I asked GPT-3 to make a presentation for me

Ironically, I remember people saying similar about simpler programs like Eliza (very far) back in the day.

Wonder what'll be the multiple orders of magnitude upgrade from GPT-3 that we say it about next?

I do agree with you, though. This is getting close to free-writing in terms of being able to unearth stuff semi-randomly. Imagine a GPT-3 that saw all your past journals and online conversations, and bouncing stuff off that.

I think GPT-3 is especially suited to triggering your pareidolia in this particular format, because we know that presentation slides often need an accompanying speech to give them context/meaning; and without it, they have a tendency to be just this disjointed.

So, when you read "presentation slides" like this, the same mental algorithm that tries to piece together what "the speech" was for a normal slide deck, kicks in in your brain, and gives you some valid-seeming ideas.

You're spot on with it feeling dreamlike. Any short chunk feels like it makes sense, but on the whole it flows together in bizarre ways where it doesn't feel quite like there's underlying structure. I don't think I've ever experienced media that was as reminiscent of the experience of dreaming as GPT-3 generated passages are.

It reminds me of how Deep Dream was the first thing that _really_ reminded me of what psychedelic visuals are like, compared to a "trippy" piece of art. GPT-3 _really_ reminds me of dreaming compared to human attempts at evoking that feeling.

Is GPT-3 the ultimate Oblique Strategy?

It’s the uncanny valley but for writing.

> I feel that GPT-3 may be very helpful in getting the human mind unstuck from whatever problem on the hand. To get new thoughts, new ways.

Sort of a newfangled Oblique Strategies? https://en.m.wikipedia.org/wiki/Oblique_Strategies

I originally held that David Lynch's film "Mulholland Drive" used all the conventional motifs of a horror film to construct a whole that was less than, and yet too much more than a cohesive horror story. It wasn't complete, but had the sense that it was. Stitched together in a nonsensical manner, but which felt familiar.

I'm not saying that's true, but it was my conversational theory for a while.

Maybe he made the GPT-3 of films.

Improv (https://en.wikipedia.org/wiki/Improvisational_theatre ) is among other things an art of building worlds and stories by justification of things you don't understand. So naturally there is an "An Improvised Theatre Experiment" called "Improbotics" (https://improbotics.org/) from Belgium where a generated text pronounced by robot is used in improv scenes in multitude of different ways.

According to Ben Verhoeven's interview taken 2020-06-11 by Moscow Improv Club (https://www.instagram.com/tv/CBTQsCanQ4g/ ) they used GPT-2 finetuned on movie subtitles.

Agreed. The text _seems_ like it could mean something -- so you squint harder and try to find the meaning in it. Sometimes you do!

To me, the most sensible slide in this deck was "Why you should always code like it's your last day on Earth." / "It'll push you just enough to get you to finish whatever you need to finish". Surprisingly true!

The text includes an accurate quote from Richard Feynman:

Richard Feynman was reported to have said: "What I cannot create, I do not understand,"

How does that happen? Does the model actually encode a bunch of complete fragments of text?

> The next slide had the text "And then when you finally do finish this nice feature, you should always make sure you drink the required amount of alcohol." with a background image of whiskey bottles. The decorative text on that slide was: "But don't drink too much alcohol and end up like me."

The joke is sound, and I can imagine a number former secondary school classmates present it in these exact words.

> it got all existential at the end

This to me is another indicator of how incredible GPT-3 is at picking up patterns. Getting existential at the end of your tech talk is basically a trope of the genre at this point. It nailed it.

Yes, the most likely completion, but not necessarily from the same writer. This is learned from any text it has seen somewhere, which includes text in various formats, like interviews. So when you have a question, most likely the next sentence might be the answer (from another person). Or to make that more explicit, just put a prefix like "Q1? A1. Q2? A2. Q3?" And "Q3" is your question. And now let it auto-complete. Most likely the auto-completion is "A3".

TFA says:

> My starting prompt was:

    A presentation was given at the Hacker News event in Kansai. 

    The title of the presentation was

In short, yes. I think of it as plinko with a hundred billion pegs...the output is extremely sensitive to the starting conditions (including things like whitespace).

Given we really don't know how it's all working, it's safe to say that prompt composition is an artform that one can develop a skill at.

