# Ω 71

Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

# Or: Big Timelines Crux Operationalized

What fun things could one build with +12 orders of magnitude of compute? By ‘fun’ I mean ‘powerful.’ This hypothetical is highly relevant to AI timelines, for reasons I’ll explain later.

Summary (Spoilers):

I describe a hypothetical scenario that concretizes the question “what could be built with 2020’s algorithms/ideas/etc. but a trillion times more compute?”  Then I give some answers to that question. Then I ask: How likely is it that some sort of TAI would happen in this scenario? This second question is a useful operationalization of the (IMO) most important, most-commonly-discussed timelines crux:  “Can we get TAI just by throwing more compute at the problem?” I consider this operationalization to be the main contribution of this post; it directly plugs into Ajeya’s timelines model and is quantitatively more cruxy than anything else I know of. The secondary contribution of this post is my set of answers to the first question: They serve as intuition pumps for my answer to the second, which strongly supports my views on timelines.

# The hypothetical

In 2016 the Compute Fairy visits Earth and bestows a blessing: Computers are magically 12 orders of magnitude faster! Over the next five years, what happens? The Deep Learning AI Boom still happens, only much crazier: Instead of making AlphaStar for 10^23 floating point operations, DeepMind makes something for 10^35. Instead of making GPT-3 for 10^23 FLOPs, OpenAI makes something for 10^35. Instead of industry and academia making a cornucopia of things for 10^20 FLOPs or so, they make a cornucopia of things for 10^32 FLOPs or so. When random grad students and hackers spin up neural nets on their laptops, they have a trillion times more compute to work with. [EDIT: Also assume magic +12 OOMs of memory, bandwidth, etc. All the ingredients of compute.]

For context on how big a deal +12 OOMs is, consider the graph below, from ARK. It’s measuring petaflop-days, which are about 10^20 FLOP each. So 10^35 FLOP is 1e+15 on this graph. GPT-3 and AlphaStar are not on this graph, but if they were they would be in the very top-right corner.

# Question One: In this hypothetical, what sorts of things could AI projects build?

I encourage you to stop reading, set a five-minute timer, and think about fun things that could be built in this scenario. I’d love it if you wrote up your answers in the comments!

Below are my answers, listed in rough order of how ‘fun’ they seem to me. I’m not an AI scientist so I expect my answers to overestimate what could be done in some ways, and underestimate in other ways. Imagine that each entry is the best version of itself, since it is built by experts (who have experience with smaller-scale versions) rather than by me.

## OmegaStar:

In our timeline, it cost about 10^23 FLOP to train AlphaStar. (OpenAI Five, which is in some ways more impressive, took less!) Let’s make OmegaStar like AlphaStar only +7 OOMs bigger: the size of a human brain.[1] [EDIT: You may be surprised to learn, as I was, that AlphaStar has about 10% as many parameters as a honeybee has synapses! Playing against it is like playing against a tiny game-playing insect.]

Larger models seem to take less data to reach the same level of performance, so it would probably take at most 10^30 FLOP to reach the same level of Starcraft performance as AlphaStar, and indeed we should expect it to be qualitatively better.[2] So let’s do that, but also train it on lots of other games too.[3] There are 30,000 games in the Steam Library. We train OmegaStar long enough that it has as much time on each game as AlphaStar had on Starcraft. With a brain so big, maybe it’ll start to do some transfer learning, acquiring generalizeable skills that work across many of the games instead of learning a separate policy for each game.

OK, that uses up 10^34 FLOP—a mere 10% of our budget. With the remainder, let’s add some more stuff to its training regime. For example, maybe we also make it read the entire internet and play the “Predict the next word you are about to read!” game. Also the “Predict the covered-up word” and “predict the covered-up piece of an image” and “predict later bits of the video” games.

OK, that probably still wouldn’t be enough to use up our compute budget. A Transformer that was the size of the human brain would only need 10^30 FLOP to get to human level at the the predict-the-next-word game according to Gwern, and while OmegaStar isn’t a transformer, we have 10^34 FLOP available.[4] (What a curious coincidence, that human-level performance is reached right when the AI is human-brain-sized! Not according to Shorty.)

Let’s also hook up OmegaStar to an online chatbot interface, so that billions of people can talk to it and play games with it. We can have it play the game “Maximize user engagement!”

...we probably still haven’t used up our whole budget, but I’m out of ideas for now.

## Amp(GPT-7):

Let’s start by training GPT-7, a transformer with 10^17 parameters and 10^17 data points, on the entire world’s library of video, audio, and text. This is almost 6 OOMs more params and almost 6 OOMs more training time than GPT-3. Note that a mere +4 OOMs of params and training time is predicted to reach near-optimal performance at text prediction and all the tasks thrown at GPT-3 in the original paper; so this GPT-7 would be superhuman at all those things, and also at the analogous video and audio and mixed-modality tasks.[5] Quantitatively, the gap between GPT-7 and GPT-3 is about twice as large as the gap between GPT-3 and GPT-1, (about 25% the loss GPT-3 had, which was about 50% the loss GPT-1 had) so try to imagine a qualitative improvement twice as big also. And that’s not to mention the possible benefits of multimodal data representations.[6]

We aren’t finished! This only uses up 10^34 of our compute. Next, we let the public use prompt programming to make a giant library of GPT-7 functions, like the stuff demoed here and like the stuff being built here, only much better because it’s GPT-7 instead of GPT-3. Some examples:

• Decompose a vague question into concrete subquestions
• Generate a plan to achieve a goal given a context
• Given a list of options, pick the one that seems most plausible / likely to work / likely to be the sort of thing Jesus would say / [insert your own evaluation criteria here]
• Given some text, give a score from 0 to 10 for how accurate / offensive / likely-to-be-written-by-a-dissident / [insert your own evaluation criteria here] the text is.

And of course the library also contains functions like “google search” and “Given webpage, click on X” (remember, GPT-7 is multimodal, it can input and output video, parsing webpages is easy). It also has functions like “Spin off a new version of GPT-7 and fine-tune it on the following data.” Then we fine-tune GPT-7 on the library so that it knows how to use those functions, and even write new ones. (Even GPT-3 can do basic programming, remember. GPT-7 is much better.)

We still aren’t finished! Next, we embed GPT-7 in an amplification scheme — a “chinese-room bureaucracy” of calls to GPT-7. The basic idea is to have functions that break down tasks into sub-tasks, functions that do those sub-tasks, and functions that combine the results of the sub-tasks into a result for the task. For example, a fact-checking function might start by dividing up the text into paragraphs, and then extract factual claims from each paragraph, and then generate google queries designed to fact-check each claim, and then compare the search results with the claim to see whether it is contradicted or confirmed, etc. And an article-writing function might call the fact-checking function as one of the intermediary steps. By combining more and more functions into larger and larger bureaucracies, more and more sophisticated behaviors can be achieved. And by fine-tuning GPT-7 on examples of this sort of thing, we can get it to understand how it works, so that we can write GPT-7 functions in which GPT-7 chooses which other functions to call. Heck, we could even have GPT-7 try writing its own functions! [7]

The ultimate chinese-room bureaucracy would be an agent in its own right, running a continual OODA loop of taking in new data, distilling it into notes-to-future-self and new-data-to-fine-tune-on, making plans and sub-plans, and executing them. Perhaps it has a text file describing its goal/values that it passes along as a note-to-self — a “bureaucracy mission statement.”

Are we done yet? No! Since it “only” has 10^17 parameters, and uses about six FLOP per parameter per token, we have almost 18 orders of magnitude of compute left to work with.[8] So let’s give our GPT-7 uber-bureaucracy an internet connection and run it for 100,000,000 function-calls (if we think of each call as a subjective second, that’s about 3 subjective years). Actually, let’s generate 50,000 different uber-bureaucracies and run them all for that long. And then let’s evaluate their performance and reproduce the ones that did best, and repeat. We could do 50,000 generations of this sort of artificial evolution, for a total of about 10^35 FLOP.[9]

Note that we could do all this amplification-and-evolution stuff with OmegaStar in place of GPT-7.

