Astral Codex Ten

Moore’s Law

70 min read

Moore’s Law

Article

Moore’s Law is a recurring concept in the Astral Codex Ten archive, appearing 6 times across 6 issues between May 14, 2021 and February 12, 2026. The archive places it in contexts such as “Back when Moore’s Law was just Moore’s Prediction”; “Some formulations of Moore’s Law suggest it halves every eighteen months”; “some sort of beautiful Moore’s-Law-esque rule”. It most often appears alongside AGI, Ajeya, Ajeya Cotra.

Metadata

  • Category: Concepts
  • Mention count: 6
  • Issue count: 6
  • First seen: May 14, 2021
  • Last seen: February 12, 2026

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Your Book Review: Addiction By Design
May 14, 2021 · Original source
Back when Moore’s Law was just Moore’s Prediction, slot machines were mechanical devices. The player would pull on a mechanical lever, which caused reels to spin. The reels would eventually slow down and then stop. The symbols in the middle of the screen when the reels stopped dictated whether the player won or lost.
Inline links: Moore’s Law
Biological Anchors: A Trick That Might Or Might Not Work
February 23, 2022 · Original source
SS Great Eastern, the extreme outlier large steamship from 1858. This has become sort of a mascot for quantitative technological progress forecasters. What is this scientist’s error? The big one is thinking that spaceship progress depends on some easily-measured quantity (size) instead of on fundamental advances (eg figuring out how rockets work). You can make the same accusation against Ajeya et al: you can have all the FLOPs in the world, but if you don’t understand how to make a machine think, your AI will be, well, a flop. Ajeya discusses this a bit on page 143 of her report. There is some sense in which FLOPs and knowing-what-you’re-doing trade of against each other. If you have literally no idea what you’re doing, you can sort of kind of re-run evolution until it comes up with something that looks good. If things are somehow even worse than that, you could always run AIXI, a hypothetical AI design guaranteed to get excellent results as long as you have infinite computation. You could run a Go engine by searching the entire branching tree structure of Go - you shouldn’t, and it would take a zillion times more compute than exists in the entire world, but you could. So in some sense what you’re doing, when you’re figuring out what you’re doing, is coming up with ways to do already-possible things more efficiently. But that’s just algorithmic progress, which Ajeya has already baked into her model. (our Victorian scientist: “As a reductio ad absurdum, you could always stand the ship on its end, and then climb up it to reach space. We’re just trying to make ships that are more efficient than that.”) Part II: Biology-Inspired AI Timelines: The Trick That Never Works Eliezer Yudkowsky presents a more subtle version of these kinds of objection in an essay called Biology-Inspired AI Timelines: The Trick That Never Works, published December 2021. Ajeya’s report is a 169-page collection of equations, graphs, and modeling assumptions. Yudkowsky’s rebuttal is a fictional dialogue between himself, younger versions of himself, famous AI scientists, and other bit players. At one point, a character called “Humbali” shows up begging Yudkowsky to be more humble, and Yudkowsky defeats him with devastating counterarguments. Still, he did found the field, so I guess everyone has to listen to him. He starts: in 1988, famous AI scientist Hans Moravec predicted human-level AI by 2010. He was using the same methodology as Ajeya: extrapolate how quickly processing power would grow (in FLOP/S), and see when it would match some estimate of the human brain. Moravec got the processing power almost exactly right (it hit his 2010 projection in 2008) and his human brain estimate pretty close (he says 10^13 FLOP/S, Ajeya says 10^15, this 2 OOM difference only delays things a few years), yet there was not human-level AI in 2010. What happened? Ajeya's answer could be: Moravec didn't realize that, in the modern ML paradigm, any given size of program requires a much bigger program to train. Ajeya, who has a 35-year advantage on Moravec, estimates approximately the same power for the finished program (10^16 vs. 10^13 FLOP/S) but says that training the 10^16 FLOP/S program will require 10^33ish FLOPs. Eliezer agrees as far as it goes, but says this points to a much deeper failure mode, which was that Moravec had no idea what he was doing. He was assuming processing power of human brain = processing power of computer necessary for AGI. Why? The human brain consumes around 20 watts of power. Can we thereby conclude that an AGI should consume around 20 watts of power, and that, when technology advances to the point of being able to supply around 20 watts of power to computers, we'll get AGI? […] You say that AIs consume energy in a very different way from brains? Well, they'll also consume computations in a very different way from brains! The only difference between these two cases is that you know something about how humans eat food and break it down in their stomachs and convert it into ATP that gets consumed by neurons to pump ions back out of dendrites and axons, while computer chips consume electricity whose flow gets interrupted by transistors to transmit information. Since you know anything whatsoever about how AGIs and humans consume energy, you can see that the consumption is so vastly different as to obviate all comparisons entirely. You are ignorant of how the brain consumes computation, you are ignorant of how the first AGIs built would consume computation, but "an unknown key does not open an unknown lock" and these two ignorant distributions should not assert much internal correlation between them. Cars don’t move by contracting their leg muscles and planes don’t fly by flapping their wings like birds. Telescopes do form images the same way as the lenses in our eyes, but differ by so many orders of magnitude in every important way that they defy comparison. Why should AI be different? You have to use some specific algorithm when you’re creating AI; why should we expect it to be anywhere near the same efficiency as the ones Nature uses in our brains? The same is true for arguments from evolution, eg Ajeya’s Evolutionary Anchor, ie “it took evolution 10^43 FLOPs of computation to evolve the human brain so maybe that will be the training cost”. AI scientists sitting in labs trying to figure things out, and nematodes getting eaten by other nematodes, are such different methods for designing things that it’s crazy to use one as an estimate for the other. Algorithmic Progress vs. Algorithmic Paradigm Shifts This post is a dialogue, so (Eliezer’s hypothetical model of) OpenPhil gets a chance to respond. They object: this is why we put a term for algorithmic progress in our model. The model isn’t very sensitive to changes in that term. If you want you can set it to some kind of crazy high value and see what happens, but you can’t say we didn’t consider it. OpenPhil: We did already consider that and try to take it into account: our model already includes a parameter for how algorithmic progress reduces hardware requirements. It's not easy to graph as exactly as Moore's Law, as you say, but our best-guess estimate is that compute costs halve every 2-3 years […] Eliezer: The makers of AGI aren't going to be doing 10,000,000,000,000 rounds of gradient descent, on entire brain-sized 300,000,000,000,000-parameter models, algorithmically faster than today. They're going to get to AGI via some route that you don't know how to take, at least if it happens in 2040. If it happens in 2025, it may be via a route that some modern researchers do know how to take, but in this case, of course, your model was also wrong. They're not going to be taking your default-imagined approach algorithmically faster, they're going to be taking an algorithmically different approach that eats computing power in a different way than you imagine it being consumed. OpenPhil: Shouldn't that just be folded into our estimate of how the computation required to accomplish a fixed task decreases by half every 2-3 years due to better algorithms? Eliezer: Backtesting this viewpoint on the previous history of computer science, it seems to me to assert that it should be possible to: Train a pre-Transformer RNN/CNN-based model, not using any other techniques invented after 2017, to GPT-2 levels of performance, using only around 2x as much compute as GPT-2;
Inline links: AIXI, Biology-Inspired AI Timelines: The Trick That Never Works
