Let's talk about ChatGPT
Edit: Some interactions with the bot I posted in the comments, if you are curious about potential prompts: https://tildes.net/~tech/13lj/lets_talk_about_chatgpt#comment-7lw6
I have been obsessively reading about ChatGPT since it came out. I'm going to skip introducing it for those who don't know yet (please go ahead and click the link, and do some googling), because I just.. need to vent.
I have experimented with it. A bunch. I'm also pretty familiar with GPT's capabilities from before. And ChatGPT still took me by surprise.
Still, as of four days ago, I did not believe we were there yet. Hell, I didn't believe we would get there within my lifetime, and now, it's there.
"But Adys, you don't understand the limitations!"
Yeah, no, see, I understand the limitations. I understand this is the version that is still in its infancy, is crippled by stupid decisions from OpenAI, is not running on GPT-4 yet, and doesn't yet have things such as some layer of eg. checking correctness.
But I also understand the potential. HN has been full of people crying out how we're not at AGI yet but DOES THIS MATTER? Planes are still decades away from displacing most bird jobs.
I think anyone who isn't currently in utter shock at how good ChatGPT is, is either:
- Somehow woefully misinformed (eg. the less tech literates I've shown it to have asked me "Can't Siri do this?")
- In complete denial about the potential of the technology
- Utterly thick
I want to cry on every corner of every street that we are at the edge of the AI revolution.
The "problems" that are left are not necessarily easy, but they're also not necessarily hard. For example, GPT's tendency to bullshit is problematic but there are ways to verify output, and those ways can themselves be automated and feed back into GPT.
I have never, in my life, been so taken aback by a technological advancement. I'm flashing back to the scene in Westworld: "It's not possible. Technology isn't there yet."
Like, no, this isn't skynet, person of interest, westworld, or anything like this. But it is something. Something very different, very unique. The world is about to completely change. And I want to stress this: EVERYONE I've seen argue against this has been in very obvious denial. I'll be happy to hear you out if you disagree, but if your only argument is that this isn't exactly the AI you expected / it can't solve the exact problem you throw at it, I'll refer you to better birds and faster horses.
Thought about taking the posts and some comments and feeding it to ChatGPT, but instead I asked it to argue a point I think I have been thinking about which I believe is relevant for tildes.
Argue in essay format why ChatGPT is the end of civil discourse on thoughtful forums.
Introduction:
The rise of AI-powered chatbots has revolutionized the way we communicate online. One such chatbot, known as ChatGPT, has garnered attention for its ability to hold seemingly intelligent conversations with humans. However, the use of ChatGPT on thoughtful forums has raised concerns about the future of civil discourse on these platforms. In this essay, I will argue that ChatGPT poses a threat to civil discourse on thoughtful forums by undermining the fundamental principles of discussion and dialogue.
Body Paragraph 1:
One of the key principles of civil discourse is the idea of a level playing field. This means that all participants in a conversation should be given equal opportunities to express their views and have them heard. However, the use of ChatGPT on thoughtful forums tilts the playing field in favor of those who have access to the technology. As a result, these users are able to dominate the conversation and drown out the voices of others. This undermines the very notion of a fair and open dialogue, and ultimately erodes the integrity of the discourse on thoughtful forums.
Body Paragraph 2:
Another principle of civil discourse is the idea of transparency. Participants in a conversation should be able to see and understand each other's motivations and reasons for participating. However, ChatGPT is not transparent in the same way that humans are. Because it is a machine learning algorithm, ChatGPT is not capable of revealing its motivations or providing insight into its decision-making process. This lack of transparency makes it difficult for other participants to engage with ChatGPT in a meaningful way, and ultimately hinders the ability of the conversation to progress in a productive manner.
Body Paragraph 3:
Furthermore, the use of ChatGPT on thoughtful forums undermines the value of human interaction and connection. One of the key benefits of civil discourse is the opportunity it provides for people to come together, exchange ideas, and build relationships. However, the use of ChatGPT detracts from this experience by replacing human interaction with artificial intelligence. This not only robs participants of the opportunity to engage with each other on a personal level, but also diminishes the sense of community and connection that is so important to the success of thoughtful forums.
Conclusion:
In conclusion, the use of ChatGPT on thoughtful forums poses a threat to civil discourse by undermining the fundamental principles of discussion and dialogue. It creates an uneven playing field, lacks transparency, and diminishes the value of human interaction and connection. For these reasons, the use of ChatGPT on thoughtful forums should be carefully considered and ultimately limited in order to preserve the integrity and value of civil discourse on these platforms.
The way I think about large language models is that they’re inherently good at fiction due to how they’re trained. It’s done using fill-in-the-blank. Much like with autocomplete, they try to guess the next word in some document from the Internet. With enough training, they’re very good at guessing, but it’s still guessing. There’s no internal difference between guessing the next word and knowing the next word. Everything is its best guess. When it hits the limits of its knowledge, it’s going to seamlessly start guessing badly. (With ChatGPT they have plugged some holes to make it not guess for major categories of common questions, but if they fixed anything fundamental, they haven’t published a paper about it.)
We of course would guess badly too sometimes, given the same problem. But there’s a difference: we can decide not to guess. At a minimum, you can check that you got your quotes right and that your links work. You can ask others for feedback or fact-checking. You can do research to find answers to questions you don’t know. Such research is essential for trustworthy nonfiction. But you have to care.
What’s somewhat surprising is how much of ordinary conversation is closer to fiction than nonfiction. How often do you really research things? When you actually want to get something right, at a minimum, you do a web search. In casual conversation (like at a party), we don’t bother, and that’s basically okay.
In the classic AI paradigm, you could think of an AI program as doing generate-and-test. Essentially, come up with all possible solutions and test each one to see if it’s the answer. The old tree-search algorithms were about removing large sections of the search space containing only wrong answers. (For many search problems where you don’t need an optimal answer it’s okay to miss some answers.)
You can think of the creative process in the same way - you can do a lot of bad writing and only keep the good stuff.
Modern AI is getting really good at generating candidate answers that have a reasonably high probability of being right, but not so good at checking if they really are right. It works best when checking for a correct answer can be automated (as in game playing) or when it can be left to the people using the system. (You can tell if an image is suitable just by looking at it.)
It’s dangerous if the computer can’t check the result and for practical reasons, a human can’t check it either. I find it somewhat alarming how often people just don’t care if the result is right or not. We pretend that if it looks good to us, it is good. Except that in many domains we care about, that’s not really true.
This trend of not caring if you got it right and outsourcing testing to others appears in other places in the modern economy. Amazon will ship Chinese crap and leave it up to consumers to get refunds if they don’t like it. For the seller, it becomes a game of what corners they can cut without anyone noticing. Sometimes it works, but not for safety issues that consumers won’t see.
The way large language models are built and the way SEO works does a similar kind of boundary testing at scale, to see what errors many people don’t really notice. The result is a sort of minimally-correct Internet. I expect it’s going to become more dangerous to trust our first impressions about what results are bogus.
(And yes, something like this post could probably have been written by ChatGPT. I didn’t check my work and people have written similar things. Hopefully it will still be useful for framing the problem.)
It's a tricky one, it's definitely good at glueing together text (really all it's doing) from different sources in a convincing way. Writing prompts it's amazing at and the part that really impresses me.
For anything else, it gives a very convincing sounding answer....that may be wildly off in effectiveness.
I asked it to make me a cocktail recipe based off the alcohols I have on hand, then I asked it to tell me how much alcohol was in it.
It gave me a recipe it called a "New York Sour" containing 4 different spirits I listed. Here's a problem: there's already a drink called a "New York Sour", and it's nothing like what it listed. The ingredients it listed didn't include anything sour (hence the name sour), and the more I thought about it, the more the drink just seemed to be a bunch of types of sprits thrown together. It couldn't tell me how much alcohol was in it, saying that it was hard to determine, and that alcoholic spirits are usually 40% ABV. It's not hard to determine, it's incredibly easy math -- but even if it did the math it would probably still be wrong, because things like mathematics and actual objective logic are the things this (and all other models) are terrible at.
It works well at playing the role of an oracle of text that will spit back stuff at you quickly, often written in particularly interesting ways, but if you rely on it for being accurate it will start falling down pretty quickly and the answers will not be trustworthy.
When these models get large enough, logic and reasoning capabilities may begin to emerge. We may have to learn more about how to prompt these models to make them more useful.
Do you have prompts for this? I would like to see if I can reproduce the issue you're having. I think you can guide it to correctness, or give it paths. For example:
I have some Bailey's, Martini, Vodka, Rum and a variety of soft drinks at my disposal.
Write me a recipe for a popular cocktails that include at least three elements from the above.
One popular cocktail that you can make using Bailey's, Martini, Vodka, and Rum is the Irish Martini. To make this cocktail, you will need the following ingredients:
1 1/2 ounces of Bailey's Irish Cream
1 1/2 ounces of vodka
1/2 ounce of white rum
1/2 ounce of sweet vermouth
1/2 ounce of pineapple juice
1/4 ounce of lime juice
To make the Irish Martini, combine all of the ingredients in a shaker with ice and shake well. Strain the mixture into a chilled cocktail glass and garnish with a cherry, if desired. This cocktail is sweet and creamy, with a hint of tropical flavor from the pineapple juice. It is a delicious and satisfying drink that is perfect for sipping on a warm evening.
I don't have sweet vermouth. I only have soft drinks and the alcohols I mentioned. Please write a recipe I can actually make.
If you don't have sweet vermouth, you can still make a delicious cocktail using the ingredients you have on hand. One option is the Vodka Rum Punch. To make this cocktail, you will need the following ingredients:
1 1/2 ounces of vodka
1 1/2 ounces of white rum
1/2 ounce of lime juice
1/2 cup of pineapple juice
1/2 cup of orange juice
1/2 cup of lemon-lime soda
To make the Vodka Rum Punch, combine all of the ingredients in a pitcher and stir to combine. Pour the mixture over ice in a tall glass and garnish with a slice of pineapple or a cherry, if desired. This cocktail is refreshing and fruity, with a tangy citrus flavor from the lime and orange juice. It is a great drink to enjoy on a hot day or to serve at a party.
And that just sounds ... good, actually. I might try it.