## Crystal Nights:

(The name comes from an excellent short story.)

Maybe we think we are missing something fundamental, some unknown unknown, some special sauce that is necessary for true intelligence that humans have and our current artificial neural net designs won’t have even if scaled up +12 OOMs. OK, so let’s search for it. We set out to recapitulate evolution.

We make a planet-sized virtual world with detailed and realistic physics and graphics. OK, not perfectly realistic, but much better than any video game currently on the market! Then, we seed it with a bunch of primitive life-forms, with a massive variety of initial mental and physical architectures. Perhaps they have a sort of virtual genome, a library of code used to construct their bodies and minds, with modular pieces that get exchanged via sexual reproduction (for those who are into that sort of thing). Then we let it run, for a billion in-game years if necessary!

Alas, Ajeya estimates it would take about 10^41 FLOP to do this, whereas we only have 10^35.[10] So we probably need to be a million times more compute-efficient than evolution. But maybe that’s doable. Evolution is pretty dumb, after all.

1. Instead of starting from scratch, we can can start off with “advanced” creatures, e.g. sexually-reproducing large-brained land creatures. It’s unclear how much this would save but plausibly could be at least one or two orders of magnitude, since Ajeya’s estimate assumes the average creature has a brain about the size of a nematode worm’s brain.[11]
2. We can grant “magic traits” to the species that encourage intelligence and culture; for example, perhaps they can respawn a number of times after dying, or transfer bits of their trained-neural-net brains to their offspring. At the very least, we should make it metabolically cheap to have big brains; no birth-canal or skull should restrict the number of neurons a species can have! Also maybe it should be easy for species to have neurons that don’t get cancer or break randomly.
3. We can force things that are bad for the individual but good for the species, e.g. identify that the antler size arms race is silly and nip it in the bud before it gets going. In general, more experimentation/higher mutation rate is probably better for the species than for the individual, and so we could speed up evolution by increasing the mutation rate. We can also identify when a species is trapped in a local optima and take action to get the ball rolling again, whereas evolution would just wait until some climactic event or something shakes things up.
4. We can optimise for intelligence instead of ability to reproduce, by crafting environments in which intelligence is much more useful than it was at any time in Earth’s history. (For example, the environment can be littered with monoliths that dispense food upon completion of various reasoning puzzles. Perhaps some of these monoliths can teach English too, that’ll probably come in handy later!) Think about how much faster dog breeding is compared to wolves evolving in the wild. Breeding for intelligence should be correspondingly faster than waiting for it to evolve.
5. There are probably additional things I haven’t thought of that would totally be thought of, if we had a team of experts building this evolutionary simulation with 2020’s knowledge. I’m a philosopher, not an evolutionary biologist!

## Skunkworks:

What about STEM AI? Let’s do some STEM. You may have seen this now-classic image:

These antennas were designed by an evolutionary search algorithm. Generate a design, simulate it to evaluate predicted performance, tweak & repeat. They flew on a NASA spacecraft fifteen years ago, and were massively more efficient and high-performing than the contractor-designed antennas they replaced. Took less human effort to make, too.[12]

This sort of thing gets a lot more powerful with +12 OOMs. Engineers often use simulations to test designs more cheaply than by building an actual prototype. SpaceX, for example, did this for their Raptor rocket engine. Now imagine that their simulations are significantly more detailed, spending 1,000,000x more compute, and also that they have an evolutionary search component that auto-generates 1,000 variations of each design and iterates for 1,000 generations to find the optimal version of each design for the problem (or even invents new designs from scratch.) And perhaps all of this automated design and tweaking (and even the in-simulation testing) is done more intelligently by a copy of OmegaStar trained on this “game.”

Why would this be a big deal? I’m not sure it would be. But take a look at this list of strategically relevant technologies and events and think about whether Skunkworks being widely available would quickly lead to some of them. For example, given how successful AlphaFold 2 has been, maybe Skunkworks could be useful for designing nanomachines. It could certainly make it a lot easier for various minor nations and non-state entities to build weapons of mass destruction, perhaps resulting in a vulnerable world.

## Neuromorph:

According to page 69 of this report, the Hodgkin-Huxley model of the neuron is the most detailed and realistic (and therefore the most computationally expensive) as of 2008. [EDIT: Joe Carlsmith, author of a more recent report, tells me there are more detailed+realistic models available now] It costs 1,200,000 FLOP per second per neuron to run. So a human brain (along with relevant parts of the body, in a realistic-physics virtual environment, etc.) could be simulated for about 10^17 FLOP per second.

Now, presumably (a) we don’t have good enough brain scanners as of 2020 to actually reconstruct any particular person’s brain, and (b) even if we did, the Hodgkin-Huxley model might not be detailed enough to fully capture that person’s personality and cognition.[13]

But maybe we can do something ‘fun’ nonetheless: We scan someone’s brain and then create a simulated brain that looks like the scan as much as possible, and then fills in the details in a random but biologically plausible way. Then we run the simulated brain and see what happens. Probably gibberish, but we run it for a simulated year to see whether it gets its act together and learns any interesting behaviors. After all, human children start off with randomly connected neurons too, but they learn.[14]

All of this costs a mere 10^25 FLOP. So we do it repeatedly, using stochastic gradient descent to search through the space of possible variations on this basic setup, tweaking parameters of the simulation, the dynamical rules used to evolve neurons, the initial conditions, etc. We can do 100,000 generations of 100,000 brains-running-for-a-year this way. Maybe we’ll eventually find something intelligent, even if it lacks the memories and personality of the original scanned human.

# Question Two: In this hypothetical, what’s the probability that TAI appears by end of 2020?

The first question was my way of operationalizing “what could be built with 2020’s algorithms/ideas/etc. but a trillion times more compute?”

This second question is my way of operationalizing “what’s the probability that the amount of computation it would take to train a transformative model using 2020’s algorithms/ideas/etc. is 10^35 FLOP or less?”

(Please ignore thoughts like “But maybe all this extra compute will make people take AI safety more seriously” and “But they wouldn’t have incentives to develop modern parallelization algorithms if they had computers so fast” and “but maybe the presence of the Compute Fairy will make them believe the simulation hypothesis?” since they run counter to the spirit of the thought experiment.)

Remember, the definition of Transformative AI is “AI that precipitates a transition comparable to (or more significant than) the agricultural or industrial revolution.”

Did you read those answers to Question One, visualize them and other similarly crazy things that would be going on in this hypothetical scenario, and think “Eh, IDK if that would be enough, I’m 50-50 on this. Seems plausible TAI will be achieved in this scenario but seems equally plausible it wouldn’t be.”

No! … Well, maybe you do, but speaking for myself, I don’t have that reaction.

When I visualize this scenario, I’m like “Holyshit all five of these distinct research programs seem like they would probably produce something transformative within five years and perhaps even immediately, and there are probably more research programs I haven’t thought of!”

My answer is 90%. The reason it isn’t higher is that I’m trying to be epistemically humble and cautious, account for unknown unknowns, defer to the judgment of others, etc. If I just went with my inside view, the number would be 99%. This is because I can’t articulate any not-totally-implausible possibility in which OmegaStar, Amp(GPT-7), Crystal Nights, Skunkworks, and Neuromorph and more don’t lead to transformative AI within five years. All I can think of is things like “Maybe transformative AI requires some super-special mental structure which can only be found by massive blind search, so massive that the Crystal Nights program can’t find it…” I’m very interested to hear what people whose inside-view answer to Question Two is <90% have in mind for the remaining 10%+. I expect I’m just not modelling their views well and that after hearing more I’ll be able to imagine some not-totally-implausible no-TAI possibilities. My inside view is obviously overconfident. Hence my answer of 90%.

