The LC0 pure policy is most certainly not superhuman. To test this, I just had it (network 791556, i.e. standard network of the current LC0 release) play a game against a weak computer opponent (Shredder Chess Online). SCO plays maybe at the level of a strong expert/weak candidate master at rapid chess time controls (but it plays a lot faster, thereby making generation of a test game more convenient than trying to beat policy-only lc0 myself, which I think should be doable). Result was draw, after lc0 first completely outplayed SCO positionally, and then blundered tactically in a completely won position, with a strong-looking move that had a simple tactical refutation. It then probably still had a very slight advantage, but opted to take a draw by repetition.
I think policy-only lc0 plays worse relative to strong humans than Katago/LeelaZero in Go. I would attribute this to chess being easier to lose by tactical blunder than Go.
I would disagree with the notion that the cost of mastering a world scales with the cost of the world model. For instance, the learning with errors problem has a completely straightforward mathematical description, and yet strong quantum-resistant public-key cryptosystems can be built on it; there is every possibility that even a superintelligence a million years from now will be unable to read a message encrypted today using AES-256 encapsulated using a known Kyber public key with conservatively chosen security parameters.
Similarly, it is not clear to me at all what is even meant by saying that a tiny neural network can perfectly predict the "world" of Go. I would expect that even predicting the mere mechanics of the game, for instance determining that a group has just been captured by the last move of the opponent, will be difficult for small neural networks when examples are adversarially chosen (think of a group that snakes around the whole board, overwhelming the small NN capability to count liberties). The complexity of determining consequences of actions in Go is much more dependent on the depth of the required search than on the size of the game state, and it is easy to find examples on the 19x19 standard board size that will overwhelm any feed-forward neural network of reasonable size (but not necessarily networks augmented with tree search).
With regards to FOOM, I agree that doom from foom seems like an unlikely prospect (mainly due to diminishing returns on the utility of intelligence in many competitive settings) and I would agree that FOOM would require some experimental loop to be closed, which will push out time scales. I would also agree that the example of Go does not show what Yudkowsky thinks it does (it does help that this is a small world where it is feasible to do large reinforcement learning runs, and even then, Go programs have mostly confirmed human strategy, not totally upended it). But the possibility that if an unaided large NN achieved AGI or weak ASI, it would then be able to bootstrap itself to a much stronger level of ASI in a relatively short time (similar to the development cycle timeframe that led to the AGI/weak ASI itself; but involving extensive experimentation, so neither undetectable nor done in minutes or days) by combining improved algorithmic scaffolding with a smaller/faster policy network does not seem outlandish to me.
Lastly, I would argue that foom is in fact an observable phenomenon today. We see self-reinforcing, rapid, sudden onset improvement every time a neural network during training discovers a substantially new capability and then improves on it before settling into a new plateau. This is known as grokking and well-described in the literature on neural networks; there are even simple synthetic problems that produce a nice repeated pattern of grokking at successive levels of performance when a neural network is trained to solve them. I would expect that fooming can occur at various scales. However, I find the case that a large grokking step automatically happens when a system approaches human-level competence on general problem unconvincing (on the other hand, of course a large grokking step could happen in a system already at human-level competence by chance or happenstance and push into the weak ASI regime in a short time frame).
This seems clearly wrong:
Go is extremely simple: the entire world of Go can be precisely predicted by trivial tiny low depth circuits/programs. This means that the Go predictive capability of a NN model as a function of NN size completely flatlines at an extremely small size. A massive NN like the brain's cortex is mostly wasted for Go, with zero advantage vs the tiny NN AlphaZero uses for predicting the tiny simple world of Go.
Top go-playing programs utilize neural networks, but they are not neural networks. Monte-Carlo Tree Search boosts their playing strength immensely. The underlying pure policy networks would be strong amateur level when playing against opponents who are unaware that they are playing a pure neural network, but they would lose quite literally every game against top humans. It seems very likely that a purely NN-based player without search would have to be based on a far more complex neural network than the ones we see in, say, Leela Zero or Katago. In addition, top programs like Katago use some handcrafted features (things about the current game state that can be efficiently computed by traditional hand-written code, but would be difficult to learn or compute for a neural network), so they deviate to a significant extent from the paradigm of pure reinforcement learning via self-play from just the rules that AlphaZero proved viable. This, too, significantly improves their playing strength.
