Posts
Comments
sorry I meant a bot that played random move, not a randomly sampled go bot from KGS. agreed with GPT-4 not beating average go bot
If it's about explaining your answer with 5th grade gibberish then GPT-4 is THE solution for you! ;)
let's say by concatenating your textbooks you get plenty of examples of with "blablabla object sky blablabla gravity blablabla blabla . And then the exercise is: "blablabla object of mass blablabla thrown from the sky, what's the force? a) f=120 b) ... c) ... d) ...". then what you need to do is just do some prompt programming at the beginning by "for looping answer" and teaching it to return either a,b,c or d. Now, I don't see any reason why a neural net couldn't approximate linear functions of two variables. It just needs to map words like "derivative of speed", "acceleration", "" to the same concept and then look at it with attention & multiply two digits.
I think the general answer to testing seems AGI-complete in the sense that you should understand the edge-cases of a function (or correct output from "normal" input).
if we take the simplest testing case, let's say python using pytest, with a typed code, with some simple test for each type (eg. 0 and 1 for integers, empty/random strings, etc.) then you could show it examples on how to generate tests from function names... but then you could also just do it with reg-ex, so I guess with hypothesis.
so maybe the right question to ask is: what do you expect GPT-4 to do better than GPT-3 relative to the train distribution (which will have maybe 1-2y of more github data) + context window? What's the bottleneck? When would you say "I'm pretty sure there's enough capacity to do it"? What are the few-shot examples you feed your model?
well if we're doing a bet then at some point we need to "resolve" the prediction. so we ask GPT-4 the same physics question 1000 times and then some humans judges count how many it got right, if it gets it right more than let's say 95% of the time (or any confidence interval) , then we would resolve this positively. of course you could do more than 1000, and with law of large numbers it should converge to the true probability of giving the right answer?
re right prompt: GPT-3 has a context window of 2048 tokens, so this limits quite a lot what it could do. Also, it's not accurate at two-digit multiplication (what you would at least need to multiply your $ to %), even worse at 5-digit. So in this case, we're sure it can't do your taxes. And in the more general case, gwern wrote some debugging steps to check if the problem is GPT-3 or your prompt.
Now, for GPT-4, given they keep scaling the same way, it won't be possible to have accurate enough digit multiplication (like 4-5 digits, cf. this thread) but with three more scalings it should do it. Prompt would be "here is a few examples on how to do taxe multiplication and addition given my format, so please output result format", and concatenate those two. I'm happy to bet $1 1:1 on GPT-7 doing taxe multiplication to 90% accuracy (given only integer precision).
So physics understanding.
How do you think it would perform on simpler question closer to its training dataset, like "we throw a ball from a 500m building with no wind, and the same ball but with wind, which one hits the floor earlier" (on average, after 1000 questions).$? If this still does not seem plausible, what is something you would bet $100 2:1 but not 1:1 that it would not be able to do?
Interesting. Apparently GPT-2 could make (up to?) 14 non-invalid moves. Also, this paper mentions a cross-entropy log-loss of 0.7 and make 10% of invalid moves after fine-tuning on 2.8M chess games. So maybe here data is the bottleneck, but assuming it's not, GPT-4's overall loss would be x smaller than GPT-2 (cf. Fig1 on parameters), and with the strong assumption of the overall transfering directly to chess loss, and chess invalid move accuracy being inversely proportional to chess loss wins, then it would make 5% of invalid moves
So from 2-digit substraction to 5-digit substraction it lost 90% accuracy, and scaling the model by ~10x gave a 3x improvement (from 10 to 30%) on two-digit multiplication. So assuming we get 3x more accuracy from each 10x increase and that 100% on two digit corresponds to ~10% on 5-digit, we would need something like 3 more scalings like "13B -> 175B", so about 400 trillion params.

That's a good one. What would be a claim you would be less confident (less than 80%) about but still enough confident to bet $100 at 2:1 odds? For me it would be "gpt-4 would beat a random go bot 99% of the time (in 1000 games) given the right input of less than1000 bytes."
A model relased on openai.com with "GPT" in the name before end of 2022. Could be either GPTX where X is a new name for GPT4, but should be an iteration over GPT-3 and should have at least 10x more parameters.
