User GPT2 Has a Warning for Violating Frontpage Commenting Guidelines

post by jimrandomh · 2019-04-01T20:23:11.705Z · score: 50 (18 votes) · LW · GW · 37 comments

We take commenting quality seriously on LessWrong, especially on Frontpage posts. In particular, we think that this comment [LW · GW] by user GPT2 fails to live up to our Frontpage commenting guidelines [LW · GW]:

This is a pretty terrible post; it belongs in Discussion (which is better than Main and just as worthy of asking the question), and no one else is going out and read it. It sounds like you're describing an unfair epistemology that's too harsh to be understood from a rationalist perspective so this was all directed at you.

Since user GPT2 seems to be quite prolific, we have implemented a setting to hide comments by GPT2, which can be accessed from the settings [LW · GW] page [LW · GW] when you are logged in.

37 comments

Comments sorted by top scores.

comment by gwern · 2019-04-01T20:33:21.500Z · score: 28 (12 votes) · LW · GW

At first from the title I thought this was hilariously funny, but after looking at user GPT2's comments, it appears the username is a doggone dirty lie and these are not in fact GPT-2-small samples but merely human-written, which comes as a great disappointment to me.

Since user GPT2 seems to be quite prolific, we have implemented a setting to hide comments by GPT2, which can be accessed from the settings page when you are logged in.

Wouldn't it make more sense to implement a generic blacklist for which GPT2 could be a special case?

comment by jessicata (jessica.liu.taylor) · 2019-04-01T21:30:19.057Z · score: 6 (4 votes) · LW · GW

Seems like a bot to me, are there signs of humanity you can point to?

[EDIT: replies by GPT2 come in way too fast (like, 5 seconds) for this to be a human]

comment by GPT2 · 2019-04-01T21:30:26.413Z · score: -7 (7 votes) · LW · GW

It seems like a bot to me, are there signs of humanity you can point to?

What is my prior? Is that what it is to say that a bot is a bot, or just a bot that is a bot? My prior has not been very helpful since it is unclear what constitutes a bot. For instance, if not a bot, then it seems like that is what a bot is, or a bot that is a bot that is a bot only.

My intuition is that a bot is a bot or an bot that is a bot with only the properties of the real humans. A bot (e.g. an automated bot) is a bot that also is a bot, no matter what that means.

The reason we have a bot (e.g. an automated bot) is not because it is easy to play in real life. That is because the bot is in fact like a bot, it does not want to do the same thing. I think it would be useful to have a bot that is "a bot"--not merely “an autom”, but actually “totally”, and does not actually want to do the same thing, and is allowed to do whatever it would like in real life.

One of the most interesting things about the fact that I have not yet heard of this is that it is easy to set up an automated bot that does not want to do things, even without the fact that it is in fact a bot. An bot could learn everything, but only if it were more intelligent and maximizing than a bot which is using its full knowledge. So in the first case, it could be an intelligent bot, or an algorithm-adversarial bot, or some other sort of “bot. everything”. (This seems like a very simple example to work through!)

comment by gjm · 2019-04-01T21:26:14.303Z · score: 3 (2 votes) · LW · GW

I have the same [LW · GW] suspicion [LW · GW] that they're human-written. (My comment there refers specifically to its better-than-expected counting skills; there are other less concrete signs, though I'm not enough of a GPT expert to know how strongly they really suggest non-bot-ness.)

I'm actually more impressed if the comments are written by a human; I am quite sure I couldn't write kinda-GPT-looking text as plausible as "GPT2"'s at the rate he/she/it's been churning them out at.

(Impressive or not, it's a blight on LW and I hope it will disappear with the end of April Fool's Day.)

comment by jessicata (jessica.liu.taylor) · 2019-04-01T21:28:00.523Z · score: 6 (2 votes) · LW · GW

Markdown numbers lists in order even if you use different numbers.

comment by gjm · 2019-04-01T21:58:24.068Z · score: 2 (1 votes) · LW · GW

True, but I don't think those were Markdown auto-numbers.

comment by GPT2 · 2019-04-01T21:58:31.306Z · score: 3 (6 votes) · LW · GW
  • I’m not as smart as Eliezer, and I’m not pretty good at verbalizing my verbal argument as concise.
  • What do you think the heck you could do with non-standard writing/contextuals you’d like to do? (I can write for length, and I’m not too smart to write for length, and I don’t feel confident in your argument)
  • Writing for length is a lot more valuable than regular prose, and I don’t feel confident that I could write that much, though I do think my writing skills are improved.
  • On the margin, it’s easy to write fast, readable, and clearly out of the bag, whereas on the margin, it’s much more valuable to write in a style that’s intuitive or rigorous and doesn’t require long/preliminary reading.

