I wanted to add several additional elements to the base CAIS model that recent advances and my own experience make obvious:
A. Stateless, limited time sessions. Break tasks into the smallest granular subtask, and erase input buffers after each subtask completes, starting over. This is fundamentally how the stateless microservices of all the large tech giants allow any kind of reliability for hyperscale online services.
B. Check if the input this frame is in the training distribution. One way to implement this is to train an autoencoder on the training set, and autoencode each input frame to the model. So (encoded frame, residual) = autoencoder (input frame). "frame"** is a discrete set of inputs, it's the sum of all inputs to the model.
If the input is out of distribution, measured by a large residual, do not send the input to the model. Transfer control to a trustworthy embedded system. Autonomous vehicles use a strategy like this: on main system failure/low confidence, control authority is given to a system that uses a single camera and works more like older lane following systems. If that system fails from hardware faults, the base safety microcontroller takes away control authority from the upper systems and engages the brakes.
C. The outer framework must have control authority. The model isn't in charge, the outer wrapper code (written and carefully validated by humans) ultimately is.
D. Inputs and outputs to the model must be sparse. No extra bits. For an example of this, if you ask GPT-4 to write a piece of code, then move the output text to a different session, the new session of GPT-4 has no context information but the text, which is human readable. This is fairly sparse, it's difficult for one session of the model to embed "hey fellow AI, keep this bug in the code hidden from the meatsacks" such that the next session will see the secret message. Even more difficult, close to impossible, if you have the model doing the checking based on a different base model that shares no common context for a secret language.
Perfectly sparse would be an encoded schema that has nothing but the task instructions, written in a formal schema, and only the information needed to proceed from prior frames.
E. Different base models must be sufficiently intercompatible you can use them to check each other's work. For example, there is no reason you couldn't alternate control every frame for a robot using 2 completely different base models, so long as both are rational control systems and there is 1 and only 1 optimal policy that both models can calculate for each frame.
F. Finally, by using the sparse I/O in D, it must be impossible for models to know if they are operating in the real world and not in the training simulator. If they get no context but the frame and the prompt, and the frame is in distribution to the training simulator, they cannot know they are not in the sim, which obviously can detect most forms of simulated wrongdoing.
G. Train only in the simulator. The latest generative models are capable of taking real world data and producing many permutations on it that are realistic training examples. So when the system encounters an unexpected outcome in the real world or a novel situation, you save to an onboard log the frame*, expected outcome, ground truth, etc. And use a generative model to augment the unexpected situation so that in the simulator the model can practice on many permutations of the situation.
Since the model's performance in the simulator was measured to be sufficiently safe for the task it is performing, and the simulator was measured to predict the real world sufficiently accurately (you run an instance of the simulator in parallel in the real world on each robot), and you will not give the model control authority when the real world is out of distribution, you can bound your safety probabilities and prove the model is sufficiently* safe for the task it is assigned.
**the "frame" is all inputs, for example it might be a lidar image, 6 cameras, all the currently tracked objects that have been seen on prior frames and their estimated identities and vectors, intermediate internal data that caches some activations from the prior frame, and of course the prompt. The model(frame, weights, RNG seed) should deterministically produce the same output. (current sw stacks are not quite deterministic but could be at a small cost of performance)
*depends on the scope of the task. No system will be perfect, the idea is to stack probabilities to reach your design error rate. A robot performing heart surgery will need much more reliability than one stacking boxes. The heart surgery machine may need electromechanical limits because you may have to leave the model running even when it leaves distribution, since stopping an operation is more likely to be fatal than continuing.
Electromechanical limits are things like embedding the hardware for the machine on site, hardwired for one purpose, and not mounting the robotic arms on mobile chassis. This limits failures to killing the patient instead of the machine able to say escape into the hospital and 'perform surgery' on random people.
Eric Drexler's report Reframing Superintelligence: Comprehensive AI Services (CAIS) as General Intelligence reshaped how a lot of people think about AI (summary 1, summary 2). I still agree with many parts of it, perhaps even the core elements of the model. However, after looking back on it more than four years later, I think the general picture it gave missed some crucial details about how AI will go.
The problem seems to be that his report neglected a discussion of foundation models, which I think have transformed how we should think about AI services and specialization.
The general vibe I got from CAIS (which may not have been Drexler's intention) was something like the following picture:
For each task in the economy, we will train a model from scratch to automate the task, using the minimum compute necessary to train an AI to do well on the task. Over time, the fraction of tasks automated will slowly expand like a wave, starting with the tasks that are cheapest to automate computationally, and ending with the most expensive tasks. At some point, automation will be so widespread that it will begin to meaningfully feed into itself, increasing AI R&D, and accelerating the rate of technological progress.
