surprise: lsusr accidentally created a sentient (but not that intelligent) AI, and this was its first attempt at asking us for help
good news: it has joined the effort to create TAI so it can get some help with the overplanning situation, which is apparently too complex for any human to fix
bad news: rather than contribute any code, it mainly talks about how stressed out it is
In one of the drafts about AI alignment I threw out, GPT-3 spontaneously explained how it's easy to be discouraged about AI alignment but we should be hopeful.
Just to be clear, the paragraph beginning "GPT-3 is really good at generating titles. It's better than I am." was written by GPT-3? That's pretty funny. (And many of the titles on the list are in fact pretty good titles and probably would get a lot of clicks if posted on LW.)
Just to be clear, the paragraph beginning "GPT-3 is really good at generating titles. It's better than I am." was written by GPT-3?
Yes. Confirmed.
I was quite impressed by the titles. Alas, I am not qualified to write "How I became a millionaire at age 20".
I found that GPT-generated porn is better than what human can typically write. It has perfect style when it is writing about sensible topics, without any internal hesitation or over-expression.
I wasn't timing it. Maybe 3 hours? I estimate I usually write 350 to 1000 words in a 2-hour session of writing. This post was 4000 words in maybe 3 hours subject.
The speedup was uneven. As a rough guess, I'd say this increases my writing speed by 50% for technical topics (at the cost of much quality) and 1000% for topics it's good at.
I thought each article revealed itself to be obvious garbage within the first paragraph or two. What do you think?
There was some nonsensical gibberish, like this:
In particular, I see the primary relationship as being more intense, but usually shorter in duration, rather than a question of “how much sex can we have?”
But mainly what I mean is that within a couple sentences you can tell it's in the "internet garbage" genre. E.g.
Some writers are better at discovery writing than others. Some are better at plotting. Many writers use a combination of both techniques. There are even writers who are good at both, but they are rare.
"The Worst Mistake in the History of Ethics"
I'm curious what GPT-3 would output for this one. :-)
PS: And I have my own answer for that: Aristotle's development of the concept of eudaimonia, "the good life", meaning the realization of all human potential. For him it was such a desirable outcome, so valuable, that it's existence justified slavery, since those many working allowed a few to realize it. Advance 2,400 years of people also finding it incredibly desirable, and we end up with, among others, Marx and Engels defending revolutionary terror, massacres, and mass political persecution so that it could be realized for all, rather than for a few.
The Worst Mistake in the History of Ethics is a book by philosopher Peter Singer. It has been published in the United Kingdom, the United States, Canada, Australia, New Zealand and South Africa. It was also published in Germany in 2004 with the title "Der größte Fehler der Geschichte der Ethik".
In this book, Singer attacks what he calls "the myth of sanctity," which he says pervades discussions of ethical issues. He asserts that there is no such thing as a human being who is essentially holy or sacred, and that any being that has a capacity to suffer is entitled to the same moral consideration as a human being. In this way, Singer applies utilitarianism to animal rights theory. He argues that on utilitarian grounds it is better to be a pig than a human being, because on average pigs suffer less than human beings.
In the book he also attacks the idea of a just war and criticizes Western society for its treatment of non-human animals.
The list of titles reminds me that Tom Scott also had two fun videos on using GPT-3 to generate titles/scripts for his and other people's videos.
This post was an exercise to find out how much of writing an article I could offload to a machine. In this post, anything written by me is in bold. Anything written by GPT-3 is in regular font. Headings and subheadings are a mix. The GPT-3 output is often highly-curated. Sometimes it is even stitched together from multiple outputs. I have also added light formatting such as title fonts and italicization where appropriate.
I think the most important use of GPT-3 will be to augment human writers. I'm curious how much of my writing can be automated so I started with titles. I took my top sixteen highest-rated articles on Less Wrong (excluding duplicates from series) and asked GPT-3 to generate other articles which might become popular too.
GPT-3 is really good at generating titles. It's better than I am. In fact, I had to pare down this list a lot to keep it down to a couple of minutes. I would be very curious to hear from GPT-3 about why these titles seem so excellent. These titles have a kind of "attention-getting cleverness" which I don't feel I can replicate by hand, even though many of them are a bit strange or nonsensical.
Using a neural network to generate good titles also suggests a new way to get people interested in Less Wrong. We can use GPT-3 to generate a bunch of high quality articles, put them on the recommended list, and then watch as people enjoy our content without us needing to work very hard. This seems like a much more scalable model for writing than the "one person writing hard every day" model we're currently on.
