Summary:

• In the current paradigm, training is much more expensive than inference. So whenever we finish end-to-end training a language model, we can run a lot of them in parallel.
  • If a language model was trained with Chinchilla scaling laws on the FLOP-equivalent of a large fraction of the world’s current GPU and TPUs: I estimate that the training budget could produce at least ~20 million tokens per second.
  • Larger models trained on more data would support more tokens per second.
• Language models can also run faster than humans. Current models generate 10-100 tokens per second. It’s unclear whether future models will be slower or faster.
• This suggests that, before AI changes the world via being broadly superior to human experts, it will change the world via providing a lot of either mediocre (by the standard of human experts) or specialized thinking.
• This might make the early alignment problem easier. But the full alignment problem will come soon thereafter, in calendar-time, so this mainly matters if we can use the weaker AI to buy time or make progress on alignment.

More expensive AI → you can run more AIs with your training budget

The more expensive it is to train AI an AI, the more copies of that AI system can be run in parallel using your training budget. At least, that's the case if we’re making them more expensive by increasing parameter-count and training data.

We’re currently in a paradigm where:

• Training isn’t very sample-efficient.
• When increasing capabilities, training costs increase faster (~squared) than inference costs.
• Training is massively parallelizable.^[1] While this paradigm holds, it implies that the most capable models will be trained using massively parallelized training schemes, equivalent to running a large number of models in parallel. The larger the model, the more data it needs, and so more copies of them will have to be run in parallel during training, in order to finish within a reasonable time-frame.^[2]

This means that, once you have trained a highly capable model, you are guaranteed to have the resources to run a huge number of them in parallel. And the bigger and more expensive the model was — the more of them can run in parallel on your training cluster.

Here’s a rough calculation of how many language models you can run in parallel using just your training cluster:

• Let’s say you use p parameters.
• Running the model for one token takes kp FLOP, for some k.
• Chinchilla scaling laws say training data is proportional to parameters, implying that the model is trained for mp tokens.
  • For Chinchilla, m=20 tokens / parameter.
• Total training costs are 3kmp^2.
  • The 3 is there because backpropagation is ~2x as expensive as forward propagation.
• You spend N seconds training your model.
• During training, you use (3kmp^2/N) FLOP/s, and at inference you can run one model for kp FLOP/s. So using just your training compute, you can run (3kmp^2/N)/(kp) = 3mp/N tokens per second, just by reallocating your training compute to inference.

If you take a horizon-length framework seriously, you might expect that we’ll need more training data to handle longer-horizon tasks. Let’s introduce a parameter H that describes how many token-equivalents correspond to one data-point.

• Total training costs are now 3kmHp^2.
• So with the compute you used to train your models, you can process 3mpH/N token-equivalents per second.

Some example numbers (bolded ones are changed from the top one):

• For p=1e14, N=1y, H=1, m=20, the above equation says you can process 200 million token-equivalents per second, with just your training budget.
• For p=1e15, N=1y, H=1, m=20, it’s ~2 billion token-equivalents/second.
• For p=1e14, N=3 months, H=1 hour, m=20, it’s ~1 trillion token-equivalents/second..

In addition, there are various tricks for lowering inference costs. For example, reducing precision (which is less important during training than inference) and knowledge distillation; see here for more discussion. These would further increase the number of models you can run in parallel.

A rough lower bound for number of AIs the world could run

The bigger the training run, the more AIs you can run with your training cluster. Conversely, if human-level AI comes earlier, with smaller training runs, you’ll be able to run fewer of them with your training cluster.

On the other hand, if a training run is very small, then it’s only using a small fraction of the world’s compute. This means that there’s a lot of room to run many models in parallel just by acquiring more compute. (It would certainly be economically efficient for a large fraction of the world’s compute to run AI systems, if we did have human-level AI — whether that happens via the developers+investors buying more compute, the developers selling their software, a government seizing the software, or some other way.)