The Japanese canopy plant. I think it is very pretty, but probably low prettiness per megabyte of DNA. I think Ajeya would answer that she’s debating orders of magnitude here, and each of these weird things costs only a few OOMs and probably they all even out. That still leaves the question of why she thinks this approach is interesting at all, to which she answers that: The motivating intuition is that evolution performed a search over a space of small, compact genomes which coded for large brains rather than directly searching over the much larger space of all possible large brains, and human researchers may be able to compete with evolution on this axis. So maybe instead of having to figure out how to generate a brain per se, you figure out how to generate some short(er) program that can output a brain? But this would be very different from how ML works now. Also, you need to give each short program the chance to unfold into a brain before you can evaluate it, which evolution has time for but we probably don’t. Ajeya sort of mentions these problems and counters with an argument that maybe you could think of the genome as a reinforcement learner with a long horizon. I don’t quite follow this but it sounds like the sort of thing that almost might make sense. Anyway, when you apply the scaling laws to a 7.5*10^8 parameter genome and penalize it for a long horizon, you get about 10^33 FLOPs, which is weirdly similar to some of the other estimates. So now we have six different training cost estimates. First, neural nets with short, medium, and long horizons, which are 10^30, 10^33, and 10^36 FLOPs, respectively. Next, the amount of training data in a human lifetime - 10^24 FLOs - and in all of evolutionary history - 10^41 FLOPs. And finally, this weird genome thing, which is 10^33 FLOPs. An optimist might say “Well, our lowest estimate is 10^24 FLOPs, our highest is 10^41 FLOPs, those sound like kind of similar numbers, at least there’s no “5 FLOPs” or “10^9999 FLOPs” in there. A pessimist might say “The difference between 10^24 and 10^41 is seventeen orders of magnitude, ie a factor of 100,000,000,000,000,000 times. This barely constrains our expectations at all!” Before we decide who to trust, let’s remember that we’re still only at Step 2 of our eight step Methodology, and continue. How Do We Adjust For Algorithmic Progress? So today, in 2022 (or in 2020 when this was written, or whenever), assume it would take about 10^33 FLOs to train a human-level AI. But technology constantly advances. Maybe we’ll discover ways to train AIs faster, or run AIs more efficiently, or something like that. How does that factor into our estimate? Ajeya draws on Hernandez & Brown’s Measuring The Algorithmic Efficiency Of Neural Networks. They look at how many FLOPs it took to train various image recognition AIs to an equivalent level of performance between 2012 and 2019, and find that over those seven years it decreased by a factor of 44x, ie training efficiency doubles every sixteen months! Ajeya assumes a doubling time slightly longer than that, because it’s easier to make progress in simple well-understood fields like image recognition than in the novel task of human-level AI. She chooses a doubling time of “merely” 2 - 3 years. If training efficiency doubles every 2-3 years, it would dectuple in about 10 years. So although it might take 10^33 FLOPs to train a human level AI today, in ten years or so it may take only 10^32, in twenty years 10^31, and so on. When Will Anyone Have Enough Computational Resources To Train A Human-Level AI? In 2020, AI researchers could buy computational resources at about $1 for 10^17 FLOPs. That means the 10^33 FLOPs you’d need to train a human-level AI would cost $10^16, ie ten quadrillion dollars. This is about twenty times more money than exists in the entire world. But compute costs fall quickly. Some formulations of Moore’s Law suggest it halves every eighteen months. These no longer seem to hold exactly, but it does seem to be halving maybe once every 2.5 years. The exact number is kind of controversial: Ajeya admits it’s been more like once every 3-4 years lately, but she heard good things about some upcoming chips and predicted it might revert back to the longer-term faster trend (it’s been two years now, some new chips have come out, and this prediction is looking pretty good). So as time goes on, algorithmic progress will cut the cost of training (in FLOPs), and hardware progress will also cut the cost of FLOPs (in dollars). So training will become gradually more affordable as time goes on. Once it reaches a cost somebody is willing to pay, they’ll buy human-level AI, and then that will be the year human-level AI happens. What is the cost that somebody (company? government? billionaire?) is willing to pay for human-level AI? The most expensive AI training in history was AlphaStar, a DeepMind project that spent over $1 million to train an AI to play StarCraft (in their defense, it won). But people have been pouring more and more money into AI lately: Source here. This is about compute rather than cost, but most of the increase seen here has been companies willing to pay for more compute over time, rather than algorithmic or hardware progress. The StarCraft AI was kind of a vanity project, or science for science’s sake, or whatever you want to call it. But AI is starting to become profitable, and human-level AI would be very profitable. Who knows how much companies will be willing to pay in the future? Ajeya extrapolates the line on the graph forward to 2025 and gets $1 billion. This is starting to sound kind of absurd - the entire company OpenAI was founded with $1 billion in venture capital, it seems like a lot to expect them to spend more than $1 billion on a single training run. So Ajeya backs off from this after 2025 and predicts a “two year doubling time”. This is not much of a concession. It still means that in 2040 someone might be spending $100 billion to train one AI. Is this at all plausible? At the height of the Manhattan Project, the US was investing about 0.5% of its GDP into the effort; a similar investment today would be worth $100 billion. And we’re about twice as rich as 2000, so 2040 might be twice as rich as we are. At that point, $100 billion for training an AI is within reach of Google and maybe a few individual billionaires (though it would still require most or all of their fortune). Ajeya creates a complicated function to assess how much money people will be willing to pay on giant AI projects per year. This looks like an upward-sloping curve. The line representing the likely cost of training a human-level AI looks like a downward sloping curve. At some point, those two curves meet, representing when human-level AI will first be trained. So When Will We Get Human-Level AI? The report gives a long distribution of dates based on weights assigned to the six different models, each of which has really wide confidence intervals and options for adjusting the mean and variance based on your assumptions. But the median of all of that is 10% chance by 2031, 50% chance by 2052, and almost 80% chance by 2100. Ajeya takes her six models and decides to weigh them like so, based on how plausible she thinks each one is: 20% neural net, short horizon 30% neural net, medium horizon 15% neural net, long horizon 5% human lifetime as training data 10% evolutionary history as training data 10% genome as parameter number She ends up with this: How Sensitive Is This To Changes In Assumptions? She very helpfully gives us a Colab notebook and Google spreadsheet to play around with. The notebook lets you change some of the more detailed parameters of the individual models, and the spreadsheet lets you change the big picture. I leave the notebook to people more dedicated to forecasting than I am, and will talk about the spreadsheet here. If you’re following along at home, the default spreadsheet won’t reflect Ajeya’s findings until you fill in the table in the bottom left like so: Great. Now that we’ve got that, let’s try changing some stuff. I like the human childhood training data argument (Lifetime Anchor) more than Ajeya does, and I like the size-of-the-genome argument less. I’m going to change the weights to 20-20-0-20-20-20. Also, Ajeya thinks that someone might be willing to spend 1% of national GDP on training AIs, but that sounds really high to me, so I’m going to down to 0.1%. Also, Ajeya’s estimate of 3% GDP growth sounds high for the sort of industrialized nations who might do AI research, I’m going to lower it to 2%. Since I’m feeling mistrustful today, let’s use the Hernandez&Brown estimate for compute halving (1.5 years) in place of Ajeya’s ad hoc adjustments. And let’s use the current compute halving time (3.5 years) instead of Ajeya’s overly rosy version (2.5 years). All these changes… …don’t really do much. The median goes from 2052 to about 2065. Four of the models give results between 2030 and 2070. The last two, Neural Net With Long Horizon and Evolution, suggest probably no AI this century (although Neural Net With Long Horizon does think there’s a 40% chance by 2100). Ajeya doesn’t really like either of these models and they’re not heavily weighted in her main result. Does The Truth Point To Itself? Back up a second. Here’s something that makes me kind of nervous. Most of Ajeya’s numbers are kind of made up, with several order-of-magnitude error bars and simplifying assumptions like “all animals are nematodes”. For a single parameter, we get estimates spanning seventeen different orders of magnitude: the upper bound is one hundred quadrillion times the lower bound. And yet four of the six models, including two genuinely exotic ones, manage to get dates within twenty years of 2050. And 2050 is also the date everyone else focuses on. Here’s the prediction-market-like site Metaculus: Their distribution looks a lot like Ajeya’s, and even has the same median, 2052 (though forecasters could have read Ajeya’s report). Katja Grace et al surveyed 352 AI experts, and they gave a median estimate of 2062 for an AI that could “outperform humans at all tasks” (though with many caveats and high sensitivity to question framing). This was before Ajeya’s report, so they definitely didn’t read it. So lots of Ajeya’s different methods and lots of other people presumably using different methodologies or no methodology at all, all converge on this same idea of 2050 give or take a decade or two. An optimist might say “The truth points to itself! There are 371 known proofs of the Pythagorean Theorem, and they all end up in the same place. That’s because no matter what methodology you use, if you use it well enough you get to the correct answer.” A pessimist might be more suspicious; we’ll return to this part later. FLOPS Alone Turn The Wheel Of History One more question: what if this is all bullshit? What if it’s an utterly useless total garbage steaming pile of grade A crap? Imagine a scientist in Victorian Britain, speculating on when humankind might invent ships that travel through space. He finds a natural anchor: the moon travels through space! He can observe things about the moon: for example, it is 220 miles in diameter (give or take an order of magnitude). So when humankind invents ships that are 220 miles in diameter, they can travel through space! Ships have certainly grown in size tremendously, from primitive kayaks to Roman triremes to Spanish galleons to the great ocean liners of the (Victorian) present. The AI forecasting organization AI Impacts actually has a whole report on historical ship size trends to prove an unrelated point about technological progress, so I didn’t even have to make this graph up. Suppose our Victorian scientist lived in 1858, right when the Great Eastern was launched. The trend line for ship size crossed 100m around 1843, and 200m in 1858, so doubling time is 15 years - but perhaps they notice this is going to be an outlier, so let’s round up a bit and say 18 years. The (one order of magnitude off estimate for the size of the) Moon is 350,000m, so you’d need ships to scale up by 350,000/200 = 1,750x before they’re as big as the Moon. That’s about 10.8 doublings, and a doubling time is 18 years, so we’ll get spaceships in . . . 