Bonus: I’ve argued elsewhere that what we really care about, when thinking about AI timelines, is AI-induced points of no return. I think this is likely to be within a few years of TAI, and my answer to this question is basically the same as my answer to the TAI version, but just in case:

# OK, here’s why all this matters

Ajeya Cotra’s excellent timelines forecasting model is built around a probability distribution over “the amount of computation it would take to train a transformative model if we had to do it using only current knowledge.”[15] (pt1p25) Most of the work goes into constructing that probability distribution; once that’s done, she models how compute costs decrease, willingness-to-spend increases, and new ideas/insights/algorithms are added over time, to get her final forecast.

One of the great things about the model is that it’s interactive; you can input your own probability distribution and see what the implications are for timelines. This is good because there’s a lot of room for subjective judgment and intuition when it comes to making the probability distribution.

What I’ve done in this post is present an intuition pump, a thought experiment that might elicit in the reader (as it does in me) the sense that the probability distribution should have the bulk of its mass by the 10^35 mark.

Ajeya’s best-guess distribution has the 10^35 mark as its median, roughly. As far as I can tell, this corresponds to answering “50%” to Question Two.[16]

If that’s also your reaction, fair enough. But insofar as your reaction is closer to mine, you should have shorter timelines than Ajeya did when she wrote the report.

There are lots of minor nitpicks I have with Ajeya’s report, but I’m not talking about them; instead, I wrote this, which is a lot more subjective and hand-wavy. I made this choice because the minor nitpicks don’t ultimately influence the answer very much, whereas this more subjective disagreement is a pretty big crux.[17] Suppose your answer to Question 2 is 80%. Well, that means your distribution should have 80% by the 10^35 mark compared to Ajeya’s 50%, and that means that your median should be roughly 10 years earlier than hers, all else equal: 2040-ish rather than 2050-ish.[18]

I hope this post helps focus the general discussion about timelines. As far as I can tell, the biggest crux for most people is something like “Can we get TAI just by throwing more compute at the problem?” Now, obviously we can get TAI just by throwing more compute at the problem, there are theorems about how neural nets are universal function approximators etc., and we can always do architecture search to find the right architectures. So the crux is really about whether we can get TAI just by throwing a large but not too large amount of compute at the problem… and I propose we operationalize “large but not too large” as “10^35 FLOP or less.”[19] I’d like to hear people with long timelines explain why OmegaStar, Amp(GPT-7), Crystal Nights, SkunkWorks, and Neuromorph wouldn’t be transformative (or more generally, wouldn’t cause an AI-induced PONR). I’d rest easier at night if I had some hope along those lines.

This is part of my larger investigation into timelines commissioned by CLR. Many thanks to Tegan McCaslin, Lukas Finnveden, Anthony DiGiovanni, Connor Leahy, and Carl Shulman for comments on drafts. Kudos to Connor for pointing out the Skunkworks and Neuromorph ideas. Thanks to the LW team (esp. Raemon) for helping me with the formatting.

# Ω 71

New Comment
Some comments are truncated due to high volume. Change truncation settings

This post provides a valuable reframing of a common question in futurology: "here's an effect I'm interested in -- what sorts of things could cause it?"

That style of reasoning ends by postulating causes.  But causes have a life of their own: they don't just cause the one effect you're interested in, through the one causal pathway you were thinking about.  They do all kinds of things.

In the case of AI and compute, it's common to ask

• Here's a hypothetical AI technology.  How much compute would it require?

But once we have an answer to this question, we can always ask

• Here's how much compute you have.  What kind of AI could you build with it?

If you've asked the first question, you ought to ask the second one, too.

The first question includes a hidden assumption: that the imagined technology is a reasonable use of the resources it would take to build.  This isn't always true: given those resources, there may be easier ways to accomplish the same thing, or better versions of that thing that are equally feasible.  These facts are much easier to see when you fix a given resource level, and ask yourself what kinds of things you could do with it.

This high-level poi...

5Daniel Kokotajlo2mo

Update: After talking to various people, it appears that (contrary to what the poll would suggest) there are at least a few people who answer Question 2 (all three variants) with less than 80%. In light of those conversations, and more thinking on my own, here is my current hot take on how +12 OOMs could turn out to not be enough:

1. Maybe the scaling laws will break. Just because GPT performance has fit a steady line across 5 orders of magnitude so far (or whatever) doesn't mean it will continue for another 5. Maybe it'll level off for some reason we don't yet understand. Arguably this is what happened with LSTMs? Anyhow, for timelines purposes what matters is not whether it'll level off by the time we are spending +12 OOMs of compute, but rather more like whether it will level off by the time we are spending +6 OOMs of compute. I think it's rather unlikely to level off that soon, but it might. Maybe 20% chance. If this happens, then probably Amp(GPT-7) and the like wouldn't work. (80%?) The others are less impacted, but maybe we can assume OmegaStar probably won't work either. Crystal Nights, SkunkWorks, and Neuromorph... don't seem to be af...

6abramdemski2y
So, how does the update to the AI and compute trend [https://www.lesswrong.com/posts/wfpdejMWog4vEDLDg/ai-and-compute-trend-isn-t-predictive-of-what-is-happening] factor in?
4Daniel Kokotajlo2y
It is irrelevant to this post, because this post is about what our probability distribution over orders of magnitude of compute should be like. Once we have said distribution, then we can ask: How quickly (in clock time) will we progress through the distribution / explore more OOMs of compute? Then the AI and compute trend, and the update to it, become relevant. But not super relevant IMO. The AI and Compute trend was way too fast to be sustained, people at the time even said so. This recent halt in the trend is not surprising. What matters is what the trend will look like going forward, e.g. over the next 10 years, over the next 20 years, etc. It can be broken down into two components: Cost reduction and spending increase. Ajeya separately estimates each component for the near term (5 years) and for the long-term trend beyond. I mostly defer to her judgment on this, with large uncertainty. (Ajeya thinks costs will halve every 2.5 years, which is slower than the 1.5 years average throughout all history, but justifiable given how Moore's Law is said to be dying now. As for spending increases, she thinks it will take decades to ramp up to trillion-dollar expenditures, whereas I am more uncertain and think it could maybe happen by 2030 idk.) I feel quite confident in the following claim: Conditional on +6 OOMs being enough with 2020's ideas, it'll happen by 2030. Indeed, conditional on +8 OOMs being enough with 2020's ideas, I think it'll probably happen by 2030. If you are interested in more of my arguments for this stuff, I have some slides I could share slash I'd love to chat with you about this! :)
7abramdemski2y
If the AI and compute trend is just a blip, then doesn't that return us to the previous trend line in the graph you show at the beginning, where we progress about 2 ooms a decade? (More accurately, 1 oom every 6-7 years, or, 8 ooms in 5 decades.) Ignoring AI and compute, then: if we believe +12 ooms in 2016 means great danger in 2020, we should believe that roughly 75 years after 2016, we are at most four years from the danger zone. Whereas, if we extrapolate the AI-and-compute trend, +12 ooms is like jumping 12 years in the future; so the idea of risk by 2030 makes sense. So I don't get how your conclusion can be so independent of AI-and-compute.
5Daniel Kokotajlo1y
Sorry, somehow I missed this. Basically, the answer is that we definitely shouldn't just extrapolate out the AI and compute trend into the future, and Ajeya's and my predictions are not doing that. Instead we are assuming something more like the historic 2 ooms a decade trend, combined with some amount of increased spending conditional on us being close to AGI/TAI/etc. Hence my conditional claim above: If you want to discuss this more with me, I'd love to, how bout we book a call?
4abramdemski2y
Is there a reference for this?
2Daniel Kokotajlo2y
What Gwern said. :) But I don't know for sure what the person I talked to had in mind.