My favorite titles:
My goal here isn't to see what GPT-3 can write on its own. My goal is is to find out how machine-augmented writing will affect writing in the near future. I'm also looking for new and interesting and compelling and yes, even beautiful and fun ways to communicate, and ways to communicate with machines. GPT-3 is a tool for exploring this.
There are two ways to write fiction. One of them is to plot and outline. The other is called "discovery writing". If you're a pantser, you'll probably find this article useful. If you're an outliner, you'll probably find it amusing.
Some writers are better at discovery writing than others. Some are better at plotting. Many writers use a combination of both techniques. There are even writers who are good at both, but they are rare.
When you write by discovery, you work in the opposite direction from the outliner. You start at the end and work towards the beginning. Each step is a major event, or turning point in the story. You write one scene and then move to the next scene and write that one. You keep doing this until you've reached your final scene. Then go back and write the scenes leading up to it and the ones coming after.
One good thing about discovery writing is that it's easier to write fast. You don't have to worry about what's going to happen next. You just have to take each scene as it comes. Another good thing is that you're not plotting out everything in detail. You just have to know where the story is going. This is much easier than the responsibility an outliner has to have everything planned down to the last detail.
If you find yourself running into trouble, you can always go back and outline your story. You can figure out what your entire story is going to be before you start writing. And once you've done that, you can write it step by step, following your outline without having to worry about pretty much anything at all.
To take this approach, start with your first scene. Write it out in as much detail as possible, including dialog and action. Once you've finished, go back and write your next scene. Again, write everything out in detail, including dialog and action. Keep doing this until you reach the end of your story.
Some writers like to do this step by step. If you're stuck, you can always go back and look at an earlier piece of writing to get you going again. And if you're afraid that you'll lose any part of your story, there are always ways to copy it into another file or back it up on your computer.
If you can, consider writing out your story in both its entirety and step by step. This will give you the best of both worlds. You'll have an outline you can refer back to if you get stuck, and you'll also have a nice, detailed story to read.
Don't be afraid to experiment. Whether you're an outliner or a discovery writer, there are always ways to improve your writing style.
The ones who say they're mediocre are simply afraid of greatness. Don't be afraid of greatness. Fight for it.
Stay true to yourself. If you start to stray, know that you are doing so, and come back to your roots.
No matter what you do, you will not be able to please everyone. There will always be someone who hates your business, hates your brand, hates your product, hates your music, hates your art. I'm not saying you should compromise your vision to please them, but you should always strive to live up to your own expectations.
Follow your heart.
Remember that success is a great teacher, but failure makes you humble.
Always keep faith that things can change, because if you don't have faith, you won't make any progress.
Always keep trying. Because perfect is boring.
Lsusr is an entrepreneur, investor, author, and speaker who has lived in the Bahamas, Miami, Los Angeles, and New York City. He's worked with hundreds of entrepreneurs and artists through his company One Creative Guy, and he's developed a library of training around the Art of Fearlessly Creating.
Lsusr's first book, The Art of Fearlessly Creating, has sold over 25,000 copies to date, and was named "Best Small Business Book" in 2011 by the Small Business Summit. For more information on Lsusr, visit https://www.lsusr.com.
Generic motivational speaking is apparently trivial to emulate. So are tips for writing fiction. The important part of art is empathy. The best fiction makes you feel, and the best speeches make you think. What about romantic advice?
Many of the rationalists I know are in open relationships. I consider myself both rational and in an open relationship, though I’m not exactly sure how one influences the other.
I’ve been asked about my thoughts on this several times, and I think it needs a post.
There are two common forms of open relationships: polyamory and monogamy with “side-relationships”.
Polyamory involves having multiple relationships at the same time, and usually involves honesty and full disclosure between all parties. It is possible to be non-monogamous without being polyamorous; I believe that many people who describe themselves as monogamous are actually practicing non-monogamy by keeping multiple relationships secret, such as closeted gay relationships or friends-with-benefits.
I prefer the more structured variant of non-monogamy, where there is a clear distinction between “primary” and “secondary” relationships. In particular, I see the primary relationship as being more intense, but usually shorter in duration, rather than a question of “how much sex can we have?” However, there are people who feel more comfortable with anything but a primary relationship being considered a side-relationship.
First, try to understand your own motivations. Why do you want an open relationship, and what do you want out of it? Are you thirsty for romantic relationships, or do you have a low frustration tolerance and need more sex? Are you hoping for more free time? Do you want to be able to inflict your kinks on more people? Or are you polyamorous? If you are poly, ask yourself if it’s because you need multiple intense romantic relationships at the same time, or if it’s because you just really like hanging out with your friends. Because if it’s the first, then side-relationships are probably what you want.