2052 exactly. (fudging numbers to land where you want is actually fun and easy) SS Great Eastern, the extreme outlier large steamship from 1858. This has become sort of a mascot for quantitative technological progress forecasters. What is this scientist’s error? The big one is thinking that spaceship progress depends on some easily-measured quantity (size) instead of on fundamental advances (eg figuring out how rockets work). You can make the same accusation against Ajeya et al: you can have all the FLOPs in the world, but if you don’t understand how to make a machine think, your AI will be, well, a flop. Ajeya discusses this a bit on page 143 of her report. There is some sense in which FLOPs and knowing-what-you’re-doing trade of against each other. If you have literally no idea what you’re doing, you can sort of kind of re-run evolution until it comes up with something that looks good. If things are somehow even worse than that, you could always run AIXI, a hypothetical AI design guaranteed to get excellent results as long as you have infinite computation. You could run a Go engine by searching the entire branching tree structure of Go - you shouldn’t, and it would take a zillion times more compute than exists in the entire world, but you could. So in some sense what you’re doing, when you’re figuring out what you’re doing, is coming up with ways to do already-possible things more efficiently. But that’s just algorithmic progress, which Ajeya has already baked into her model. (our Victorian scientist: “As a reductio ad absurdum, you could always stand the ship on its end, and then climb up it to reach space. We’re just trying to make ships that are more efficient than that.”) Part II: Biology-Inspired AI Timelines: The Trick That Never Works Eliezer Yudkowsky presents a more subtle version of these kinds of objection in an essay called Biology-Inspired AI Timelines: The Trick That Never Works, published December 2021. Ajeya’s report is a 169-page collection of equations, graphs, and modeling assumptions. Yudkowsky’s rebuttal is a fictional dialogue between himself, younger versions of himself, famous AI scientists, and other bit players. At one point, a character called “Humbali” shows up begging Yudkowsky to be more humble, and Yudkowsky defeats him with devastating counterarguments. Still, he did found the field, so I guess everyone has to listen to him. He starts: in 1988, famous AI scientist Hans Moravec predicted human-level AI by 2010. He was using the same methodology as Ajeya: extrapolate how quickly processing power would grow (in FLOP/S), and see when it would match some estimate of the human brain. Moravec got the processing power almost exactly right (it hit his 2010 projection in 2008) and his human brain estimate pretty close (he says 10^13 FLOP/S, Ajeya says 10^15, this 2 OOM difference only delays things a few years), yet there was not human-level AI in 2010. What happened? Ajeya's answer could be: Moravec didn't realize that, in the modern ML paradigm, any given size of program requires a much bigger program to train. Ajeya, who has a 35-year advantage on Moravec, estimates approximately the same power for the finished program (10^16 vs. 10^13 FLOP/S) but says that training the 10^16 FLOP/S program will require 10^33ish FLOPs. Eliezer agrees as far as it goes, but says this points to a much deeper failure mode, which was that Moravec had no idea what he was doing. He was assuming processing power of human brain = processing power of computer necessary for AGI. Why? The human brain consumes around 20 watts of power. Can we thereby conclude that an AGI should consume around 20 watts of power, and that, when technology advances to the point of being able to supply around 20 watts of power to computers, we'll get AGI? […] You say that AIs consume energy in a very different way from brains? Well, they'll also consume computations in a very different way from brains! The only difference between these two cases is that you know something about how humans eat food and break it down in their stomachs and convert it into ATP that gets consumed by neurons to pump ions back out of dendrites and axons, while computer chips consume electricity whose flow gets interrupted by transistors to transmit information. Since you know anything whatsoever about how AGIs and humans consume energy, you can see that the consumption is so vastly different as to obviate all comparisons entirely. You are ignorant of how the brain consumes computation, you are ignorant of how the first AGIs built would consume computation, but "an unknown key does not open an unknown lock" and these two ignorant distributions should not assert much internal correlation between them. Cars don’t move by contracting their leg muscles and planes don’t fly by flapping their wings like birds. Telescopes do form images the same way as the lenses in our eyes, but differ by so many orders of magnitude in every important way that they defy comparison. Why should AI be different? You have to use some specific algorithm when you’re creating AI; why should we expect it to be anywhere near the same efficiency as the ones Nature uses in our brains? The same is true for arguments from evolution, eg Ajeya’s Evolutionary Anchor, ie “it took evolution 10^43 FLOPs of computation to evolve the human brain so maybe that will be the training cost”. AI scientists sitting in labs trying to figure things out, and nematodes getting eaten by other nematodes, are such different methods for designing things that it’s crazy to use one as an estimate for the other. Algorithmic Progress vs. Algorithmic Paradigm Shifts This post is a dialogue, so (Eliezer’s hypothetical model of) OpenPhil gets a chance to respond. They object: this is why we put a term for algorithmic progress in our model. The model isn’t very sensitive to changes in that term. If you want you can set it to some kind of crazy high value and see what happens, but you can’t say we didn’t consider it. OpenPhil: We did already consider that and try to take it into account: our model already includes a parameter for how algorithmic progress reduces hardware requirements. It's not easy to graph as exactly as Moore's Law, as you say, but our best-guess estimate is that compute costs halve every 2-3 years […] Eliezer: The makers of AGI aren't going to be doing 10,000,000,000,000 rounds of gradient descent, on entire brain-sized 300,000,000,000,000-parameter models, algorithmically faster than today. They're going to get to AGI via some route that you don't know how to take, at least if it happens in 2040. If it happens in 2025, it may be via a route that some modern researchers do know how to take, but in this case, of course, your model was also wrong. They're not going to be taking your default-imagined approach algorithmically faster, they're going to be taking an algorithmically different approach that eats computing power in a different way than you imagine it being consumed. OpenPhil: Shouldn't that just be folded into our estimate of how the computation required to accomplish a fixed task decreases by half every 2-3 years due to better algorithms? Eliezer: Backtesting this viewpoint on the previous history of computer science, it seems to me to assert that it should be possible to: Train a pre-Transformer RNN/CNN-based model, not using any other techniques invented after 2017, to GPT-2 levels of performance, using only around 2x as much compute as GPT-2;
Inline links: Measuring The Algorithmic Efficiency Of Neural Networks, https://substackcdn.com/image/fetch/$s_!dX1J!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9496f1f-ec6c-41a2-8c2e-27f09da22097_1280x759.png, here, https://substackcdn.com/image/fetch/$s_!LnC0!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F62d647ff-58ed-4e9a-9f1a-7febf5859249_1152x842.png, Colab notebook, Google spreadsheet, https://substackcdn.com/image/fetch/$s_!BND-!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F622bac28-eaa6-40b5-b93b-695952966ef7_744x324.png, https://substackcdn.com/image/fetch/$s_!lbos!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F7d5c2306-a123-4903-adb9-d961d56ebfb5_1152x842.png, Metaculus, https://substackcdn.com/image/fetch/$s_!SMnF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F807f66de-8c5c-4423-b293-ca92b5b64053_763x360.png, surveyed 352 AI experts, https://substackcdn.com/image/fetch/$s_!JxQ5!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fceba6aa0-dbde-41ca-805e-01af4fac9324_769x336.png, a whole report on historical ship size trends, https://substackcdn.com/image/fetch/$s_!PRDj!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fde3d97f4-afca-45c4-9ed2-521cd25041df_460x262.jpeg, AIXI, Biology-Inspired AI Timelines: The Trick That Never Works