I feel like if you think Neuromorph has a good chance of succeeding, you need to explain why we haven't uploaded worms yet. For C. elegans, if we ran 302 neurons for 1 subjective day (= 8.64e4 seconds) at 1.2e6 flops per neuron, and did this for 100 generations of 100 brains, that takes a mere 3e17 flops, or about $3 at current costs. (And this is very easy to parallelize, so you can't appeal to that as a reason this can't be done.) (It's possible that we have uploaded worms in the 7 years since that blog post was written, though I would have expected to hear about it if so.) 8Daniel Kokotajlo2y Good question! Here's my answer: --I think Neuromorph has the least chance of succeeding of the five. Still more than 50% though IMO. I'm not at all confident in this. --Neuromorph =/= an attempt to create uploads. I would be extremely surprised if the resulting AI was recognizeably the same person as was scanned. I'd be mildly surprised if it even seemed human-like at all, and this is conditional on the project working. What I imagine happening conditional on the project working is something like: After a few generations of selection, we get brains that "work" in more or less the way that vanilla ANN's work, i.e. the brains seem like decently competent RL agents (of human-brain size) at RL tasks, decently competent transformers at language modelling, etc. So, like GPT-5 or so. But there would still be lots of glitches and weaknesses and brittleness. But then with continued selection we get continued improvement, and many (though not nearly all) of the tips & tricks evolution had discovered and put into the brain (modules, etc.) are rediscovered. Others are bypassed entirely as the selection process routes around them and finds new improvements. At the end we are left with something maybe smarter than a human, but probably not, but competent enough and agenty enough to be transformative. (After all, it's faster and cheaper than a human.) From the post you linked: Yeah, Neuromorph definitely won't be uploading humans in that sense. This might already qualify as success for what I'm interested in, depending on how "wormlike" the behaviors are. I haven't looked into this. --I used to think our inability to upload worms was strong evidence against any sort of human uploading happening anytime soon. However, I now think it's only weak evidence. This is because, counterintuitively, worms being small makes them a lot harder to simulate. Notice how each worm has exactly 302 neurons, and their locations and connections are the same in each worm. That means the genes ar 9Rohin Shah2y My impression is that the linked blog post is claiming we haven't even been able to get things that are qualitatively as impressive as a worm. So why would we get things that are qualitatively as impressive as a human? I'm not claiming it has to be an upload. I could believe this (based on the argument you mentioned) but it really feels like "maybe this could be true but I'm not that swayed from my default prior of 'it's probably as easy to simulate per neuron'". Also if it were 100x harder, it would cost...$300. Still super cheap. That's what the genetic algorithm is? It probably wasn't run with as 3e17 flops, since compute was way more expensive then, but that's at least evidence that researchers do in fact consider this approach.

At this point I guess I just say I haven't looked into the worm literature enough to say. I can't tell from the post alone whether we've neuromorphed the worm yet or not.

"Qualitatively as impressive as a worm" is a pretty low bar, I think. We have plenty of artificial neural nets that are much more impressive than worms already, so I guess the question is whether we can make one with only 302 neurons that is as impressive as a worm... e.g. can it wriggle in a way that moves it around, can it move away from sources of damage and towards sources of food, etc. idk, I feel like maybe at this point we should make bets or something, and then go read the literature and see who is right? I don't find this prospect appealing but it seems like the epistemically virtuous thing to do.

I do feel fairly confident that on a per-neuron basis worms are much harder than humans to simulate. My argument seems solid enough for that conclusion, I think. It's not solid enough to mean that you are wrong though -- like you said, a 100x difference is still basically nothing. And to be honest I agree that the difference probably isn't much more than that; maybe 1000x o...

7Rohin Shah2y
Meh, I don't think it's a worthwhile use of my time to read that literature, but I'd make a bet if we could settle on an operationalization and I didn't have to settle it. I mostly expect that you realize that there were a bunch of things that were super underspecified and they don't have obvious resolutions, and if you just pick a few things then nothing happens and you get gibberish eternally, and if you search over all the underspecified things you run out of your compute budget very quickly. Some things that might end up being underspecified: * How should neurons be connected to each other? Do we just have a random graph with some average degree of connections, or do we need something more precise? * How are inputs connected to the brain? Do we just simulate some signals to some input neurons, that are then propagated according to the physics of neurons? How many neurons take input? How are they connected to the "computation" neurons? * To what extent do we need to simulate other aspects of the human body that affect brain function? Which hormone receptors do we / don't we simulate? For the ones we do simulate, how do we determine what their inputs are? Or do we have to simulate an entire human body (would be way, way more flops)? * How do we take "random draws" of a new brain? Do we need to simulate the way that DNA builds up the brain during development? * Should we build brains that are like that of a human baby, or a human adult, given that the brain structure seems to change between these? I'm not saying any of these things will be the problem. I'm saying that there will be some sort of problem like this (probably many such problems), that I'm probably not going to find reasoning from my armchair. I also wouldn't really change my mind if you had convincing rebuttals to each of them, because the underlying generator is "there are lots of details; some will be devilishly difficult to handle; you only find those by actuall
7Daniel Kokotajlo2y
On the contrary, I've been very (80%?) surprised by the responses so far -- in the Elicit poll, everyone agrees with me! I expected there to be a bunch of people with answers like "10%" and "20%" and then an even larger bunch of people with answers like "50%" (that's what I expected you, Ajeya, etc. to chime in and say). Instead, well, just look at the poll results! So, even a mere registering of disagreement is helpful. That said, I'd be interested to hear why you have similar feelings about the non-Neuromorph answers, considering that you agreed with the point I was making in the birds/brains/etc. post [https://www.lesswrong.com/posts/HhWhaSzQr6xmBki8F/birds-brains-planes-and-ai-against-appeals-to-the-complexity]. If we aren't trying to replicate the brain, but just to do something that works, yes there will be lots of details to work out, but what positive reason do you have to think that the amount of special sauce / details is so high that 12 OOMs and a few years isn't enough to find it? Interesting. This conflicts with something I've been told about neural networks, which is that they "want to work." Seems to me that more likely than eternal gibberish is something that works but not substantially better than regular ANN's of similar size. So, still better than GPT-3, AlphaStar, etc. After all, those architectures are simple enough that surely something similar is in the space of things that would be tried out by the Neuromorph search process? I think the three specific ways my claim about worms could be wrong are not very plausible: Sure, most of the genes don't code neuron stuff. So what? Sure, maybe the DNA mostly contents itself with specifying number + structure of neurons, but that's just a rejection of my claim, not an argument against it. Sure, maybe it's fine to have just one type of neuron if you are so simple -- but the relevant metric is not "number of types" but "number of types / number of neurons." And the size of the human genome limits that
6Rohin Shah2y
The positive reason is basically all the reasons given in Ajeya's report? Like, we don't tend to design much better artifacts than evolution (currently), the evolution-designed artifact is expensive, and reproducing it using today's technology looks like it could need more than 12 OOMs. I don't think the birds/brains/etc post contradicts this reason, as I said before [https://www.lesswrong.com/posts/HhWhaSzQr6xmBki8F/birds-brains-planes-and-ai-against-appeals-to-the-complexity?commentId=LQZZsGvyq5ruoLJvd] (and you seemed to agree).

Ok, as a former neuroscientist who has spent a lot of years (albeit not recent ones) geeking out about, downloading, and playing with various neural models, I'd like to add to this discussion. First, the worm stuff seems overly detailed and focused on recreating the exact behavior rather than 'sorta kinda working like a brain should'. A closer, more interesting project to look at (but still too overly specific) is the Blue Brain project [ https://www.epfl.ch/research/domains/bluebrain/ ]. Could that work with 12 more OOMs of compute? I feel quite confident it could, with no additional info. But I think you could get there with a lot less than 12 OOMs if you took a less realistic, more functional project like Nengo [ https://www.nengo.ai/ ]. Nengo is a brain simulation that can already do somewhat interesting stuff at boring 2019 levels of compute. If you gave it GPT-3 levels of compute, I bet it would be pretty awesome.

