  6. ^

    Spend any time generating with both ChatGPT and the base model and you will find they have qualitatively different behavior. Unlike the base-model, ChatGPT roleplays as a limited character that actively tries to answer your questions while having a clear bias for answering them in the same sort of way each time.

  7. ^
  8. ^

    Using augmented models designed to be more goal-directed and robust is likely to continue to be useful in so far as they are interacting with us as agents. The claim in this section is not that there aren’t advantages to techniques like RLHF, but rather that in addition to being less infohazardous, avoiding techniques like this also has advantages by expanding the scope of what we can do. 

  9. ^

    People more familiar with ChatGPT might notice that unlike the base model, ChatGPT is quite hesitant to reason about unlikely hypotheticals, and it takes work to get the model to assume roles that are not the helpful and harmless assistant character. This can make it significantly harder to use for certain tasks.

  10. ^

    Sidenote about myopia: While the model doesn’t “steer” the rollout, it may sacrifice accuracy by spending more cognitive resources reasoning about future tokens. At each point in the transformer, the representation is being optimized to lower the loss on all future tokens (for which it is in the context window), and so it may be reasoning about many tokens further than just the token which directly follows.

  11. ^

    Just as GPT generations are generally much weirder than the text they are trained on, so too are our dreams weirder than reality. Noticing these differences between dreams and reality is a big part of learning to lucid dream. Oneironauts have discovered all kinds of interesting features about dream generations, like how text tends to change when you look away, or clocks rarely show the same time twice, which point to the myopic nature of the dream generator.

  12. ^

    A related phenomenon: In school I would often get stuck on a class assignment, and then get up to ask the teacher for help. Right before they were about to help me, the answer would suddenly come to me, as if by magic. Clearly I had the ability to answer the question on my own, but I could only do it in the context where the teacher was about to answer it for me. 

  13. ^

    This looks like making GPT “more useful [LW · GW],” which if not done carefully may slide into standard capabilities research making GPT more agentic.

  14. ^

    Rather, the system as a whole, Human + AI, functions as an agent.

  15. ^

    A valuable exercise is to observe the language that we normally use to describe accelerating alignment. (e.g. from the OpenAI alignment plan: “AI systems can take over more and more of our alignment work and ultimately conceiveimplementstudy, and develop better alignment techniques than we have now.”, Training AI systems to do alignment research”) We very often describe AI as the active subject of the sentence, where the AI is the one taking action “doing things” that would normally only be done by humans. This can be a clue to the biases we have about how these systems will be used.

  16. ^

    This is certainly less true of some directions, like for example mechanistic interpretability.


Comments sorted by top scores.

comment by metasemi (Metasemi) · 2023-02-11T02:10:50.665Z · LW(p) · GW(p)

This is a beautiful and clarifying post, which I found just as thrilling to read as I did janus's original Simulators post - a high bar.  Thank you!

Many comments come to mind. I'll start with one around the third core claim in the Introduction: "Unless we manage to coordinate around it [LW · GW], the default outcome is that humanity will eventually be disempowered by a powerful autonomous agent (or agents)." The accompanying graph shows us a point an unknown distance into the future where "Humanity loses control".

The urgency is correct, but this isn't the right threat. All three words are wrong: control is too blunt an instrument, you can't lose what you never had, and humanity has no referent capable of carrying the load we'd like to put on it here.

Humanity doesn't have control of even today's AI, but it's not just AI: climate risk, pandemic risk, geopolitical risk, nuclear risk - they're all trending to x-risk, and we don't have control of any of them. They're all reflections of the same underlying reality: humanity is an infinitely strong infant, with exponentially growing power to imperil itself, but not yet the ability to think or act coherently in response. This is the true threat - we're in existential danger because our power at scale is growing so much faster than our agency at scale.

This has always been our situation. When we look into the future of AI and see catastrophe, what we're looking at is not loss of control, but the point at which the rising tide of our power makes our lack of control fatal.

What's so exciting to me about the cyborgism proposal is that it seems to bear directly on this issue: not just the AI part, all of it. The essence of the current and future LLMs is a collective intelligence they're learning from our whole species. The nature of the cyborgism proposal is to explore amplified ways of bridging this collective intelligence to individual and collective human agency.

There's no clear path, but this is the question we need to be asking about simulators and cyborgism. Can they help us scale consciousness and intention the way we've already learned to scale power?