Today, there’s about 4e21 FLOP/s out there in the form of GPUs and TPUs (source). Let’s assume that the world would want to run ~human-level AI systems on at least 25% of that (1e21 FLOP/s), given the option. If so, we can get a rough lower bound on how many ~human-level AIs could be run shortly after training by looking at the number of AIs you could run after training an AI on 1e21 FLOP/s, run for a year:

• Let’s say…
  • k = 4 FLOP/parameter/token.
    • This suggests 2 FLOP/parameter.
    • I increase to 4 to account for GPUs only having 50% utilization.
  • m = 20 datapoint/parameter. (Based on Chinchilla.)
  • H = 1 token-equivalent/datapoint.
  • N = 3e7 seconds. (A year.)
• This means that…
  • 3kmHp^2 = 1e21N ⇔ p = sqrt(1e21N/(3kmH)) ~= 1.1e13
• And the number of models you can run in parallel is:
  • 1e21/kp = 1e21/(4*1.6e13) ~= 23 million token-equivalents per second.

Some caveats in the footnote.^[3]

Serial vs parallel

It’s not clear that you can parallelize tasks well enough to make efficient use of 23 million parallel models. To what degree is it possible to run these AIs fast, so that we get them in series after each other?

I don’t understand this very well. Some relevant information:

• Jacob Steinhardt suggests 1400 tokens per second for the Chinchilla model (assuming at least 40% GPU utilization), and that increased depth would make this linearly slower, but that width wouldn’t change it at all.
  • Has an erratum saying “I believe that the overall asymptotics below are correct, but the final numbers could plausibly be off by up to an order of magnitude.“
• I think Pope et al. (2022) is the public state of the art in inference speed, with minimum reported latency for PaLM 540B being 29ms ~= 34 tokens per second.

  • The speed is mainly bottlenecked by bandwidth. I’m unsure if the analysis says that latency would only increase with depth or also somewhat with width.^[4]

  • Palm only has 1.5x as many layers as Chinchilla,^[5] so this is much slower than Steinhardt’s analysis suggests.

  • Anecdotal reports about the GPT API are consistent with these slower speeds. The GPT-4 API typically delivers 20 tokens or less per second. (Though potentially up to 40 sometimes?) Though GPT-3.5 Turbo is much faster.

• How much will depth increase in the future?
  • According to Levine et al. (2021), transformers can be scaled a lot without getting much deeper, e.g. it would be fine to increase parameter-count by a factor of 100x while increasing depth by less than 2x. (I’ve done no due diligence on whether the paper is good, but its results are used by the Chinchilla authors.)
  • Kaplan (2020) says “width/depth should remain fixed” which would imply that depth is proportional to the p^(1/3), because parameters are proportional to the depth*width^2.
    • However, it continues: “But more importantly, we find that the precise architectural hyperparameters are unimportant compared to the overall scale of the language model”, which suggests that people could hold off on scaling depth if they were concerned about latency.
  • So depth will probably increase somewhere between “not at all” and as p(1/3).
• Better hardware will probably lead to lower latency. E.g. the newest generation of NVIDIA hardware has increased bandwidth as well as some other potentially speed-increasing improvements. (E.g. supporting FP-8 computation.)
• The above-mentioned tricks for reducing inference cost could also give you faster inference speeds. In addition to those, there’s also the option of running faster models to predict easy tokens, and then running larger models on multiple tokens at once. And if you’re willing, you can reduce hardware utilization to get further speed-ups in latency. (The Steinhardt post claims that reducing utilization by k gets you a k^2.)

In short: We’re currently at 30-40 tokens per second, which will be reduced by bigger model sizes, increased by future hardware, and increased by better techniques.

This is all for generating tokens. Reading content into the context window doesn’t add latency, since the entire context window can be processed in parallel. (Combining this with parallelism is interesting. An AI could split into 10 copies, investigate 10 different lines of thoughts, and then instantly merge and read all thoughts so-far — and then repeat.)