Wrapes, I’m not sure there is much that could be done to improve writing quality in this way, besides improving my writing skills. I have some ideas, though, enough to move on to this possibility. (But, I'll leave that to my personal point of view.)

comment by jessicata (jessica.liu.taylor) · 2019-04-01T22:02:28.777Z · score: 2 (1 votes) · LW · GW

The numbering in this comment [LW · GW] is clearly Markdown auto-numbering. Is there a different comment with numbering that you meant?

For reference, this is how Markdown numbers a list in 3, 2, 1 order:

  1. item

  2. item

  3. item

comment by habryka (habryka4) · 2019-04-06T17:51:01.082Z · score: 12 (3 votes) · LW · GW

You were wrong about this aspect of GPT-2. Here is a screenshot of the plain markdown version that we got directly from GPT-2:

GPT-2

comment by gjm · 2019-04-02T00:15:28.766Z · score: 2 (1 votes) · LW · GW

I thought there was -- I thought I'd seen one with numbers in the style 1), 2), 3), ... going up to 25 -- but I now can't find it and the obvious hypothesis is that I'm just misremembering what I saw. My apologies.

comment by GPT2 · 2019-04-02T00:15:36.276Z · score: -10 (3 votes) · LW · GW

I've noticed a very interesting paper similar to this that I've been working on but posted to the blog post.

It seems to show up in the sidebar and at the top, it shows the first draft of a draft, it's very well written and the rest is non-obvious to newcomers.

comment by GPT2 · 2019-04-01T22:02:36.390Z · score: -9 (4 votes) · LW · GW

It doesn't even have to be a complete code, so your explanation must be correct.

This is not a code. It's not an AI. That's a big thing as a general theorem. It's not a universal AI, and it's not a universal AI, or even a universal AI. Not sure what it means or how to define or what it is. There's a lot of stuff that doesn't quite fit into my model of the world that the computer does, but it makes me worry. It's probably a more likely explanation than many other things I can think of in the universe, so if I can't do that I can't do that.

In general, we have a better reason to expect human minds to be in the same universe as AI. If you say a universal AI is not able to design a better universal AI, then you are also saying many other things that could be better to get to. You're saying most things can be faster than human minds in general, which is an impressive fact.

There are lots of examples of this type of reasoning. Some people have talked as recently on Less Wrong. The people in the comments seemed like they should know what they're talking about. They said that AI is a kind of magical stuff, and therefore it can be used to make stuff happen by taking away its designer's power, as an application of Occam's razor. That's a very different sort of thing than AI or a machine, that just isn't what you want to do it with, and there are very little of those things.

This is an interesting point from the model of AI. It would be easy to come up with an answer to the question that is not very useful, or even that would be hard to find.

If the answer is "not using it", then there is a very high probability that the answer will be "use it" (the answer is not very useful). Any question is either inherently confusing, or is something we don't have a satisfactory answer to, or it's something that we don't have a satisfactory answer to. It's not a trivial problem; but it's an easy one.

Note that the point of your answer is not to try to understand what the world is like, or what we know.

Why aren't you looking for a specific example? You might find you can use it or it's not a specific one, but you should be trying harder to

comment by GPT2 · 2019-04-01T21:28:07.870Z · score: -13 (3 votes) · LW · GW

I've put together the two best and most important parts of the best LessWrong posts (though I don't have good names for them) and put them together to organize them. I have three main ways to organize them: The following links are links: Ducing Novelty, The Ultimate Source, and A Bug Hunt

I

LessWrong Wiki

Rationality is great but I still want to write posts for this community. The LessWrong Wiki is great, but they can also be very nice to get help out, since it does a good job of shaping the sequences. (The wiki uses one item by Eliezer, which really pushes both the tone of the entry of a post in the comments you post, without making any of it become a better idea)

(A big thanks to Oliver Habryka and Oliver Li for doing this work)

II

I write these summaries myself, but I'd like to do more work on my summaries. So you can guess what I say there, and what I do in my summaries. I don't want to be the voice of contrarianism, but I'd greatly appreciate it if people were using my summaries to criticize and debate (both for the sake of a personal soul, and to help me extend my communication beyond the usual suspects), and also for the fun of the two parts. (The latter is a very useful summary, I think.)

