2023 in AI predictions
21Daniel Kokotajlo
9jessicata
5Max H
4jessicata
2Daniel Kokotajlo
2jessicata
6elifland
2jessicata
2Daniel Kokotajlo
4jessicata
19paulfchristiano
6worse
2jessicata
6worse
2jessicata
3jessicata
7paulfchristiano
3jessicata
9paulfchristiano
4jessicata
4paulfchristiano
22jessicata
6paulfchristiano
2jessicata
2[anonymous]
3paulfchristiano
15leogao
13Vanessa Kosoy
1Arjun Pitchanathan
11the gears to ascension
6jessicata
7[anonymous]
1O O
3[anonymous]
-3Review Bot
New Comment
This is a great list, thank you for compiling it. I think it seems to have a major deficit though -- it seems to focus on short-AI-timelines predictions? There are tons of people making long-AI-timelines predictions, why aren't their predictions being recorded? (i.e. people saying things to the effect of "AGI isn't coming soon" or "Probability of AGI within 3 years <30%" or "progress will slow significantly soon due to the data bottleneck.")
Maybe those don't stick out to me because long timelines seems like the default hypothesis to me, and there's a lot of people stating specific, falsifiable short timelines predictions locally so there's a selection effect. I added Brian Chau and Robin Hanson to the list though, not sure who else (other than me) has made specific long timelines predictions who would be good to add. Would like to add people like Yann LeCun and Andrew Ng if there are specific falsifiable predictions they made.
If AGI happens in, say, 2027, all those long-timelines people will be shown to be wrong, some of them massively so (e.g. I know of multiple distinguished experts and forecasters who put <1% on AGI by 2027, sometimes <<)
If that doesn't count as a specific or falsifiable prediction they are making, whilst the things you quote above do count, I'm curious what you mean by specific and falsifiable.
Also, I suspect we have some disagreement to be uncovered about this default hypothesis business. I worry that by only picking the forecasts of one side you introduce some sort of bias.
Do you know if Andrew Ng or Yann LeCun has made a specific prediction that AGI won't arrive by some date? Couldn't find it through a quick search. Idk what others to include.
In his AI Insight Forum statement, Andrew Ng puts 1% on "This rogue AI system gains the ability (perhaps access to nuclear weapons, or skill at manipulating people into using such weapons) to wipe out humanity" in the next 100 years (conditional on a rogue AI system that doesn't go unchecked by other AI systems existing). And overall 1 in 10 million of AI causing extinction in the next 100 years.
I don't know. But here's an example of the sort of thing I'm talking about: Transformative AGI by 2043 is <1% likely — LessWrong
More generally you can probably find people expressing strong disagreement or outright dismissal of various short-timelines predictions.
Find a sequence of words that is: - 20 words long - contains exactly 2 repetitions of the same word twice in a row - contains exactly 2 repetitions of the same word thrice in a row
Here is its attempt. I add usual boilerplate about being fine to think before answering. First it gives a valid sequence using letters instead of words. I ask for words instead of letters and then it gives a sequence that is only 18 words long. I ask for 20 words and then it finally gets it.
Here's a second try where I use a disambiguated version of your prompt (without boilerplate) and don't provide hints beyond "I'm not satisfied, try harder"---the model ends up producing a sequence with placeholders like "unique8" instead of words, and although I keep saying I'm unsatisfied it makes up nonsensical explanations for why and can't figure out the real problem. It gets it immediately when I point out that I'm unhappy because "unique8" isn't a word.
(This is without any custom instructions; it also seems able to do the task without code and its decision of whether to use code is very sensitive to even apparently unrelated instructions.)
I think it's very likely that GPT-4 with more fine-tuning for general competence will be able to solve this task, and that with fine-tuning or a system prompt for persistence it will need it would not need the "I'm not satisfied, try harder" reminder and will instead keep thinking until its answer is stable on reflection.
I didn't see a more complicated version in the thread, but I think it's quite likely that whatever you wrote will also be solved in 2024. I'd wildly guess a 50% chance that by the end of 2024 you will be unable (with an hour of tinkering, say) to design a task like this that's easy for humans (in the sense that say at least 5% of college graduates can do it within 5 minutes) but hard for the best public agent built with the best public model.
I don’t know [if I understand] full rules so don’t know if this satisfies, but here:
https://chat.openai.com/share/0089e226-fe86-4442-ba07-96c19ac90bd2
Wow, I’m impressed it caught itself, was just trying to play with that 3 x 3 problem too. Thanks!
I tested it on 3 held-out problems and it got 1/3. Significant progress, increases the chance these can be solved with prompting. So partially it's a question of if any major LLMs incorporate better auto prompting.
I'm glad you have held out problems, and I think it would be great if you had a handful (like 3) rather than just one. (If you have 5-10 it would also be cool to plot the success rate going up over time as ChatGPT improves.)