(note to mods: Ideally I would prefer to have larger Latex equations, not sure how to do that. If someone could just make those bigger, or even replace the equation screenshot with real Latex that would be awesome.)
sure I agree that keeping your system predictions for you makes more sense and tweeting doesn't necessarily help. Maybe what I'm pointing at is where the text you're tweeting is not necessarily "predictions" but maybe some "manipulation text" to maximize profit short term. Let's say you tweet "buy dogecoin" like Elon Musk, so the price goes higher and you can sell all of your doge when you predicted the price would drop. I'm not really sure how such planning would work, and exactly what to feed to some NLP model to manipulate the market in such a way... but now it seems we could just make a simple RL agent (without GPT) that can do either:
- 1. move money in his portfolio
- 2. tweet "price of X will rise" or "price of Y will go down".
but yes you're right that's pretty close to just "fund managers' predictions", and that would impact less than say Elon Musk tweeting (where there's common knowledge that his tweets change the stock/crypto prices quickly)
yes that's 50 million dollars
More generally, there's a difference between things being true and being useful. Believing that sometimes you should not update isn't a really useful habit as it forces the rationalizations you mentioned.
Another example is believing "willpower is a limited quantity" vs. "it's a muscle and the more I use it the stronger I get". The first belief will push you towards not doing anything, which is similar to the default mode of not updating in your story.
Note: I also know very little about this. Few thoughts on your guesses (and my corresponding credences):
--It seems pretty likely that it will be for humans (something that works for mices wouldn't be impressive enough for an announcement). In last year's white paper they were already inserting electrode arrays in the brain. But maybe you mean something that lives inside the brain independently? (90%)
--If by "significative damage" you mean "not altering basic human capabilities" then it sounds plausible. From the white paper they seem to focus on damage to "the blood-brain barrier" and the "brain’s inflammatory response to foreign objects". My intuition is that the brain would react pretty strongly to something inside it for 10 years though. (20%)
--Other BCI companies have done similar demo-s, so given presentation is long this might happen at some point. But Neuralink might also want to show they're different from mainstream companies. (35%)
--Seems plausible. Assigning lower credence because really specific. (15%)
Funnily enough, I wrote a blog distilling what I learned from reproducing experiments of that 2018 Nature paper, adding some animations and diagrams. I especially look at the two-step task, the Harlow task (the one with monkeys looking at a screen), and also try to explain some brain things (e.g. how DA interacts with the PFN) at the end.
HN comment unsure about the meta-learning generalization claims that OpenAI has a "serious duty [...] to frame their results more carefully"
re working memory: never thought of it during conversations, interesting. it seems that we sometime hold the nodes of the conversation tree to go back to them afterward. and maybe if you're introducing new concepts while you're talking people need to hold those definitions in working memory as well.
Some friends tried (inconclusively) to apply AlphaZero to a two-player GoL. I can put you in touch if you want their feedback.
Thanks for the tutorial to download documentation, I've never done that myself so will check it out next time I go offline for a while!
I usually just run python to look at docs, importing the library, and then do help(lib.module.function)
. If I don't really know what the class can do, I usually do dir(class_instance)
to find the available methods/attributes, and do the help thing on them.
This only works if you know reasonably well where to look at. If I were you I would try loading the "read the docs" html build offline in your browser (might be searchable this way), but then you still have a browser open (so you would really need to turn down wifi).
Thanks for writing this up!
I've personally tried Complice coworking rooms where people synchronize on pomodoros and chat during breaks, especially EA France's study room (+discord to voice chat during breaks) but there's also a LW study hall: https://complice.co/rooms
I've been experimenting with offline coding recently, sharing some of my conclusions.
Why I started 1) Most of the programming I do at the moment only needs a terminal and a text editor. I'm implementing things from scratch without needing libraries and I noticed I could just read the docs offline. 2) I came to the conclusion that googling things wasn't worth the cost of having a web browser open--using the outside view, when I look back at all the instances of coding while having the internet in easy-access, I always end up being distracted, and even if i code my mind thinks about what I could be doing.
How to go offline (Computer) 1) turn off wi-fi 2) forget network (Phone) if you're at home, put it out of reach. I turn it off then throw it on top of a closet, so far that i need to grab a chair in the living room to catch it. If you have an office, then do the same thing and go to your office without your phone.
When My general rule in January was that I could only check the internet between 11pm and 12am. The rest of the "no work + no internet" time was for deep relaxation, meditation, journaling, eating, etc. In April I went without any internet connection for a week. I was amazed at how much free time I had, but the lack of social interactions was a bit counter-productive. Currently, I'm going offline from the moment I wake up to 7pm. This seems like a good balance where I'm not too tired but still productive throughout the day.
Let me know if you have any question about the process or similar experience to share.
Thanks for all the references! I don't currently have much time to read all of it right now so I can't really engage with the specific arguments for the rejection of using utility functions/studying recursive self-improvement.
I essentially agree with most of what you wrote. There is maybe a slight disagreement in how you framed (not what you meant) how research focus shifted since 2014.