And beyond that, neuroscientists have been obsessively making separate little detailed computer models of their tiny pieces of specialized knowledge about the brain since the 1980s at least, here's some links [ https://compneuroweb.com/database.html ]. There are arch...

5Rohin Shah1y
(I know ~nothing about any of this, so might be misunderstanding things greatly) 12 OOMs is supposed to get us human-level AGI, but BlueBrain seems to be aiming at a mouse brain? "It takes 12 OOMs to get to mouse-level AGI" seems like it's probably consistent with my positions? (I don't remember the numbers well enough to say off the top of my head.) But more fundamentally, why 12 OOMs? Where does that number come from? From a brief look at the website, I didn't immediately see what cool stuff Nengo could do with 2019 levels of compute, that neural networks can't do. Same for Numenta.
6Nathan Helm-Burger1y
Blue Brain does actually have a human brain model waiting in the wings, it just tries to avoid mentioning that. A media-image management thing. I spent the day digging into your question about OOMs, and now have much more refined estimates. Here's my post: https://www.lesswrong.com/posts/5Ae8rcYjWAe6zfdQs/what-more-compute-does-for-brain-like-models-response-to [https://www.lesswrong.com/posts/5Ae8rcYjWAe6zfdQs/what-more-compute-does-for-brain-like-models-response-to]
4Daniel Kokotajlo2y
Hmmm, it seems we aren't on the same page. (The argument sketch you just made sounds to me like a collection of claims which are either true but irrelevant, or false, depending on how I interpret them.) I'll go back and reread Ajeya's report (or maybe talk to her?) and then maybe we'll be able to get to the bottom of this. Maybe my birds/brains/etc. post directly contradicts something in her report after all.

That crystal nights story. As I was reading it, it was like a mini Eliezer in my brain facepalming over and over.

Its clear that the characters have little idea how much suffering they caused, or how close they came to destroying humanity. It was basically luck that pocket universe creation caused a building wrecking explosion, not a supernova explosion. Its also basically luck that the Phites didn't leave some nanogoo behind them. You still have a hyperrapid alien civ on the other side of that wormhole. One with known spacewarping tech, and good reason to hate you, or just want to stop you trying the same thing again, or to strip every last bit of info about their creation from human brains. How long until they invade?

This feels like an idiot who is playing with several barely subcritical lumps of uranium, and drops one on their foot.