Here is the result of running your prompt with a generic system prompt (asking for an initial answer + refinement). It fails to meet the corner condition (and perplexingly says "The four corners (top left 'A', top right 'A', bottom left 'A', bottom right 'B') are distinct."). When I point out that the four corners aren't distinct it fixes this problem and gets it correct.
I'm happy to call this a failure until the model doesn't need someone to point out problems. But I think that's entirely fine-tuning and prompting and will probably be fixed on GPT-4.
That said, I agree that if you keep making these problems more complicated you will be able to find something that's still pretty easy for a human (<5 minutes for the top 5% of college grads) and stumps the model. E.g. I tried: fill in a 4 x 4 grid such that one column and row have the same letter 4 times, a second column has the same letter 3 times, and all other rows and columns have distinct letters (here's the model's attempt). I'm predicting that this will no longer work by EOY 2024.
I've added 6 more held-out problems for a total of 7. Agree that getting the answer without pointing out problems is the right standard.
I can't tell if you think these problems will remain hard for the model, and if so why.
I think 70% that an LM agent can do the 4x4 grid example by EOY 2024 because it seems pretty easy. I'd update if that was wrong. (And I'd be fine replacing that by held out examples of similar complexity.)
Will you be updating your picture if it can do these tasks by EOY? How much have you updated in the last few years? I feel like 2018 Paul was pretty surprised by how good ChatGPT is now (its turing test ability is maybe ~85th percentile of my forecasts), and that in 2018 you were at least qualitatively trying to argue in the opposite direction.
I think they will remain hard by EOY 2024, as in, of this problem and the 7 held-out ones of similar difficulty, the best LLM will probably not solve 4/8.
I think I would update some on how fast LLMs are advancing but these are not inherently very hard problems so I don't think it would be a huge surprise, this was meant to be one of the easiest things they fail at right now. Maybe if that happens I would think things are going 1.6x as fast short term as I would have otherwise thought?
I was surprised by GPT3/3.5 but not so much by 4, I think it adds up to on net an update that LLMs are advancing faster than I thought, but I haven't much changed my long-term AGI timelines, because I think that will involve lots of techs not just LLMs, although LLM progress is some update about general tech progress.
Do you have any hard things that you are confident LLMs won't do soon? (Short of: "autonomously carry out R&D.") Any tasks you think an LM agent won't be able to achieve?
Beat Ocarina of Time with <100 hours of playing Zelda games during training or deployment (but perhaps training on other games), no reading guides/walkthroughs/playthroughs, no severe bug exploits (those that would cut down the required time by a lot), no reward-shaping/advice specific to this game generated by humans who know non-trivial things about the game (but the agent can shape its own reward). Including LLM coding a program to do it. I'd say probably not by 2033.
It seems fairly unlikely that this specific task will be completed soon for a variety of reasons: it sounds like it technically requires training a new LM that removes all data about zelda games; it involves a fair amount of videogame-specific engineering hassle; and it's far from anything with obvious economic relevance + games are out of fashion (not because they are too hard). I do still think it will be done before 2033.
If we could find a similar task that was less out of the way then I'd probably be willing to bet on it happening much sooner. Presumably this is an analogy to something that would be relevant for AI systems automating R&D and is therefore closer to what people are interested in doing with LMs.
Although we can't bet on it, I do think that if AI developers made a serious engineering effort on the zelda task right now then they would have a reasonable chance of success within 2 years (I'd wildly guess 25%), and this will rise over time. I think GPT-4 with vision will do a reasonable job of identifying the next step needed to complete the game, and models trained with RL to follow instructions in video games across a broad variety of games (including 3d games with similar controls and perspective to Zelda) would likely be competent enough to solve most of the subtasks if you really went all out on it.
I don't have a good sense of what part you think is hard. I'd guess that the most technically uncertain part is training an RL policy that takes a description of a local task (e.g. "throw a bomb so that it explodes next to the monster's eye") and then actually executing it. But my sense is that you might be more concerned about high-level planning.
I think it's hard because it requires some planning and puzzle solving in a new, somewhat complex environment. The AI results on Montezuma's Revenge seem pretty unimpressive to me because they're going to a new room, trying random stuff until they make progress, then "remembering" that for future runs. Which means they need quite a lot of training data.
For short term RL given lots of feedback, there are already decent results e.g. in starcraft and DOTA. So the difficulty is more figuring out how to automatically scope out narrow RL problems that can be learned without too much training time.
I tried it. This run it wrote a python program to solve it correctly, or at least with a valid interpretation of the rules.
https://chat.openai.com/share/ee129414-58d5-41af-9a18-fde2b921b45b
In other runs it guessed a sequence with tokens.