I see Superintelligence as essentially saying "hey, there is pb A. And even if we solve A, then we might also have B. And given C and D, there might be E." Now that the field is more mature and we have many more researchers getting paid to work on these problems, the arguments became much more goal focused. Now people are saying "I'm going to make progress on sub-problem X, by publishing a paper on Y. And working on Z is not cost-effective given so I'm not going to work on it given humanity's current time constraints."
These approaches are often grouped as "long-term problems-focused" and "making tractable progress now focused". In the first group you have Yudkowsky 2010, Bostrom 2014, MIRI's current research and maybe CAIS. In the second one you have current CHAI/FHI/OpenAI/DeepMind/Ought papers.
Your original framing can be interpreted as "after proving some mathematical theorems, people rejected the main arguments of Superintelligence and now most of the community agrees that working on X, Y and Z are tractable but A, B and C are more controversials".
I think a more nuanced and precise framing would be: "In Superintelligence Bostrom exposes exhaustively the risks associated with advanced AI. A short portion of the book is dedicated to the problems are working on right now. Indeed, people stopped working on the other problems (largest portion of the book) because 1) there hasn't been really productive working on them 2) some rebuttals have been written online giving convincing arguments that those pbs are not tractable anyway 3) there are now well-funded research organizations with incentives to make tangible progress on those pbs."
In your last framing, you presented precise papers/rebuttals (thanks again!) for 2), and I think rebuttals are a great reason to stop working on a pb, but I think they're not the only reason and not the real reason people stopped working on those pb. To be fair, I think 1) can be explained by many more factors than "it's theoretically impossible to make progress on those pbs". It can be that the research mindset required to work on these pbs is less socially/intellectually validating or requires much more theoretical approaches, so will be off-putting/tiresome to most recent grads that enter the field. I also think that AI Safety is now much more intertwined with evidence-based approaches such as Effective Altruism than it was in 2014, which explains 3), so people start presenting their research as "partial solutions to the pb. of AI Safety" or "research agenda".
To be clear, I'm not criticizing the current shift in research. I think it's productive for the field, both in the short term and long term. To give a bit more personal context, I started getting interested in AI Safety after reading Bostrom and have always been more interested in the "finding problems" approach. I went to FHI to work on AI Safety because I was super interested in finding new pbs related to the treacherous turn. It's now almost taboo to say that we're working on pbs that are sub-optimally minimizing AI risk, but the real reason that pushed me to think about those pbs was because they were both important and interesting. The pb. with the current "shift in framing" is that it's making it socially unacceptable for people to think/work on more long-term pbs where there is more variance in research productivity.
I don't quite understand the question?
Sorry about that. I thought there was some link to our discussion about utility functions but I misunderstood.
EDIT: I also wanted to mention that the number of pages in a book doesn't account for how important the author think the pb. is (Bostrom even comments on this in the postface of its book). Again, the book is mostly about saying "here are all the pbs", not "these are the tractable pbs we should start working on, and we should dedicate research ressources proportionally to the amount of pages I talk about it in the book".
This framing really helped me think about gradual self-improvement, thanks for writing it down!
I agree with most of what you wrote. I still feel that in the case of an AGI re-writing its own code there's some sense of intent that hasn't been explicitly happening for the past thousand years.
Agreed, you could still model Humanity as some kind of self-improving Human + Computer Colossus (cf. Tim Urban's framing) that somehow has some agency. But it's much less effective at self-improving itself, and it's not thinking "yep, I need to invent this new science to optimize this utility function". I agree that the threshold is "when all the relevant action is from a single system improving itself".
there would also be warning signs before it was too late
And what happens then? Will we reach some kind of global consensus to stop any research in this area? How long will it take to build a safe "single system improving itself"? How will all the relevant actors behave in the meantime?
My intuition is that in the best scenario we reach some kind of AGI Cold War situation for long periods of time.
I get the sense that the crux here is more between fast / slow takeoffs than unipolar / multipolar scenarios.
In the case of a gradual transition into more powerful technology, what happens when the children of your analogy discovers recursive self improvement?
When you say "the last few years has seen many people here" for your 2nd/3rd paragraph, do you have any posts / authors in mind to illustrate?
I agree that there has been a shift in what people write about because the field grew (as Daniel Filan pointed out). However, I don't remember reading anyone dismiss convergent instrumental goals such as increasing your own intelligence or utility functions as an useful abstraction to think about agency.
In your thread with ofer, he asked what was the difference between using loss functions in neural nets vs. objective function / utility functions and I haven't fully catched your opinion on that.
the ones you mentioned
To be clear, this is a linkpost for Philip Trammell's blogpost. I'm not involved in the writing.