Finally, Go has a very narrow (or, with half-integer komi and rulesets that prevent long cycles, non-existent) path to draw, and games last for about 250 moves. That means that even small differences in skill can be reliably converted to wins. I would guess that the skill ceiling for Go (and thereby, the advantage that a superintelligence would have in Go over humans or current go-playing machines) is higher than in most real-life problems. Go is as complicated as the opponent makes it. I would, for these reasons, in fact not be too surprised if the best physically realizable go-playing system at tournament time controls with hardware resources, say, equivalent to a modern-day data center would include a general intelligence (that would likely adjust parameters or code in a more specialized go-player on the fly, when the need arises).
The machines playing chess and go, are a mixed example. I suck at chess, so the machines better than me have already existed decades ago. But at some moment they accelerated and surpassed the actual experts quite fast. More interestingly, they surpassed the experts in a way more general than the calculator does; if I remember it correctly, the machine that is superhuman at go is very similar to the machine that is superhuman at chess.
I think the story of chess- and Go-playing machines is a bit more nuanced, and that thinking about this is useful when thinking about takeoff.
The best chess-playing machines have been fairly strong (by human standards) since the late 1970s (Chess 4.7 showed expert-level tournament performance in 1978, and Belle, a special-purpose chess machine, was considered a good bit stronger than it). By the early 90s, chess computers at expert level were available to consumers at a modest budget, and the best machine built (Deep Thought) was grandmaster-level. It then took another six years for the Deep Thought approach to be scaled up and tuned to reach world-champion level. These programs were based on manually designed evaluation heuristics, with some automatic parameter tuning, and alpha-beta search with some manually designed depth extension heuristics. Over the years, people designed better and better evaluation functions and invented various tricks to reduce the amount of work spent on unpromising branches of the game tree.
Long into the 1990s, many strong players were convinced that this approach would not scale to world championship levels, because they believed that play competitive at the world champion level required correctly dealing with various difficult strategic problems, and that working within the prevailing paradigm would only lead to engines that were even more superhuman at tactics than had been already observed, while still failing against the strongest players due to lack of strategic foresight. This proved to be wrong: classical chess programs reached massively superhuman strength on the traditional approach to chess programming, and this line of programs was completely dominant and still improving up to about the year 2019.
In 2019, a team at DeepMind showed that throwing reinforcement learning and Monte Carlo Tree Search at chess (and various other games) could produce a system playing at an even higher level than the then-current version of Stockfish running on very strong hardware. Today, the best engines use either this approach or the traditional approach to chess programming augmented by incorporation of a very lightweight neural network for accurate positional evaluation.
For Go, there was hardly any significant progress from about the early 90s to the early 2010s: programs were roughly at the level of a casual player who had studied the game for a few months. A conceptual breakthrough (the invention of Monte-Carlo Tree Search) then brought them to a level equivalent in chess maybe to a master by the mid-2010s. DeepMind's AlphaGo system then showed in 2016 that reinforcement learning and MCTS could produce a system performing at a superhuman level when run on a very powerful computer. Today, programs based on the same principles (with some relatively minor go-specific improvements) run at substantially higher playing strength than AlphaGo on consumer hardware. The vast majority of strong players was completely convinced in 2016 that AlphaGo would not win its match against Lee Sedol (a world-class human player).
Chess programs had been superhuman at the things they were good at (spotting short tactics) for a long time before surpassing humans in general playing strength, arguably because their weaknesses improved less quickly than their strengths. Their weaknesses are in fact still in evidence today: it is not difficult to construct positions that the latest versions of LC0 or Stockfish don't handle correctly, but it is very difficult indeed to exploit this in real games. For Go programs, similar remaining weak spots have recently been shown to be exploitable in real games (see https://goattack.far.ai/), although my understanding is that these weaknesses have now largely been patched.
I think the general lesson that AI performance at a task will be determined by the aspects of that task that the AI handles best when the AI is far below human levels and by the aspects of the task that the AI handles worst when it is at or above human level, and that this slows down perceived improvement relative to humans once the AI is massively better than humans at some task-relevant capabilities, does in my expectation carry over to some extent from narrow AI (like chess computers) to general AI (like language models). In terms of the transition from chimpanzee-level intelligence to Einstein, this means that the argument from the relatively short time span evolution took to bridge that gap is probably not as general as it might look at first sight, as chimpanzees and humans probably share similar architecture-induced cognitive gaps, whereas the bottlenecks of an AI could be very different.
This would suggest (maybe counterintuitively) that fast takeoff scenarios are more likely with cognitive architectures that are similar to humans than with very alien ones.
Sure, the AI probably can't use all the mass-energy of the solar system efficiently within the next week or something, but that just means that it's going to want to store that mass-energy for later (...)