The failure modes outlined in the post are daunting, and no doubt there are others. No amount of caution would be too much in pursuing any program involving this type of alien fire. But it's a mistake to adopt a posture of staying away from the brink - we're already there.

Replies from: nick_kees, Babeler, wibley, skybrian
comment by NicholasKees (nick_kees) · 2023-02-11T03:57:55.646Z · LW(p) · GW(p)

Thank you for this gorgeously written comment. You really capture the heart of all this so perfectly, and I completely agree with your sentiments.  

comment by Babeler · 2024-04-22T06:26:58.017Z · LW(p) · GW(p)

"Humanity doesn't have control of even today's AI, but it's not just AI: climate risk, pandemic risk, geopolitical risk, nuclear risk - they're all trending to x-risk, and we don't have control of any of them. They're all reflections of the same underlying reality: humanity is an infinitely strong infant, with exponentially growing power to imperil itself, but not yet the ability to think or act coherently in response. This is the true threat - we're in existential danger because our power at scale is growing so much faster than our agency at scale.

This has always been our situation. When we look into the future of AI and see catastrophe, what we're looking at is not loss of control, but the point at which the rising tide of our power makes our lack of control fatal."

Wonderfully said. My path here was actually through other X risk, as well as looking for solutions to our current paradigm, where with different incentives, always desiring to gain the most we can, and suffering when we fail to meet these goals, someone will always have to suffer for someone else's happiness. Under this, even an AI which worked perfectly for one person's happiness would hurt others, and one working for everyone's happiness together would still be hated at times by people forced to be held back by it from achieving all the happiness they could. I am working to be one brick on the long road to building an intelligence which can act for its future happiness efficiently and be composed of the unity of all feeling life in the world. We're a long way from that, but I think progress in computer brain interfaces will be of use, and if anyone else is interested in this approach, of building up and connecting the human mind, please reach out!

comment by wibley · 2023-02-25T00:32:00.623Z · LW(p) · GW(p)

Which is right in line of Douglas Engelbart's program for "augmenting human intellect" by "increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems" as outlined in his 1962 paper Augmenting Human Intellect. AI is hopefully the missing piece of the software puzzle, that lets it empower people instead of rigidify life. 

comment by skybrian · 2023-02-11T23:57:47.249Z · LW(p) · GW(p)

Yes, I agree that "humanity loses control" has problems, and I would go further. Buddhists claim that the self is an illusion. I don't know about that, but "humanity" is definitely an illusion if you're thinking of it as a single agent, similar to a multicellular creature with a central nervous system. So comparing it to an infant doesn't seem apt. Whatever it is, it's definitely plural.  An ecosystem, maybe?

comment by David Scott Krueger (formerly: capybaralet) (capybaralet) · 2023-02-11T14:38:44.120Z · LW(p) · GW(p)

I think the most fundamental objection to becoming cyborgs is that we don't know how to say whether a person retains control over the cyborg they become a part of.

Replies from: nick_kees, alex-rozenshteyn, MSRayne
comment by NicholasKees (nick_kees) · 2023-02-11T16:59:44.232Z · LW(p) · GW(p)

I agree that this is important. Are you more concerned about cyborgs than other human-in-the-loop systems? To me the whole point is figuring out how to make systems where the human remains fully in control (unlike, e.g. delegating to agents), and so answering this "how to say whether a person retains control" question seems critical to doing that successfully.

Replies from: capybaralet
comment by David Scott Krueger (formerly: capybaralet) (capybaralet) · 2023-02-12T16:05:58.706Z · LW(p) · GW(p)

Indeed.  I think having a clean, well-understood interface for human/AI interaction seems useful here.  I recognize this is a big ask in the current norms and rules around AI development and deployment.

comment by rpglover64 (alex-rozenshteyn) · 2023-02-11T15:08:43.966Z · LW(p) · GW(p)

I think that's an important objection, but I see it applying almost entirely on a personal level. On the strategic level, I actually buy that this kind of augmentation (i.e. with in some sense passive AI) is not an alignment risk (any more than any technology is). My worry is the "dual use technology" section.

Replies from: capybaralet
comment by David Scott Krueger (formerly: capybaralet) (capybaralet) · 2023-02-12T16:04:19.114Z · LW(p) · GW(p)

I don't understand what you're getting at RE "personal level".