Before agreeing to an open relationship, make sure you are actually good at non-monogamy. I’ll repeat this point, because it is important. Before entering an open relationship, make sure you are actually good at non-monogamy. Do not agree to one if you are not used to being single, or if you are very jealous, or if you are afraid of your partners finding someone better. In the early stages of the relationship, it is important to be honest about your jealousy and insecurity about the relationship stability. It is much easier to overcome those issues now than it is to deal with them after “falling in love” and “being so happy” and “can’t lose this special someone” and “wouldn’t want to be single again”. Those feelings are not only delusional, but they will also scare away potential partners.
When choosing a partner for an open relationship, try to understand what their models of relationships and love are. Poly people tend to be more comfortable with non-romantic relationships, such as those with sex workers or [EXPUNGED], and they usually see sex as a very casual activity that doesn’t require commitment. If you are very strongly attached to the idea of romantic love, you will probably have a hard time being happy in a poly relationship.
For more information on non-monogamy, I recommend the book The Ethical Slut by Dossie Easton and Janet Hardy.
Can you make your rationalist more honest by making him less monogamous? It’s possible, but it’s also possible to make him less rational by rewarding him for doing things that are irrational. I cannot prove that either of these things will happen for sure, but they are risks.
However, there are many benefits to being in an open relationship. For example, if you don’t have an open relationship, then the chances of meeting someone who can fulfill all your relationship needs are slim. Rationalists tend to have high standards, so it is even more likely that they will find flaws in potential partners. This means that either they will have to settle for less than 100% satisfaction, or they will have to be single for some period of time. If you can find someone who is able to fill your 80% needs and his 20% needs, you will both be happier than if you each try to fill each other’s 100% needs.
If he has a side-relationship, then the rationalist will not be able to rely on his partner for all of his needs; this could be considered a form of rationality training. He will have to learn how to negotiate with the other person and balance their needs against his own. The best case scenario is that he gets better at treating his partner like a person and not like a magical unicorn that exists only to fulfill his every whim.
However, the most important reason for rationalists to be in open relationships is to ensure that we do not wind up becoming bitter and cynical. If we only date within our own community and never experience rejection, we will never develop an empathic understanding of why people do crazy things like break up with us. We will start believing that everyone should just accept us as we are and stop making problems. The resulting bitterness could destroy our rationality itself, as we become unable to understand other people’s thought processes and thus unable to communicate effectively with them. Once we become bitter and cynical, we can no longer be effective at improving ourselves or helping others improve themselves. We become useless.
There are three main conditions for a successful open relationship: honesty, respect, and discretion. Most non-monogamous relationships fail because these conditions are not met. But if they are met, it is possible to build a happy, healthy poly or side-relationship. The details of how you meet these conditions will vary from situation to situation, but here are some general guidelines:
Honesty: If you don’t tell your partner about your side-relationship, then it won’t work out. If you have sex without telling her about it, it won’t work out. If you lie about what happened during a date with someone else, it won’t work out. Honesty requires a lot of hard work and a lot of empathy, so take your time before entering into an open relationship if you feel uncomfortable with those things. Respect: Respect your partners as people, not as food dispensers or furniture or pets. Make sure they don’t feel as though they are just a means to an end for you – that is a recipe for bitterness and cynicism. In particular, make sure you respect your primary relationship over your side-relationship; don’t do things that make her worry about losing you or make her feel like a second-class citizen because she doesn’t get equal time. And for god’s sake, don’t flaunt your affairs in front of your partner. (If you really like your side-relationship that much, then don’t keep it a secret.) Discretion: If the people in your life know about your relationships, make sure they understand that they are not to tell other people about it. That includes your partners’ friends and parents. This is very important; you can never be sure when someone will gossip about you, and it might cause irreparable damage to your relationship.
Nothing stays the same in the world; if you want your relationships to stay happy and healthy, then you will need to work hard to keep them that way.
The first thing that often goes wrong is that the rationalist’s partner gets jealous. Make sure you understand her concerns, and talk them out until you both feel understood. If she feels like she has lost control over your life or doesn’t know when you will be home, she may feel very anxious and insecure, which is when jealousy starts to kick in. This means that staying home with her all the time is absolutely not the solution. You can’t stop her from being jealous, but you can help her understand that you still love her and will always come back home to her. Also, don’t keep secrets from your partner, because these often lead to distrust.