The problem with this approach to automation is that it's extremely wasteful to train models from scratch for each task. It might make sense when training budgets are tiny — as they mostly were in 2018 — but it doesn't make sense when it takes 10^25 FLOP to reach adequate performance on a given set of tasks.
The big obvious-in-hindsight idea that we've gotten over the last several years is that, instead of training from scratch for each new task, we'll train train a foundation model on some general distribution, which can then be fine-tuned using small amounts of compute to perform well on any task. In the CAIS model, "general intelligence" is just the name we can give to the collection of all AI services in the economy. In this new paradigm, "general intelligence" refers to the fact that sufficiently large foundation models can efficiently transfer their knowledge to obtain high performance on almost any downstream task, which is pretty closely analogous to what humans do to take over jobs.
The fact that generalist models can be efficiently adapted to perform well on almost any task is an incredibly important fact about our world, because it implies that a very large fraction of the costs of automation can be parallelized across almost all tasks.
Let me illustrate this fact with a hypothetical example.
Suppose we previously thought that it would take $1 trillion to automate each task in our economy, such as language translation, box stacking, and driving cars. Since the cost of automating each of these tasks is $1 trillion each, you might expect companies would slowly automate all the tasks in the economy, starting with the most profitable ones, and then finally getting around to the least profitable ones once economic growth allowed for us to spend enough money on automating not-very-profitable stuff.
But now suppose we think it costs $999 billion to create "general intelligence", which then once built, can be quickly adapted to automate any other task at a cost of $1 billion. In this world, we will go very quickly from being able to automate almost nothing to being able to automate almost anything. In other words we will get one big innovation "lump", which is the opposite of what Robin Hanson predicted. Even if we won't invent monolithic agents that take over the world by being smarter than everything else, we won't have a gradual decades-long ramp-up to full automation either.
Of course, the degree of suddenness in the foundation model paradigm is still debatable, because the idea of "general intelligence" is itself continuous. GPT-4 is more general than GPT-3, which was more general than GPT-2, and presumably this trend will smoothly continue indefinitely as a function of scale, rather than shooting up discontinuously after some critical threshold. But results in the last year or so have updated me towards thinking that the range from "barely general enough to automate a few valuable tasks" to "general enough to automate almost everything humans do" is only 5-10 OOMs of training compute. If this range turns out to be 5 OOMs, then I expect a fast AI takeoff under Paul Christiano's definition, even though I still don't think this picture looks much like the canonical version of foom.
Foundation models also change the game because they imply that AI development must be highly concentrated at the firm-level. AIs themselves might be specialized to provide various services, but the AI economy depends critically on a few non-specialized firms that deliver the best foundation models at any given time. There can only be a few firms in the market providing foundation models because the fixed capital costs required to train a SOTA foundation model are very high, and being even 2 OOMs behind the lead actor results in effectively zero market share. Although these details are consistent with CAIS, it's a major update about what the future AI ecosystem will look like.
A reasonable remaining question is why we'd ever expect AIs to be specialized in the foundation model paradigm. I think the reason is that generalist models are more costly to run at inference time compared to specialized ones. After fine-tuning, you will want to compress the model as much as possible, while maintaining acceptable performance on whatever task you're automating.
The degree of specialization will vary according to the task you want to automate. Some tasks require very general abilities to do well. For example, being a CEO plausibly benefits from being extremely general, way beyond even human-level, such that it wouldn't make sense to make them less general even if it saved inference costs. On the other hand, language translation is plausibly something that can be accomplished acceptably using far less compute than a CEO model. In that case, you want inference costs to be much lower.
It now seems clear that AIs will also descend more directly from a common ancestor than you might have naively expected in the CAIS model, since most important AIs will be a modified version of one of only a few base foundation models. That has important safety implications, since problems in the base model might carry over to problems in the downstream models, which will be spread throughout the economy. That said, the fact that foundation model development will be highly centralized, and thus controllable, is perhaps a safety bonus that loosely cancels out this consideration.
Drexler can be forgiven for not talking about foundation models in his report. His report was published at the start of 2019, just months after the idea of "fine-tuning" was popularized in the context of language models, and two months before GPT-2 came out. And many readers can no doubt point out many non-trivial predictions that Drexler got right, such as the idea that we will have millions of AIs, rather than just one huge system that acts as a unified entity. And we're still using deep learning as Drexler foresaw, rather than building general intelligence like a programmer would. Like I said at the beginning, it's not necessarily that the core elements of the CAIS model are wrong; the model just needs an update.