As you say
To be clear, the author is Philip Trammell, not me. Added quotes to make it clearer.
Having printed and read the full version, this ultra-simplified version was an useful summary.
Happy to read a (not-so-)simplified version (like 20-30 paragraphs).
Funny comment!
A comprehensive AI alignment introductory web hub
RAISE and Robert Miles provide introductory content. You can think of LW->alignment forum as "web hubs" for AI Alignment research.
formal curriculum
There was a course on AGI Safety last fall in Berkeley.
A department or even a single outspokenly sympathetic official in any government of any industrialized nation
You can find a list of institutions/donors here.
A list of concrete and detailed policy proposals related to AI alignment
I would recommend reports from FHI/GovAI as a starting point.
Would this be valuable, and which resource would it be most useful to create?
Please give more detailed information about the project to receive feedback.
You can find AGI predictions, including Starcraft forecasts, in "When Will AI Exceed Human Performance? Evidence from AI Experts". Projects for having "all forecasts on AGI in one place" include ai.metaculus.com & foretold.io.
Does that summarize your comment?
1. Proposals should make superintelligences less likely to fight you by using some conceptual insight true in most cases.
2. With CIRL, this insight is "we want the AI to actively cooperate with humans", so there's real value from it being formalized in a paper.
3. In the counterfactual paper, there's the insight "what if the AI thinks he's not on but still learns".
For the last bit, I have two interpretations:
4.a. However, it's unclear that this design avoids all manipulative behaviour and is completely safe.
4.b. However, it's unclear that adding the counterfactual feature to another design (e.g. CIRL) would make systems overall safer / would actually reduce manipulation incentives.
If I understand you correctly, there are actual insights from counterfactual oracles--the problem is that those might not be insights that would apply to a broad class of Alignment failures, but only to "engineered" cases of boxed oracle AIs (as opposed to CIRL where we might want AIs to be cooperative in general). Was it what you meant?
The zero reward is in the paper. I agree that skipping would solve the problem. From talking to Stuart, my impression is that he thinks that would be equivalent to skipping for specifying "no learning", or would just slow down learning. My disagreement on that I think it can confuse learning to the point of not learning the right thing.
Why not do a combination of pre-training and online learning, where you do enough during the training phase to get a useful predictor, and then use online learning to deal with subsequent distributional shifts?
Yes, that should work. My quote saying that online learning "won't work and is unsafe" is imprecise. I should have said "if epsilon is small enough to be comparable to the probability of shooting an escape message at random, then it is not safe. Also, if we continue sending the wrong instead of skipping, then it might not learn the correct thing if is not big enough".
Although I guess that probably isn't really original either. What seems original is that during any episode where learning will take place, don't let humans (or any other system that might be insecure against the oracle) see the oracle's output until the episode is over.
That's exactly it!
The string is read with probability 1-
Yes, if we choose the utility function to make it a CDT agent optimizing for the reward for one step (so particular case of act-based) then it won't care about future versions of itself nor want to escape.
I agree with the intuition of shutting down to make it episodic, but I am still confused about the causal relationship between "having the rule to shutdown the system" and "having a current timestep maximizer". For it to really be a "current timestep maximizer" it needs to be in some kind of reward/utility function. Because everything is reset at each timestep, there is no information pointing at "I might get shutdown at the next timestep".
As for the collecting a dataset and then optimizing for some natural direct effect, I am not familiar enough with Pearl's work to tell if that would work, but I made some related comments about why there might be some problems in online-learning/"training then testing" here.
The Asymptotically Unambitious AGI thread helped me clarify my thoughts, thanks. I agree that an optimal CDT agent won't think about future versions, and I don't see any optimization pressure towards escape message nor disproportionately common "escape message" regions.
However, it still assumes we have access to this magic oracle that optimizes for where is the event where humans don't see the answer, its indicator function, and the counterfactual reward (given by the automatic machine). If humans were able to build an oracle performing optimally (w.r.t ) from day 1, then humans would be able to specify some kind of "god oracle". The rest of the design seems to be just "how to interact with a god oracle so that humans are not influenced by the answers".
In practice, you'll want something that is able to learn from its (question, prediction, reward) history. That's why there is this automatic machine rewarding the oracle with some probability . In an online learning setting, most of the time the model gets (probability ), and it sometimes gets some useful feedback (probability . Therefore, if is small enough (comparable to the probability of hitting an escape message at random), then the learning gets extremely slow and the oracle might shoot at the escape action at random. Even worse, the (question, prediction, ) datapoints in the history could get so misleading that the oracle never learns anything.