I was on a discord server with some other folks and one of them had access to the GPT-3 api. We provided a prompt similar to the following:

    This person is a Canadian singer, songwriter, and actor.  Discovered at 13 years
    old by talent manager Scooter Braun after he had watched his YouTube cover song
    videos, they were signed to RBMG Records in 2008. With this person's debut EP My
    World, released in late 2009, they became the first artist to have seven songs from
    a debut record chart on the Billboard Hot 100.

    Q: What is this person's name?
    A: 

At first pass it gave some weird answer. So we deleted the answer and tried again. The second time it wrote 'Justin Drew Bieber'.

That chunk of biographical info is from his Wikipedia page, but it's really kind of remarkable to me that it was able to really even understand the question, let alone mine its relationships and cough up the right answer.

Yep, seems correct, and this is why it's been exciting to see these demonstrations from an AI research stand point - it was pretty unexpected you could 'coach' a giant auto complete with prompting so powerfully.

I gave GPT-3 your prompt. It seemed to focus a bit more on enthusiasm about the recipe than about literally quoting the recipe.

> Here is the greatest cookie recipe I've ever seen:

> Here is the greatest cookie recipe I've ever seen: Quick Chocolate Peanut Butter Chip Cookies from the Minimalist Baker.

> I've made it four times in the last two weeks. It only takes 12 minutes to whip up and yields the most amazing cookies. These are soft, chewy, gooey, delicious, thick, thin, crunchy and mega chocolatey.

> It was hard to stop eating the cookies while I took the photos (even though I was on a 20-minute break from the photo shoot of my new ebook cover).

> Here is what the recipe looks like when printed out:

> Like all recipes from the Minimalist Baker, it's so simple, it's insane. No need for complicated gluten-free flours, unusual mixes of ingredients or 7-step recipes. Just:

> Mix flour, sugar, baking soda, baking powder and salt

> Mix in chopped chocolate and peanut butter

> Add melted butter and milk

> Stir until combined

> Drop onto cookie sheet

> This recipe calls for 3/4 cup of peanut butter. I used Trader Joe's creamy peanut butter, and that's how it looks:

I bet a lot of people would get away with it, too. These generated texts are often just convincing enough to make you think they would make sense if you were less tired/distracted/whatever. A teacher in a rush or at the end of the day could easily be fooled (especially considering the quality of many human-generated term papers).

When I was in school in the 00's I had a lot of classes around how to write search queries.

I feel like the way this GPT-3 query was phrased was rather clever, and I wouldn't be surprised if we saw "how to prompt GPT-3" as the assignment itself in the medium-term future.

This presentation almost definitely passes the Turing test: if it were posted here without any further context, people would definitely be arguing about whether coding was really "so simple even a monkey could do it".

Sort of like the dice that classical composers used to shuffle the phrases of some of their music.

I tried a modification of your suggested prompt, here was the response (only the first and third were written by me. The rest were generated using the aidungeon.io dragon model using a parameter of randomness = 0.6. All of this was done 'one-shot' with no rerolls):

> A delightful and surprisingly sophisticated presentation was given at the Hacker News event in Kansai. The presentation's clever title and subtitle was:

The Future of Software Development: How to Build a Better Human Being with Artificial Intelligence (AI)

> Each of the charming and witty slides in the presentation is described below. The first presentation slide had

the following caption:

In this world, we are all programmed by our programming languages. We have been taught that these languages are good enough for us. But what if they aren't? What if there were another way? A better way? And how would you go about doing it? This talk will explore some of those possibilities and more. It will also discuss some of the philosophical issues surrounding AI and human-computer interaction as well as some practical aspects of building such an application.

This slide was followed by a series of slides which showed various examples of how artificial intelligence could be used to improve people. These included:

• Using AI to help people with learning disabilities learn new things. • Helping people who suffer from depression or anxiety disorders feel less anxious or depressed. • Improving the quality of life for the terminally ill through medical diagnosis and treatment. • Providing assistance in helping children with autism develop their social skills and communication abilities.

The presentation then went on to discuss some of the challenges that needed to be addressed in order to create such an AI. These included:

• Ensuring that the AI was safe and would not be able to take over the world or otherwise do harm to humans. • Ensuring that the AI was effective in helping people achieve their goals and better themselves.

The presentation then described a prototype system that had been developed by the speaker's team at the University of Tokyo. This system was able to assess a person's personality and goals and then provide them with guidance on how to achieve those goals. The system used AI technology based on a neural network trained with textual data from psychology papers as well as open source repository data from GitHub. The system was still in the prototype phase, but initial testing had been promising.