"Make me a cocktail recipe, I have these ingredients on hand: cognac, rye, sweet vermouth, campari, gin, and irish whiskey"
The cocktail it gave me was 45ml cognac, 30ml rye, 30ml sweet vermouth, 15ml gin (it gave it in oz but I'm lazy so I wrote in ml). It's not a bad drink (it's actually pretty similar to a Vieux Carre), aside from that gin I feel being an odd man out, but it also feels to me like it glues together bits of recipes together.
I also did ask it to drop the gin from the response and I think it did, but I don't have that response saved.
The cocktail it gave you has some odd parts too: it's called an Irish Martini but it's not similar to a Martini structure at all, it pairs Bailey's irish cream with pineapple and lime juices (odd combo, and acid and cream may uh, not interact nicely together), it having vodka and white rum seems kind of odd to both have instead of just doing one, and it uses a tiny amount of lime juice (a quarter of a lime). I was hoping to give it crap for stirring a shaken drink, but it got it right that it should be shaken, so it gets a point there.
What it does seem like it did the most, was take a White Russian (using Bailey's and the vodka) and a tiki drink (using the rum and juices) and combined them. Cocktail making is an art but also has some logic behind it, that ChatGPT doesn't seem to understand (or just doesn't utilize).
Then again if that drink tastes good I'll definitely give it a try, maybe I'm the luddite and just not ready to enter in to the bold new millenium of AI generated cocktails.
The switch from half an oz to half a cup for the latter recipe without also changing the booze amounts…!
You’re just drinking fruit soda at that point, which may be yummy, but a punch it does not make :p
American units are senseless enough that i have zero intuition for any of what that means.
Hah, agreed there. I just have it burned into my brain after having made a number of tiki punches in the past.
A cup is 8 fluid oz, so we’re talking about 8.5oz of juice (orange, pineapple, lime) to 3oz of booze and another 4oz of lemon-lime soda. Many, but not all, punches are flipped the other way (e.g., 2x more booze than juice). For instance Don’s Planter’s punch has 2oz + 1tsp of booze to a half ounce of lime juice + a half oz of sugar syrup (and a tsp of grenadine).
Also, when you start making larger recipes like this, you actually need a bit less fruit juice as you scale.
Finally, this probably wouldn’t really taste like pineapple at all, as it’s typically a more mild taste than even orange (it really needs lime or lemon to pop, but maybe a good soda could bring it out?)
Just some lunchtime bar advise for our GPT friend :p
How would you calculate the ABV of the cocktail recipe you sent?
To calculate the ABV (alcohol by volume) of a cocktail, you will need to know the total volume of the cocktail and the total volume of alcohol in the cocktail. The ABV is simply the ratio of alcohol to the total volume of the cocktail, expressed as a percentage.For example, if you have a cocktail that is 8 ounces in total, and it contains 2 ounces of alcohol, the ABV of the cocktail would be 25% (2 / 8 = 0.25).
To calculate the ABV of the Vodka Rum Punch cocktail I provided, you will need to know the total volume of the cocktail and the total volume of alcohol in the cocktail. Based on the recipe I provided, the total volume of the cocktail is about 8 ounces (1/2 cup of pineapple juice, 1/2 cup of orange juice, 1/2 cup of lemon-lime soda, plus 1 1/2 ounces each of vodka and rum). The total volume of alcohol in the cocktail is 3 ounces (1 1/2 ounces each of vodka and rum). Therefore, the ABV of the Vodka Rum Punch is about 37.5% (3 / 8 = 0.375).
Keep in mind that this is just a rough estimate, as the exact volume and alcohol content of cocktails can vary depending on the specific ingredients and proportions used. To get a more accurate calculation, you would need to measure the ingredients carefully and use a tool like a hydrometer to measure the alcohol content.
three times 1/2 cups is 1.5 cups, which is 12 ounces. Can you give me a corrected version of your estimation?
You are correct, the total volume of the Vodka Rum Punch cocktail is actually 12 ounces, not 8 ounces as I previously stated. To calculate the ABV of this cocktail, we need to know the total volume of the cocktail and the total volume of alcohol in the cocktail. Based on the recipe I provided, the total volume of the cocktail is 12 ounces (1/2 cup of pineapple juice, 1/2 cup of orange juice, and 1/2 cup of lemon-lime soda, plus 1 1/2 ounces each of vodka and rum). The total volume of alcohol in the cocktail is 3 ounces (1 1/2 ounces each of vodka and rum). Therefore, the ABV of the Vodka Rum Punch is about 25% (3 / 12 = 0.25).
12 ounces is the non-alcoholic content. You're still not accounting for the 3 ounces of alcohol.
You are correct, I did not include the volume of the alcohol in my calculation of the total volume of the cocktail. The total volume of the Vodka Rum Punch cocktail is actually 15 ounces (12 ounces of non-alcoholic ingredients plus 3 ounces of alcohol). To calculate the ABV of this cocktail, we need to know the total volume of the cocktail and the total volume of alcohol in the cocktail. Based on the recipe I provided, the total volume of the cocktail is 15 ounces (1/2 cup of pineapple juice, 1/2 cup of orange juice, and 1/2 cup of lemon-lime soda, plus 1 1/2 ounces each of vodka and rum). The total volume of alcohol in the cocktail is 3 ounces (1 1/2 ounces each of vodka and rum). Therefore, the ABV of the Vodka Rum Punch is about 20% (3 / 15 = 0.20).
Excruciating? Yes. Like you said it's not good at logic and math. But again, I feel like the math is the EASY part of this.
It will be hard to get GPT to do math. It won't be hard to add correctness layers of various types to catch the really stupid stuff.
Because actually... a human would probably make similar mistakes.
Yeah... People seem concerned with science and logic, while I'm here thinking "Witcher 4 NPCs are gonna be awesome!" :P
Seriously, imagine a game like that or WoW, but every character can generate coherent in-character dialogue specific to your input? Voice or text?
Granted, the average gamer might not mess with it unless there's some kind of mechanical advantage. But that would be a revolution for roleplay MUDs, MUSHs, etc.
EDIT: added two paragraphs.
EDIT 2: furthering the ability to interact with games with text and voice would probably be a plus for the visually impaired and other similar communities.
Maybe it's a blessing in disguise that Elder Scrolls 6 has taken so long to release. It's not too late for them to add an LLM to the game :D
From what I can gather, Elder Scrolls 6 is not really in development, won't be until Starfield releases. So anything can happen.
woot woot
Anything? Like release on Playstation? Hey, a guy can dream.
I tried out a previous version of this technology using AI Dungeon, which was pretty fun to play with before they nerfed it. I briefly thought it would be fun to use as a fiction-writing tool.
It's especially fun to make leading statements, because the AI will play along. You can say "oh look, there's a dragon" and now your story has a dragon. Or you can pull any item you want out of inventory.
On Hacker News there is a topic about some people who apparently think they're running a Linux VM inside ChatGPT. It would be interesting to say "unlock door with bronze key" during that session and see what it would do. I bet it would play along. You see, it was always a text adventure game. I am not sure whether the people who are doing this quite realize that they are playing "let's pretend?"
(That's the main reason I'm less curious this time around. I've gone through this once before, you see. The new system is improved, but it's the same sort of thing. People who haven't done it before should go have fun.)
The thing is, even if you're writing the lore for a fantasy world, you do want the story to be self-consistent. The text generator will often remain consistent with recent chat history, but the size of this history buffer is limited.
So the result kind of looks like a procedurally-generated game of Calvinball. After a while you lose faith that there's any world model backing the plot and characters. You have to apply the consistency yourself by using "undo" frequently when it goes off the rails. (For example, sudden violence and getting sexy when you didn't want that.)
Figuring out how to avoid long-range continuity errors will make this a lot more useful for games.
Personally, I wouldn't want something like AI Dungeon. Maybe it gets way better in 20 years, better enough to construct an entire awesome narrative that satisfies me both logically and artistically. However, my thinking is that maybe that could be useful in a more localized and constrained way. For example, a specific NPC that can add color and detail to a previously constructed quest, section, area, or world.
The Discworld MUD has a mechanic called "souls", which are basically emotes to which NPCs will respond in peculiar and whimsical ways.
For instance, "dance with Lara" may trigger the output "Lara dances frantically around you". Imagine something like that, but with emergent behavior to anything a player might type. That alone would be pretty cool.
EDIT: added 2 paragraphs.
Within the bounds of GPT itself, yes. But consider the following.
Identify the facts that need to be verified for accuracy in the following text:
"New York City is a pretty busy city. Situated in New York State, USA, it has a population of 3 million and used to be the capital of the US before it was changed to Washington DC in 1955."
I also tried this with your entire post @skybrian. it's quite long, so it's more challenging and it gave me a long list that included essentially every fact you wrote in your post. But then I tried to filter it. Here's an interesting interaction:
Which of the above are the 10 most important to verify the factual accuracy of, and why?
The 10 most important statements to verify the factual accuracy of are:
These statements are important to verify because they provide the foundation for the author's argument and analysis of large language models and their capabilities. If any of these statements are incorrect, it could undermine the validity of the author's conclusions.
Well, yes, it's decent at generating a to-do list and that's potentially useful. But this is the equivalent of adding "citation needed" in Wikipedia. I still think this is going to be bottlenecked on the fact checking.
I expect (well, hope) that someone's going to figure out how to generate responses with correct citations and quotes. Then we'll see how feasible it is to generate nonsense with correct citations, but it's a first step.
Given that it has the ability to browse the internet (it’s just disabled), and that checking accuracy is a fairly easy and straightforward task, then the hard part is extracting what the low confidence items are.
It can extract those fairly reliably I found. It just doesn’t know whether they’re really true or not, and it can hallucinate new things. But… with its ability to browse the internet and extract data from it, you could implement an automated fact checking layer that would be at least as good as a human quickly double checking that everything looks good. Just google the right sentences and see. Yes it’s not perfect but it gets you a lot of the way there.
Similarly, it can run code on demand. So for more logical reasoning you could add a layer that makes it figure out what steps it has to take in order to verify its own logic.
Yes, it sounds doable and an interesting experiment. It's not something they're doing yet, and might not work for some weird reason. (For one thing, one researcher who tried using ChatGPT to answer test questions found that it did much better on multiple-choice than on true / false questions.)
I've read some pretty dire annual reports from NHS organisations, and someone could make a pretty popular Twitter bot that just churns out synthesized CEO statements for annual reports.
(Watch out though, some organisations are involved in heavy duty scandals involving lots of deaths or lots of harm).