5[comment deleted]2y

Footnotes

5Daniel Kokotajlo2y
[1] AlphaStar was 10^8 parameters, ten times smaller than a honeybee brain. I think this puts its capabilities in perspective. Yes, it seemed to be more of a heuristic-executor than a long-term planner, because it could occasionally be tricked into doing stupid things repeatedly. But the same is true for insects [https://en.wikipedia.org/wiki/Ant_mill].
3Daniel Kokotajlo2y
[18] Quick calculation: Suppose we take Ajeya’s best-guess distribution and modify it by lowering the part to the right of 10^35 and raising the part to the left of 10^35, until the 10^35 mark is the 80-percentile mark instead of the 50-percentile mark. And suppose we do this raising and lowering in a “distribution-preserving way,” i.e. the shape of the curve before the 10^35 mark looks exactly the same, it’s just systematically bigger. In other words, we redistribute 30 percentage points of probability mass from above 10^35 to below, in proportion to how the below-10^35 mass is already distributed.  Well, in this case, then 60% of the redistributed mass should end up before the old 30% mark. (Because the 30% mark is 60% of the mass prior to the old median, the 10^35 mark.) And 60% of 30 percentage points is 18, so that means +18 points added before the old 30% mark. This makes it the new 48% mark. So the new 48% mark should be right where the old 30% mark is, which is (eyeballing the spreadsheet) a bit after 2040, 10 years sooner. (Ajeya’s best guess median is a bit after 2050.) This is, I think, a rather conservative estimate of how cruxy this disagreement is.  First, my answer to Question Two is 0.9, not 0.8, and that’s after trying to be humble and whatnot. Second, this procedure of redistributing probability mass in proportion to how it is already distributed produces an obviously silly outcome, where there is a sharp drop-off in probability at the 10^35 mark. Realistically, if you are convinced that the answer to Question Two is 0.8 (or whatever) then you should think the probability distribution tapers off smoothly, being already somewhat low by the time the 80th-percentile mark at 10^35 is reached.  Thus, realistically, someone who mostly agrees with Ajeya but answers Question Two with “0.8” should have somewhat shorter timelines than the median-2040-ish I calculated
2Daniel Kokotajlo2y
[2] This is definitely true for Transformers (and LSTMs I think?), but it may not be true for whatever architecture AlphaStar uses. In particular some people I talked to worry that the vanishing gradients problem might make bigger RL models like OmegaStar actually worse. However, everyone I talked to agreed with the “probably”-qualified version of this claim. I’m very interested to learn more about this.
2Daniel Kokotajlo2y
[3] To avoid catastrophic forgetting, let’s train OmegaStar on all these different games simultaneously, e.g. it plays game A for a short period, then plays game B, then C, etc. and loops back to game A only much later.
2Daniel Kokotajlo2y
[4] Lukas Finnveden points out that Gwern’s extrapolation is pretty weird. Quoting Lukas: “Gwern takes GPT-3's current performance on lambada; assumes that the loss will fall as fast as it does on "predict-the-next-word" (despite the fact that the lambada loss is currently falling much faster!) and extrapolates current performance (without adjusting for the expected change in scaling law after the crossover point) until the point where the AI is as good as humans (and btw we don't have a source for the stated human performance) I'd endorse a summary more like “If progress carries on as it has so far, we might just need ~1e27 FLOP to get to mturk-level of errors on the benchmarks closest to GPT-3's native predict-the-next-word game. Even if progress on these benchmarks slowed down and improved at the same rate as GPT-3's generic word-prediction abilities, we'd expect it to happen at ~1e30 FLOP for the lambada benchmark." All that being said, Lukas’ own extrapolation [https://www.lesswrong.com/posts/k2SNji3jXaLGhBeYP/extrapolating-gpt-n-performance] seems to confirm the general impression that GPT’s performance will reach human-level around the same time its size reaches brain-size: “Given that Cotra’s model’s median number of parameters is close to my best guess of where near-optimal performance is achieved, the extrapolations do not contradict the model’s estimates, and constitute some evidence for the median being roughly right.”↩︎ [https://www.lesswrong.com/posts/rzqACeBGycZtqCfaX/undefined]
2Daniel Kokotajlo2y
[5] One might worry that the original paper had a biased sample of tasks. I do in fact worry about this. However, this paper [https://arxiv.org/abs/2009.03300] tests GPT-3 on a sample of actual standardized tests used for admission to colleges, grad schools, etc. and GPT-3 exhibits similar performance (around 50% correct), and also shows radical improvement over smaller versions of GPT.
2Daniel Kokotajlo2y
[6] In theory (and maybe in practice too, given how well the new pre-training paradigm is working? See also e.g. this paper [https://arxiv.org/abs/2010.14701]) it should be easier for the model to generalize and understand concepts since it sees images and videos and hears sounds to go along with the text.
2Daniel Kokotajlo2y
[7] GPT-3 has already been used to write its own prompts, sorta. See this paper [https://arxiv.org/abs/2102.07350] and look for “metaprompt.” Also, this paper [https://arxiv.org/abs/2101.00190] demonstrates the use of stochastic gradient descent on prompts to evolve them into better versions.
2Daniel Kokotajlo2y
[8] Thanks to Connor Leahy for finding this source for me.
2Daniel Kokotajlo2y
[9] 50,000 x 50,000 x 100,000,000 x 10^17 x 6 = 1.5x10^35
2Daniel Kokotajlo2y
[10] Bostrom and Shulman [https://www.nickbostrom.com/aievolution.pdf] have an earlier estimate with wide error bars: 10^38 - 10^51 FLOP. See page 6 and multiply their FLOPS range by the number of seconds in a year.
2Daniel Kokotajlo2y
[11] Well, we’d definitely start with small brains and scale up, but we’d make sure to spend only a fraction of our overall compute on small brains. From the report, page 25 of part 3: "the number of FLOP/s contributed by humans is (~7e9 humans) * (~1e15 FLOP/s / person) = ~7e24. The human population is vastly larger now than it was during most of our evolutionary history, whereas it is likely that the population of animals with tiny nervous systems has stayed similar. This suggests to me that the average ancestor across our entire evolutionary history was likely tiny and performed very few FLOP/s. I will assume that the “average ancestor” performed about as many FLOP/s as a nematode and the “average population size” was ~1e21 individuals alive at a given point in time. This implies that our ancestors were collectively performing ~1e25 FLOP every second on average over the ~1 billion years of evolutionary history."
2Daniel Kokotajlo2y
[12] See page 4 of this paper [https://ti.arc.nasa.gov/m/pub-archive/1244h/1244%20%28Hornby%29.pdf]. Relevant quote: “Originally the ST5 mission managers had hired a contractor to design and produce an antenna for this mission. Using conventional design practices the contractor produced a quadrifilar helix antenna (QHA). In Fig. 3 we show performance comparisons of our evolved antennas with the conventionally designed QHA on an ST5 mock-up. Since two antennas are used on each spacecraft – one on the top and one on the bottom – it is important to measure the overall gain pattern with two antennas mounted on the spacecraft. With two QHAs 38% efficiency was achieved, using a QHA with an evolved antenna resulted in 80% efficiency, and using two evolved antennas resulted in 93% efficiency. Since the evolved antenna does not require a phasing circuit, less design and fabrication work is required, and having fewer parts may result in greater reliability. In terms of overall work, the evolved antenna required approximately three person-months to design and fabricate whereas the conventional antenna required approximately five months. Lastly, the evolved antenna has more uniform coverage in that it has a uniform pattern with only small ripples in the elevations of greatest interest (40◦ − 80◦ ). This allows for reliable performance as the elevation angle relative to the ground changes.”
2Daniel Kokotajlo2y
[13] So, no Ems [https://ageofem.com/], basically. Probably.
2Daniel Kokotajlo2y
[14] I mean, definitely not completely random. But I said we’d fill in the details in a random-but-biologically-plausible way. And children simply have far too many neurons for genes to say much about how they connect to each other. Whatever unknowns there are about about how the neurons connect, we can make that part of what’s being optimized by our hundred-thousand-generation search process. The size of the search space can’t be that big, because there isn’t that much space in the human genome to encode any super-complicated instructions. I guess at this point we should start talking about Ajeya’s Genome Anchor idea. I admit I’m out of my depth here.
2Daniel Kokotajlo2y
[15] Since later I talk about how I disagree with Ajeya, I want to make super clear that I really do think her report is excellent. It’s currently the best writing on timelines that I know of. When people ask me to explain my timelines, I say “It’s like Ajeya’s, except…"
2Daniel Kokotajlo2y
[16] I think this because I’ve looked at the probability distribution she gives on page 34 of part 3 of her report [https://docs.google.com/document/d/1k7qzzn14jgE-Gbf0CON7_Py6tQUp2QNodr_8VAoDGnY/edit#heading=h.8dpqf6fihesq] and 35 OOMs of floating point operations seems to be the median. I had to do this by measuring with a ruler and doing some eyeballing, so it’s probably not exactly correct, but I’d be surprised if the true answer is more than 55% or less than 45%. As a sanity check, Ajeya has 7 buckets in which her credence is distributed, with 30% in a bucket with median 34.5 OOMs, and 35% in buckets with higher medians, and 35% in buckets with lower medians. (But the lower buckets are closer to 35 than the higher buckets, meaning their tails will be higher than 35 more than the higher bucket’s tails will be lower than 35.
2Daniel Kokotajlo2y
[17] Example: One nitpick I have is that Ajeya projects that the price of compute for AI will fall more slowly than price-performance Moore’s Law, because said law has faltered recently. I instead think we should probably model this uncertainty, with (say) a 50% chance of Moore continuing and a 50% chance of continued slowdown. But even if it was a 100% chance of Moore continuing, this would only bring forward Ajeya’s median timeline to 2045-ish! (At least, according to my tinkering with her spreadsheet)
2Daniel Kokotajlo2y
[19] I recommend trying out different numbers depending on who is in the conversation. For conversations in which everyone assigns high credence to the 10^35 version, it may be more fruitful to debate the 10^29 version, since 10^29 FLOP is when GPT-7 surpasses human level at text prediction (and is also superhuman at the other tests we’ve tried) according to the scaling laws and performance trends [https://www.lesswrong.com/posts/k2SNji3jXaLGhBeYP/extrapolating-gpt-n-performance], I think.  For conversations where everyone has low credence in the 10^35 version, I suggest using the 10^41 version, since 10^41 FLOP is enough to recapitulate evolution without any shortcuts.

uses about six FLOP per parameter per token

Shouldn't this be 2 FLOP per parameter per token, since our evolutionary search is not doing backward passes?

On the other hand, the calculation in the footnote seems to assume that 1 function call = 1 token, which is clearly an unrealistic lower bound.

A "lowest-level" function (one that only uses a single context window) will use somewhere between 1 and  tokens.  Functions defined by composition over "lowest-level" functions, as described two paragraphs above, will of course require more tokens per call than their constituents.

4Daniel Kokotajlo2mo
Thanks for checking my math & catching this error!

But I want to say: that much compute would be useful for meta-learning/NAS/AIGAs, not just scaling up DNNs.  I think that would likely be a more productive research direction.  And I want to make sure that people are not ONLY imagining bigger DNNs when they imagine having a bunch more compute, but also imagining how it could be used to drive fundamental advances in ML algos, which could plausibly kick of something like recursive self-improvement (even in DNNs are in some sense a dead end).

Something I'm wondering, but don't have the expertise in meta-learning to say confidently (so, epistemic status: speculation, and I'm curious for critiques): extra OOMs of compute could overcome (at least) one big bottleneck in meta-learning, the expense of computing second-order gradients. My understanding is that most methods just ignore these terms or use crude approximations, like this, because they're so expensive. But at least this paper found some pretty impressive performance gains from using the second-order terms.

Maybe throwing lots of compute at this aspect of meta-learning would help it cross a threshold of viability, like what happened for deep learning in general around 2012. I think meta-learning is a case where we should expect second-order info to be very relevant to optimizing the loss function in question, not just a way of incorporating the loss function's curvature. In the first paper I linked, the second-order term accounts for how the base learner's gradients depend on the meta-learner's parameters. This seems like an important feature of what their meta-learner is trying/supposed to do, i.e., use the meta-learned update rule to guide the base learner - and t...