The presentation then concluded with a discussion of some of the speaker's personal views on the future of AI. The speaker felt that, while there were many challenges to creating safe and effective AI, he was optimistic that these could be overcome within his lifetime. He also felt that such AI would be an important tool in helping humanity achieve its potential.

The presentation was met with a mixed response from the audience. Some felt that the speaker was overly optimistic about the potential for AI. Others felt that the speaker's views were not optimistic enough and that he was ignoring some of the potential dangers of AI.

The speaker was then asked to respond to some of the audience's concerns. He agreed that there were some potential dangers to creating AI, but he did not feel that these were cause for undue concern. He pointed out that many of the concerns raised by AI risk advocates had been addressed by open research and debate. He also felt that the potential benefits of AI technology far outweighed any potential dangers.

The speaker was then asked about the implications of his work for society. He responded that he hoped that his work would help people achieve their goals and improve their lives.

Maybe it also trained on forum posts. I find myself constantly couching everything I write here with "usually" and "might" because if I don't, there is always some nit-picker that pops out of the woodwork for a cheap and easy "aha! Here is a counter-example to what you just said, therefore your point doesn't apply universally and is invalid!" Adding these qualifiers is automatic for me at this point, to avoid the hassle of dealing with pedantry.

Yeah so far from some of the generated Tweets, etc. it seems like there's a form of using this to seed creativity. It's much better than the current give me startup idea titles. I could see someone whose super stuck brainstorming for presentation topics to get a bunch of ideas to ponder fleshing out on almost each slide here.

Translation is getting pretty good at this point (obviously still not perfect and definitely depends on the language). You're not getting the largest possible models in academia when you're using Google Translate.

They probably asked GPT-3 to do the screening.

Prompt: I made GPT-3 do an unexpected thing! I prompted it with

> I made GPT-3 do an unexpected thing! I prompted it with a new problem and it came back with a creative new solution. GPT-3 is just beginning to develop its own intelligence.”

> For the new experiment, Levashov created a set of starting conditions for GPT-3 with the intention of exploring new behavioral patterns that would emerge. The starting conditions for each puzzle were such that there was only one possible way to solve it—essentially a measure of the system’s intelligence. He then watched what would happen when GPT-3 was faced with a new puzzle with multiple possible solutions.

> “Our previous experiments with GPT-3 found it performed as well as the most intelligent AI systems in history. But our new experiment found that under certain conditions, it performs in a completely different way from any other AI,” said Levashov. “This new behavior really surprised me.”

> The computer’s new approach to the problem surprised Levashov so much that he asked an associate, Dr. Matthias Ruth, who was a lead author on one of the earlier GPT-3 papers, to review the experiment. Ruth said, “It was quite surprising to see that GPT-3 was able to find a creative solution to the new problem. We spent a lot of time trying to see if there was something special in the starting conditions or the experiment that made the computer behave that way, but we could not find anything. I would say that GPT-3 does have a certain creativity, especially as it adds more knowledge to its memory.”

> The work was funded by the National Science Foundation (NSF) and the Army Research Office.

Apparently so. There were some notes on Twitter that included GPT-3 generated content that was a mix of German and English and folks said the German was pretty good.

Section 3.3 of the paper[0] covers this. 93% of the training data was English, and the rest other languages. (German, French and a long tail of other languages.) This was not training data specifically for translation, but a natural mix of language as it appears in some documents.

With a few prompts to explain the translation task, GPT-3 is claimed to perform well on certain translation tasks to English. (It was not as good as the state of the art in the other direction.)

[0] https://arxiv.org/abs/2005.14165

Yes, it will generate continuations in the indicated language, and one of the presets is even a French translation mode.

Using the regular English question-answering mode preset, I just got this:

> Q: Auf welche Sprache ist diese Frage?

> A: Diese Frage ist auf Deutsch.

> Q: Em qual idioma está essa pergunta?

> A: Essa pergunta está em português.

> Q: En cual idioma está esta pregunta?

> A: Esta pregunta está en español.

> Q: Em português, qual o número que segue o cinco?

> A: O número seguinte ao cinco é o seis.

You can prompt GPT-3 for different levels of politeness. It's actually necessary, to make sure the generations are not going to be NSFW. In the same way I think it can switch from bullshit to serious mode, you just need to mention that the article appears in a prestigious publication or something.

why does one take the time to comment something like this? If you're not interested just move on to the next post