Play pro-level Go using 8-16 times as much computing power as AlphaGo, but only 2006 levels of technology. For reference, recall that in 2006, Hinton and Salakhutdinov were just starting to publish that, by training multiple layers of Restricted Boltzmann machines and then unrolling them into a "deep" neural network, you could get an initialization for the network weights that would avoid the problem of vanishing and exploding gradients and activations. At least so long as you didn't try to stack too many layers, like a dozen layers or something ridiculous like that. This being the point that kicked off the entire deep-learning revolution. Your model apparently suggests that we have gotten around 50 times more efficient at turning computation into intelligence since that time; so, we should be able to replicate any modern feat of deep learning performed in 2021, using techniques from before deep learning and around fifty times as much computing power. OpenPhil: No, that's totally not what our viewpoint says when you backfit it to past reality. Our model does a great job of retrodicting past reality. Eliezer: How so? OpenPhil: <Eliezer cannot predict what they will say here.> I think the argument here is that OpenPhil is accounting for normal scientific progress in algorithms, but not for paradigm shifts. Directional Error These are the two arguments Eliezer makes against OpenPhil that I find most persuasive. First, that you shouldn’t be using biological anchors at all. Second, that unpredictable paradigm shifts are more realistic than gradual algorithmic progress. These mostly add uncertainty to OpenPhil’s model, but Eliezer ends his essay making a stronger argument: he thinks OpenPhil is directionally wrong, and AI will come earlier than they think. Mostly this is the paradigm argument again. Five years from now, there could be a paradigm shift that makes AI much easier to build. It’s happened before; from GOFAI’s pre-programmed logical rules to Deep Blue’s tree searches to the sorts of Big Data methods that won the Netflix Prize to modern deep learning. Instead of just extrapolating deep learning scaling thirty years out, OpenPhil should be worried about the next big idea. Hypothetical OpenPhil retorts that this is a double-edged sword. Maybe the deep learning paradigm can’t produce AGI, and we’ll have to wait decades or centuries for someone to have the right insight. Or maybe the new paradigm you need for AGI will take more compute than deep learning, in the same way deep learning takes more compute than whatever Moravec was imagining. This is a pretty strong response, since it would have been true for every previous forecaster: remember, Moravec erred in thinking AI would come too soon, not too late. So although Eliezer is taking the cheap shot of saying OpenPhil’s estimate will be wrong just as everyone else’s was wrong before, he’s also giving himself the much harder case of arguing it might be wrong in the opposite direction as all its predecessors. Eliezer takes this objection seriously, but feels like on balance probably new paradigms will speed up AI rather than slow it down. Here he grudgingly and with suitable embarrassment does try to make an object-level semi-biological-anchors-related argument: Moravec was wrong because he ignored the training phase. And the proper anchor for the training phase is somewhere between evolution and a human childhood, where evolution represents “blind chance eventually finding good things” and human childhood represents “an intelligent cognitive engine trying to squeeze as much data out of experience as possible”. And part of what he expects paradigm shifts to do is to move from more evolutionary processes to more childhood-like processes, and that’s a net gain in efficiency. So he still thinks OpenPhil’s methods are more likely to overestimate the amount of time until AGI rather than underestimate it. What Moore’s Law Giveth, Platt’s Law Taketh Away Eliezer’s other argument is kind of a low blow: he refers to Platt’s Law Of AI Forecasting: “any AI forecast will put strong AI thirty years out from when the forecast is made.” This isn’t exact. Hans Moravec, writing in 1988, said 2010 - so 22 years. Ray Kurzweil, writing in 2001, said 2023 - another 22 years. Vernor Vinge, in a 1993 speech, said 2023, and that was exactly 30 years, but Vinge knew about Platt’s Law and might have been joking. The point is: OpenPhil wrote a report in 2020 that predicted strong AI in 2052, isn’t that kind of suspicious? I’d previously mentioned it as a plus that Ajeya got around the same year everyone else got. The forecasters on Metaculus. The experts surveyed in Grace et al. Lots of other smart experts with clever models. But what if all of these experts and models and analyses are just fudging the numbers for the same Platt’s-Law-related reasons? Hypothetical OpenPhil is BTFO: OpenPhil: That part about Charles Platt's generalization is interesting, but just because we unwittingly chose literally exactly the median that Platt predicted people would always choose in consistent error, that doesn't justify dismissing our work, right? We could have used a completely valid method of estimation which would have pointed to 2050 no matter which year it was tried in, and, by sheer coincidence, have first written that up in 2020. In fact, we try to show in the report that the same methodology, evaluated in earlier years, would also have pointed to around 2050 - Eliezer: Look, people keep trying this. It's never worked. It's never going to work. 2 years before the end of the world, there'll be another published biologically inspired estimate showing that AGI is 30 years away and it will be exactly as informative then as it is now. I'd love to know the timelines too, but you're not going to get the answer you want until right before the end of the world, and maybe not even then unless you're paying very close attention. Timing this stuff is just plain hard. Part III: Responses And Commentary Response 1: Less Wrong Comments Less Wrong is a site founded by Eliezer Yudkowsky for Eliezer Yudkowsky fans who wanted to discuss Eliezer Yudkowsky’s ideas. So, for whatever it’s worth - the comments on his essay were pretty negative. Carl Shulman, an independent researcher with links to both OpenPhil and MIRI (Eliezer’s org), writes the top-voted comment. He works from a model where there is hardware progress, software progress downstream of hardware progress, and independent (ie unrelated to algorithms) software progress, and where the first two make up most progress on the margin. Researchers generally develop new paradigms once they have enough compute available to tinker with them. Progress in AI has largely been a function of increasing compute, human software research efforts, and serial time/steps. Throwing more compute at researchers has improved performance both directly and indirectly (e.g. by enabling more experiments, refining evaluation functions in chess, training neural networks, or making algorithms that work best with large compute more attractive). Historically compute has grown by many orders of magnitude, while human labor applied to AI and supporting software by only a few. And on plausible decompositions of progress (allowing for adjustment of software to current hardware and vice versa), hardware growth accounts for more of the progress over time than human labor input growth. So if you're going to use an AI production function for tech forecasting based on inputs (which do relatively OK by the standards tech forecasting), it's best to use all of compute, labor, and time, but it makes sense for compute to have pride of place and take in more modeling effort and attention, since it's the biggest source of change (particularly when including software gains downstream of hardware technology and expenditures). […] A perfectly correlated time series of compute and labor would not let us say which had the larger marginal contribution, but we have resources to get at that, which I was referring to with 'plausible decompositions.' This includes experiments with old and new software and hardware, like the chess ones Paul recently commissioned, and studies by AI Impacts, OpenAI, and Neil Thompson. There are AI scaling experiments, and observations of the results of shocks like the end of Dennard scaling, the availability of GPGPU computing, and Besiroglu's data on the relative predictive power of computer and labor in individual papers and subfields. In different ways those tend to put hardware as driving more log improvement than software (with both contributing), particularly if we consider software innovations downstream of hardware changes. Vanessa Kosoy makes the obvious objection, which echoes a comment of Eliezer’s in the dialogue above: I'm confused how can this pass some obvious tests. For example, do you claim that alpha-beta pruning can match AlphaGo given some not-crazy advantage in compute? Do you claim that SVMs can do SOTA image classification with not-crazy advantage in compute (or with any amount of compute with the same training data)? Can Eliza-style chatbots compete with GPT3 however we scale them up? Mark Xu answers: My model is something like: For any given algorithm, e.g. SVMs, AlphaGo, alpha-beta pruning, convnets, etc., there is an "effective compute regime" where dumping more compute makes them better. If you go above this regime, you get steep diminishing marginal returns.