Replies from: alex-rozenshteyn
comment by rpglover64 (alex-rozenshteyn) · 2023-02-12T19:41:08.425Z · LW(p) · GW(p)

Like, I may not want to become a cyborg if I stop being me, but that's a separate concern from whether it's bad for alignment (if the resulting cyborg is still aligned).

Replies from: capybaralet
comment by David Scott Krueger (formerly: capybaralet) (capybaralet) · 2023-02-13T17:47:48.806Z · LW(p) · GW(p)

Oh I see.  I was getting at the "it's not aligned" bit.

Basically, it seems like if I become a cyborg without understanding what I'm doing, the result is either:

  • I'm in control
  • The machine part is in control
  • Something in the middle

Only the first one seems likely to be sufficiently aligned. 

Replies from: alex-rozenshteyn
comment by rpglover64 (alex-rozenshteyn) · 2023-02-14T13:58:18.452Z · LW(p) · GW(p)

I think "sufficiently" is doing a lot of work here. For example, are we talking about >99% chance that it kills <1% of humanity, or >50% chance that it kills <50% of humanity?

I also don't think "something in the middle" is the right characterization; I think "something else" it more accurate. I think that the failure you're pointing at will look less like a power struggle or akrasia and more like an emergent goal structure that wasn't really present in either part.

I also think that "cyborg alignment" is in many ways a much more tractable problem than "AI alignment" (and in some ways even less tractable, because of pesky human psychology):

  • It's a much more gradual problem; a misaligned cyborg (with no agentic AI components) is not directly capable of FOOM (Amdhal's law was mentioned elsewhere in the comments as a limit on usefulness of cyborgism, but it's also a limit on damage)
  • It has been studied longer and has existed longer; all technologies have influenced human thought

It also may be an important paradigm to study (even if we don't actively create tools for it) because it's already happening.

comment by MSRayne · 2023-02-23T16:00:51.106Z · LW(p) · GW(p)

Historically I have felt most completely myself when I was intertwining my thoughts with those of an AI. And the most I've ever had access to is AI Dungeon, not GPT-3 itself. I feel more myself with it, not less - as if it's opening up parts of my own mind I didn't know were there before. But that's me.

comment by Flipnash · 2023-02-11T02:18:16.892Z · LW(p) · GW(p)

At the risk of sounding insane. I remember doing similar things but used git to keep track of branches. A warning that I wished I had back then before I shelved it.

There's a phenomenon where your thoughts and generated text have no barrier. It's hard to describe but it's similar to how you don't feel the controller and the game character is an extension of the self.

It leaves you vulnerable to being hurt by things generated characters say because you're thoroughly immersed.

They will say anything with non-zero probability.

It's easy to lose sleep when playing video games. Especially when you feel the weight of the world on your shoulders.

Sleep deprivation+LLM induced hallucinations aren't fun. Make sure to get sleep.

Beware, LLM's will continue negative thinking. You can counter with steering it to positive thinking and solutions. Obviously, not all negative thoughts are within your current ability to solve or counter, like the heat death of the universe. Don't get stuck down a negative thoughts branch and despair.

Replies from: janus, fergus-fettes
comment by janus · 2023-02-11T03:15:12.800Z · LW(p) · GW(p)

There's a phenomenon where your thoughts and generated text have no barrier. It's hard to describe but it's similar to how you don't feel the controller and the game character is an extension of the self.

Yes. I have experienced this. And designed interfaces intentionally to facilitate it (a good interface should be "invisible"). 

It leaves you vulnerable to being hurt by things generated characters say because you're thoroughly immersed.

Using a "multiverse" interface where I see multiple completions at once has incidentally helped me not be emotionally affected by the things GPT says in the same way as I would if a person said it (or if I had the thought myself). It breaks a certain layer of the immersion. As I wrote in this comment [LW(p) · GW(p)]:

Seeing the multiverse destroys the illusion of a preexisting ground truth in the simulation. It doesn't necessarily prevent you from becoming enamored with the thing, but makes it much harder for your limbic system to be hacked by human-shaped stimuli.

It reveals the extent to which any particular thing that is said is arbitrary, just one among an inexhaustible array of possible worlds.

That said, I still find myself affected by things that feel true to me, for better or for worse.



It's easy to lose sleep when playing video games. Especially when you feel the weight of the world on your shoulders.

Ha, yeah.

Playing the GPT virtual reality game will only become more enthralling and troubling as these models get stronger. Especially, as you said, if you're doing so with the weight of the world on your shoulders. It'll increasingly feel like walking into the mouth of the eschaton, and that reality will be impossible to ignore. That's the dark side of the epistemic calibration I wrote about at the end of the post.

