Posts
Comments
Perhaps if you needed a larger number of ternary weights, but the paper claims to achieve the same performance with ternary weights as one gets with 16-bit weights using the same parameter count.
I think this could be a big boon for mechanistic interpretability, since it's can be a lot more straightforward to interpret a bunch of {-1, 0, 1}s than reals. Not a silver bullet by any means, but it would at least peel back one layer of complexity.
Wouldn't the granularity of the action space also impact things? For example, even if a child struggles to pick up some object, you would probably do an even worse job if your action space was picking joint angles, or forces for muscles to apply, or individual timings of action potentials to send to separate nerves.
This is a cool model. I agree that in my experience it works better to study sentence pairs than single words, and that having fewer exact repetitions is better as well. Probably paragraphs would be even better, as long as they're tailored to be not too difficult to understand (e.g. with a limited number of unknown words/grammatical constructions).
One thing various people recommend for learning languages quickly is to talk with native speakers, and I also notice that this has an extremely large effect. I generally think of it as having to do with more of one's mental subsystems involved in the interaction, though I only have vague ideas as to the exact mechanics of why this should be so helpful.
Do you think this could somehow fit parsimoniously into your model?
A few others have commented about how MSFT doesn't necessarily stifle innovation, and a relevant point here is that MSFT is generally pretty good at letting its subsidiaries do their own thing and have their own culture. In particular GitHub (where I work), still uses Google Workspace for docs/email, slack+zoom for communication, etc. GH is very much remote-first whereas that's more of an exception at MSFT, and GH has a lot less suffocating bureaucracy, and so on. Over the years since the acquisition this has shifted to some extent, and my team (Copilot) is more exposed to MSFT than most, but we still get to do our own thing and at worst have to jump through some hoops for compute resources. I suspect if OAI folks come under the MSFT umbrella it'll be as this sort of subsidiary with almost complete ability to retain whatever aspects of its previous culture that it wants.
Standard disclaimer: my opinions are my own, not my employer's, etc.
It'd be great if one of the features of these "conversation" type posts was that they would get an LLM-genererated summary or a version of it not as a conversation. Because at least for me this format is super frustrating to read and ends up having a lower signal to noise ratio.
You have a post about small nanobots being unlikely, but do you have similar opinions about macroscopic nanoassemblers? Non-microscopic ones could have a vacuum and lower temperatures inside, etc.
Strong upvote for the core point of brains goodhearting themselves being a relatively common failure mode. I honestly didn't read the second half of the post due to time constraints, but the first rang true to me. I've only experienced something like social media addiction at the start of the Russian invasion last year since most of my family is still back in Ukraine. I curated a Twitter list of the most "helpful" authors, etc., but eventually it was taking too much time and emotional energy and I stopped, although it was difficult.
I think this is related to a more helpful, less severe version of the same phenomenon. When I get frustrated, sometimes it's helpful to accomplish some small household todo like cleaning the table or taking out the trash, and that helps me feel more in control/accomplished and helps me get back into a reasonable mood in which I can be happier and more productive.
Brief remarks:
- For AIs we can use the above organizational methods in concert with existing AI-specific training methodologies, which we can't do with humans and human organizations.
- It doesn't seem particularly fair to compare all human organizations to what we might build specifically when trying to make aligned AI. Human organizations have existed in a large variety of forms for a long time, they have mostly not been explicitly focused on a broad-based "promotion of human flourishing", and have had to fit within lots of ad hoc/historically conditional systems (like distributions between for profit vs non profit entities) that have significant influence on the structure of newer human organizations.
I grew up in Arizona and live here again now. It has had a good system of open enrollment for schools for a long time, meaning that you could enroll your kid into a school in another district if they have space (though you'd need to drive them, at least to a nearby school bus stop). And there are lots of charter schools here, for which district boundaries don't matter. So I would expect the impact on housing prices to be minimal.
Godzilla strategies now in action: https://simonwillison.net/2022/Sep/12/prompt-injection/#more-ai :)
No super detailed references that touch on exactly what you mention here, but https://transformer-circuits.pub/2021/framework/index.html does deal with some similar concepts with slightly different terminology. I'm sure you've seen it, though.
Is the ordering intended to reflect your personal opinions, or the opinions of people around you/society as a whole, or some objective view? Because I'm having a hard time correlating the order to anything in my wold model.
This is the trippiest thing I've read here in a while: congratulations!
If you'd like to get some more concrete feedback from the community here, I'd recommend phrasing your ideas more precisely by using some common mathematical terminology, e.g. talking about sets, sequences, etc. Working out a small example with numbers (rather than just words) will make things easier to understand for other people as well.
My mental model here is something like the following:
- a GPT-type model is trained on a bunch of human-written text, written within many different contexts (real and fictional)
- it absorbs enough patterns from the training data to be able to complete a wide variety of prompts in ways that also look human-written, in part by being able to pick up on implications & likely context for said prompts and proceeding to generate text consistent with them
Slightly rewritten, your point above is that:
The training data is all written by authors in Context X. What we want is text written by someone who is from Context Y. Not the text which someone in Context X imagines someone in Context Y would write but the text which someone in Context Y would actually write.