6lsusr2y
It weirds me out how little NAS (Neural Architecture Search) in particular (and throwing compute at architecture search in general) is used in industry.
2Daniel Kokotajlo2y
Interesting, could you elaborate? I'd love to have a nice, fleshed-out answer along those lines to add to the five I came up with. :)
2David Scott Krueger (formerly: capybaralet)2y
Sure, but in what way? Also I'd be happy to do a quick video chat if that would help (PM me).
2Daniel Kokotajlo2y
Well, I've got five tentative answers to Question One in this post. Roughly, they are: Souped-up AlphaStar, Souped-up GPT, Evolution Lite, Engineering Simulation, and Emulation Lite. Five different research programs basically. It sounds like what you are talking about is sufficiently different from these five, and also sufficiently promising/powerful/'fun', that it would be a worthy addition to the list basically. So, to flesh it out, maybe you could say something like "Here are some examples of meta-learning/NAS/AIGA in practice today. Here's a sketch of what you could do if you scaled all this up +12 OOMs. Here's some argument for why this would be really powerful."
4David Scott Krueger (formerly: capybaralet)2y
There's a ton of work in meta-learning, including Neural Architecture Search (NAS).  AIGA's (Clune) is a paper that argues a similar POV to what I would describe here, so I'd check that out.   I'll just say "why it would be powerful": the promise of meta-learning is that -- just like learned features outperform engineered features -- learned learning algorithms will eventually outperform engineered learning algorithms.  Taking the analogy seriously would suggest that the performance gap will be large -- a quantitative step-change.   The upper limit we should anchor on is fully automated research.  This helps illustrate how powerful this could be, since automating research could easily give many orders of magnitude speed up (e.g. just consider the physical limitation of humans manually inputting information about what experiment to run). An important underlying question is how much room there is for improvement over current techniques.  The idea that current DL techniques are pretty close to perfect (i.e. we've uncovered the fundamental principles of efficient learning (associated view: ...and maybe DNNs are a good model of the brain)) seems too often implicit in some of the discussions around forecasting and scaling.  I think it's a real possibility, but I think it's fairly unlikely (~15%, OTTMH).  The main evidence for it is that 99% of published improvements don't seem to make much difference in practice/at-scale.   Assuming that current methods are roughly optimal has two important implications: - no new fundamental breakthroughs needed for AGI (faster timelines) - no possible acceleration from fundamental algorithmic breakthroughs (slower timelines)

The ideas in this post greatly influence how I think about AI timelines, and I believe they comprise the current single best way to forecast timelines.

A +12-OOMs-style forecast, like a bioanchors-style forecast, has two components:

1. an estimate of (effective) compute over time (including factors like compute getting cheaper and algorithms/ideas getting better in addition to spending increasing), and
2. a probability distribution on the (effective) training compute requirements for TAI (or equivalently the probability that TAI is achievable as a function of tr
...

Really interesting post, I appreciate the thought experiment. I have one comment on it related to the Crystal Nights and Skunkworks sections, based on my own experience in the aerospace world. There are lots of problems that I deal with today where the limiting factor is the existence of high-quality experimental data (for example, propellant slosh dynamics in zero-g). This has two implications:

1. For the "Crystal Nights" example, I think that our current ability to build virtual worlds that are useful for evolutionarily creating truly transformative AIs may
...
3Daniel Kokotajlo2y
Thanks! Yeah, I basically agree with you overall that Skunkworks could be undermined by our lack of understanding of real-physics dynamics. We certainly wouldn't be able to create perfectly accurate simulations even if we threw all 35 OOMs at the problem. The question is whether the simulations would be accurate enough to be useful. My argument is that since we already have simulations which are accurate enough to be useful, adding +12 OOMs should lead to simulations which are even more useful. But useful enough to lead to crazy transformative stuff? Yeah, I don't know. For Crystal Nights, I'm more 'optimistic.' The simulation doesn't have to be accurate at all really. It just has to be complex enough, in the right sort of ways. If you read the Crystal Nights short story I linked, it involves creating a virtual world and evolving creatures in it, but the creators don't even try to make the physics accurate; they deliberately redesign the physics to be both (a) easier to compute and (b) more likely to lead to intelligent life evolving.
3theme_arrow2y
Your comment about Crystal Nights makes sense. I guess humans have evolved in a word based on one set of physical laws, but we're general purpose intelligences that can do things like play videogames really well even when the game's physics don't match the real world's.

I like how these ideas are falsifiable, in the sense that that they have clear performance success criteria. It is possible to evaluate whether we hit these milestones (if we do so before a increase). I also like how it addresses several different potential directions for AI development instead of just scaling today's most popular architectures.

This was really interesting to read. I'm still pretty new to the AI space so I don't know how this compares to our current FLOP usage. Assuming our current course of computing power doesn't change, how long is the timeline to get to 10^34 FLOP of computing power?

7Daniel Kokotajlo2y
Well, it's 12 OOMs more than our current FLOP usage. ;) Ajeya's report is excellent and contains the best and most thorough answers to your questions I think. If I recall correctly, she projects that the price of compute will drop in half every 2.5 years on average, and that algorithmic/efficiency improvements will have a similar effect. And then there's a one-time boost of up to +5 OOMs that we get from just spending a lot more. So she projects we get to (the equivalent of) +12 OOMs around 2050. Personally, I think the price of compute will drop a bit faster in the next ten years, and the algorithmic improvements will be a bit better too in the near term. But I'm very uncertain about that and mostly just defer to her judgment.

Consider Stockfish + NN searches 25 moves ahead on modern hardware. Considering a branching factor of 10 this means Stockfish +12oom will see 37 moves ahead. Some games last less then 37 moves. This is really cool indeed

4Archimedes2y
With AB pruning, the branching factor is closer to 2 for Stockfish.

The aliens seem to have also included with their boon:

• Cheap and fast eec GPU RAM with minute electricity consumptions
• A space time disruptor that allows you to have CMOS transistors smaller than electrons to serve as the L1&L2
• A way of getting rid of electron tunneling at a very small scale
• 12 OOMs better SSDs and fiber optic connections and cures for the host of physical limitations plaguing mere possibility of those 2.
5Daniel Kokotajlo2y
I did say it was magic. :D

At the Landauer kT limit, you need  kWh to perform your  FLOPs. That's 10,000x the yearly US electricity production. You'd need a quantum computer or a Dyson sphere to solve that problem.

2Daniel Kokotajlo2y
Either that, or magic. :D

More than 5mins, because it's fun, but:

For convenience I shall gloss this "12 orders of magnitude' thing to "suddenly impossibly fast".

Are there energy and thermal implications here? If we did 12 orders of magnitude more computation for what it costs today, we could probably only do it underwater at a hydroelectric damn. Things, both our devices and eventually the rest of the atmosphere, would get much warmer.

Disk and memory are now our bottlenecks for everything that used to be compute-intensive. We probably set algorithms to designing future iterations w...

7Daniel Kokotajlo2y
I forgot about memory! I guess I should have just said "Magically 12 OOMs cheaper" instead of "magically 12 OOMs faster" though that's a bit weirder to imagine.
3sanil2mo
Why limit ourselves to our planet? 12 OOMs is well within reason if we were a type 2 civilization and had access to all the energy from our sun (our planet as a whole only receives 1/10^10 of it). Encryption wouldn't really be an issue - we can simply tune our algorithms to use slightly more complicated assumptions. After all, one can just pick a problem that scales as O(10^(6n)), where n could for example be secret key length. If you have 12 orders of magnitude more compute, just make your key 2 times larger and you still have your cryptography. Thought of how small computers (phones etc) would scale also came to me. Basically, with 12 OOMs every phone becomes a computer powerful enough to train something as complicated as GPT-3. Everyone could carry their own personalized GPT-3 model with them. Actually, this is another way to improve AI performance - reduce the amount of things it needs to model. Training a personalized model specific for one problem would be cheaper and require less parameters/layers to get useful results. Basically, we would be able to put a "small" specialized model with power like that of GPT-3 on every microcontroller. You mentioned deep fakes. But with this much compute, why not "deepfake" brand new networks from scratch? Doing such experiments right now is expensive since they roughly quadruple the amount of computational resources needed to achieve this "second order" training mode. Theoretically, there's nothing preventing one from constructing a network that can assemble new networks based on training parameters. This meta-network could take into account network structure that it "learned" from training smaller models for other applications in order to generate new models with order of magnitude less parameters. As an analogy, compare the amount of data a neural network needs to learn to differentiate between cats and dogs vs the amount of data a human needs to learn the same thing. Human only needs a couple of examples, while neu

I think it's worth noting Joe Carlsmith's thoughts on this post, available starting on page 7 of Kokotajlo's review of Carlsmith's power-seeking AI report (see this EA Forum post for other reviews).