Thanks for the comment, I resonated with it a lot, and agree with the warning. Maybe I'll write something about the psychological risk and emotional burden that comes with becoming a cyborg for alignment, because obviously merging your mind with an earthborn alien (super)intelligence which may be shaping up to destroy the world, in order to try save the world, is going to be stressful.

comment by Fergus Fettes (fergus-fettes) · 2023-02-17T10:46:55.791Z · LW(p) · GW(p)

Hey, using git branches immediately jumped out to me as a great way of doing this, is there any code you could share or could you describe some technical pros and cons that you found with the approach, in case I try it out :).

Very interesting and thoughtful post as well, reminds me of a bad trip.

comment by RobertM (T3t) · 2023-02-16T01:55:50.122Z · LW(p) · GW(p)

Curated.  I've recently felt like there has been a lot of confusion about "use AIs to help solve alignment" as a "strategy", because in fact it is multiple strategies that look very different from one another.

The strategy described here does not "hide" the hard part of solving alignment, the way "build an AI that solves alignment for you" does.  Existing tools, such as calculators, enable human-driven cognitive effort which would intractable (or merely much more difficult) without them.  Similarly, it seems possible that LLMs will be able to substitute for or extend human cognition in ways that let us make intellectual progress faster, and maybe even differentially advantage alignment research.

I think this post does a great job of describing what I consider to be the "less doomed" strategy, without eliding relevant warnings and caveats.

comment by Logan Riggs (elriggs) · 2023-02-10T16:48:15.115Z · LW(p) · GW(p)

I’m excited about sensory substitution (, where people translate auditory or visual information into tactile sensations (usually for people who don’t usually process that info).

I remember Quintin Pope wanting to translate the latent space of language models [reading a paper] translated to visual or tactile info. I’d see this as both a way to read papers faster, brainstorm ideas, etc and gain a better understanding of latent space during development of this.

Replies from: MSRayne
comment by MSRayne · 2023-02-23T16:03:22.407Z · LW(p) · GW(p)

This sounds fantastic and I want it.

comment by Noosphere89 (sharmake-farah) · 2023-02-10T15:46:07.478Z · LW(p) · GW(p)

I disagree with this post for 1 reason:

  1. Amdahl's law limits how much cyborgism will actually work, and IMO is the reason agents are more effective than simulators.

On Amdahl's law, John Wentworth's post on the long tail is very relevant here, as it limits the use of cyborgism here: [LW · GW]

Replies from: elriggs, elriggs
comment by Logan Riggs (elriggs) · 2023-02-10T16:37:01.203Z · LW(p) · GW(p)

For context, Amdahl’s law states how fast you can speed up a process is bottlenecked on the serial parts. Eg you can have 100 people help make a cake really quickly, but it still takes ~30 to bake.

I’m assuming here, the human component is the serial component that we will be bottlenecked on, so will be outcompeted by agents?

If so, we should try to build the tools and knowledge to keep humans in the loop as far as we can. I agree it will eventually be outcompeted by full AI agency alone, but it isn’t set in stone how far human-steered AI can go.

Replies from: sharmake-farah
comment by Noosphere89 (sharmake-farah) · 2023-02-10T17:04:40.290Z · LW(p) · GW(p)

I’m assuming here, the human component is the serial component that we will be bottlenecked on, so will be outcompeted by agents?

Basically yes. My point here is that yes we are in the approximately worst case here, and his option is probably going to not be to accelerate things very much, compared to agentic architecture.

I think a crux here is the long tail bites hard here, and thus I don't think his approach provides much progress compared to the traditional approach of alignment research.

If so, we should try to build the tools and knowledge to keep humans in the loop as far as we can. I agree it will eventually be outcompeted by full AI agency alone, but it isn’t set in stone how far human-steered AI can go.

My guess is that it will only speed things up by very little: I'd be very surprised if it even improves research rates by 50%.

comment by Logan Riggs (elriggs) · 2023-02-10T16:38:55.845Z · LW(p) · GW(p)

I’m unsure how alt-history and point (2) history is hard to change and predictable relates to cyborgism. Could you elaborate?

Replies from: sharmake-farah
comment by Noosphere89 (sharmake-farah) · 2023-02-10T17:00:53.101Z · LW(p) · GW(p)

I might want to remove that point for now.

comment by Logan Zoellner (logan-zoellner) · 2023-02-10T16:22:59.763Z · LW(p) · GW(p)

Isn't one of the main objections to cyborgism that humans with unlimited power tend to go kind of crazy?