If the AI can indeed engineer black-hole powered matter-to-energy converters, it will have so much fuel that the mass stored in human bodies will be a rounding error to it. Indeed, given the size of other easily accessible sources, this would seem to be the case even if it has to resort to more primitive technology and less abundant fuel as its terminal energy source, such as hydrogen-hydrogen fusion reactors. Almost irrespective of what its terminal goals are, it will have more immediate concerns than going after that rounding error. Likewise, it would in all likelihood have more pressing worries than trying to plan out its future to the heat death of the universe (because it would recognize that no such plan will survive its first billion years, anyway).
I think we're imagining slightly different things by "superintelligence", because in my mind the obvious first move of the superAI is to kill literally all humans (...) The oft-quoted way around these parts that the AI can kill us all without us knowing is by figuring out which DNA sequences to send to a lab to have them synthesized into proteins, (...creating...) a virus much more lethal than anything we've ever seen, or a new species of bacteria with diamond skin, or some other thing that can be made from DNA-coded proteins.
I am imagining by "superintelligence" an entity that is for general cognition approximately what Stockfish is for chess: globally substantially better at thinking than any human expert in any domain, although possibly with small cognitive deficiencies remaining (similar to how it is fairly easy to find chess positions that Stockfish fails to understand but that are not difficult for humans). It might be smarter than that, of course, but anything with these characteristics would qualify as an SI in my mind.
I don't find the often-quoted diamondoid bacteria very convincing. Of course it's just a placeholder here, but still I cannot help but note that producing diamondoid cell membranes would, especially in a unicellular organism, more likely be an adaptive disadvantage (cost, logistics of getting things into and out of the cell) than a trait that is conducive to grey-gooing all naturally evolved organisms. More generally, it seems to me that the argument from bioweapons hinges on the ability of the superintelligence to develop highly complex biological agents without significant testing. It furthermore needs to develop them in such a way, again without testing, that they are quickly and quietly lethal after spreading through all or most of the human population without detection. In my mind, that combination of properties borders on assuming the superintelligence has access to magic, at least in a world that has reasonable controls against access to biological weapons manufacturing and design capabilities in place.
When setting in motion such a murderous plan, the AI would also, on its first try, have to be extremely certain that it is not going to get caught if it is playing the long game we assume it is playing. Otherwise cooperation with humans followed by expansion beyond Earth seems like a less risky strategy for long-term survival than hoping that killing everyone will go right and hoping that there is indeed nothing left to learn for it from living organisms.
At the limits of technology you can just convert any form of matter into energy by dumping it into a small black hole. Small black holes are actually really hot and emit appreciable fractions of their total mass per second through hawking radiation, so if you start a small black hole by concentrating lasers in a region of space, and you then feed it matter with a particle accelerator, you have essentially a perfect matter -> energy conversion. This is all to say that a superintelligence would certainly have uses for the kinds of atoms our bodies (and the earth) are made of.
I don't think this follows. Even if there is engineering that overcomes the practical obstacles towards building and maintaining a black hole power plant, it is not clear a priori that converting a non-negligible percentage of available atoms into energy would be required or useful for whatever an AI might want to do. At some scale, generating more energy does not advance one's goals, but only increases the waste heat emitted into space.
Obviously, things become lethal anyway (both for life and for the AI) long before anything more than an tiny fraction of the mass-energy of the surface layers of a planet has been converted by the local civilization's industries, due exactly to the problem of waste heat. But building hardware massive enough to cause problems of this kind takes time, and causes lesser problems on the way. I don't see why normal environmental regulations couldn't stop such a process at that point, unless the entities doing the hardware-building are also in control of hard military power.
An unaligned superintelligence would be more efficient than humans at pursuing its goals on all levels of execution, from basic scientific work to technical planning and engineering to rallying social support for its values. It would therefore be a formidable adversary. In a world where it would be the only one of its kind, its soft power would in all likelihood be greater than that of a large nation-state (and I would argue that, in a sense, something like GPT-4 would already wield an amount of soft power rivalling many nation-states if its use were as widespread as, say, that of Google). It would not, however, be able to work miracles and its hard power could plausibly be bounded if military uses of AI remain tightly regulated and military computing systems are tightly secured (as they should be anyway, AGI or not).
Obviously, these assumptions of controllability do not hold forever (e.g. into a far future setting, where the AI controls poorly regulated off-world industries in places where no humans have any oversight). But especially in a near-term, slow-takeoff scenario, I do not find the notion compelling that the result will be immediate intelligence explosion unconstrained by the need to empirically test ideas (most ideas, in human experience, don't work) followed by rapid extermination of humanity as the AI consumes all resources on the planet without encountering significant resistance.