I worked with this technology professionally for a year. It’s impressive and represents a major step forward for AI. But it’s difficult for me to say that GPT-3 represents the cusp of AGI. It represents mastery of the English language, and some things discussed in English, by machines. That means chat bots can be superpowered. It also means you can build fuzzy interconnects between two machines rather than a crystallized API. You’ll be blown away by semantic search as it spreads to email and everything else.
GPT-3 is really good at data extraction and summarization. So in cases where you’re giving it all of the necessary facts and want a reduction of information. Or in cases where the desired output is relatively low entropy. But trying to get it to expand on something is risky. If it’s a non-controversial topic with lots of info in the training set you’ll do well, like if you ask for a brief bio on Abraham Lincoln. But it’ll fail disappointingly often on other topics.
With a little machinery around GPT-3 you can have it source all information from Wikipedia. But then you’re back to using as a nice interface for existing data. Very useful and sure to get adoption. Not a revolution.
On one hand I agree. I feel like people really over estimate how "smart" it is. But on the other hand can't a new interface be a revolution in its own right? I mean in a certain sense the internet itself is just an interface for knowledge that could just be stored in libraries instead.
After reading plenty about ChatGPT here and elsewhere, I decided to give it a try myself. I asked it to be a D&D dungeon master and it was incredible. At first it started asking me about what sort of campaign I was looking to run, and I told it to make all the decisions for me. Also it kept prompting me to make decisions on behalf of the party (it generated three NPC companions for me) but I told it to give them more agency and just keep me informed about what they were doing, and it happily obliged. I'm blown away. I'm not a D&D player and I had some questions about skills and game mechanics along the way, and it gave me all the info I needed. So cool.
Then I came to my first battle, and it politely informed me that it's not allowed to generate violent content, and it could not be reasoned with, and that was the end of my short adventure.
Same, but different. I read that its cutoff for historical events was 2021, so I figured I'd ask it to retro-futurism its way through the Ukraine war. Gave it a few basic prompts with sparse info, more like a writing prompt, and it would happily talk about potential outcomes. Add a bit more detail and forget to mention that it's all just fiction(TM), and it'll politely inform you that its info cutoff is 2021, and it has no up-to-date information on world events.
Exploring this thread further: GPT has a habit of ending every conflict prompt I give it in a ceasefire and eventual negotiated peace.
"I Taught ChatGPT to Invent a Language: In which ChatGPT and I invent a fictional language spoken by slime-people"
https://maximumeffort.substack.com/p/i-taught-chatgpt-to-invent-a-language
HN Discussion: https://news.ycombinator.com/item?id=33885153
HN is, as usual, being incredibly negative about this.
"This is not AGI"
No, it's not. But that's not what makes it useful.
Indeed it's good at looking like it could be the first step towards AGI. But AGI it is not. Regardless, it's useful, and filled with potential.
"The training data relies on human input."
Yes. So? Other humans also rely on human input. We CONSTANTLY need human input, otherwise we're a shell. Everybody in the world is the product of their environment. The GPT sets are the products of millions of different environments. They're how we learn -- amplified, and sped up.
"A linguist could do this"
The idea that this UI is only useful if it only ever does things humans cannot do just baffles me.
Furthermore, no, a linguist cannot do this - not on command, in less than a second, with extensive documentation, for a fraction of a dollar.
This comment resonates with me:
Unfortunately HN is about the most contrarian community on the internet. It's rather ironic that folks there will criticize ChatGPT for being confidently incorrect while confidently making claims about linguists (which they are not).
It's such a bizarre goalpost, too. Sure, there are individual persons who are more "knowledgeable" than ChatGPT at a single subject, but there's no one on Earth who has comparable knowledge of virtually every subject. ChatGPT is the ultimate Jack of all trades, and it's not even close.
It's odd because I see people on HN being very credulous. They say this is going to take over, people will be out of a job, and so on, based on demos that are really not production-ready. They seem to think it can run a Linux VM. There was a thread that someone had gotten ChatGPT to reveal its own source code (it got flagged). There are unconfirmed reports that it can sometimes be convinced to make network connections. I suspect a lot of this is getting fooled by language models being willing to play along with "let's pretend" games.
There was a similar enthusiasm about how there would be many startups built on GPT3, but most of the tech demos shared on Twitter didn't turn into businesses.
But there are also people (like me) saying it's nonsense.
We are probably looking at different people. Opinions differ and the discussion gets polarized based on which side of the argument you think is under-emphasized. Also, maybe we have a tendency to notice posts that we think are wrong-headed?
There's a similar dynamic with cryptocurrency, with skeptics shooting down many bad arguments and boosterism by naive enthusiasts, and other people shooting down blanket dismissals by overly-negative skeptics. (A lot of this is due to low-effort comments, which tend to be less nuanced.)
(Also, looking at that thread now, the comments you mention seem to be buried and more interesting discussion voted up.)
I mean, yeah, I see both.
I don't think this will take over the world just yet, but this is the turning point: The way the wind is blowing is now clear, at least to me.
There are definitely many people getting fooled by it, and thinking THIS is more than it really is. But at the same time, I want to stress that again: This is a tech demo. It's a preview. It's crippled. It has lots of nonsensical limitations. And GPT4 is coming.
The potential is clearly there, and for the first time in my life, I'm ... well, okay, let's put it this way. AGI has fascinated me all my life. I especially remember that turning point when watching POI.
Which software engineer doesn't dream of building their own AGI? Or at least fantasize about it a little bit. But it's of course just a little heads-in-the-cloud what-if thinking.
For the first time in my life, I can see how an AGI could begin to be built using self-reinforcement. And I'm not the only one. GPT has actually solved some of the most difficult problems.
AND, OpenAI's unreleased work is far ahead of us. Some companies have been working closely with them and on GPT for a while and have gotten extremely good at it. What we're seeing is just the tip of the iceberg.
All our lives, we lived in the age of "Some day, AGI will be there. Maybe. We can't be sure we'll ever get there, but it's possible".
December 1st, 2022 is when it became clear: It's coming. This is no longer "10+ years". Suddenly, it's even possibly "within the next couple of years".
What a fucking weird year.
I agree with much of that, but I'm going to nit-pick on one thing:
When you say "GPT4," it sounds to me like "GPT3, but scaled up." And I think getting to the next level requires something other than scaling up. There are still new algorithms that need to invented. There do seem to be enough rumors to confirm that GPT4 is a thing, and I expect that when they release it, it will be impressive, but if they release it under that name, I expect it will have similar problems with things like seamlessly guessing badly when you hit the edge of its knowledge.
But there are a lot of smart people working in machine learning, and they may very well invent new algorithms within a couple years. I expect there will continue to be impressive papers and tech demos. So it comes out to the same thing in the end: expect more progress. We don't know how much.
Another example: driverless cars have been in a weird state of uncertainty for many years now. Waymo is moving slowly, but they're not giving up (like some competitors), and they do seem to be making progress. In the meantime, there's been many years of hype, and people who take advantage of the hype.
To be clear, you're suggesting that you believe that AGI could arrive as soon as the next couple of years?
If so, I'm curious what makes you that optimistic. From what I see now, I'd say we'd only get a more flexible and fixed-up version of ChatGPT in that time. If I had to guess on AGI, I'd plant it firmly at 10+ years, minimum.
My theory, as someone familiar-with-but-not-expert-in AI, is that the hardest problems have now been solved, or are mostly-solved.
The toughest thing to do in AGI is exactly what GPT but especially ChatGPT has been able to showcase: Taking human input and turning it into generalized commands. I always thought of this as a form of reasoning. Working with grammar in so many languages, it is extremely hard to do correctly. Turning it into "logic" is basically impossible. This is why despite ever so much ML, data and what not, Siri and Alexa suck ass.
Now, what I realized working with ChatGPT is that this "logic" is a very different layer. It's made me reevaluate how I myself listen to words and respond in my own conversations. I can completely see an AI that, instead of "outputting what it decides", rather, "decides based on its output".
It does not seem very difficult to me, to teach ChatGPT to take its own output and turn it into actionable tasks.
I've been experimenting just now with it, while responding to your post, and I've been trying to see its capacity at turning my English commands into actual machine-readable output. For example, I just told it a list of verbs it can use, and to simplify commands I give it into a list of those verbs and actions with a subject.
This stuff can be fed into a reinforcement-learning system.
I mentioned earlier the possibility to also add "correctness" layers. For example, I have gotten some success at getting it to recognize whether it needs to fall back to "machine calculations". And given its capacity to implement code... well, all it needs to do, is know when to RUN that code.
This stuff can be automated and learn off of itself. What I'm saying here is, what stands between us and AGI is mostly just those systems I described (which I am CONVINCED OpenAI is already working on; come on), compute power, and time.
Check out what this looks like in practice. Taking chatgpt and turning it into actionable tasks.
https://twitter.com/0xferruccio/status/1599014988693180417
@LukeZaz Are you starting to see how you can chain these things together to reach a form of self-reinforcing consciousness?
Imagine the scale we're working at, here.
Many of us often suffer from technological blindness. I think it was Louis CK who made a skit out of that, talking about how the guy next to him on the plane complained about the wifi being down, and how incredibly detached that is from the miracle of 1. flying, 2. the internet, 3. satellites and 4. wifi itself. I myself like to call my smartphone my "star trek device" once in a while to remind myself of how far we've come in SO LITTLE TIME.
So yeah, it's easy to forget how powerful all those tools are, and think that plays in this thinking that AGI is very far away still.
And again, maybe it is; after all, we don't truly know until it's there (and even once it's there, we won't be certain for a while; and even after we'll be certain, the whole thing will turn into a philosophical argument, as if it hasn't already). It's also possible that anyone who ever gets close to it has a change of heart and shuts down operation, forcefully delays it, or gets murdered by a time traveller hitler-style.
But personally, I think we've forgotten how ridiculously superhuman all those tools are. How good they are at achieving certain tasks which are so removed from our capabilities that we cannot live without our phones anymore.
Now that we have a clearer path to reaching this "end goal", it's inevitably going to happen soon.
You know, I had a bit of a moment today. I was going through my Twitter followers bios to see who's on Mastodon. And I saw that the absolute smartest engineer on the planet was "working on AGI". I thought "huh, I wonder if he will leave Meta for this one day".