After all, those of us writing in Context X don't actually know what someone in Context Y would write; that's why simulating/predicting someone in Context Y is useful in the first place.
If I understand the above correctly, the difference you're referring to is the difference between:
- Fictional
- prompt = "A lesswrong post from a researcher in 2050:"
- GPT's internal interpretation of context = "A fiction story, so better stick to tropes, plot structure, etc. coming from fiction"
- Non-fictional
- prompt = "A lesswrong post from a researcher in 2050:"
- GPT's internal interpretation of context = "A lesswrong post (so factual/researchy, rather than fiction) from 2050 (so better extrapolate current trends, etc. to write about what would be realistic in 2050)"
Similar things could be done re: the "stable, research-friendly environment".
The internal interpretation is not something we can specify directly, but I believe sufficient prompting would be able to get close enough. Is that the part you disagree with?
Alas, querying counterfactual worlds is fundamentally not a thing one can do simply by prompting GPT.
Citation needed? There's plenty of fiction to train on, and those works are set in counterfactual worlds. Similarly, historical, mistaken, etc. texts will not be talking about the Current True World. Sure right now the prompting required is a little janky, e.g.:
But this should improve with model size, improved prompting approaches or other techniques like creating optimized virtual prompt tokens.
And also, if you're going to be asking the model for something far outside its training distribution like "a post from a researcher in 2050", why not instead ask for "a post from a researcher who's been working in a stable, research-friendly environment for 30 years"?
Please consider aggregating these into a sequence, so it's easier to find the 1/2 post from this one and vice versa.
Sounds similar to what this book claimed about some mental illnesses being memetic in certain ways: https://astralcodexten.substack.com/p/book-review-crazy-like-us
If you do get some good results out of talking with people, I'd recommend trying to talk to people about the topics you're interested in via some chat system and then go back and extract out useful/interesting bits that were discussed into a more durable journal. I'd have recommended IRC in the distant past, but nowadays it seems like Discord is the more modern version where this kind of conversation could be found. E.g. there's a slatestarcodex discord at https://discord.com/invite/RTKtdut
YMMV and I haven't personally tried this tactic :)
Well written post that will hopefully stir up some good discussion :)
My impression is that LW/EA people prefer to avoid conflict, and when conflict is necessary don't want to use misleading arguments/tactics (with BS regulations seen as such).
I agree I've felt something similar when having kids. I'd also read the relevant Paul Graham bit, and it wasn't really quite as sudden or dramatic for me. But it has had a noticeable effect long term. I'd previously been okay with kids, though I didn't especially seek out their company or anything. Now it's more fun playing with them, even apart from my own children. No idea how it compares to others, including my parents.
Love this! Do consider citing the fictional source in a spoiler formatted section (ctrl+f for spoiler in https://www.lesswrong.com/posts/2rWKkWuPrgTMpLRbp/lesswrong-faq)
Also small error "from the insight" -> "from the inside"
The most similar analysis tool I'm aware of is called an activation atlas (https://distill.pub/2019/activation-atlas/), though I've only seen it applied to visual networks. Would love to see it used on language models!
As it is now, this post seems like it would fit in better on hacker new, rather than lesswrong. I don't see how it addresses questions of developing or applying human rationality, broadly interpreted. It could be edited to talk more about how this is applying more general principles of effective thinking, but I don't really see that here right now. Hence my downvote for the time being.
Came here to post something along these lines. One very extensive commentary with reasons for this is in https://twitter.com/kamilkazani/status/1497993363076915204 (warning: long thread). Will summarize when I can get to laptop later tonight, or other people are welcome to do it.
Have you considered lasik much? I got it about a decade ago and have generally been super happy with the results. Now I just wear sunglasses when I expect to benefit from them and that works a lot better than photochromatic glasses ever did for me.
The main real downside has been slight halos around bright lights in the dark, but this is mostly something you get used to within a few months. Nowadays I only noticed it when stargazing.
This seems like something that would be better done as a Google form. That would make it easier for people to correlate questions + answers (especially on mobile) and it can be less stressful to answer questions when the answers are going to be kept private.
How is it that authors get reclassified as "harmful, as happened to Wright and Stross"? Do you mean that later works become less helpful? How would earlier works go bad?
Given that you didn't actually paste in the criteria emailed to Alcor, it's hard to tell how much of a departure the revision you pasted is from it. Maybe add that in for clarity?
My impression of Alcor (and CI, who I used to be signed up with before) is that they're a very scrappy/resource-limited organization, and thus that they have to stringently prioritize where to expend time and effort. I wish it weren't so, but that seems to be how it is. In addition, they have a lot of unfortunate first-hand experience with legal issues arising during cryopreservation due to family intervention, which I suspect is influencing their proposed wording.
I would urge you to not ascribe to malice or incompetence what can be explained by time limitations and different priors. My suspicion is that if you explain where you're coming from and why you don't like their proposed wording (and maybe ask why they wanted to change some of the specific things you were suggesting) then they would be able to give you a more helpful response.