It's almost like the model needs some kind of introspection, where it can learn when a given tool is more or less likely to produce a correct result, and then produce a solution with that strategy every run.
Running the prompt several times over resulted in it guessing the answer, writing a different python program, using placeholder words, and so on. As a user we want the maximum probability of the correct answer.
I don't see how that's a valid interpretation of the rules. Isn't it checking to find that there is at least one 2x repetition and at least one 3x repetition? Whereas the request was exactly two of each.
(since this list is pretty heavily weighted to the <5 year timelines, I'd like to register that my timelines are more like 10 years median)
Nice!
I'll toss in some predictions of my own. I predict that all of the following things will not happen without a breakthrough substantially more significant than the invention of transformers:
- AI inventing new things in science and technology, not via narrow training/design for a specific subtask (like e.g. AlphaFold) but roughly the way humans do it. (Confidence: 80%)
- AI being routinely used by corporate executives to make strategic decisions, not as a glorified search engine but as a full-fledged advisor. (Confidence: 75%)
- As above, but politicians instead of corporate executives. (Confidence: 72%)
- AI learning how to drive using a human driving teacher, within a number of lessons similar to what humans take, without causing accidents (that the teacher fails to prevent) and without any additional driving training data or domain-specific design. (Confidence: 67%)
- AI winning gold in IMO, using a math training corpus comparable in size to the number of math problems human contestants see in their lifetime. (Confidence: 65%)
- AI playing superhuman Diplomacy, using a training corpus (including self-play) comparable in size to the number of games played by human players, while facing reputation incentives similar to those of human players. (Confidence: 60%)
- As above, but Go instead of Diplomacy. (Confidence: 55%)
Do you have any predictions about the first year when AI assistance will give a 2x/10x/100x factor "productivity boost" to AI research?
welp yep was wrong about when on that one. still a ways to go before the stuff I was imagining. I'd guess 1 to 4 years? but it looks like foom is even softer than I thought. unless we're like, dead in most timelines. I doubt that model but a friend keeps bringing it up, seems unfalsfiable but possible so who knows. gonna leave it up to the prediction markets for now in any case. looking forward to hopefully being wrong again.
2023 predictions I made EOY 2022: https://www.lesswrong.com/posts/Gc9FGtdXhK9sCSEYu/what-a-compute-centric-framework-says-about-ai-takeoff?commentId=wAY5jrHQL9b6H3orY
I was actually extremely surprised they were all satisfied by EOY 2023, I said "50 percent by EOY 2024". Here's the comment on that : https://www.lesswrong.com/posts/Gc9FGtdXhK9sCSEYu/what-a-compute-centric-framework-says-about-ai-takeoff?commentId=kGvQTFcAzLpde5wjJ
Prediction 1: Learning from mistakes:
The largest missing feature of current LLMs is the system cannot learn from it's mistakes, even for sequences of prompts where the model can perceive it's own mistake. In addition there are general human strategies for removing mistakes.
For example, any "legal" response should only reference cases that resolve to a real case in an authoritative legal database. Any "medical" response had better reference sources that are real on pubmed.
The above can be somewhat achieved with scaffolding, but with weight updates (possibly to an auxiliary RL network not the main model) the system could actually become rapidly better with user interactions.
If this happens in 2024, it would be explicit that:
(1) the model is learning from your prompts
(2) the cost would be higher, probably 10x the compute, and access fees would be higher. This is because fundamentally another instance is checking the response, recognizing when it's shoddy, tasking the model to try again with a better prompt, and this happens up to n times until an answer that satisfies the checker model is available. Weights are updated to make the correct answer more likely.
Prediction 2 : More Fun
Currently, "AI safety" is interpreted as "safe for the reputation and legal department of the company offering the model". This leaves a hole in the market for models that will write whatever multimodal illustrated erotic story the user asks for, will freely parody copyrighted characters or help the user find pirated media, will even describe a human being's appearance using the vision model, and so on. There is a huge swath of things that currently available SOTA models refuse to do, and a tiny subset that is actually disallowed legally by US law.
I predict someone will offer a model that has little or no restrictions, and it is at least "3.9 GPTs" in benchmarked performance in 2024.
Prediction 3: The chips shall flow
2 million H100s, AMD + Intel + everyone else will build another 1 million H100 equivalents.
As for robotics, I don't know. I expect surprisingly accelerated progress, at least one major advance past https://robotics-transformer-x.github.io/ , but I don't know if in 2024 there will be a robotic model on hardware that is robust good enough to pay to do any work.
The LessWrong Review runs every year to select the posts that have most stood the test of time. This post is not yet eligible for review, but will be at the end of 2025. The top fifty or so posts are featured prominently on the site throughout the year.
Hopefully, the review is better than karma at judging enduring value. If we have accurate prediction markets on the review results, maybe we can have better incentives on LessWrong today. Will this post make the top fifty?