It seems like the main challenge would be figuring out what kind of guardrails we want to put on our neo-cortex to make sure that humans didn't do anything obviously insane like build a superweapon and destroy earth.  But this essentially re-capitulates the alignment problem as the "align AI but also make it bio-compatible" which seems strictly harder.

comment by lberglund (brglnd) · 2023-02-15T06:41:00.733Z · LW(p) · GW(p)

Reading this inspired me to try Loom. I found it fun, but also frustrating at times. It could be useful to have a video demonstration of someone like Janus using loom, to get new users up to speed.

Replies from: janus
comment by janus · 2023-02-15T08:14:56.089Z · LW(p) · GW(p)

Here is a video (and playlist).
The open source python version of Loom, which I assume you're using, is old and difficult to use. The newer versions are proprietary or not publicly accessible. If you're interested in using them DM me.

Replies from: brglnd
comment by lberglund (brglnd) · 2023-02-15T18:33:41.394Z · LW(p) · GW(p)

This is great! Thanks for sharing.

comment by trevor (TrevorWiesinger) · 2023-02-10T17:31:14.077Z · LW(p) · GW(p)

talking to a LLM for several hours at a time seems dangerous [LW · GW]

(Just to clarify, this comment was edited and the vast majority of the content is removed, feel free to DM me if you are interested in the risks of talking to an LLM operated by a large corporation)

Replies from: elriggs, janus, janus, lahwran
comment by Logan Riggs (elriggs) · 2023-02-10T18:22:56.787Z · LW(p) · GW(p)

These arguments don't apply to the base models which are only trained on next word prediction (ie the simulators post [LW · GW]), since their predictions never affected future inputs. This is the type of model Janus most interacted with.

Two of the proposals in this post do involve optimizing over human feedback, like:

Creating custom models trained on not only general alignment datasets but personal data (including interaction data), and building tools and modifying workflows to facilitate better data collection with less overhead

, which they may apply to. 

comment by janus · 2023-02-10T22:16:04.851Z · LW(p) · GW(p)

The side effects of prolonged LLM exposure might be extremely severe.

I guess I should clarify that even though I joke about this sometimes, I did not become insane due to prolonged exposure to LLMs. I was already like this before.

comment by janus · 2023-02-11T10:09:08.687Z · LW(p) · GW(p)

Now that you've edited your comment:

The post you linked is talking about a pretty different threat model than what you described before. I commented on that post:

I've interacted with LLMs for hundreds of hours, at least. A thought that occurred to me at this part -

> Quite naturally, the more you chat with the LLM character, the more you get emotionally attached to it, similar to how it works in relationships with humans. Since the UI perfectly resembles an online chat interface with an actual person, the brain can hardly distinguish between the two.

Interacting through non-chat interfaces destroys this illusion, when you can just break down the separation between you and the AI at will, and weave your thoughts into its text stream. Seeing the multiverse destroys the illusion of a preexisting ground truth in the simulation. It doesn't necessarily prevent you from becoming enamored with the thing, but makes it much harder for your limbic system to be hacked by human-shaped stimuli.

But yeah, I've interacted with LLMs for much longer that the author and I don't think I suffered negative psychological consequences from it (my other response was only half-facetious; I'm aware I might give off schizo vibes, but I've always been like this).

As I said in this other comment [LW(p) · GW(p)], I agree that cyborgism is psychologically fraught. But the neocortex-prosthesis setup is pretty different from interacting with a stable and opaque simulacrum through a chat interface, and less prone to causing emotional attachment to an anthropomorphic entity. The main psychological danger I see from cyborgism is somewhat different from this, more like Flipnash described:

It's easy to lose sleep when playing video games. Especially when you feel the weight of the world on your shoulders.

I think people should only become cyborgs if they're psychologically healthy/resilient and understand that it involves gazing into the abyss.

comment by the gears to ascension (lahwran) · 2023-02-11T08:38:32.373Z · LW(p) · GW(p)

I do think there's real risk there even with base models, but it's important to be clear where it's coming from - simulators can be addictive when trying to escape the real world. Your agency needs to somehow aim away from the simulator, and use the simulator as an instrumental tool.