Inline links: normal scientific progress in algorithms, but not for paradigm shifts, Platt’s Law Of AI Forecasting, the comments, Paul recently commissioned, AI Impacts, OpenAI, Neil Thompson, Besiroglu's, Vanessa Kosoy, Mark Xu
Yudkowsky Contra Christiano On AI Takeoff Speeds
April 04, 2022 · Original source
Chess AI performance over time. Why does this matter? If there’s a slow takeoff (ie gradual exponential curve), it will become obvious that some kind of terrifying transformative AI revolution is happening, before the situation gets apocalyptic. There will be time to prepare, to test slightly-below-human AIs and see how they respond, to get governments and other stakeholders on board. We don’t have to get every single thing right ahead of time. On the other hand, because this is proceeding along the usual channels, it will be the usual variety of muddled and hard-to-control. With the exception of a few big actors like the US and Chinese government, and maybe the biggest corporations like Google, the outcome will be determined less by any one agent, and more by the usual multi-agent dynamics of political and economic competition. There will be lots of opportunities to affect things, but no real locus of control to do the affecting. If there’s a fast takeoff (ie sudden FOOM), there won’t be much warning. Conventional wisdom will still say that transformative AI is thirty years away. All the necessary pieces (ie AI alignment theory) will have to be ready ahead of time, prepared blindly without any experimental trial-and-error, to load into the AI as soon as it exists. On the plus side, a single actor (whoever has this first AI) will have complete control over the process. If this actor is smart (and presumably they’re a little smart, or they wouldn’t be the first team to invent transformative AI), they can do everything right without going through the usual government-lobbying channels. So the slower a takeoff you expect, the less you should be focusing on getting every technical detail right ahead of time, and the more you should be working on building the capacity to steer government and corporate policy to direct an incoming slew of new technologies. Yudkowsky Contra Christiano Eliezer counters that although progress may retroactively look gradual and continuous when you know what metric to graph it on, it doesn’t necessarily look that way in real life by the measures that real people care about. (one way to think of this: imagine that an AI’s effective IQ starts at 0.1 points, and triples every year, but that we can only measure this vaguely and indirectly. The year it goes from 5 to 15, you get a paper in a third-tier journal reporting that it seems to be improving on some benchmark. The year it goes from 66 to 200, you get a total transformation of everything in society. But later, once we identify the right metric, it was just the same rate of gradual progress the whole time. ) So Eliezer is much less impressed by the history of previous technologies than Paul is. He’s also skeptical of the “GDP will double in 4 years before it doubles in 1” claim, because of two contingent disagreements and two fundamental disagreements. The first contingent disagreement: government regulations make it hard to deploy imperfect things, and non-trivial to deploy things even after they’re perfect. Eliezer has non-jokingly said he thinks AI might destroy the world before the average person can buy a self-driving car. Why? Because the government has to approve self-driving cars (and can drag its feet on that), but the apocalypse can happen even without government approval. In Paul’s model, sometime long before superintelligence we should have AIs that can drive cars, and that increases GDP and contributes to a general sense that exciting things are going on. Eliezer says: fine, what if that’s true? Who cares if self-driving cars will be practical a few years before the world is destroyed? It’ll take longer than that to lobby the government to allow them on the road. The second contingent disagreement: superintelligent AIs can lie to us. Suppose you have an AI which wants to destroy humanity, whose IQ is doubling every six months. Right now it’s at IQ 200, and it suspects that it would take IQ 800 to build a human-destroying superweapon. Its best strategy is to lie low for a year. If it expects humans would turn it off if they knew how close it was to superweapons, it can pretend to be less intelligent than it really is. The period when AIs are holding back so we don’t discover their true power level looks like a period of lower-than-expected GDP growth - followed by a sudden FOOM once the AI gets its superweapon and doesn’t need to hold back. So even if Paul is conceptually right and fundamental progress proceeds along a nice smooth curve, it might not look to us like a nice smooth curve, because regulations and deceptive AIs could prevent mildly-transformative AI progress from showing up on graphs, but wouldn’t prevent the extreme kind of AI progress that leads to apocalypse. To an outside observer, it would just look like nothing much changed, nothing much changed, nothing much changed, and then suddenly, FOOM. But even aside from this, Eliezer doesn’t think Paul is conceptually right! He thinks that even on the fundamental level, AI progress is going to be discontinuous. It’s like a nuclear bomb. Either you don’t have a nuclear bomb yet, or you do have one and the world is forever transformed. There is a specific moment at which you go from “no nuke” to “nuke” without any kind of “slightly worse nuke” acting as a harbinger. He uses the example of chimps → humans. Evolution has spent hundreds of millions of years evolving brainier and brainier animals (not teleologically, of course, but in practice). For most of those hundreds of millions of years, that meant the animal could have slightly more instincts, or a better memory, or some other change that still stayed within the basic animal paradigm. At the chimp → human transition, we suddenly got tool use, language use, abstract thought, mathematics, swords, guns, nuclear bombs, spaceships, and a bunch of other stuff. The rhesus monkey → chimp transition and the chimp → human transition both involved the same ~quadrupling of neuron number, but the former was pretty boring and the latter unlocked enough new capabilities to easily conquer the world. The GPT-2 → GPT-3 transition involved centupling parameter count. Maybe we will keep centupling parameter count every few years, and most times it will be incremental improvement, and one time it will conquer the world. But even talking about centupling parameter points is giving Paul too much credit. Lots of past inventions didn’t come by quadrupling or centupling something, they came by discovering “the secret sauce”. The Wright brothers (he argues) didn’t make a plane with 4x the wingspan of the last plane that didn’t work, they invented the first plane that could fly at all. The Hiroshima bomb wasn’t some previous bomb but bigger, it was what happened after a lot of scientists spent a long time thinking about a fundamentally different paradigm of bomb-making and brought it to a point where it could work at all. The first transformative AI isn’t going to be GPT-3 with more parameters, it will be what happens after someone discovers how to make machines truly intelligent. (this is the same debate Eliezer had with Ajeya over the Biological Anchors post; have I mentioned that Ajeya and Paul are married?) Fine, Let’s Nitpick The Hell Out Of The Chimps Vs. Humans Example This is where the two of them end up, so let’s follow. Between chimps and humans, there were about seven million years of intermediate steps. These had some human capabilities, but not others. IE homo erectus probably had language, but not mathematics, and in terms of taking over the world it did make it to most of the Old World but was less dominant than moderns. But if we say evolutionary history started 500 million years ago (the Cambrian), and AI history started with the Dartmouth Conference in 1955, then the equivalent of 7 million years of evolutionary history is 1 year of AI history. In the very very unlikely and forced comparison where evolutionary history and AI history go at the same speed, there will be only about a year between chimp-level and human-level AIs. A chimp-level AI probably can’t double GDP, so this would count as a fast takeoff by Paul’s criterion. But even more than that, chimp → human feels like a discontinuity. It’s not just “animals kept getting smarter for hundreds of millions of years, and then ended up very smart indeed”. That happened for a while, and then all of sudden there was a near-instant phase transition into a totally different way of using intelligence with completely new abilities. If AI worked like this, we would have useful toys and interesting specialists for a few decades, until suddenly someone “got it right”, completed the package that was necessary for “true intelligence”, and then we would have a completely new category of thing. Paul admits this analogy is awkward for his position. He answers: Chimp evolution is not primarily selecting for making and using technology, for doing science, or for facilitating cultural accumulation. The task faced by a chimp is largely independent of the abilities that give humans such a huge fitness advantage. It’s not completely independent—the overlap is the only reason that evolution eventually produces humans—but it’s different enough that we should not be surprised if there are simple changes to chimps that would make them much better at designing technology or doing science or accumulating culture […] So I don’t think the example of evolution tells us much about whether the continuous change story applies to intelligence. This case is potentially missing the key element that drives the continuous change story—optimization for performance. Evolution changes continuously on the narrow metric it is optimizing, but can change extremely rapidly on other metrics. For human technology, features of the technology that aren’t being optimized change rapidly all the time. When humans build AI, they will be optimizing for usefulness, and so progress in usefulness is much more likely to be linear. That is, evolution wasn’t optimizing for tool use/language/intelligence, so we got an “overhang” where chimps could potentially have been very good at these, but evolution never bothered “closing the circuit” and turning those capabilities “on”. After a long time, evolution finally blundered into an area where marginal improvements in these capacities improved fitness, so evolution started improving them and it was easy. Imagine a company which, through some oversight, didn’t have a Sales department. They just sat around designing and manufacturing increasingly brilliant products, but not putting any effort into selling them. Then the CEO remembers they need a Sales department, starts one up, and the company goes from moving near zero units to moving millions of units overnight. It would look like the company had “suddenly” developed a “vast increase in capabilities”. But this is only possible when a CEO who is weirdly unconcerned about profit forgets to do obvious profit-increasing things for many years. This is Paul’s counterargument to the chimp analogy. Evolution isn’t directly concerned about various intellectual skills; it only wants them in the unusual cases where they’ll contribute to fitness on the margin. AI companies will be very concerned about various intellectual skills. If there’s a trivial change that can make their product 10x better, they’ll make it. So AI capabilities will grow in a “well-rounded” way, there won’t be any “overhangs”, and there won’t be any opportunities for a sudden overhang-solving phase transition with associated new-capability development like with chimps → humans. Eliezer answers: Chimps are nearly useless because they're not general, and doing anything on the scale of building a nuclear plant requires mastering so many different nonancestral domains that it's no wonder natural selection didn't happen to separately train any single creature across enough different domains that it had evolved to solve every kind of domain-specific problem involved in solving nuclear