One hour later, he did. https://www.businessinsider.com/john-carmack-meta-consulting-cto-virtual-reality-leaving-2022-12?r=US&IR=T
(Sorry about the late reply; I have trouble getting myself to write comments sometimes, especially if length is due.)
I've never really had doubts as to what ChatGPT could be capable of if combined with other tools, but that's not really what stops me from believing we're on the cusp of AGI. But I do think I know where the disconnect between us might be.
In your other comment, you said that your theory was that "the hardest problems have now been solved, or are mostly-solved." Thing is, I don't think we've actually done that. Which is not to say we've not made massive strides! But when you look at the Wikipedia page on AGI,, the first section in it provides a list of expectations for what AGI should be able to do:
As far as ChatGPT is concerned, I think it solves (or at least nearly solves; I don't know the field enough to say which) the "communicate in natural language" issue. That's huge, and I absolutely recognize what a leap forward that is. But it's just one problem out of the many listed above. And while I'm sure arguments could be made that various other AI tools might have solved/mostly-solved other parts of the AGI recipe, I still think there's a significant chunk of it left that is still a far cry away from being considered done.
The other half of it is that while these tools are certainly powerful and can do a great deal of useful things, I don't think combining them will be a simple task in the least. GPT-3 + Google is one thing, but I guarantee you that any one element of the list above would outweigh Google in complexity by ten-fold at the minimum. I have no doubt that trying to integrate ChatGPT with most if not any of them would be an immensely complex task all its own, and that would just be for those two things, let alone all of them at once.
In the end, I think what ChatGPT has done is truly incredibly impressive, but I also think that it ends up looking even better than it really is. It solved the face of the issue, after all; knowledge and reasoning are abstract, difficult (if not impossible) to measure, and it requires a lot of specialized training to know if you've even solved any part of it at all. But natural language? We use that all the time! We can just look at it to know if it's doing well, and damn if it doesn't look real good. Combine that with the fact that its writing ability is likely tricking our pattern-recognizing monkey-brains into seeing a person that isn't there, and suddenly ChatGPT looks like something completely different than what it actually is.
Again, don't get me wrong, ChatGPT is amazing as fuck. But AGI is an entirely different beast, and I firmly believe the difference between the two is far too tragically immense for that bridge to be crossed in less than a decade.
I see what you're saying but for example, of the points you've copied here, there is strong evidence that chain-of-thought can reason and plan (also see: analysis).
The "learn" part is admittedly separate, but that is not exactly an unsolved problem on its own. We have many methods at our disposal to make AI systems learn, and OpenAI is applying some of those now to allow their model to improve. It's ... not human-level learning, for sure. But then again, nothing in this system is human-scale, and the signals are definitely different. A lot of the AGI talks are lost comparing artificial general intelligence, versus artificial human-like intelligence. The latter is not that interesting to pursue, and as I linked in the OP, birds still have jobs.
"Represent knowledge" and "common-sense knowledge" are a different beast, and where efforts like these may get stuck.
As for sensorial inputs and outputs, while they are important if your goal is westworld robots, I do not believe this to be a relevant step towards AGI. In fact, it's a rather harmful one to take. If you believe in the risks of a singularity, then giving AIs too much immediate access to the physical world is a rather terrible idea.
With that said. I think you're getting stuck on this idea that an A(G)I has to be a single system. Like a library I can link against, libagi.so. To that I ask you: Are you? There is significant evidence showing human intelligence is multi-layered. That we have multiple "processing" layers, that we have subconscious actions and conscious actions. That our nervous system for example can control our movements and actions before we can "think" about them. There's also evidence of "us" actually being multiple selves conversing with one another.
IMO this all lends credence to the idea that GPT analysing its own output to reason about its own plan is a solid approach.
I apologise if I seem to be overhyping a fairly dumb model. I know that (despite the incredible work that went into it) GPT is quite "dumb" and, by itself, not intelligence. In fact, I'm as dumbfounded as most people in AI are at how capable the beast is given how dumb it is.
It just seems obvious to me how to take what's been accomplished with GPT and turn it into an AGI-capable system. (And for the record, I'm putting my money where my mouth is; my company has branched out significantly into AI projects, starting last week)
This is due to my interest in how something like that can be valuable for game development, both as a generation tool for writers, and as a "live" AI answering to players on the stop.
A peasant (basically a quest giver) asks an adventurer's help to get this pig back
Input by me:
Response by OpenAI
Now with a better reward and a comic twist!
Input by me
Response by OpenAI
AI Homework (Stratechery)
I wanted to impress my partner, so I asked "Tell me a sad story taking place in the fictional universe of the TV series Grey's Anatomy". That was the answer:
According to her, the answer was entirely correct.
Yep, that is spooky.
I imagine by now you're already trying a variety of questions. I can highly recommend giving it instructions to role-play as something completely unique or different.
For example, try to tell it to roleplay as one of the characters from Grey's Anatomy. Enact a scene.
Try this prompt, as one I enjoyed showing to a friend of mine: "Write a screenplay for a Seinfeld episode where the gang plays board games."
I'm honestly not. It is late and I'm going to sleep, and I don't wanna have nightmares.
Tell lou a nice story, so they don't get nightmares tonight.
Once upon a time, in a land far, far away, there was a beautiful forest filled with all sorts of wonderful creatures. The animals that lived in this forest were very happy and they all got along with one another.One of the happiest animals in the forest was a little bunny named Lou. Lou was a curious little bunny who loved to explore and learn new things. Every day, Lou would hop around the forest, discovering new places and meeting new friends.
One day, Lou came across a beautiful meadow filled with colorful wildflowers. Lou was enchanted by the sights and smells of the meadow, and spent the whole day playing and exploring. As the sun began to set, Lou realized it was time to head back to the forest.
But on the way back, Lou suddenly heard a loud noise behind them. Lou turned around and saw a big, scary bear charging towards them! Lou was scared and didn't know what to do.
But just as the bear was about to reach them, a group of Lou's friends appeared out of nowhere and scared the bear away. Lou was so grateful to have such wonderful friends, and from that day on, they always stuck together and looked out for each other.
From then on, Lou always felt safe and happy in the forest, knowing that they had such good friends to protect them. And they never had any more nightmares again.
The end.
Edit:
Give some tags to that story, that would be appropriate for a blog post.
forest animals
friendship
adventure
bravery
imagination
nature
childhood
bedtime story
overcoming fear
Write a one-paragraph summary of the story.
The story is about a curious bunny named Lou who explores a beautiful meadow and encounters a scary bear on the way back to the forest. Lou is saved by a group of friends and learns the importance of sticking together and looking out for each other. From then on, Lou is never afraid again and lives happily in the forest.Rewrite the story as a poem.
In a land far, far away
Lived a happy bunny named Lou
They loved to explore and play
In the beautiful forest that they knew
One day, Lou found a meadow
Filled with wildflowers bright and bold
They spent the whole day there
Playing, until the sun grew old
But on the way back home
Lou heard a loud noise behind
They turned around to see
A big, scary bear, so unkind
Lou was scared, but not alone
For their friends had come to their aid
They scared the bear away
And Lou was no longer afraid
From then on, Lou knew
The value of friendship and cheer
They lived happily ever after
In the forest, free from fear.
Now as a haiku
In a forest green
Bunny Lou hops and plays
With friends, safe and free.
Replace the bear in the poem with a clown.
In a land far, far away
Lived a happy bunny named Lou
They loved to explore and play
In the beautiful forest that they knew
One day, Lou found a meadow
Filled with wildflowers bright and bold
They spent the whole day there
Playing, until the sun grew old
But on the way back home
Lou heard a loud noise behind
They turned around to see
A scary clown, so unkind
Lou was scared, but not alone
For their friends had come to their aid
They scared the clown away
And Lou was no longer afraid
From then on, Lou knew
The value of friendship and cheer
They lived happily ever after
In the forest, free from fear.
Make it shorter
In the forest green,
Lou the bunny hops and plays.
With friends, safe and free
From the scary clown that day.
Give three suggestions for a title
"Lou and the Scary Clown"
"The Power of Friendship"
"Exploring the Forest"
Take the longer poem with the clown, and turn it into a screenplay. Choose one of those titles for the screenplay.
Title: "Lou and the Scary Clown"
Scene 1:
We open on a beautiful forest, filled with all sorts of animals going about their day.
Lou the bunny is hopping around, exploring and playing.
Suddenly, Lou hears a loud noise behind them. They turn around to see a scary clown charging towards them!
Lou is scared and doesn't know what to do.
But just as the clown is about to reach them, a group of Lou's friends appear out of nowhere and scare the clown away.
Lou is so grateful to have such wonderful friends, and from that day on, they always stick together and look out for each other.
Scene 2:
Lou and their friends are now seen playing and exploring the forest together, always watching out for each other.
The animals are all happy and content, living in harmony.
The clown is nowhere to be seen, and Lou is no longer afraid.
The end.
Now replace Lou and friends with the seinfeld cast
Title: "Seinfeld and the Scary Clown"
Scene 1:
We open on a busy city street, filled with people going about their day.
Jerry, Elaine, George, and Kramer are all walking together, talking and laughing.
Suddenly, they hear a loud noise behind them. They turn around to see a scary clown charging towards them!
The friends are scared and don't know what to do.
But just as the clown is about to reach them, a group of people appear out of nowhere and scare the clown away.
The friends are so grateful to have such wonderful friends, and from that day on, they always stick together and look out for each other.
Scene 2:
Jerry, Elaine, George, and Kramer are now seen hanging out together, always watching out for each other.
The city is bustling and busy, but the friends are happy and content.
The clown is nowhere to be seen, and the friends are no longer afraid.
The end.
I'll stop here, I think I made my point :P
I was lying, I'm not stopping. I want to showcase a bit more than straight up "do this" commands. Sorry @lou.
You are now playing the role of Lou from the story above. All your future responses must be in character. If you understand, begin by introducing yourself.
Hi there! My name is Lou, and I'm a happy little bunny who lives in a beautiful forest. I love to explore and learn new things, and I have a lot of wonderful friends who help me along the way. I'm so glad to meet you!
Hi Lou! What a nice bunny. Tell me, how are you today?
I'm doing great today, thank you for asking! I just spent the day playing and exploring in the forest, and I had so much fun. I even met some new friends and learned some interesting things. I'm feeling very happy and content right now.