Given other sketchy things I've read about them (there is plenty of debate on this site and elsewhere calling them out for bad behavior)
I don't follow things too closely but would be interested in what you're referring to, if you could provide any links.
Downvoted for lack of standard punctuation, capitalization, etc., which makes the post unnecessarily hard to read.
Do you mean these to apply at the level of the federal government? At the level of that + a majority of states? Majority of states weighted by population? All states?
Thanks! Reversed :)
Downvoted for burying the lede. I assumed from the buildup this was something other than what it was, e.g. how a model that contains more useful information can still be bad, e.g. if you run out of resources for efficiently interacting with it or something. But I had to read to the end of the second section to find out I was wrong.
Came here to suggest exactly this, based on just the title of the question. https://qntm.org/structure has some similar themes as well.
Re: looking at the relationship between neuroscience and AI: lots of researchers have found that modern deep neural networks actually do quite a good job of predicting brain activation (e.g. fmri) data, suggesting that they are finding some similar abstractions.
Examples: https://www.science.org/doi/10.1126/sciadv.abe7547 https://www.nature.com/articles/s42003-019-0438-y https://cbmm.mit.edu/publications/task-optimized-neural-network-replicates-human-auditory-behavior-predicts-brain
I'll make sure to run it when I get to a laptop. But if you ever get a chance to set the distill.pub article up to run on heroku or something, that'll increase how accessible this is by an order of magnitude.
Sounds intriguing! You have a GitHub link? :)
The biggest rationalist-ish issue for me has been my partners not being interested (or actively disinterested) in signing up for cryonics. This has been the case in three multi-year relationships.
You'd be more likely to get a meaningful response if you sold the article a little bit more. E.g. why would we want to read it? Does it seem particularly good to you? Does it draw a specific interesting conclusion that you particularly want to fact-check?
I really loved the thorough writeup and working of examples. Thanks!
I would say I found the conclusion section the least generally useful, but I can see how it is the most personal (that kinda why it has a YMMV feel to it for me).
Reverse osmosis filters will already be more common in some places that have harder water (and decided that softening it at the municipal level wouldn't be cost-effective). If there was fine grained data available about water hardness and obesity levels, that might provide at least a little signal.
There's a more elaborate walkthrough of the last argument at https://web.stanford.edu/~peastman/statmech/thermodynamics.html#the-second-law-of-thermodynamics
It's part of a statistical mechanics textbook, so a couple of words of jargon may not make sense, but this section is highly readable even without those definitions. To me it's been the most satisfying resolution to this question.
Nice video reviewing this paper at https://youtu.be/-buULmf7dec
In my experience it's reasonably easy to listen to such videos while doing chores etc.
https://youtu.be/QMqPAM_knrE is a video by one of the authors presenting on this research
The problem definition talks about clusters in the space of books, but to me it’s cleaner to look at regions of token-space, and token-sequences as trajectories through that space.
GPT is a generative model, so it can provide a probability distribution over the next token given some previous tokens. I assume that the basic model of a cluster can also provide a probability distribution over the next token.
With these two distribution generators in hand, you could generate books by multiplying the two distributions when generating each new token. This will bias the story towards the desired cluster, while still letting GPT guide the overall dynamic. Some hyperparameter tuning for weighting these two contributions will be necessary.
You could then fine-tune GPT using the generated books to break the dependency on the original model.
Seems like a fun project to try, with GPT-3, though probably even GPT-2 would give some interesting results.
Ok, I misread one of gwern's replies. My original intent was to extract money from the fact that gwern gave (from my vantage point) too high a probability of this being a scam.
Under my original version of the terms, if his P(scam) was .1:
- he would expect to get $1000 .1 of the time
- he would expect to lose $100 .9 of the time
- yielding an expected value of $10
Under my original version of the terms, if his P(scam) was .05:
- he would expect to get $1000 .05 of the time
- he would expect to lose $100 .95 of the time
- yielding an expected value of -$45
In the second case, he would of course not want to take that bet. I'd thus like to amend my suggested conditions to have gwern only put $52 at stake against my $1000. For any P(scam) > .05 this is a positive expected value, so I would expect it to have been satisfactory to gwern[19 August 2012 01:53:58AM].
Well I still accept, since now it's a much better deal for me!
Done. $100 from you vs $1000 from me. If you lose, you donate it to her fund. If I lose, I can send you the money or do with it what you wish.
There are a lot of things I'd like to say, but you have put forth a prediction
It's probably a scam
I would like to take up a bet with you on this ending up being a scam. This can be arbitrated by some prominent member of CI, Alcor, or Rudi Hoffman. I would win if an arbiter decides that the person who posted on Reddit was in fact diagnosed with cancer essentially as stated in her Reddit posts, and is in fact gathering money for a her own cryonics arrangements. If none of the proposed arbiters can vouch for the above within one month (through September 18), then you will win the bet.
What odds would you like on this, and what's the maximum amount of money you'd put on the line?