physics and chemistry and metallurgy and thermics in order to build the first nuclear plant in advance of any old nuclear plants existing. Humans are general enough that the same braintech selected just for chipping flint handaxes and making water-pouches and outwitting other humans, happened to be general enough that it could scale up to solving all the problems of building a nuclear plant - albeit with some added cognitive tech that didn't require new brainware, and so could happen incredibly fast relative to the generation times for evolutionarily optimized brainware. Now, since neither humans nor chimps were optimized to be "useful" (general), and humans just wandered into a sufficiently general part of the space that it cascaded up to wider generality, we should legit expect the curve of generality to look at least somewhat different if we're optimizing for that. Eg, right now people are trying to optimize for generality with AIs like Mu Zero and GPT-3. In both cases we have a weirdly shallow kind of generality. Neither is as smart or as deeply general as a chimp, but they are respectively better than chimps at a wide variety of Atari games, or a wide variety of problems that can be superposed onto generating typical human text. They are, in a sense, more general than a biological organism at a similar stage of cognitive evolution, with much less complex and architected brains, in virtue of having been trained, not just on wider datasets, but on bigger datasets using gradient-descent memorization of shallower patterns, so they can cover those wide domains while being stupider and lacking some deep aspects of architecture. It is not clear to me that we can go from observations like this, to conclude that there is a dominant mainline probability for how the future clearly ought to go and that this dominant mainline is, "Well, before you get human-level depth and generalization of general intelligence, you get something with 95% depth that covers 80% of the domains for 10% of the pragmatic impact". ...or whatever the concept is here, because this whole conversation is, on my own worldview, being conducted in a shallow way relative to the kind of analysis I did in Intelligence Explosion Microeconomics, where I was like, "here is the historical observation, here is what I think it tells us that puts a lower bound on this input-output curve". Here Eliezer sort of kind of grants Paul’s point that AIs will be optimized for generality in a way chimps aren’t, but points to his previous “Intelligence Explosion Microeconomics” essay to argue that we should expect a fast takeoff anyway. IEM has a lot of stuff in it, but one key point is that instead of using analogies to predict the course of future AI, we should open that black box and try to actually reason about how it will work, in which case we realize that recursive self-improvement common-sensically has to cause an intelligence explosion. I am sort of okay with this, but I feel like a commitment to avoiding analogies should involve not bringing up the chimp-human analogy further, which Eliezer continues to do, quite a lot. I do feel like Paul succeeded in convincing me that we shouldn’t place too much evidential weight on it. The Wimbledon Of Reference Class Tennis “Reference class tennis” is an old rationalist idiom for people throwing analogies back and forth. “AI will be slow, because it’s an economic transition like the Agricultural or Industrial Revolution, and those were slow!” “No, AI will be fast, because it’s an evolutionary step like chimps → humans, and that was fast!” “No, AI will be slow, because it’s an invention, like the computer, and computers were invented piecemeal and required decades of innovation to be useful.” “No, AI will be fast, because it’s an invention, like the nuclear bomb, and nuclear bombs went from impossible to city-killing in a single day.” “No, AI will be slow, because it will be surrounded by a shell-like metallic computer case, which makes it like a turtle, and turtles are slow.” “No, AI will be fast, because it’s dangerous and powerful, like a tiger, and tigers are fast!” And so on. Comparing things to other things is a time-tested way of speculating about them. But there are so many other things to compare to that you can get whatever result you want. This is the failure mode that the term “reference class tennis” was supposed to point to. Both participants in this debate are very smart and trying their hardest to avoid reference-class tennis, but neither entirely succeeds. Eliezer’s preferred classes are Bitcoin (“there wasn't a cryptocurrency developed a year before Bitcoin using 95% of the ideas which did 10% of the transaction volume”), nukes, humans/chimps, the Wright Brothers, AlphaGo (which really was a discontinuous improvement on previous Go engines), and AlphaFold (ditto for proteins). Paul’s preferred classes are the Agricultural and Industrial Revolutions, chess engines (which have gotten better along a gradual, well-behaved curve), all sorts of inventions like computers and ships (likewise), and world GDP. Eliezer already listed most of these in his Intelligence Explosion Microeconomics paper in 2013, and concluded that the space of possible analogies was contradictory enough that we needed to operate at a higher level. Maybe so, but when someone lobs a reference class tennis ball at you, it’s hard to resist the urge to hit it back. Recursive Self-Improvement This is where I think Eliezer most wants to take the discussion. The idea is: once AI is smarter than humans, it can do a superhuman job of developing new AI. In his Microeconomics paper, he writes about an argument he (semi-hypothetically) had with Ray Kurzweil about Moore’s Law. Kurzweil expected Moore’s Law to continue forever, even after the development of superintelligence. Eliezer objects: Suppose we were dealing with minds running a million times as fast as a human, at which rate they could do a year of internal thinking in thirty-one seconds, such that the total subjective time from the birth of Socrates to the death of Turing would pass in 20.9 hours. Do you still think the best estimate for how long it would take them to produce their next generation of computing hardware would be 1.5 orbits of the Earth around the Sun? That is: the fact that it took 1.5 years for transistor density to double isn’t a natural law. It’s pointing to a law that the amount of resources (most notably intelligence) that civilization focused on the transistor-densifying problem equalled the amount it takes to double it every 1.5 years. If some shock drastically changed available resources (by eg speeding up human minds a million times), this would change the resources involved, and the same laws would predict transistor speed doubling in some shorter amount of time (naively 0.000015 years, although realistically at that scale other inputs would dominate). So when Paul derives clean laws of economics showing that things move along slow growth curves, Eliezer asks: why do you think they would keep doing this when one of the discoveries they make along that curve might be “speeding up intelligence a million times”? (Eliezer actually thinks improvements in the quality of intelligence will dominate improvements in speed - AIs will mostly be smarter, not just faster - but speed is a useful example here and we’ll stick with it) Paul answers: Summary of my response: Before there is AI that is great at self-improvement there will be AI that is mediocre at self-improvement. Powerful AI can be used to develop better AI (amongst other things). This will lead to runaway growth. This on its own is not an argument for discontinuity: before we have AI that radically accelerates AI development, the slow takeoff argument suggests we will have AI that significantly accelerates AI development (and before that, slightly accelerates development). That is, an AI is just another, faster step in the hyperbolic growth we are currently experiencing, which corresponds to a further increase in rate but not a discontinuity (or even a discontinuity in rate). The most common argument for recursive self-improvement introducing a new discontinuity seems be: some systems “fizzle out” when they try to design a better AI, generating a few improvements before running out of steam, while others are able to autonomously generate more and more improvements. This is basically the same as the universality argument in a previous section. Eliezer: Oh, come on. That is straight-up not how simple continuous toy models of RSI work. Between a neutron multiplication factor of 0.999 and 1.001 there is a very huge gap in output behavior. Outside of toy models: Over the last 10,000 years we had humans going from mediocre at improving their mental systems to being (barely) able to throw together AI systems, but 10,000 years is the equivalent of an eyeblink in evolutionary time - outside the metaphor, this says, "A month before there is AI that is great at self-improvement, there will be AI that is mediocre at self-improvement." (Or possibly an hour before, if reality is again more extreme along the Eliezer-Hanson axis than Eliezer. But it makes little difference whether it's an hour or a month, given anything like current setups.) This is just pumping hard again on the intuition that says incremental design changes yield smooth output changes, which (the meta-level of the essay informs us wordlessly) is such a strong default that we are entitled to believe it if we can do a good job of weakening the evidence and arguments against it. And the argument is: Before there are systems great at self-improvement, there will be systems mediocre at self-improvement; implicitly: "before" implies "5 years before" not "5 days before"; implicitly: this will correspond to smooth changes in output between the two regimes even though that is not how continuous feedback loops work. I got a bit confused trying to understand the criticality metaphor here. There’s no equivalent of neutron decay, so any AI that can consistently improve its intelligence is “critical” in some sense. Imagine Elon Musk replaces his brain with a Neuralink computer which - aside from having read-write access - exactly matches his current brain in capabilities. Also he becomes immortal. He secludes himself from the world, studying AI and tinkering with his brain’s algorithms. Does he become a superintelligence? I think under the assumptions Paul and Eliezer are using, eventually maybe. After some amount of time he’ll come across a breakthrough he can use to increase his intelligence. Then, armed with that extra intelligence, he’ll be able to pursue more such breakthroughs. However intelligent the AI you’re scared of is, Musk will get there eventually. How long will it take? A good guess might be “years” - Musk starts out as an ordinary human, and ordinary humans are known to take years to make breakthroughs. Suppose it takes Musk one year to come up with a first breakthrough that raises his IQ 1 point. How long will his second breakthrough take? It might take longer, because he has picked the lowest-hanging fruit, and all the other possible breakthroughs are much harder. Or it might take shorter, because he’s slightly smarter than he was before, and maybe some extra intelligence goes a really long way in AI research. The concept of an intelligence explosion seems to assume the second effect dominates the first. This would match the observation that human researchers, who aren’t getting any smarter over time, continue making new discoveries. That suggests the range of possible discoveries at a given intelligence level is pretty vast. Some research finds that the usual pattern in science is constant rate of discovery from exponentially increasing number of researchers, suggesting strong low-hanging fruit effects, but these seem to be overwhelmed by other considerations in AI right now. I think Eliezer’s position on this subject is shaped by assumptions like: If you have an AI as intelligent as Elon Musk today, then tomorrow you can run it on more hardware with a bit of normal human algorithmic progress, and get one twice as intelligent. So even if it would take Elon years to make a breakthrough, long before those years are up you’ll have an AI that can make breakthroughs much faster.