I feel pretty unsure about how that adds up. But if well-optimized future models (running on future hardware) could operate at, say, ~50 tokens per second, then 23 million tokens per second would correspond to ~500,000 separate streams of 50 tokens/second.

Implications

The above numbers suggest that (as long as sample efficiency doesn’t significantly improve) the world will always have enough compute to produce at least 23 million token-equivalents per second from any model that the world can afford to train (end-to-end, chinchilla-style). Notably, these are many more token-equivalents per second than we currently have human-AI-researcher-seconds per second. (And the AIs would have the further advantage of having much faster serial speeds.)

So once an AI system trained end-to-end can produce similarly much value per token as a human researcher can produce per second, AI research will be more than fully automated. This means that, when AI first contributes more to AI research than humans do, the average research progress produced by 1 token of output will be significantly less than an average human AI researcher produces in a second of thinking.^[6] Instead, the collective’s intelligence will largely come from a combination of things like:

• Individual systems “thinking” for a long time, churning through many more explicit thoughts than a skilled human would need to solve a problem.^[7]

• Splitting up things in more granular subtasks, delegating them to other AI systems.

• Generating huge numbers of possible solutions, and evaluating them all before picking one.

Assuming that much of this happens “behind the scenes”, a human interacting with this system might just perceive it as a single super-smart AI. Nevertheless, I think this means that AI will be more alignable at a fixed level of productivity. (Eventually, we’ll face the full alignment problem — but “more alignable at a fixed level of productivity” helps if we can use that productivity for something useful, such as giving us more time or helping us with alignment research.)

Most obviously, the token-by-token output of a single AI system should be quite easy for humans to supervise and monitor for danger. It will rarely contain any implicit cognitive leaps that a human couldn’t have generated themselves. (C.f. visible thoughts project and translucent thoughts hypothesis.)

But what about collectives of AIs, or AIs thinking for a long period of time? If people get capability-boosts by fine-tuning such systems end-to-end, then the situation looks quite different. Perhaps it will prove beneficial to finetune such systems to communicate with each other using uninterpretable vector embeddings. Or even if they keep using English, they might start using steganography.

There are still a few reasons for why this situation seems safer (at a fixed level of AI capability) than it could have been:

• Perhaps end-to-end SGD won’t have a big advantage over process-based methods, where humans fine-tune networks individually and glue them together in a way where each network’s output remains interpretable. After all, you can’t afford to do a lot of end-to-end training on the large collectives, since they’re so expensive to run.

  • Supervised learning is generally more sample-efficient than RL, which is a good sign.
  • The AI systems themselves might be able to help with designing such collectives in a maximally efficient way.^[8]

• Even if people do end-to-end training, the representations passed between models need not immediately become useless. Perhaps there are ways to fight steganography. Intuitively, it at least seems like interpreting the almost-English should be easier than mechanistic interpretability of the neural networks. (Though that isn’t a high bar.)

• Even if you ignore the internals of the collectives, it seems like process-based feedback might work unusually well in this regime. This one requires a bit more explanation.

  • Above, I gestured at “process-based” as distinct from end-to-end training. But a weaker definition of process-based feedback (as distinct from outcomes-based feedback) is: You only ever train your AI to recommend suggested actions, and when deciding what feedback to give, you never test its suggestions in the real world. Instead, you make a decision by thinking carefully, potentially informed by a long investigation, including AI advice. (On episodes when you’re not providing feedback, you can implement the suggested actions without such detailed oversight.)^[9]

    • Importantly, this outer objective doesn’t incentivize the AI collective to optimize the world in any way (other than via incentivizing solutions that look good to humans, who have human preferences about the world).
    • Ideally, it would get you a myopic/act-based agent. But it doesn’t come with a solution to inner alignment, so it definitely doesn’t guarantee safety.

  • The downside of this strategy is that it isn’t very competitive — e.g. if you’re serious about it, you might have to evaluate AI pull requests without testing the code, which is a serious downside.