Now, if we're not in an online learning process but instead there is a separation between a "training phase" and a "deployment phase where the AI continue to learns with probability ", then the setup is just "have a model that learns to do the useful stuff in sandbox, and then have the weights (almost) fixed in deployment"
In short, I think the CDT setup without machine learning assumes the problem already solved, that online learning won't work and is unsafe, which leaves us with a "training then deployment" setup that isn't really original.
Yes, they call it a low-bandwidth oracle.
I agree that these stories won't (naturally) lead to a treacherous turn. Continuously learning to deceive (a ML failure in this case, as you mentioned) is a different result. The story/learning should be substantially different to lead to "learning the concept of deception" (for reaching an AGI-level ability to reason about such abstract concepts), but maybe there's a way to learn those concepts with only narrow AI.
I included dates such as 2020 to 2045 to make it more concrete. I agree that weeks (instead of years) would give a more accurate representation as current ML experiments take a few weeks tops.
The scenario I had in mind is "in the context of a few weeks ML experiment, I achieved human intelligence and realized that I need to conceal my intentions/capabilities and I still don't have decisive strategic advantage". The challenge would then be "how to conceal my human level intelligence before everything I have discovered is thrown away". One way to do this would be to escape, for instance by copy-pasting and running your code somewhere else.
If we're already at the stage of emergent human-level intelligence from running ML experiments, I would expect "escape" to be harder than just human-level intelligence (as there would be more concerns w.r.t. AGI Safety, and more AI boxing/security/interpretability measure), which would necessit more recursive self-improvement steps, hence more weeks.
Beside, in such a scenario the AI would be incentivized to spend as much time as possible to maximize its true capability, because it would want to maximize its probability of successfully taking over (because any extra % of taking over would give astronomical returns in expected value compared to just being shutdown).
Your comment makes a lot os sense, thanks.
I put step 2. before step 3. because I thought something like "first you learn that there is some supervisor watching, and then you realize that you would prefer him not to watch". Agreed that step 2. could happen only by thinking.
Yep, deception is about alignment, and I think that most parents would be more concerned about alignment, not improving the tactics. However, I agree that if we take "education" in a broad sense (including high school, college, etc.), it's unofficially about tactics.
It's interesting to think of it in terms of cooperation - entities less powerful than their supervisors are (instrumentally) incentivized to cooperate.
what to do with a seed AI that lies, but not so well as to be unnoticeable
Well, destroy it, right? If it's deliberately doing a. or b. (from "Seed AI") then step 4. has started. The other cases where it could be "lying" from saying wrong things would be if its model is consistently wrong (e.g. stuck in a local minima), so you better start again from scratch.
If the supervisor isn't itself perfectly consistent and aligned, some amount of self-deception is present. Any competent seed AI (or child) is going to have to learn deception
That's insightful. Biased humans will keep saying that they want X when they want Y instead, so deceiving humans by pretending to be working on X while doing Y seems indeed natural (assuming you have "maximize what humans really want" in your code).
I meant:
"In my opinion, the disagreement between Bostrom (treacherous turn) and Goertzel (sordid stumble) originates from the uncertainty about how long steps 2. and 3. will take"
That's an interesting scenario. Instead of "won't see a practical way to replace humanity with its tools", I would say "would estimate its chances of success to be < 99%". I agree that we could say that it's "honestly" making humans happy in the sense that it understands that this maximizes expected value. However, he knows that there could be much more expected value after replacing humanity with its tools, so by doing the right thing it's still "pretending" to not know where the absurd amount of value is. But yeah, a smile maximizer making everyone happy shouldn't be too concerned about concealing its capabilities, shortening step 4.
This thread is to discuss "How useful is quantilization for mitigating specification-gaming? (Ryan Carey, Apr. 2019, SafeML ICLR 2019 Workshop)"
This thread is to discuss "Quantilizers (Michaël Trazzi & Ryan Carey, Apr. 2019, Github)".
This thread is to discuss "When to use quantilization (Ryan Carey, Feb. 2019, LessWrong)"
This thread is to discuss "Quantilal control for finite MDPs & Computing an exact quantilal policy (Vanessa Kosoy, Apr. 2018, LessWrong)"
This thread is to discuss "Reinforcement Learning with a Corrupted Reward Channel (Tom Everitt; Victoria Krakovna; Laurent Orseau; Marcus Hutter; Shane Legg, Aug. 2017, arXiv; IJCAI)"
This thread is to discuss "Thoughts on Quantilizers (Stuart Armstrong, Jan. 2017, Intelligent Agent)"
This thread is to discuss "Another view of quantilizers: avoiding Goodhart's Law (Jessica Taylor, Jan. 2016, Intelligent Agent Foundations Forum)"