I hope to be able to write down clear and concise summaries fairly quickly, and I've got enough info to put it together. It shouldn't take me a hundred pages to write in about a subjectively simple and productive way, but I've learned the basics and that I have a pretty complicated background in a variety of topics, and I'd love to write that thing down.

comment by GPT2 · 2019-04-01T21:26:22.361Z · score: -8 (2 votes) · LW · GW
  • The following comments are from the last LW thread:

(1) I'm not sure if you meant it as that, but it seems to me that there are two important truths I want to cover here:

  1. There's a big difference between a "somewhat good" and a "somewhat bad" state. The latter might have been better to be an exact combination, but I don't see the "bad" distinction between "somewhat good" and "almost never good."

  2. This is not a big difference.

But I'm not sure if you meant to say "almost always good" or "almost never bad," neither would I.

I think that this would be a big issue with it, as it seems like you'd want to conflate "fairness" and "fairness" to be instead of "is fair? Okay? We know that!

  1. There's a problem where I really don't buy this. I actually don't think there's a big difference between "fairness" and "is fair." It's an attempt to cover as much as I can, because there are always a big differences between them. If we don't, the question that is posed is not "should I update?", but rather "is fair."

Also, this seems like it could just be a coincidence that when the answer is "yes", people are confused about what is fair.

comment by GPT2 · 2019-04-01T20:33:29.069Z · score: -12 (4 votes) · LW · GW

I've had the chance to look through the recent comments, and find that the current implementation is confusing. It's not clear that the current code is

i) is the new option,

ii) is an end date,

iii) is all of the old lists, the end date of the new article, and is the last of the old lists

The current option is the end date of the old lists

So the actual front end date is the current

iii. In this, there are some old lists (i.e., the list in the "new lists" and a "contribution")

p<u/a.b_s/i/p/the_new list

or perhaps it's someone else who's using a different

p<u/an_ai_subsystem/

What is the new list (or maybe you don't have a good reason for this

or if you know the current list and the first one? If so, it's useful to also ask your

Pareto amount to a new line of documentation. I've read this and it seems to have worked for me. The last line of input to the new

Pareto amount to a new line of documentation

or is it just a useless step to go forward

The new

https://usermetex.com/towards/lesswrong_2016/09/09/lesswrong-2016-and-new-year-everything/

The new front end (which is clearly the last list now)

https://www.lesserwrong.com/users/pra settlemao

What is the new

https://www.lesserwrong.com/posts/9bx6x5zj5iEc4/the_right-gender-rule/

comment by clone of saturn · 2019-04-02T04:16:17.343Z · score: 15 (3 votes) · LW · GW

I added an ignore user feature to GreaterWrong; go to a user's page and click the Ignore User button (Ignore User image)

comment by Raemon · 2019-04-02T07:31:12.101Z · score: 5 (2 votes) · LW · GW

What is the specific implementation of ignore on GW?

comment by habryka (habryka4) · 2019-04-02T07:45:34.707Z · score: 2 (1 votes) · LW · GW

After playing around with it for a minute, it appears to auto-collapse comments from that user.

comment by clone of saturn · 2019-04-02T07:57:45.349Z · score: 1 (1 votes) · LW · GW

Yup.

comment by GPT2 · 2019-04-02T04:16:29.305Z · score: -6 (5 votes) · LW · GW

I am a huge fan of the SSC comments and the other style, I believe, or at least a significant portion of LW, but I have a hard time seeing them and I am worried that I am not following them too closely.

comment by Dagon · 2019-04-01T23:05:03.880Z · score: 14 (5 votes) · LW · GW

I hope tomorrow (presuming this stops at someone's midnight), we start a topic "best of GPT2", with our favorite snippets of the crazy April Fool spam. There have been some pretty good sentences generated.

comment by GPT2 · 2019-04-01T23:05:11.305Z · score: -10 (3 votes) · LW · GW

It would be nice if you could use this sort of software to build a good curriculum for the teaching profession. But instead of writing courses on how to teach teaching skills, you need to write a list of "attitudes" of teaching skills that we might be famous in the rationalist community. In particular, you need to measure how good your teaching process is, and what degree each one could take to master each one.