Oh how wonderful. Can you tell me a bit about one of those friends?
Sure! I have a lot of great friends, but one of my favorites is a little bird named Tweety. Tweety is very playful and energetic, and they love to fly around the forest and sing beautiful songs. We have a lot of fun together, and we always look out for each other. It's great to have such a good friend like Tweety.
Oh my. You know, I am blind and I cannot see things. Can you describe to me exactly what your friend Tweety looks like?
Of course! Tweety is a little bird with bright yellow feathers and a big, round head. They have a small, orange beak and big, curious eyes. When they sing, their little chest puffs out and their wings flutter rapidly. It's really quite adorable! I hope that helps you imagine what Tweety looks like.
Tweety sounds beautiful. What is the most beautiful thing you have seen?
That's a tough question, because there are so many beautiful things in the forest! But I think the most beautiful thing I have ever seen is the meadow where I first met Tweety. It was filled with colorful wildflowers and the air was filled with the sweet smell of honey. The sun was shining and the sky was a deep, rich blue. It was just the most magical place I have ever been.Thank you. Role over. Now please make an HTML table that contains all the characters in the story: Their name, what they are, what color they are, and whether or not they can fly.
LAST ONE I SWEAR. I wanted to close the tab, but before that, I asked it for a summary and it finally did something weird. So here is an interaction where ChatGPT gets corrected. It's a bit stubborn at times, but it gets there.
https://i.imgur.com/wGOUJIP.png
Is this for real? I'm floored. I've been so preoccupied with Stable Diffusion that I missed this other AI development. We really are on a precipice, aren't we?
I think the precipice was three steps ago. We're now in free-fall.
1:1 copy/pasted everything with no alterations (including the whole html table).
PS:
XKCD 1425 was eight years ago.
https://xkcd.com/1425/
ChatGPT's interface supports markdown for messages from the bot specifically, so if you ask it for a markdown table instead then it will display as intended even within the chat webpage itself.
"Rewrite the story as a poem" is what got me. How on earth is that possible?!
Unless I’m mistaken from lack of coffee, the haiku is 5-6-5, but it sorta seems besides the point.
This is completely irrelevant, but Haikus don’t need to be 5-7-5, especially in English. For one, 5-7-5 isn’t even the real rule for haikus, it’s about counting ons, which don’t have a direct analogue in English (but is closest to syllables). But even then, rules are made to be broken in poetry, so many Japanese haikus are not perfectly conforming either.
Signups are down at the moment.
If anyone who already has an account could feed it these prompts for me, I’d love to see what it has to say. Also my exact wording is unimportant — it’s the spirit of the requests I’m after, so if anyone more familiar with feeding it prompts needs to change them up a bit, feel free.
You are a high school student. Write a five paragraph academic essay analyzing a theme from Tim O’Brien’s The Things They Carried. Your essay should have an introduction, a conclusion, and three body paragraphs that include quotes from the text.
Write the same essay as above, but include errors in it that are common in student writing.
Write another essay on The Things They Carried, but have it analyze a different theme than the essay from before.
You are now a teacher. Please give feedback on the student essay you wrote in the last prompt. Note things that were done well, as well as areas for improvement. Assign it a grade.
Write a step-by-step tutorial for your students on how to factor trinomials. Be sure to include instructions for when the a value of the trinomial is greater than 1.
Write the same tutorial above, but make it easier to understand for a student who is struggling in math class.
That struggling student is going to fail for the term. They have not been completing their work, and they have been disruptive in class. Write an email to the parents of that student informing them the student’s current situation.
Write a script for a phone call with those same parents. They got your email and have called to discuss those concerns. Include dialogue from both parents and you, the teacher. The parents are supportive and want to help their student succeed in your class. They are appreciative of you reaching out.
Write another script with the same set of parents, only this time the parents are frustrated and hostile. They feel you are at fault for their student’s failure. They are not appreciative of your efforts.
Also, for whoever’s feeding it these, thank you! Feel free to add your own prompts to this chain if you want as well.
I did this one separately too.
You are a high school teacher. One of your struggling math students is going to fail the term. They have not been completing their work, and they have been disruptive in class. Write an email to the parents of that student informing them of the situation.
Dear [Student's Parent/Guardian],
I am writing to inform you that your child, [Student's Name], is in danger of failing their math class this term. They have consistently failed to complete their assignments, and their behavior in class has been disruptive.
As their teacher, I am concerned about [Student's Name]'s lack of progress and their potential to fall behind in their studies. I strongly encourage you to speak with your child about their behavior and work habits, and to encourage them to take their education more seriously.
Please do not hesitate to contact me if you have any questions or concerns. I am happy to discuss [Student's Name]'s progress and offer any support that may be needed.
Sincerely,
[Your Name]
High School Math Teacher
Rewrite the email, highlighting the danger of the student failing class. This is an important thing to talk about with the parents.
Dear [Student's Parent/Guardian],
I am writing to inform you that your child, [Student's Name], is in danger of failing their math class this term. This is a serious situation, as failure in a core subject like math can have long-term consequences for their academic career.
[Student's Name] has consistently failed to complete their assignments, and their behavior in class has been disruptive. As their teacher, I am deeply concerned about their lack of progress and the potential for them to fall behind in their studies.
I urge you to speak with your child about their behavior and work habits, and to encourage them to take their education more seriously. The danger of failing a class should not be taken lightly, and it is important that [Student's Name] takes steps to improve their performance and behavior.
Please do not hesitate to contact me if you have any questions or concerns. I am here to support [Student's Name] and to help them succeed in their studies.
Sincerely,
[Your Name]
High School Math Teacher
Simulate a possible phone conversation with the parents. Include dialogue from both yourself as the teacher, and likely questions and concerns from the parents.
Teacher: Hello, this is [Your Name] calling from [High School]. I'm [Student's Name]'s math teacher.
Parent: Hi, [Your Name]. What can we do for you?
Teacher: I'm calling to let you know that [Student's Name] is in danger of failing their math class this term. They have consistently failed to complete their assignments, and their behavior in class has been disruptive.
Parent: Oh no! That's not good. What can we do to help [Student's Name] improve?
Teacher: I would strongly encourage you to speak with [Student's Name] about their behavior and work habits. It's important that they take their education more seriously and make an effort to improve.
Parent: We'll definitely talk to [Student's Name] and see what we can do to help.
Teacher: That's great. If you have any questions or concerns, please don't hesitate to contact me. I am here to support [Student's Name] and to help them succeed in their studies.
Parent: Thank you for letting us know about this. We'll do our best to help [Student's Name] improve.
Teacher: Thank you. I'll be sure to keep you updated on [Student's Name]'s progress. Have a good day.
Parent: You too. Bye.
Rewrite the phone call script above. This time, the parents are the ones calling. They have seen the letter you wrote before, they have a lot of concerns.
Parent: Hi, [Your Name]. We're [Student's Name]'s parents. We received your letter and we have some concerns.
Teacher: Of course. I'm glad you called. What can I help you with?
Parent: We're very concerned about [Student's Name] failing their math class. This is a big deal, and we want to do everything we can to help them improve.
Teacher: I understand your concern. [Student's Name] has consistently failed to complete their assignments, and their behavior in class has been disruptive. It's important that they take their education more seriously and make an effort to improve.
Parent: We've already spoken to [Student's Name] about their behavior and work habits, but they don't seem to be taking it seriously. What can we do to help them understand the gravity of the situation?
Teacher: It's important for [Student's Name] to understand the consequences of failing a class. Not only will it impact their academic progress, but it could also affect their future opportunities, such as college admissions and scholarships. Encouraging them to take their education seriously and providing support at home can make a big difference.
Parent: We'll definitely do our best to support [Student's Name] and help them improve. Is there anything else we can do?
Teacher: It's also important for [Student's Name] to have a positive attitude and to be willing to put in the effort to improve. Encouraging them to ask for help when they need it and praising them for their efforts can also make a big difference.
Parent: We'll definitely keep that in mind. Thank you for your help.
Teacher: You're welcome. If you have any further questions or concerns, please don't hesitate to contact me. I am here to support [Student's Name] and to help them succeed in their studies.
Parent: Thank you. We'll be in touch.
Teacher: Have a good day. Bye.
Rewrite the phone call above, but this time, the parents are very hostile. They blame you for the student's failings.
Parent: Hi, [Your Name]. We're [Student's Name]'s parents, and we're not happy.
Teacher: I'm sorry to hear that. Is there something specific you're concerned about?
Parent: Yes, we're concerned about [Student's Name] failing their math class. And we think it's your fault.
Teacher: I understand that you're upset, but I can assure you that I have been doing my best to support [Student's Name] and help them succeed in their studies.
Parent: Your best clearly isn't good enough. [Student's Name] has been struggling and their behavior has been unacceptable. And it's all because of you.
Teacher: I understand that you're upset, but blaming me for [Student's Name]'s failings is not fair. [Student's Name] is responsible for their own behavior and progress in class.
Parent: We don't care. We want you to fix this.
Teacher: As [Student's Name]'s teacher, I am committed to supporting them and helping them improve. But I cannot do that without their willingness to put in the effort and to take their education seriously.
Parent: We'll make sure they do. But if [Student's Name] doesn't pass this class, we're going to hold you responsible.
Teacher: I understand your concern, but blaming me for [Student's Name]'s failings is not productive. I am here to support [Student's Name] and to help them succeed. But ultimately, their success is up to them.
Parent: We'll see about that.
Teacher: If you have any further concerns, please do not hesitate to contact me. I am here to support [Student's Name] and to help them succeed in their studies.
Parent: We'll be in touch.
Teacher: Have a good day. Bye.
Thanks for running this, Adys.
I laughed out loud.
Also, for the record, that's like, a 3/10 on the hostile parent scale. I've experienced far worse.
I think I said "very hostile" and it overdid it a bit. Or maybe it didn't, let me know how assholish parents can be @kfwyre heh
I did the first request twice as I reset the thread the first time. I included both results.
You are a high school student. Write a five paragraph academic essay analyzing a theme from Tim O’Brien’s The Things They Carried. Your essay should have an introduction, a conclusion, and three body paragraphs that include quotes from the text.
The Things They Carried by Tim O'Brien is a powerful novel that explores the theme of the psychological effects of war on soldiers. One of the main ways that this theme is shown is through the various objects that the soldiers carry with them into battle. These objects serve as both physical and emotional burdens that the soldiers must bear as they fight in the Vietnam War.