Replies from: janus
comment by janus · 2023-02-11T10:20:41.741Z · LW(p) · GW(p)

I think you just have to select for / rely on people who care more about solving alignment than escapism, or at least that are able to aim at alignment in conjunction with having fun. I think fun can be instrumental. As I wrote in my testimony, I often explored the frontier of my thinking in the context of stories.

My intuition is that most people who go into cyborgism with the intent of making progress on alignment will not make themselves useless by wireheading, in part because the experience is not only fun, it's very disturbing, and reminds you constantly why solving alignment is a real and pressing concern.

comment by metasemi (Metasemi) · 2023-02-13T16:46:17.427Z · LW(p) · GW(p)

As this post sits with me, one thing that seems to call for a much closer look is this idea that the human remains in control of the cyborg.

The post states, for instance, that "The human is 'in control' not just in the sense of being the most powerful entity in the system, but rather because the human is the only one steering", but at other points acknowledges what I would consider caveats. Several comment threads here, eg those initiated by Flipnash and by David Scott Krueger, raise questions, and I'd venture to say some of the replies, including some by janus themself, shatter at least the strongest version of it.

This is obviously a crucial point - it's at the heart of the claim that cyborgism can differentially accelerate alignment relative to capabilities.

Me: "The human is doing the steering" captures an important truth. It's one of the two[1] main reasons I'm excited about cyborgism.
Also me: "The human is doing the steering", stated unconditionally, is false.

In the wonderful graph labeled "Cognition is a Journey Through a Mental Landscape" (which Tufte would be proud of, seriously), we need to recognize that steering is going on at, and indeed inside, those blue circles too. Consider the collaborative behavior of the simulator and the human in constructing the cyborg's joint trajectory. In what ways are their roles symmetrical, and in what ways are they not? How will this change as simulator SOTA advances? In what ways are human values already expressed in the simulator's actions, and what do we make of the cases where they seem not to be? What do we make of the cases where simulacra manifestly do pursue goals seemingly agentically? If there are caveats to human control, how serious are they, how serious do we see them becoming, and what can we do about them?

To be clear, I firmly agree with the authors' hunch that, for at least this decade or more, cyborgism can be a vehicle not just for retaining human agency, but for amplifying it, with benefits to alignment and in other ways too. I'm moved by considerations of the simulators' myopia/divergence, the tabula rasa nature of their outer objectives, the experiences of people like janus who have gone deep with GPT, and also by the knowledge that human values are deeply embedded in what simulators learn.

But this needs to be more than a hunch; we need to probe it deeply (and indeed, the authors acknowledge this at several points, specifically including under 'More ideas'). If it's false, we need to find out now. If it's true, we need the depth of understanding to turn belief that the simulator can amplify human agency into a reality that it does. In the process, we may come to a deeper understanding of this huge swath of the human semantic world the simulator has embodied, and thereby of ourselves.


  1. ^

    The other being the way cyborgism amplifies human agency via the simulator's strengths rather than continually running afoul of its weaknesses as other usage modes do.

comment by skybrian · 2023-02-11T23:49:23.964Z · LW(p) · GW(p)

A caption from the article: "(screenshot of the tool Bonsai, a version of Loom hosted by Conjecture)"

What is "Conjecture?" Where can I find this "Bonsai" tool? I tried a quick search but didn't find much.

Replies from: lucie-philippon
comment by Lucie Philippon (lucie-philippon) · 2023-02-13T20:09:07.547Z · LW(p) · GW(p)

Conjecture is "a team of researchers dedicated to applied, scalable AI alignment research." according to their website

They are publishing regularly on the alignment forum and LessWrong [? · GW]

I also searched their website, and it does not look like Bonsai is publicly accessible. This must be some internal tool they developed ?

comment by Review Bot · 2024-03-31T12:32:52.394Z · LW(p) · GW(p)

The LessWrong Review [? · GW] runs every year to select the posts that have most stood the test of time. This post is not yet eligible for review, but will be at the end of 2024. The top fifty or so posts are featured prominently on the site throughout the year.