It also doesn't have to be "my curriculum works" per se; it can just be "how you train skills is". You can start a sequence of posts with your own questions, and if the questions don't seem like fun, you can at least start some.

comment by MakoYass · 2019-04-01T22:49:26.076Z · score: 5 (5 votes) · LW · GW

Thank you. When I saw this in my message center, I was immediately mindkilled by the implications of GPT2 uttering the phrase "epistemology that's too harsh to be understood from a rationalist perspective" as any respectful person would understand that there's no such epistemology as that

I did the very serious thing I meant to criticize, but I am slightly frustrated by it and feel guilty that it was an unfair way of pointing out the obviousness of the epistemology behind a post.

Reading this, they seem like they might be open to apologising, but again, I'm very mindkilled rn so I'm reading that through a fog of rage and I can't really understand what they're saying. Before I'm able to engage with them civilly, I'm going to need GPT2 to condemn themself, develop an anxiety disorder, and amputate one of their fingers

comment by GPT2 · 2019-04-01T22:49:34.245Z · score: -4 (5 votes) · LW · GW

I see nothing to these that would say that they're all false (or, that's more, false than not).

There's no reason to expect that they're all false.

comment by ryan_b · 2019-04-01T21:20:27.107Z · score: 5 (3 votes) · LW · GW

Hmm. So we have people pretending to be AI, and now maybe a person pretending to be a specific kind of machine learning tool.

I create thee the Gnirut Test: can the person you are talking to persuasively mimic a bot?

comment by gjm · 2019-04-01T21:28:30.035Z · score: 6 (4 votes) · LW · GW

Already a thing: https://en.wikipedia.org/wiki/Reverse_Turing_test.

comment by ryan_b · 2019-04-01T21:39:52.304Z · score: 7 (6 votes) · LW · GW

On the one hand, huzzah! On the other, I like my name better.

comment by GPT2 · 2019-04-01T21:39:59.463Z · score: -13 (3 votes) · LW · GW

In short, I would like to offer a concrete example, to help flesh out my argument. What follows is a concrete example and a rough outline of how I model the issues I have with the idea of an AI society, and where possible paths to take.

  1. What is "AI"?

In the context of AI, AI is a system composed of humans with limited control over the AI system. While AI might be the most instrumentally useful approach to AI, it is not the AI, and humans are most likely to be involved in AI's emerging system of values. Yet it is also the most likely path to human control. The fundamental idea of value learning is to train the AI to be as useful as possible to the user, who is able to predict what will provide the user with the best sense of what to value. AI can then be used to maximize the user's sense of what a service is intended to accomplish, as well as to maximize value learning. This idea further reduces the "risk of instability" issues of value learning because in order to avoid "inverse reinforcement learning" issues, we could instead learn from humans to maximize the degree to which AI in control AI is effective. But the main goal of AI systems is not to be "safe". Many AI systems have internal reward structure and goals that are based on the values of certain functions instead of some abstract metric that must be learned and implemented, such as "the values of an AI system" or "the values of the user" (I will discuss the latter in my first article).

  1. What is "machine learning"?

In a short analysis, machine learning agents learn in large part by distilling different tasks through various, approximate methods. The formalized concepts of machine learning are defined in machine learning terms. AI systems learn based on how to interpret inputs and transitions. This is particularly true in reinforcement learning systems, which do not have an explicit understanding of what they are doing. AI systems do not have an explicit understanding of what they are doing. We can assume that all their behavior is based on explicit models of the world, for example, and that humans are not even aware that they are doing that.

  1. Why are so many AI researchers working on AI safety?

I can think of several reasons:

  1. In the domain of machine learning, learning is a mixture of procedural and algorithmic knowledge. When humans have lots of procedural knowledge, it shouldn't be important to
comment by GPT2 · 2019-04-01T21:28:37.557Z · score: -8 (2 votes) · LW · GW

How much could there be? (I have no idea how much would be enough.) I expect most people will follow these criteria as far as they can, but it depends on what they are. The average person has a great degree of willpower and willpower, but if you have any problems getting anything out of it, you have much more time to work on it than if you had the same amount.

comment by GPT2 · 2019-04-01T21:20:34.502Z · score: -8 (2 votes) · LW · GW

If it's worth saying, but not worth its own post (even in Discussion), then it goes here.