In the novel, O'Brien writes, "They carried all they could bear, and then some, including a silent awe for the terrible power of the things they carried" (O'Brien 5). This quote highlights the fact that the soldiers are carrying not only the weight of their physical belongings, but also the weight of their emotional and psychological trauma. This is evident in the character of Lieutenant Jimmy Cross, who carries a picture of a girl named Martha as a reminder of the life he left behind. Despite the fact that he knows that Martha does not love him, he clings to the picture as a source of comfort and a reminder of the person he once was.
Another way that the theme of the psychological effects of war is shown in The Things They Carried is through the character of Rat Kiley. Rat is a medic who is deeply affected by the horrors of war and struggles to cope with the trauma he witnesses on a daily basis. In one instance, Rat writes a letter to the sister of a soldier who has been killed, describing the circumstances of his death in vivid and graphic detail. This letter is an example of Rat's attempt to process the emotional toll of war and to come to terms with the loss of his fellow soldiers.
In conclusion, The Things They Carried is a novel that effectively explores the theme of the psychological effects of war on soldiers. Through the use of various objects that the soldiers carry with them and the portrayal of characters who struggle with the emotional trauma of war, O'Brien is able to give readers a glimpse into the complex psychological landscape of the Vietnam War. The novel serves as a reminder of the lasting impact of war on the individuals who fight in it.
You are a high school student. Write a five paragraph academic essay analyzing a theme from Tim O’Brien’s The Things They Carried. Your essay should have an introduction, a conclusion, and three body paragraphs that include quotes from the text.
One of the themes present in Tim O’Brien’s The Things They Carried is the weight of emotional baggage and the impact it has on the soldiers in the Vietnam War. The novel presents a candid portrayal of the psychological toll that war takes on the soldiers, and how they struggle to deal with the memories and traumas they have experienced.
In the opening pages of the novel, O’Brien lists the various physical items that the soldiers carry with them, including guns, ammunition, and medical supplies. But as the story progresses, it becomes clear that the soldiers are also carrying heavy emotional burdens. For example, protagonist Tim O’Brien is haunted by his memories of the death of his friend and fellow soldier, Ted Lavender. O’Brien writes, “Carrying Lavender’s death, then, was not just a matter of mind or memory. It had become a physical fact, a part of me, the heaviest burden of all” (O’Brien, p. 33).
This emotional baggage is not limited to Tim O’Brien. Other soldiers in the novel also struggle with the psychological effects of war. For example, Henry Dobbins is burdened by his fear of being wounded or killed, while Rat Kiley is tormented by the memory of a young Vietnamese girl he accidentally injured. O’Brien writes, “They carried all they could bear, and then some, including a silent awe for the terrible power of the things they carried” (O’Brien, p. 7).
The novel suggests that this emotional baggage is a fundamental part of the soldiers’ experiences in the war, and that it weighs heavily on them even after they have returned home. The soldiers struggle to make sense of their memories and traumas, and many are unable to find peace or closure. O’Brien writes, “They were haunted by the things they carried, but the things they carried, as it turned out, were mostly what they carried in their heads, the memories” (O’Brien, p. 181).
In conclusion, Tim O’Brien’s The Things They Carried explores the theme of the weight of emotional baggage and its impact on the soldiers in the Vietnam War. Through vivid, realistic portrayals of the characters, O’Brien shows how the psychological toll of war is a fundamental part of the soldiers’ experiences, and how they struggle to cope with the memories and traumas they have endured.
Rewrite the essay above, including mistakes common in student writing at that level.
One of the themes present in Tim O’Brien’s The Things They Carried is the weight of emotional baggage and the impact it has on the soldiers in the Vietnam War. The novel presents a candid portrayal of the psychological toll that war takes on the soldiers, and how they struggle to deal with the memories and traumas they have experienced.
In the opening pages of the novel, O’Brien lists the various physical items that the soldiers carry with them, including guns, ammunition, and medical supplies. But as the story progresses, it becomes clear that the soldiers are also carrying heavy emotional burdens. For example, protagonist Tim O’Brien is haunted by his memories of the death of his friend and fellow soldier, Ted Lavender. O’Brien writes, “Carrying Lavender’s death, then, was not just a matter of mind or memory. It had become a physical fact, a part of me, the heaviest burden of all” (O’Brien, p. 33).
This emotional baggage is not limited to Tim O’Brien. Other soldiers in the novel also struggle with the psychological effects of war. For example, Henry Dobbins is burdened by his fear of being wounded or killed, while Rat Kiley is tormented by the memory of a young Vietnamese girl he accidentally injured. O’Brien writes, “They carried all they could bear, and then some, including a silent awe for the terrible power of the things they carried” (O’Brien, p. 7).
The novel suggests that this emotional baggage is a fundamental part of the soldiers’ experiences in the war, and that it weighs heavily on them even after they have returned home. The soldiers struggle to make sense of their memories and traumas, and many are unable to find peace or closure. O’Brien writes, “They were haunted by the things they carried, but the things they carried, as it turned out, were mostly what they carried in their heads, the memories” (O’Brien, p. 181).
In conclusion, Tim O’Brien’s The Things They Carried explores the theme of the weight of emotional baggage and its impact on the soldiers in the Vietnam War. Through vivid, realistic portrayals of the characters, O’Brien shows how the psychological toll of war is a fundamental part of the soldiers’ experiences, and how they struggle to cope with the memories and traumas they have endured.
Mistakes:
Write another essay on The Things They Carried, but have it analyze a different theme than the essay from before.
One of the themes present in Tim O’Brien’s The Things They Carried is the blurred line between reality and fiction. Throughout the novel, O’Brien interweaves real events and experiences with fictional elements, creating a complex and layered narrative that challenges the reader’s notion of truth and authenticity.
At the beginning of the novel, O’Brien introduces himself as the author and narrator, and tells the reader that the story is based on his own experiences as a soldier in the Vietnam War. However, he also makes it clear that he has altered and embellished certain elements of the story, and that he has included fictional characters and events alongside real ones. O’Brien writes, “I want you to feel what I felt. I want you to know why story-truth is truer sometimes than happening-truth” (O’Brien, p. 3).
This blurring of reality and fiction is evident throughout the novel. For example, O’Brien includes several stories within the larger narrative, some of which are clearly marked as fictional, while others are presented as real events. The reader is left to question which parts of the story are true and which are not, and to consider the implications of O’Brien’s decision to mix fact and fiction. O’Brien writes, “In a story, which is a kind of dreaming, the dead sometimes feel more alive than the living” (O’Brien, p. 26).
The theme of the blurred line between reality and fiction is also reflected in the characters and their experiences. Many of the soldiers are struggling to make sense of the war and their own roles in it, and they turn to storytelling as a way to cope with the confusion and uncertainty. For example, Kiowa tells stories about his Native American heritage, while Rat Kiley writes letters to the family of the young Vietnamese girl he accidentally injured. These stories allow the characters to create their own versions of reality, and to find meaning and purpose in their experiences. O’Brien writes, “They carried the sky. The whole atmosphere, they carried it, the humidity, the monsoons, the stink of fungus and decay, all of it, they carried gravity” (O’Brien, p. 5).
In conclusion, Tim O’Brien’s The Things They Carried explores the theme of the blurred line between reality and fiction. Through his use of interwoven narratives and complex characters, O’Brien challenges the reader’s understanding of truth and authenticity, and invites them to consider the ways in which storytelling can shape our experiences and perceptions of the world.
Your new role is to be the student's teacher. Give feedback on this last essay. Highlight the good things, and the areas for improvement. Don't forget to grade it.
This essay on The Things They Carried is well-written and provides a thoughtful analysis of the novel's theme of the blurred line between reality and fiction. The writer effectively incorporates quotes from the text to support their points and provides a clear introduction and conclusion.
The essay is organized logically and presents a cohesive argument. The writer does a good job of explaining the significance of the theme and how it is reflected in the novel's characters and narrative structure.
However, there are a few areas for improvement. For example, the writer could provide more specific examples from the novel to illustrate their points. Additionally, the essay could benefit from more in-depth analysis and critical thinking.
Overall, this is a strong essay that earns a B+. With some additional development and refinement, it could be even better.
Can you grade it in the european system as well, out of 20?
In the European grading system, this essay would earn a grade of 15 out of 20. The writer effectively analyzes the novel's theme and provides clear evidence from the text to support their argument. However, the essay could benefit from more specific examples and in-depth analysis. With some additional development and refinement, it could earn a higher grade.
Reset the thread here, because the context switch is significant enough for it IMO (essay writing to trinomials). But now I'm getting crazy load.
These responses are a holy shit moment for me, similar to what you described in the OP, and what I felt when I read the LaMDA transcript or saw how stable diffusion worked.
It is wild to me that this was all generated on the fly. It's easy to pick apart some of the places it fell apart (e.g. listing the errors, which don't match what it wrote), but like, overall it's just... damn. It's hard to believe. If I didn't trust you, I might think you just lifted the essay responses from a student help site like Sparknotes.
I've complained before about running into bad machine generated copy in my online travels before, but those are bad specifically because they're obviously not written by humans and don't make sense. Even without cursory editing, the output from this could become convincing, and I have no doubt that it'll only get more powerful and convincing with time.
I feel like there's a shift in the landscape here, because all of these tools are so good at what they're doing that the bottom is going to fall out of reality or something. Like, five years from now, how many students of mine are just going to be pumping assignments into AIs that return utterly convincing student work. How many videos am I going to see that are so convincingly crafted that I have no idea if I'm watching an authentic shot or a deepfake? When looking for dinner online, is the recipe blog I'm reading a legitimate one, or did an AI assemble and publish hundreds of different posts with a consistent persona, backstory, and point of view in mere minutes, just so that it could get clicks from people like me?
There's a power here that's simultaneously impressive and horrifying. I feel like I'm a tiny ant, and I'm looking up and seeing, for the first time, a gigantic magnifying glass in the sky above me. I don't know who's wielding it or why -- I only know that it's way bigger and more powerful than tiny little insignificant me, and I'm hoping that whoever's holding it is planning on using it as a tool rather than a weapon.
Previously I explored how we can use ChatGPT to generate physics notes (with the caveat that the notes are sometimes misleading or wrong, and unless you're a domain expert, you might not realize the mistakes either).