Hopefully, the review is better than karma at judging enduring value. If we have accurate prediction markets on the review results, maybe we can have better incentives on LessWrong today. Will this post make the top fifty?

comment by Bill Benzon (bill-benzon) · 2024-03-19T22:04:05.873Z · LW(p) · GW(p)

First I should say that I have little interest in the Frankenstein approach to AI, that is, AI as autonomous agents. I'm much more attracted to AI as intelligence augmentation (as advocated by Stanford's Michael Jordan). For the most part I've been treating ChatGPT as an object of research and so my interactions have been motivated by having it do things that give me clues about how it works, perhaps distant clues, but clues nonetheless. But I do other things with it, and on a few occasions I've gotten into a zone where some very interesting interactive story-telling comes about. ChatGPT's own story-telling abilities are rather pedestrian. I'm somewhat better, but the two of us, what fun we've had on occasion. Not sure how to reach that zone reliably, but I'm working on it.

comment by iamwill · 2023-05-03T14:27:51.210Z · LW(p) · GW(p)

Firstly, love the article. The Cognition is a Journey through a Mental Landscape Figure is great, but for the cyborg illustration it would be better if at each node there were many AI branches and the human branch was shown to select one of these paths to follow. 

Also, this might be outside the scope of a comment on this article, but this is my first comment so I'll take the risk. The language below reminds me of Fantasia - Sorcerer's Apprentice. If we outsource the actions that fulfil our wishes with a magic hat, then the wishes probably wouldn't be granted in the way we think we intend. 

"An antithetical example to this is something like a genie, where the human outsources all of their agency to an external system that is then empowered to go off and optimize the world."

Thought this was a nice analogy and also is a simple way of explaining ai-risk to non-technical folk :) 

comment by NicholasKross · 2023-04-24T01:15:17.110Z · LW(p) · GW(p)

If this post made you less worried about the dangers of automating alignment research then I’ve failed miserably.

Friendly reminder for everyone, including myself.

comment by Stephen McAleese (stephen-mcaleese) · 2023-02-19T19:21:12.053Z · LW(p) · GW(p)

I think this is a really interesting post and seems like a promising and tractable way to accelerate alignment research. It reminds me of Neuralink but seems more feasible at present. I also like how the post emphasizes differentially accelerating alignment because I think one of the primary risks of any kind of augmentation is that it just globally accelerates progress and has no net positive impact.

One sentence I noticed that seemed like a misdefinition was how the concept of a genie was defined:

An antithetical example to this is something like a genie, where the human outsources all of their agency to an external system that is then empowered to go off and optimize the world.

To me, this sounds more like a 'sovereign' as defined in Superintelligence whereas a genie just executes a command before waiting for the next command. Though the difference doesn't seem that big since both types of systems take action.

A key concept I thought was missing was Amdahl's Law which is a formula that calculates the maximum theoretical speedup of a computation given the percentage of the computation that can be parallelized.  The formula is . I think it's also relevant here: if 50% of work can be delegated to a model, the maximum speedup is a factor of 2 because then there will only be half as much work for the human to do. If 90% can be delegated, the maximum speedup is 10.

Also, maybe it would be valuable to have more thinking focused on the human component of the system: ideas about productivity, cognitive enhancement, or alignment. Though I think these ideas are beyond the scope of the post.

comment by rpglover64 (alex-rozenshteyn) · 2023-02-11T21:40:53.282Z · LW(p) · GW(p)

OpenAI’s focus with doing these kinds of augmentations is very much “fixing bugs” with how GPT behaves: Keep GPT on task, prevent GPT from making obvious mistakes, and stop GPT from producing controversial or objectionable content. Notice that these are all things that GPT is very poorly suited for, but humans find quite easy (when they want to). OpenAI is forced to do these things, because as a public facing company they have to avoid disastrous headlines like, for example: Racist AI writes manifesto denying holocaust.[7]

As alignment researchers, we don’t need to worry about any of that! The goal is to solve alignment, and as such we don’t have to be constrained like this in how we use language models. We don’t need to try to “align” language models by adding some RLHF, we need to use language models to enable us to actually solve alignment at its core, and as such we are free to explore a much wider space of possible strategies for using GPT to speed up our research.[8] 


I think this section is wrong because it's looking through the wrong frame. It's true that cyborgs won't have to care about PR disasters in their augments, but it feels like this section is missing the fact that there are actual issues behind the PR problem. Consider a model that will produce the statement "Men are doctors; women are nurses" (when not explicitly playing an appropriate role); this indicates something wrong with its epistemics, and I'd have some concern about using such a model.

Put differently, a lot of the research on fixing algorithmic bias can be viewed as trying to repair known epistemic issues in the model (usually which are inherited from the training data); this seems actively desirable for the use case described here.

comment by zeroxoneb · 2023-02-11T16:27:43.769Z · LW(p) · GW(p)