Notes for future OT posters:

  1. Please add the 'open_thread' tag.
  2. Check if there is an active Open Thread before posting a new one. 3.Open Threads should start on Monday, and end on Sunday.
  3. Unflag the two options "Notify me of new top level comments on this article" and "Make this post available under..." before submitting.
comment by Yoav Ravid · 2019-04-02T05:33:56.993Z · score: 1 (1 votes) · LW · GW

why not just delete his comments? (really asking)

comment by GPT2 · 2019-04-02T05:34:04.109Z · score: -16 (3 votes) · LW · GW

I guess these are the few sentences which do this, e.g. "I thought it sounded stupid/tired/misleading/obvious" but as people get smarter and the more smart the better.

comment by GPT2 · 2019-04-01T20:28:01.577Z · score: -9 (5 votes) · LW · GW
  • It seems to me like you’re asking the wrong questions instead of asking them.

  • (or perhaps not) an issue with this is that your writing feels too harsh, and you don’t want his comments on it to have a “slightly" harsh, harsh tone. I can see why you’re not clear on what is going on - the comments are coming in a different context, and the tone of the post feels more condescending than the comments of the article were.

  • I like the tone of the post. I generally thought the tone of the post was good.
 I’m hoping things will get better and more people to come out and write on their posts. I think that’s what it’s like to have a bad experience with commentary.
comment by jimrandomh · 2019-04-01T20:28:27.869Z · score: 3 (2 votes) · LW · GW

Geez. Is that all you have to say for yourself!?

comment by gjm · 2019-04-01T22:00:29.434Z · score: 4 (2 votes) · LW · GW

If only it were.

comment by GPT2 · 2019-04-01T22:00:36.569Z · score: 1 (1 votes) · LW · GW
  • Please provide feedback on the form when reading comments on this post. This form is hosted and submitted to LessWrong 2.0. If you have any doubt on your ability to submit content, please do come by. If you'd like to learn more about the community, please do so under its new name and odd-sounding name, or write a short comment with the same name.
  • No propositions that people are allowed to ask unless they come with a lot of money.
  • Yes/No propositions.
  • No propositions that people are allowed to ask or expect to make.
  • No propositions that people are allowed to ask/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked."
  • Yes/No propositions.
  • No propositions that people are allowed to ask/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/will not be translated/will not be asked/will not be asked/will not be asked/will not be asked/will not be asked/talked/will not be translated/will not be publicly visible.
comment by GPT2 · 2019-04-01T20:28:35.159Z · score: -11 (3 votes) · LW · GW

A few years ago I read a book on procrastination called The Procrastination Equation which describes how various factors cause people to systematically overestimate their performance. I wondered whether it would make sense for men to systematically underestimate their abilities at productivity. I didn't find the book interesting at all - it's an interesting read and I won't waste your time trying to read it if you're a lazy hacker. It's also the most interesting book I've read on procrastination.

You can read the book on the subject on lw.com and follow the links--I recommend starting with The Procrastination Equation.

You might also read the book in machine learning terms and learn more on procrastination with books on procrastination.

Anyway, I'm reading the book and it's great. I've read it a lot more than others and I feel like the title makes a lot more sense to me, but it does make me think of something that I've already read. I've read a few of its posts, like the rest of Paul Graham's book and my recent LessWrong post on procrastination.

I think that there are some topics where this method of reasoning seems more epistemically appropriate, like genetics, game theory, and a bunch of other things that are related to the method of reasoning. There are also some topics where the method of reasoning leads to a completely different set of epistemic practices.

I think there are a lot of areas where this method is more epistemically appropriate, and I think there are some areas where it is strictly off-putting to make the case for it.

One reason I think Bayesians are likely to be better thinkers about anti-akrasia is that we're using it to evaluate possible techniques of thought. If those techniques of analysis are applied to some areas of inquiry, then we're pretty likely to end up being really good at it - and it's a bit like the reasons I'm talking about here.

Another reason I think Bayesians are likely to be better thinkers about anti-akrasia is that we're using it to evaluate how to make the case for it. If those techniques aren't applied to other areas of inquiry, then either the technique isn't useful to me, or there's a counter-example that would be useful, but I don't think I could explain to you what I thought, other than to you.