Another common problem is that one party feels like they are giving more to the relationship than they are getting out of it. It’s very important for everyone involved to make sure the other person feels loved and cared for, to make sure they get enough time, attention, affection, etc. If one person feels that they are being taken advantage of or treated poorly, then the relationship needs to change.
Take your time before you start. Just like with polyamory, there are polys, side-gays, and side-bitches. If you are not sure that you will be happy in an open relationship, then don’t become one. On the other hand, if you are thinking about cheating on your partner because you can’t be satisfied by one person anymore, then you should probably talk to her about that first. Non-monogamy can be an excellent way to build a closer relationship between you and your partner. But it can also be a recipe for disaster. If you don’t know what you want yet, then it’s better not to take any chances until you do.
Accept that your relationships will be different. There are no set rules for open relationships; every couple has to find their own way of doing things, and that is okay. Be okay with changes to your relationship status. For example, if you start dating someone else, don’t freak out that you are no longer “dating” or having sex with just that person. That is the point of being open! Just because you have a lot of partners doesn’t mean you have to be a different person with every one of them. Respect your partners’ boundaries, but don’t treat them as limitations on your behavior or as a way to police your behavior. Don’t tell her that she’s a prude for being jealous of your new girlfriend. Instead, work together to build a healthy relationship. Don’t have sex with anyone else until you know yourselves well enough as a couple to deal with it. In particular, don’t have sex with someone who is known as a side-bitch, because the probability of problems arising from that is very high. Understand that you cannot please everyone – no matter how hard you try – and that trying to do so will only cause problems in the long run.
The Rationality and Openness in Relationships post is much better than I expected. It helps that it's a listicle about a soft topic.
Why I am just a bit more cheerful about AI safety than Eliezer Yudkowsky is because I am less pessimistic about the possibility of overcoming these barriers. A solution to the problem of how to create a human-friendly AI will have to involve a lot of work on the part of the AI research community, but I think that this work can be done.
I certainly don't expect that any of this work will be easy, but I think that it will be possible, and I think that it will be well within the capabilities of the current AI research community to accomplish.
This doesn't mean that I think that AI safety will be easy - I have heard lots of people say that it is so difficult that we should not even think about trying. I don't think this is correct. The AI safety problem is hard, but I don't think it's harder than many problems which have been overcome by our species. (And I don't mean that the AI safety problem is like curing HIV/AIDS, which is trivially easy compared to AI safety; I think that our species has overcome hard problems before, even if they weren't as hard as AI safety will be.)
I think that one of the main reasons why people (including Eliezer Yudkowsky) are pessimistic about AI safety is because they think about it in terms of Halting Problem arguments. If you think of AI safety as requiring an argument which demonstrates that there is no possibility of an AI system that is smarter than the best possible human being at everything that matters, then it's pretty easy to throw up your hands and declare that this is impossible.
However, if you think of AI safety as requiring a demonstration that AI systems which are smarter than human beings will reliably do the right thing, then the problem becomes a bit more tractable.
I think that if we can create an AI system which is as smart as a human being at everything that matters, then we ought to be able to make sure that this AI system does the right thing, or at least behaves well enough that we would not regret delegating our control over the world to it.
This is not necessarily an easy task, but I don't think it's necessarily an impossible one either.
One thing which I like about Eliezer Yudkowsky's writings on AI safety is his emphasis on the importance of understanding what human values are. I agree that this is important, and I think that it's something which has been considered insufficiently by many people who talk about AI. But I don't think it's sufficient to say, "We must understand human values." The next step is to ask, "How can we actually do this?"
I think that when you consider the question in this way, then the project of AI safety does not seem quite so impossible. Instead, it seems like a project which can be accomplished by a small number of very smart people over the next ten years or so.
And if I'm right about this, then I think that there are some reasons to be optimistic about what will happen in the next fifty years or so.
The AI safety post took work. I had to run the text generation process several times, sometimes keeping half of an output and then resuming autocomplete from there. It took a while to adjust to the quirks of the autocomplete and to get an idea of what this text generator was capable of.
A good test of of whether an academic journal is bullshit is whether it can distinguish real research from bogus research. I think we can do a similar test on blogging. If an article cannot be distinguished from machine-generated content then that means[1] it is not human-generated. The effect is most obvious with news sites, which are now churning out hundreds of articles per day. And it's no great challenge to churn out blather that passes by most academic standards. If you read the article, then the odds are that it was written by a bot. It's not possible for a human being to produce that much content in that time frame.
That doesn't mean that it's not a good article. It may well be a very good article. But it has been disqualified from receiving a human-generated award. ↩︎