If I had to think of a realistic-looking human extinction through AI scenario, I would tend to look at AI massively increasing per capita economic output, thereby generating comfortable living conditions for everyone, while quietly engineering life in a way intended to stop population explosion, but resulting in maintained below-replacement birth rates. But this class of extinction scenario does leave a lot of time for alignment and would seem to lead to continued existence of civilization.
It is worth noting that there are entire branches of science that are built around the assumption that intelligence is of zero utility for some important classes of problems. For instance, cryptographers build algorithms that are supposed to be secure against all adversaries, including superintelligences. Roughly speaking, one hopes (albeit without hard proof) for instance that the AES is secure (at least in the standard setting of single-key attacks) against all algorithms with a time-memory-data tradeoff significantly better than well-optimized exhaustive search (or quantumly, Grover search).
Turning the solar system into a Dyson sphere would enable an ASI to break AES-128 (but not AES-192 or AES-256) by brute force search, but it might well be that the intelligence of the ASI would only help with the engineering effort and maybe shave off a small factor in the required computational resources by way of better optimizing brute force search. I find it plausible that there would be many other tasks, even purely mathematical ones, where superintelligence would only yield a zero or tightly bounded planning or execution advantage over smart humans with appropriate tools.
I also find the Yudkowskian argument that an unaligned AI will disassemble everything else because it has better use for the atoms the other things are made of not massively persuasive. It seems likely that it would only have use for some kinds of atoms and not very unlikely that the atoms that human bodies are made of would not be very useful to it. Obviously, an unaligned or poorly aligned AI could still cause massive damage, even extinction-level damage, by building an industrial infrastructure that damages the environment beyond repair; rough analogues of this have happened historically, e.g. the Great Oxygenation event being an example of transformative change to Earth's ecosystems that left said ecosystems uninhabitable for most life as it was before the event. But even this kind of threat would not manifest in a foom-all-dead manner, but instead happen on a timescale similar to the current ecological crisis, i.e. on timescales where in principle societies can react.
That is odd. I certainly had a much, much higher completion rate than 1 in 40; in fact I had no games that I had to abandon with my prompt. However, I played manually, and played well enough that it mostly did not survive beyond move 30 (although my collection has a blindfold game that went beyond move 50), and checked at every turn that it reproduced the game history correctly, reprompting if that was not the case. Also, for GPT3.5 I supplied it with the narrative fiction that it could access Stockfish. Mentioning Stockfish might push it towards more precise play.
Trying again today, ChatGPT 3.5 using the standard chat interface did however seem to have a propensity to listing only White moves in its PGN output, which is not encouraging.
For exact reproducibility, I have added a game played via the API at temperature zero to my collection and given exact information on model, prompt and temperature in the PGN:
If your scripts allow testing this prompt, I'd be interested in seeing what completion rate/approximate rating relative to some low Stockfish level is achieved by chatgpt-3.5-turbo.
It seems to me that there are a couple of other reasons why LLMs might develop capabilities that go beyond the training set:
1. It could be that individual humans make random errors due to the "temperature" of their own thought processes, or systematic errors because they are only aware of part of the information that is relevant to what they are writing about. In both cases, it could be that in each instance, the most likely human completion to a text is objectively the "best" one, but that no human can consistently find the most likely continuation to a text, whereas the LLM can.
2. Humans may think for a very long time when writing a text. An LLM has to learn to predict the next token from the context provided with a relatively small fixed amount of computation (relatively meaning relative to the time spent by humans when writing the original).
3. The LLM might have different inductive biases than humans, and will therefore might fail to learn to imitate parts of human behaviour that are due to human inductive biases and not otherwise related to reality.
I think there is some evidence of these effects happening in practice. For instance, the Maiabot family of chess bots are neural networks. Each of the Maiabot networks is trained on a database of human game at a fixed rating. However, to the best of my knowledge, at least the weaker Maiabots play much stronger than the level of games they were trained on (probably mostly due to effect (1)).
Thanks for the information. I'll try out BT2. Against LazyBot I was just then able to get a draw in a blitz game with 3 seconds increment, which I don't think I could do within a few tries against an opponent of, say, low grandmaster strength (with low grandmaster strength being quite far way away from superhuman still). Since pure policy does not improve with thinking time, I think my chances would be much better at longer time controls. Certainly its lichess rating at slow time controls suggests that T80 is not more than master strength when its human opponents have more than 15 minutes for the whole game.
Self-play elo vastly exaggerates playing strength differences between different networks, so I would not expect a BT2 vs T80 difference of 100 elo points to translate to close to 100 elo playing strength difference against humans.