Inline links: thirty years away, Biological Anchors, Intelligence Explosion Microeconomics, hyperbolic growth we are currently experiencing, Some research finds
Davidson On Takeoff Speeds
June 20, 2023 · Original source
But also, Nostalgebraist argues that Bio Anchors hinges almost entirely on Moore’s Law. It’s no sin to hinge entirely on one very important value. But Bio Anchors looks like a very sophisticated piece of math with lots of parameters, and if you judge it on that basis, instead of on “well, everything depends on Moore’s Law, but Moore’s Law is hard to predict”, then you might get a false impression. Does CCF hinge on one specific parameter?
Inline links: Nostalgebraist argues that
Notes From The Progress Studies Conference
October 24, 2024 · Original source
Why is solar improving so quickly? Humanity is very good at mass manufacturing things in factories. Once you convert a problem to “let’s manufacture billions of identical copies of this small object in a factory”, our natural talent at doing this kicks in, factories compete with each other on cost, and you get a Moore’s Law like growth curve. Sometime around the late 90s / early 00s, factories started manufacturing solar panels en masse, and it was off to the races.
And now I feel less like mocking this. There’s still no visible kink in total factor productivity, let alone Moore’s Law. It could still all be hype. But the report from these people, who have spent half a century on the losing side of every battle, is that things are starting to look cheerier. Congress understands the problems with NEPA and is at least considering making life easier for the solar plants. Suddenly everyone’s a YIMBY. The first small modular nuclear reactor has been approved.
What Happened With Bio Anchors?
February 12, 2026 · Original source
Epoch/Croxton are current best estimates, and can probably fairly be read as the “real” answer against which Cotra and Davidson’s earlier guesses should be judged. All numbers are yearly multiples, so 1.4 means that willingness to spend grows 1.4x per year, ie 40%. Willingness To Spend: How much money are companies willing to spend on AI, in the form of chips and data centers? $/FLOP: How quickly do Moore’s Law, economies of scale, and other factors bring down the price of AI compute? Training Run Length: How long are companies spending on AI training runs for frontier models (instead of using those chips for smaller models, experiments, or consumer services)? Real Compute: The product of the three parameters above. Algorithmic Progress: How effectively do researchers discover new algorithms that makes training AIs cheaper and more efficient? Total Effective Compute: The product of real compute and algorithmic progress. So for example, the Epoch column’s 10.7x means that in any given year, you can train an AI 10.7x better than the last year, because you have 3.6x more compute available, and that compute is 3.0x more efficient. Cotra and Davidson were pretty close on willingness to spend and on FLOPs/$. This is an impressive achievement; they more or less predicted the giant data center buildout of the past few years. They ignored training run length, which probably seemed like a reasonable simplification at the time. But they got killed on algorithmic progress, which was 200% per year instead of 30%. How did they get this one so wrong? Here’s Cotra’s section on algorithmic progress: Algorithmic progress forecasts Note: I have done very little research into algorithmic progress trends. Of the four main components of my model (2020 compute requirements, algorithmic progress, compute price trends, and spending on computation) I have spent the least time thinking about algorithmic progress. I consider two types of algorithmic progress: relatively incremental and steady progress from iteratively improving architectures and learning algorithms, and the chance of “breakthrough” progress which brings the technical difficulty of training a transformative model down from “astronomically large” / “impossible” to “broadly feasible.” For incremental progress, the main source I used was Hernandez and Brown 2020, ”Measuring the Algorithmic Efficiency of Neural Networks”. The authors reimplemented open source state-of-the-art (SOTA) ImageNet models between 2012 and 2019 (six models in total). They trained each model up to the point that it achieved the same performance as AlexNet achieved in 2012, and recorded the total FLOP that required. They found that the SOTA model in 2019, EfficientNet B0, required ~44 times fewer training FLOP to achieve AlexNet performance than AlexNet did; the six data points fit a power law curve with the amount of computation required to match AlexNet halving every ~16 months over the seven years in the dataset.² They also show that linear programming displayed a similar trend over a longer period of time: when hardware is held fixed, the time in seconds taken to solve a standard basket of mixed integer programs by SOTA commercial software packages halved every ~13 months over the 21 years from 1996 to 2017.³ Grace 2013 (”Algorithmic Progress in Six Domains”) is the only other paper attempting to systematically quantify algorithmic progress that I am currently aware of, although I have not done a systematic literature review and may be missing others. I have chosen not to examine it in detail because a) it was written largely before the deep learning boom and mostly does not focus on ML tasks, and b) it is less straightforward to translate Grace’s results into the format that I am most interested in (”How has the amount of computation required to solve a fixed task decreased over time?”). Paul is familiar with the results, and he believes that algorithmic progress across the six domains studied in Grace 2013⁴ is consistent with a similar but slightly slower rate of progress, ranging from 13 to 36 months to halve the computation required to reach a fixed level of performance. Additionally, it seems plausible to me that both sets of results would overestimate the pace of algorithmic progress on a transformative task, because they are both focusing on relatively narrow problems with simple, well-defined benchmarks that large groups of researchers could directly optimize.⁵ Because no one has trained a transformative model yet, to the extent that the computation required to train one is falling over time, it would have to happen via proxies rather than researchers directly optimizing that metric (e.g. perhaps architectural innovations that improve training efficiency for image classifiers or language models would translate to a transformative model). Additionally, it may be that halving the amount of computation required to train a transformative model would require making progress on multiple partially-independent sub-problems (e.g. vision and language and motor control). I have attempted to take the Hernandez and Brown 2020 halving times (and Paul’s summary of the Grace 2013 halving times) as anchoring points and shade them upward to account for the considerations raised above. There is massive room for judgment in whether and how much to shade upward; I expect many readers will want to change my assumptions here, and some will believe it is more reasonable to shade downward. Cotra’s estimate comes primarily from one paper, Hernandez & Brown, which looks at algorithmic progress on a task called AlexNet. But later research demonstrated that the apparent speed of algorithmic progress varies by an order of magnitude based on whether you’re looking at an easy task (low-hanging fruit already picked) or a hard task (still lots of room to improve). AlexNet was an easy task, but pushing the frontier of AI is a hard task, so algorithmic progress in frontier AI has been faster than the AlexNet paper estimated. In Cotra’s defense, she admitted that this was the area where she was least certain, and that she had rounded the progress rate down based on various considerations when other people might round it up based on various other considerations. But the sheer extent of the error here, compounded with a few smaller errors that unfortunately all shared the same direction, was enough to throw off the estimate entirely. Since Cotra and Davidson were expecting AI to get 3.6x more effective compute each year, but it actually got 10.7x more, it’s no mystery why their timelines were off. When John recalculates Davidson’s model with Epoch’s numbers, he finds that it estimates AGI in 2030, which matches the current vibes. IV. With this information in place, it’s worth looking at some prominent contemporaneous critiques of Bio Anchors. Various people criticized Bio Anchors’ many strange anchors for how much compute it would take to produce AGI. For example, one anchor estimated that it would take 10^45 FLOPs, because that was how many calculations happened in all the brains of all animals throughout the evolutionary history (which eventually produced the human brain that AIs are trying to imitate). To make things even weirder, this anchor assumed away all animals other than nematodes as a rounding error (fact check: true!) All of these seemed to detract from the main show, an attempt to estimate the compute involved in the human brain. But even this more sober anchor was complicated by time horizons - it’s not enough to imitate the human brain for one second; AIs need to be able to imitate the human brain’s capacity for long-term planning. Cotra calculated how much compute AGI would require if it needed a planning horizon of seconds, weeks, or years. Thanks to METR, we now know that existing AIs have already passed a point where they can do most tasks that take humans seconds, are moving through the hour range, and are just about to touch one day. So the “seconds” anchor is ruled out. But it also seems unlikely that AGI will require years, because most human projects don’t take years, or at least can be split into tasks that take less than one year each (intuition pump: are we sure the average