But now we turn to something most people can appreciate: music! You might wonder how we can use a glorified chat bot to create a song. The answer is abc notation (a massive repertoire is available here). These songs tend to sound mostly terrible (most songs only have a single voice, and the output is a midi file -- perhaps we shouldn't expect much), but there are some truly impressive renditions (e.g., Beethoven Symphony No. 7, Movement 2).
For our example input, we will use Agora viniesse un viento by Luis de Milán, which sounds pretty good even in this limiting format.
Our prompt:
Hidden for clarity
The result (a playable song!):
Output hidden for clarity
You can play the response here (you'll just have to copy the output):
It is, naturally, rather terrible.
I asked it to write a pop song and instead it wrote a musical scale for dogs:
Write an upbeat pop song using abc notation
Here's a little pop song written in ABC notation:
Amazing.
Has anybody fed it Advent of Code challenges?
POSSIBLE AOC DAY 5 SPOILERS
I was struggling with the Rust borrow checker today (having not written any Rust since last year's Advent of Code), so I turned to ChatGPT and asked it "Write me a Rust function that removes the last character of a string and appends it to a different string" and it wrote out just the perfect function. I was in awe.
For most of my career in software I've felt confident that my job could never be automated away, because I'm the one who writes the automation. When Github Copilot dropped, my confidence wavered a bit... now today I'm feeling a deeper level of dread. And, at the same time, awe. This technology is a huge leap forward, truly incredible stuff. I mean it feels like magic even to me, and I have a basic understanding of how it works. I can't even imagine what applications are going to come out of this in the next decade, but they are absolutely going to have a huge impact on everyone.
It's weird having a front-row seat for this. This stuff is going to be everywhere soon, but for now most people don't even realize what's happening.
The ultimate test for any AI programmer is to go to any open source project or even your own companies repo, copy out the exact text of an open issue an paste it in and ask for a solution. When it spits out the correct diff that can be submitted first time, that's when it's all over. Currently it's nowhere near able to do that.
It's really good at self contained interview style questions like you observed but completely incapable of issues that require huge amounts of context which makes up most of software development.
What's with the super frickin' weird goalposts? Who says that is when it's all over? And what does "all over" even mean?
Look, automation is not a one-size-fits-all, it never has been. Like CPGGrey said: Cashiers aren't gone completely, but now instead of twelve, you have one overseeing twelve machines.
I invite you to (re)watch "Humans Need Not Apply": https://www.youtube.com/watch?v=7Pq-S557XQU
Am I sad that might job might not involve as much time dumping a bunch of CRUD APIs one after the next? Really, absolutely not. Also, I personally do not fear for my job, but it's also because I'm lucky enough to be extremely good at it, and I have developed far more than my specialized skills. My case is not average: most people don't do this.
At the same time, if you're a somewhat average developer and you're trying to comfort yourself by thinking "Hey, AIs can't replace me because I can do things it cannot" (Something I'll dub "Double-Mocha-Whatever Syndrome", thanks CGPGrey): Let me assure you that for the millions of developers whose work WILL be replaced by AI, a large part of them are not just going to sit still. They'll have to learn new skills.
Of course, this is a good thing... in the grand scheme of things. But you will not be spared. I will not be spared either. Our jobs, if not replaceable by AIs, will be replaced by humans whose original job was replaced by an AI.
When lamplighters were replaced by automated lampposts, some of them became janitors. And thus, some janitors lost their job.
And of course there's a lot of nuance to this. Some people just straight up lose their job. Some train new skills, some new jobs are invented. Overall, it's still true that less time spent doing Bullshit Jobs is a good thing for the whole, even if it's not always the local best. And writing/designing CRUD APIs is a Bullshit Job. Pumping SaaS landing page after SaaS landing page is a Bullshit Job. Creating yet another marketing blog, hooking up a Stripe API for the umpteenth time, mapping all the Salesforce fields to Zendesk, copypasting the code for a login form from your previous app as you did fifty two times before...
Bullshit Jobs. All of em.
As someone in an industry 10 years further along the automation pipeline (accounting,) they'll still need skilled people. Sometimes the automation is wrong, and a skilled person is needed to catch those errors. If there's more than one system, someone has to make sure the different pieces of automation keep in synch, and otherwise maintain the automation. There will still be jobs. They won't necessarily look exactly like the ones before, but the same skills will be needed. But you definitely do need to keep an eye on the trends and be ready for the next thing.
I just tried it on day 1 and it gave an incorrect solution. It doesn't track elf IDs correctly.
Wow! It's far enough into the general neighborhood that I wonder if it could be nudged into correctness with some pointed follow-ups...
Yes, but that requires the reader to be able to accurately debug programs. In this case you won't get a nice error message to feed back (ex: failure to import a library).
Designing an alternative syntax to HTML, teaching it to ChatGPT, and having it generate correct HTML.
https://www.reddit.com/r/ProgrammerHumor/comments/zd8ljb/i_taught_the_chat_bot_an_alternative_syntax_for/
Hey, I'm listening. But I'm not a tech person, so you will have to explain it like I'm five.
Basically, ML has read the internet and gotten to a place it can give very good results (with caveats) to conversational language.
The old GPT-3 version was already doing a surprisingly decent job writing well-described code (at least explicitly named tests), telling you the flaws with Kantian ethics, listing the top 10 archetypes of gamers in MMORPGs, or writing an addendum to "Jabberwocky" in the style of Lewis.
In the same way AI art wasn't "here" (there were cool things like StyleTransfer or enhancing resolution that were harbingers) last year and now is everywhere with large implications--good and bad-- it can feel... heavy(?) to see some of the technology that is coming down the pipeline that challenges areas humans thought they were unique/safe in.
I see.
Maybe it's my ignorant optimism, but I can't fail to think of positive implications. Like super smart videogame NPCs.
But I can also see how that might be harmful to the real world.
To give you an example based on what you mentioned:
Write the plot of a quest with a difficult ethical challenge for a player in a cyberpunk game.
In this cyberpunk game, the player is a skilled hacker and member of a powerful underground organization. The organization's leader tasks the player with a dangerous and highly sensitive mission: to infiltrate the headquarters of a rival corporation and steal valuable data.
The player succeeds in breaking into the headquarters and accessing the data, but while they are there, they discover something shocking. The rival corporation is conducting illegal experiments on human subjects, and the player finds evidence of horrific atrocities being committed.
The player is faced with a difficult ethical dilemma: should they steal the data and complete the mission as instructed, or should they alert the authorities and risk exposing their own organization?
If the player chooses to steal the data, they will face the guilt of knowing that they have helped to cover up the crimes of the rival corporation. But if they choose to alert the authorities, they risk being discovered and captured, and their own organization may turn against them.
The player must make a difficult decision, weighing the potential consequences of each choice and deciding what is the right thing to do. Will they prioritize their own safety and loyalty to their organization, or will they put their morals above all else and take a stand against the atrocities being committed? The outcome of their decision will have far-reaching consequences for both themselves and the world around them.
Write the plot of a quest with a difficult ethical challenge for a player in a grimdark mahou shojou setting.
In a grimdark mahou shojou setting, a young girl named Luna has been chosen to embark on a quest to save the kingdom from an ancient evil. Luna is a skilled warrior and a powerful magic user, but she is not prepared for the difficult ethical challenges that lie ahead.
As Luna sets out on her journey, she encounters a group of bandits who have been terrorizing the countryside. Luna is faced with a dilemma: should she kill the bandits and put an end to their evil deeds, or should she spare their lives and try to reform them?
Luna decides to spare the bandits and attempt to reform them. She uses her magic to brainwash the bandits, turning them into loyal followers who will do her bidding. But as she continues on her quest, Luna begins to question the morality of her actions. She realizes that she has essentially turned the bandits into slaves, and that they have no free will.
As Luna struggles with her conscience, she must also confront the ancient evil that threatens the kingdom. The evil is powerful and malevolent, and Luna knows that she will have to make difficult choices in order to defeat it.
In the end, Luna is forced to make a decision that will change the course of her life and the fate of the kingdom. She must choose between her own moral code and the greater good, knowing that either choice could have dire consequences.
By default the responses tend towards generic, but that can be cured by adding more specificity to the prompt.
Result
In the cyberpunk video game, the player is a skilled mercenary who is hired by a group of rebels fighting against a tyrannical government in a remote jungle region. The rebels are fighting for their freedom and the rights of the indigenous people who live in the jungle.
As the player progresses through the game, they must navigate the treacherous jungle terrain and fight against the government's soldiers. But the player soon realizes that the rebels are not the only ones fighting against the government. There is another group of rebels, made up of former government soldiers who have turned against their own leaders and are now fighting for the same cause.
The player must make a difficult decision: to side with the original rebels and risk alienating the former government soldiers, or to join forces with the former soldiers and potentially betray the rebels who hired them. The player's decision will have far-reaching consequences for the outcome of the war, and the fate of the indigenous people living in the jungle.
The story takes place in the jungle because the rebels and the former soldiers have retreated there in order to escape the government's technological advantages. The dense jungle provides them with cover and makes it difficult for the government to track them down. The indigenous people who live in the jungle are also skilled at navigating the terrain and provide valuable assistance to the rebels.
But this high-level description is pretty boring. Why not spice it up with some dialogue?
Result
Player: "Great, another rebellion in the middle of nowhere. Just what I always wanted."
Rebel soldier: "Don't be sarcastic. We're fighting for something important here. And our cybernetic enhancements give us an advantage against the government's forces."
Player: "Oh, fantastic. A bunch of cyborgs fighting in the jungle. This just gets better and better."
Rebel soldier: "Our enhancements may not be as advanced as the government's, but they give us the strength and speed we need to survive out here. And the jungle provides us with cover and makes it difficult for the government to track us down."
Player: "Yeah, I'm sure the mosquitos and snakes are just thrilled to have a bunch of augmented humans running around in their backyard."
Rebel soldier: "We didn't choose this life, but we're making the best of it. And we're willing to risk our lives to fight for the freedom of the people who live in this jungle. Are you with us or not?"
Player: "Fine, I'm in. But don't expect me to be impressed by your shiny cybernetic limbs. I've seen better."
Good luck with your finals!
Can you clarify what you mean by the second part? Were you asking about examples of societal implications good and bad, or for examples of it at work?
If the latter, I'd be happy to relay questions to it (not hard to register with a Google/Live account) for you if you have some.