  • But it seems like it should be unusually likely to be competitive when fine-tuning collectives of subhuman intelligences:

    • If the AI collective makes a good suggestion, there would typically exist a human-understandable decomposition of why that suggestion was good. (Or else how did the subhuman AIs generate it?)
    • The AI collective only needs fine-tuning data, so it’s not catastrophic if the human feedback is expensive to generate.
    • Most of the collective’s capabilities are already baked into the individual components. The purpose of the fine-tuning is just to make sure that those capabilities are directed in a productive direction. Intuitively, I feel like human feedback shouldn’t be much worse at this than outcomes-based feedback.

A few caveats

A big caveat to this is that AI and humans will have different distributions of capabilities.^[10] If there are some topics on which AI is much, much better than humans, then humans might not understand AI’s reasoning about that when looking at token-by-token output (even before end-to-end training). And outcomes-based feedback might be necessary to elicit AI’s full capabilities on that topic.

Indeed, it seems plausible that the story of AI automation won’t be one where many low-capability AIs combine to be human-ish. Instead, it might be that AI automates one task at a time, and that use cases where AI isn’t at least as good as humans aren’t ever that important (c.f. Tom Davidson’s takeoff speeds model and Richard Ngo’s framework). This would also have implications for the shape of early alignment, and whether early AI systems would help with later alignment — but the analysis might be quite different, and involve thinking in detail about what sort of tasks are likely to be automated in what order. I’d be interested in such analysis.

…

Acknowledgements: Thanks to Tom Davidson and Daniel Kokotajlo for comments. I work at Open Philanthropy but the views here are my own.

Notes

---

1. Non-parallelizable training wouldn’t exactly contradict the conclusions here, but it would change what arguments I’d use for them, and it would make the world into a weirder place. (E.g. extra compute wouldn’t help to make smarter models, beyond a point, and AI progress would instead be mostly driven by software, serial time (!) necessary to train models, and maybe inference-time compute, if that was more parallelizable.) ↩︎

2. According to The longest training run: “Training runs of large Machine Learning systems are likely to last less than 14-15 months. This is because longer runs will be outcompeted by runs that start later and therefore use better hardware and better algorithms. “. ↩︎

3. In practice, many of the world’s GPUs wouldn’t be able to efficiently run large models like this, e.g. because of a lack of memory. 25% of the world’s compute is probably an overestimate. On the other hand, specialized hardware is much more important for training than for inference. So if FLOP-supply keeps being dominated by non-specialized hardware, this pushes for more token-equivalents per second, because there would probably be many GPUs you could run your model on that you couldn’t train them on. ↩︎

4. See page 6 for formula T_comm = (√ BLF / √nchips) × 4E / network bandwidth. B is batch size; L is sequence length; F is the width-dimension of the feed-forward networks. E is the embedding/activation size. That’s per layer, so latency straightforwardly increases with more layers. But if you simultaneously scale the embedding dimension and the width of the feed-forward networks by 2x, I think you increase overall computation by 2^2=4x. That justifies increasing chips by 4x. But that leads to an overall change in T by (√2/√4) * 2 = √2? So maybe scaling width by 2x increases latency by √2? ↩︎

5. Chinchilla has 80 (Hoffmann et al., 2022). PaLM has 118 (Chowdhery et al., 2022). ↩︎

6. This relies on an assumption that you can make up for lack-of-intelligence by numbers or speed. Without that assumption, you could expect that AI research will be dominated by humans until AIs finally “get it”, after which they’ll take over with a huge margin. ↩︎

7. Typical reading is ~300 wpm = 5 words per second. Typical speaking might be ~half that. ↩︎

8. One framing of this is: The reason why the bitter lesson applied so strongly in the last few decades is plausibly that compute increased very quickly compared to researcher labor. If AI systems start contributing to AI research, that will correspond to a massive increase in researcher labor, which might reverse the trend. ↩︎

9. C.f. this comment. ↩︎

10. Though as long as the best pre-training task is to predict human text, they’ll be more similar than you might otherwise have expected. ↩︎