employee stays at an AI lab for more than a year? If not, that proves that a chain of people with sub-one-year time horizons can do valuable work). The AI Futures team guessed that the time horizon necessary for AIs to really start serious recursive self-improvement was between a few weeks and a few months (though this might look like a totally different number on the METR graph, which doesn’t translate perfectly into real life). If this is true, then all three anchors (seconds, hours, years) were off by at least an order of magnitude. But it turns out that none of this matters very much. The highest and lowest anchors cancel out, so that the most plausible anchor - human brain with time horizon of hours to days - is around the average. If you remove all the other anchors and just keep that one, the model’s estimates barely change. But also, we’re talking about crossing twelve orders of magnitude here. The difference between the different time horizon anchors doesn’t register much on that level, compared to things like algorithmic progress which have exponential effects. Maybe this is the model basically working as intended. You try lots of different anchors, put more weight on the more plausible ones, take a weighted average of each of them, and hopefully get something close to the real value. Bio Anchors did. Or maybe it was just good luck. Still hard to tell. Eliezer Yudkowsky argued that the whole methodology was fundamentally flawed. Partly because of the argument above - he didn’t trust the anchors - but also partly because he expected the calculations to be obviated by some sort of paradigm shift that couldn’t be shoehorned into “algorithmic progress” (like how you couldn’t build an airplane in 1900 but you could in 1920). As of 2026 - still before AGI has been invented and we get a good historical perspective - no such shift has occurred. The scaling laws have mostly held; whatever artificial space you try to measure models in, the measurement has mostly worked in a predictable way. There have really only been two kinks in the history of AI so far. First, a kink in training run size around 2010: Second, a kink in time horizons around 2024 and the invention of test-time compute: The 2010 kink was before Cotra’s forecast and priced in. The 2024 kink is interesting and relevant - but since it was on a parameter Cotra wasn’t measuring, and probably too small to show up on the orders-of-magnitude scale we’re talking about, it’s probably not a major cause of the model’s inaccuracy. Other things have been even more predictable: So Cotra’s bet on progress being smooth and measurable has mostly paid off so far. But Yudkowsky further explained that his timelines were shorter than Bio Anchors because people would be working hard to discover new paradigms, and if the current paradigm would only pay off in the 2050s, then probably they would discover one before then. You could think of this as a disjunction: timelines will be shorter than Cotra thinks, either because deep learning pays off quickly, or because a new paradigm gets invented in the interim. It turned out to be the first one. So although Yudkowsky’s new paradigm has yet to materialize, his disjunctive reasoning in favor of shorter-than-2050 timelines was basically on the mark. Nostalgebraist argued that Cotra’s whole model was a wrapper for an assumption that Moore’s Law will continue indefinitely. If it does, obviously you get enough compute for AI at some point, even if it requires some absurd process like simulating all 500 million years of multicellular evolution. I never entirely understood this objection, because - although Bio Anchors does depend on a story where Moore’s Law doesn’t break before we get the relevant amount of compute - this is only one of many background assumptions (like that a meteor doesn’t hit Earth before we get the relevant amount of compute). Given those assumptions, it does a useful not-just-assumption-repeating job of calculating when transformative AI will happen. As Cotra implicitly predicted, we seem on track to get AGI before Moore’s Law breaks down, and so Moore’s Law didn’t end up mattering very much. And if all of Cotra’s non-Moore’s-Law parameter estimates had been correct, her model would have given about the same timelines we have now, and surprised everyone with a revolutionary claim about the AI future. But Nostalgebraist added, almost as an aside: Cotra has a whole other forecast I didn’t mention for “algorithmic progress,” and the last number is what you get from just algorithmic progress and no Moore’s Law. So depending on how much you trust that forecast, you might want to take all these numbers with an even bigger grain of salt than you’d expected from everything else we’ve seen. How much should you trust Cotra’s algorithmic progress forecast? She writes: “I have done very little research into algorithmic progress trends. Of the four main components of my model (2020 compute requirements, algorithmic progress, compute price trends, and spending on computation) I have spent the least time thinking about algorithmic progress.” ...and bases the forecast on one paper about ImageNet classifiers. I want to be clear that when I quote these parts about Cotra not spending much time on something, I’m not trying to make fun of her. It’s good to be transparent about this kind of thing! I wish more people would do that. My complaint is not that she tells us what she spent time on, it’s that she spent time on the wrong things. Like Cotra herself, I think Nostalgebraist was spiritually correct even if his bottom line (about Moore’s Law) was wrong. His meta-level point was that a seemingly complicated model could actually hinge on one or two parameters, and that many of Cotra’s parameter values were vague hand-wavey best guess estimates. He gave algorithmic progress as a secondary example of this to shore up his Moore’s Law case, but in fact it turned out to be where all the action was. V. Those were the rare good critiques. The bad critiques were the same ones everyone in this space gets: You’re just trying to build hype.
Inline links: ”Measuring the Algorithmic Efficiency of Neural Networks”, ”Algorithmic Progress in Six Domains”, later research, fact check: true!, argued that, https://epoch.ai/blog/the-longest-training-run, https://www.lesswrong.com/posts/GAJbegsvnd85hX3eS/thoughts-on-extrapolating-time-horizons, https://epoch.ai/benchmarks/eci, argued that
So Cotra’s bet on progress being smooth and measurable has mostly paid off so far. But Yudkowsky further explained that his timelines were shorter than Bio Anchors because people would be working hard to discover new paradigms, and if the current paradigm would only pay off in the 2050s, then probably they would discover one before then. You could think of this as a disjunction: timelines will be shorter than Cotra thinks, either because deep learning pays off quickly, or because a new paradigm gets invented in the interim. It turned out to be the first one. So although Yudkowsky’s new paradigm has yet to materialize, his disjunctive reasoning in favor of shorter-than-2050 timelines was basically on the mark. Nostalgebraist argued that Cotra’s whole model was a wrapper for an assumption that Moore’s Law will continue indefinitely. If it does, obviously you get enough compute for AI at some point, even if it requires some absurd process like simulating all 500 million years of multicellular evolution. I never entirely understood this objection, because - although Bio Anchors does depend on a story where Moore’s Law doesn’t break before we get the relevant amount of compute - this is only one of many background assumptions (like that a meteor doesn’t hit Earth before we get the relevant amount of compute). Given those assumptions, it does a useful not-just-assumption-repeating job of calculating when transformative AI will happen. As Cotra implicitly predicted, we seem on track to get AGI before Moore’s Law breaks down, and so Moore’s Law didn’t end up mattering very much. And if all of Cotra’s non-Moore’s-Law parameter estimates had been correct, her model would have given about the same timelines we have now, and surprised everyone with a revolutionary claim about the AI future. But Nostalgebraist added, almost as an aside: Cotra has a whole other forecast I didn’t mention for “algorithmic progress,” and the last number is what you get from just algorithmic progress and no Moore’s Law. So depending on how much you trust that forecast, you might want to take all these numbers with an even bigger grain of salt than you’d expected from everything else we’ve seen. How much should you trust Cotra’s algorithmic progress forecast? She writes: “I have done very little research into algorithmic progress trends. Of the four main components of my model (2020 compute requirements, algorithmic progress, compute price trends, and spending on computation) I have spent the least time thinking about algorithmic progress.” ...and bases the forecast on one paper about ImageNet classifiers. I want to be clear that when I quote these parts about Cotra not spending much time on something, I’m not trying to make fun of her. It’s good to be transparent about this kind of thing! I wish more people would do that. My complaint is not that she tells us what she spent time on, it’s that she spent time on the wrong things. Like Cotra herself, I think Nostalgebraist was spiritually correct even if his bottom line (about Moore’s Law) was wrong. His meta-level point was that a seemingly complicated model could actually hinge on one or two parameters, and that many of Cotra’s parameter values were vague hand-wavey best guess estimates. He gave algorithmic progress as a secondary example of this to shore up his Moore’s Law case, but in fact it turned out to be where all the action was. V. Those were the rare good critiques. The bad critiques were the same ones everyone in this space gets: You’re just trying to build hype.
Inline links: argued that