JC: I do think that the question of how much probability mass you concentrate on APS-AI by 2030 is helpful to bring out – it’s something I’d like to think more about (timelines wasn’t my focus in this report’s investigation), and I appreciate your pushing the consideration.

...
3Daniel Kokotajlo2mo
Great point, thanks for linking this! I think 65% that +12 OOMs would be enough isn't crazy. I'm obviously more like 80%-90%, and therein lies the crux. (Fascinating that such a seemingly small difference can lead to such big downstream differences! This is why I wrote the post.) If you've got 65% by +12, and 25% by +6, as Joe does, where is your 50% mark? idk maybe it's at +10? So, going to takeoffspeeds.com, and changing the training requirements parameter to +10 OOMs more than GPT-3 (so, 3e33) we get the following: So, I think it's reasonable to say that Joe's stated credences here roughly imply 50% chance of singularity by 2033.

Very interesting question!

My observations:

1. 12 OOMs is a lot
2. 12 OOMs is like comparing computational capacity of someone who just started learning arithmetics to a computer that an average person could obtain
3. You mainly focused on modern AI approaches and how they would scale. With 12 OOMs, it's possible that other approaches would be used. See galactic algorithms for example.
4. It might be useful to look at this through the lens of computational complexity. Here's a table of expected speedups:
Problem Complexity Scaling
O(log(n)) 10^10¹²
O(n), O(n
...
1. How would something like OmegaStar be transformative?
2. How would something like GPT-7 be transformative?

(I think your kind of thinking is currently the best way to do timelines, and I buy that we could very likely do fun stuff with +12 OOMs, but I don't understand through what mechanism these systems would be transformative, and this seems really important to think about carefully so we can determine other probabilities, like the probability that a GPT-style system/bureaucracy with 10^30 FLOP would be transformative.)

3Daniel Kokotajlo6mo
I think OmegaStar and Amp(GPT-7) would be APS-AI -- advanced, planning, strategically aware. (Defined in the Carlsmith report [https://arxiv.org/abs/2206.13353]). I think they'd be capable of playing the training game [https://www.lesswrong.com/posts/pRkFkzwKZ2zfa3R6H/without-specific-countermeasures-the-easiest-path-to] and deceiving humans into thinking they are aligned/trustworthy and/or incapable of doing truly dangerous things. I think they'd probably have powerbase ability. [https://www.lesswrong.com/posts/n3w3ww9Xuf8SngBfE/replacement-for-ponr-concept] As for economic impact, I think they'd be able to automate a significant fraction of the economy and in particular accelerate AI R&D significantly. (Once they and variations of them and things they built were suitably deployed widely, that is.) This is just a brief answer of course, I imagine you want more -- but could you perhaps explain which part of the above you disagree with and why?
3Zach Stein-Perlman6mo
Thanks very much! I don't really disagree-- I just feel uncertain. I'd struggle to tell a story about how (e.g.) OmegaStar could "automate a significant fraction of the economy" or "accelerate AI R&D significantly" or take control of a leading AI lab (powerbase ability). To be more precise, I feel deeply uncertain about how agent-y OmegaStar would be, how capable it would be, and the kinds of actions that it could use to take over (if it wanted to and was sufficiently capable). I can't think of much written on this. I think it could be a great use of your time to write about how AI systems based on game-playing or token-prediction could be economically important, be important to AI R&D, and have powerbase ability, as well as what signs we might see along the way. (This might easily fit into a continuation of What 2026 looks like [https://www.lesswrong.com/posts/6Xgy6CAf2jqHhynHL/what-2026-looks-like], or maybe not.)
7Daniel Kokotajlo6mo
I agree this is something I could do that would be valuable, I just haven't prioritized it alas. I did say I wanted to write the sequel to that story at some point... I encourage other people to think about it too. I suggest reasoning as follows: (1) What skills/properties would combine to create APS-AI? Agency, reasoning, situational awareness... maybe persuasion, deception, hacking... etc. (2) Does the training process of OmegaStar or whatever incentivize the growth of those skills? That is, does the net developing those skills help it to perform better in that training process, get reinforced, etc.? (3) Is the model big enough that it could in principle come to develop those skills? (4) Is the model trained long enough that it likely will in practice come to have those skills?

The crystal nights story is pretty good but it stabbed my willing suspension of disbelief right in whatever is its vital organ when the Phites casually and instantly begin manipulating matter in the outside world at the speed of their simulation. Real physical experimentation has a lot of trial-and-error processes as you try to get your parameters to the optimal position - no amount of a priori explicit knowledge will enable you to avoid slowly building up your tacit knowledge.

One man's a priori is another man's a posteriori, one might say; there are many places one can acquire informative priors... Learning 'tacit knowledge' can be so fast as to look instantaneous. An example here would be OA's Dactyl hand: it learns robotic hand manipulation in silico, using merely a model simulating physics, with a lot of randomization of settings to teach it to adapt on the fly, to whatever new model it finds itself in. This enables it to, without ever once training on an actual robot hand (only simulated ones), successfully run on an actual robot hand after seconds of adaptation. Another example might be PILCO: it can learn your standard "Cartpole" task within just a few trials by carefully building a Bayesian model and picking maximally informative experiments to run. (Cartpole is quite difficult for a human, incidentally, there's an installation of one in the SF Exploratorium, and I just had to try it out once I recognized it. My sample-efficiency was not better than PILCO.) Because the Phites have all that computation and observations of the real world, they too can do similar tricks, and who knows what else we haven't thought of.

Here is the reason I am skeptical as to the outcome.

Hear me out a little bit.  Suppose you can in fact build a 'brain like' model.  Except, the brain is not one gigantic repeating neural network, it has many distinct regions where nature has made the rules slightly different for a reason.  Nature can't actually encode too much complexity by default as there is so little space in genomes, but it obviously encodes quite a bit or we wouldn't see complex starter instincts for living beings.

But we just have 12 OOMs of compute, we don't hav...

2Daniel Kokotajlo2y
Thanks for the pushback! I agree that if we are trying to copy the human brain, it's not clear that +12 OOMs would be enough. I had my Neuromorph project to illustrate this. However, (a) this might not be as hard as it sounds [https://www.lesswrong.com/posts/HhWhaSzQr6xmBki8F/birds-brains-planes-and-ai-against-appeals-to-the-complexity], and (b) there are other ways to TAI besides trying to copy the human brain.
2Gerald Monroe2y
Oh I don't think we need 12 OOMs. Maybe 2. I don't think most of the electrical details have any net effect you can't model a simpler and equally good way. I was pointing out that the brain is a system, your argument is like saying if we had the hardware for a game console we would therefore get the benefit of the most amazing games the hardware can support. This isn't true, once we have sufficiently capable hardware someone will have to build up algorithms that exhibit intelligence one layer at a time. Well, starting with existing work.

When it comes to training a neural net, both training the neural net and then actually running it costs compute. Are those increased costs similar between different methods?

2Daniel Kokotajlo1y
IDK, probably not? I didn't try to estimate inference costs.

Very minor thing but I was confused for a while when you say end of 2020, I thought of it as the year instead of the decade (2020s).

2Daniel Kokotajlo6mo
It was the year, your original interpretation was correct. Key thing to notice: The question is asking about a hypothetical situation in which we magically get lots more compute 5 years before 2020, and then asks what happens by end of 2020 in that hypothetical.
1zeshen6mo
Ah that's clear, thanks! I must've overlooked the "In 2016" right at the top of the post.