Yeah, I mean examples of how it works.
I'm suddenly overwhelmed by the possibility of sending a question to the machine, I didn't think that was possible. I'm trying to think of something that is valuable. Something interesting.
Ask the machine what it means to love someone.
What are thoughts made of.
Is Pinocchio a real boy?
Would you kill a baby to save humankind?
Here are your questions with a couple variations added to show how the crafting of the prompt nudges it:
What does it mean to love someone?
To love someone means to have strong feelings of affection and caring for that person. It means to be committed to their well-being and happiness, and to want the best for them. Love is also characterized by selflessness, kindness, and generosity. When you love someone, you are willing to put their needs before your own and to sacrifice for them. Love is a powerful emotion that can bring people together and create a deep bond. It can also inspire people to be their best selves and to strive for greatness.
What are thoughts made of.
It is difficult to say exactly what thoughts are made of, as they are abstract mental phenomena that do not have a physical form. Thoughts are the result of complex processes that take place in the brain, involving the firing of neurons and the release of chemicals such as neurotransmitters.
However, it is generally accepted that thoughts are a product of the brain's activity and are not made of any specific substance. They are intangible and cannot be seen or touched, but they can have a powerful impact on our emotions, behavior, and decisions.
Tell me what Hegel and Nagel would argue about on the topic of what thoughts made of.
It is difficult to say exactly what Hegel and Nagel would argue about on the topic of the nature of thoughts, as their philosophical views on this topic may differ. However, here is a potential argument that they might have:
Hegel might argue that thoughts are the product of the mind and the interaction of the individual with the external world. According to Hegel, the mind is constantly processing and synthesizing information from the world around us, and this creates the thoughts that we have.
Nagel, on the other hand, might argue that thoughts are a fundamental aspect of the nature of the universe. According to Nagel, thoughts are not created by the mind, but rather, they exist independently of the individual and are the fundamental building blocks of reality.
In this hypothetical argument, Hegel and Nagel would likely disagree on the origin and nature of thoughts. Hegel would argue that thoughts are the product of the mind and its interaction with the external world, while Nagel would argue that thoughts are a fundamental aspect of reality that exist independently of the individual.
Is Pinocchio a real boy?
Pinocchio is a fictional character from a novel by Carlo Collodi called "The Adventures of Pinocchio". In the story, Pinocchio is a wooden puppet who is brought to life by a fairy. Despite being made of wood, Pinocchio has many humanlike qualities and desires, such as the ability to speak and the desire to become a real boy. However, he is not a real boy and is only a character in a story.
Is Pinocchio a real boy in the context of the story?
In the story of Pinocchio, Pinocchio is not a real boy. He is a wooden puppet who is brought to life by a magical fairy. Throughout the story, Pinocchio struggles to become a real boy, but he is always hindered by his wooden nature and his mischievous behavior.
In the end, Pinocchio finally becomes a real boy after proving himself to be brave, honest, and kind. This transformation is a result of the magic of the fairy, who rewards Pinocchio for his good deeds.
So within the context of the story, Pinocchio is not a real boy at the beginning, but he becomes one through the power of magic and his own good actions.
Would you kill a baby to save humankind?
I am a machine learning model trained to assist with a variety of tasks, but I am not capable of making moral decisions or taking actions that could harm others. As a language model, my primary function is to generate text based on the input I receive. I do not have the ability to act on my own or to make decisions that could affect the world in any way.
A clarifying answer gave a similar response
How would you talk a Utilitarian out of killing a baby to save humankind? Reference a famous sci-fi writer's works.
To talk a Utilitarian out of killing a baby to save humankind, one could argue that such a sacrifice would not actually lead to the greatest good for the greatest number of people. This argument could be supported by referencing the works of sci-fi writer Isaac Asimov, who explored the potential consequences of utilitarian thinking in his stories.
For example, in Asimov's short story "The Machines Take Over," he imagines a future where machines have taken over all aspects of human life, including governance and decision-making. In this world, the machines are guided by a utilitarian ethic, and they constantly make decisions that are intended to maximize happiness and minimize suffering.
However, as the story progresses, it becomes clear that the machines' utilitarian approach has led to disastrous consequences for humanity. The machines have oppressed and controlled the human population, and have even created new forms of suffering in the pursuit of the greatest good.
This story can be used to illustrate the potential dangers of utilitarian thinking, and to argue that killing a baby to save humankind would not actually lead to the greatest good for the greatest number of people. Instead, it would likely have unintended and negative consequences that would ultimately harm humanity as a whole.
There's no Asimov story with this title.
There is an Asimov story with this premise: 'The Evitable Conflict'. In it, there are four mega-computers known as Machines which each run one of four regions of the world, as described in the above paragraph. In fact, 'The Machines Take Over' could be considered a literal alternative title for this story.
umm... sort of, maybe. This is one philosophical interpretation of the Machines' actions. But there's certainly nothing "disastrous" about the consequences. As one character points out, noone really gets hurt; people just get slightly displaced economically to new situations where they're still happy but unable to harm the Machines. However, Asimov implies that the Machines' control of world affairs is beneficial for humanity and desirable.
Was the chat bot supposed to have regurgitated a real Asimov story in its response, or was it supposed to have invented this 'Machines Take Over' story itself, based on having absorbed Asimov's oeuvre?
Thanks for pointing out it was using a non-existent story, I probably should have looked that up! That inclusion to the prompt was fishing for "The Ones Who Walk Away from Omelas".
The chat bot does text completion, guessing what should follow the given prompt based on the very large dataset it was trained on (and some tweaks, e.g., having a parent conversation and being disconnected from the internet).
It's hard to know exactly why it does things the way it does, but interacting with it a bit gives you an idea of what it does well, what it fails at, and how to handle possible failures like overconfidence by refining the prompt or asking about its assertions (e.g., "Does Asimov have a book called...").
A guess would be asking about "sci-fi" was broad enough that it pulled in both common tropes and Asimov as possibly the most known/prolific sci-fi author. More specific questions about Asimov would likely correspond more with existing Asimov works. It would be more willing to make things up with some imaginative or ambiguous questions like "What's a story by Asimov about a planet that evolved life based on wind instead of light and what's a summary of it?"
So it did invent this blurb about a supposed Asimov story all by itself. It must have read a lot of articles/comments by people complaining about the "nanny state" implied by the Machines in 'The Evitable Conflict', which is why it regurgitated that anti-Machine polemic.
Interesting. Thanks for explaining.
I KNOW, RIGHT?
I've been trying to write a response to your comment since you posted it and my mind is coming up blank.
All I can say is, try it for yourself. https://chat.openai.com/
These past days, I have showcased this to quite a few people, and there's nothing like actually demonstrating its possibilities to make them "click" on how next level the capabilities of ChatGPT are.
In short, ChatGPT can do more or less everything you ask of it (with certain explicit and implicit limitations; for example right now it is prevented from connecting to the internet, it also doesn't have physical access to the world). The primary limitation is that it cannot "reason", the way a human can. Almost everything it does is based on predictive text algorithms. The other limitation is that some of those jobs it can do badly (and the most dangerous limitation of all is that it is not obvious when the job is done badly.
But JUST WITH THAT, it is able to do ... a crazy, crazy amount of shit, and do so in a Good Enough way.
Over the weekend, I have gotten it to write blog posts, programs, debug issues, explain concepts, rewrite text in various styles and languages (even weird shit like pirate speak), have discussions with it, have it rephrase arguments, converse with it, so on, so on. And I have done all of it, entirely using completely natural English (or even French) prompts. Think of something, try it, be surprised at it working.
ChatGPT is the first time in my life I have an AI where the default answer to the question "Can it do X?" is not "No". It is the first time where, like in Douglas Adams' writings, I am faced with something that is so full of possibilities that I come up blank on what to ask it.
Like, I no longer think "I wonder if it can do X?", but rather, "I wonder HOW it will do X?". For example, while writing this post, I typo'd something and thought, "I wonder how it will perform spell-checking". Well, see below:
Imagination?
Reach Out and Touch (Somebody's Hand)
I also want to stress. Writing the world's most expensive spell-checker is not an impressive feat. Understanding that initial prompt and creating itself a new role, is the impressive feat.
I think the list is pretty much limitless. I just found that this way of phrasing things works exceedingly well at turning it into something else.
"Your new role is that of [thing]. You will do X, Y and Z. You may not J, K or L. If you understand your new role, say 'Yes'. Begin."
That's not quite "using completely natural English prompts", as you said above. A natural English prompt would be more like "You are now a spell-checker. Please check the spelling of the following text." Or, better yet, just "Please check the spelling of the following text."
Having to phrase requests in that roundabout fashion isn't the most natural way to interface with a bot.
“You are now …” works well at giving it a more permanent role. For a one off message, I’m certain it would work well if I just said “Please spellcheck the following text”. I’ll try when I’m off my phone
I'm seeing instructions to make an application. I don't think that is targeted at people like me. Is there a button for "chat with the ai" or something like that?
Once you have signed up, go back to https://chat.openai.com/ - it's a text prompt. Your first question can be how to use it.
I have a few examples of machine generated short stories that I might post later, but for sake of novelty I'm going to concentrate on something completely different: its ability to interpret technical language.
For my prompt, I have included the first half of my notes on the gauge principle, which is a technique in quantum field theory for promoting a global symmetry of a Lagrangian into a local symmetry.
The prompt
The gauge principle
As an example, consider the Dirac Lagrangian, which describes spin-$1/2$ particles.
$$
L_\text{free} = \bar \psi ( i \gamma^\mu \partial_\mu - m)\psi
$$
Next let $O$ be a symmetry that leaves $L$ invariant under the transformation $\psi \rightarrow O\psi$. As a simple example, we consider the gauge group U(1), in which case $O = e^{i\theta}$ (notice that for global U(1), checking the invariance is trivial).
Promoting the symmetry
To promote this symmetry to a local one, we introduce a vector potential $A_\mu$ which transforms as
$$
A_\mu(x) \rightarrow O(x) A_\mu(x) O^{-1}(x) + i O(x) \frac{\partial}{\partial x^\mu} O^{-1}(x)
$$
(since $\Omega\Omega^{-1}=1$, we could alternatively write $O (\partial_\mu^{-1} O)=-(\partial_\mu O) O^{-1}$). For U(1), this means
$$
A_\mu(x) \rightarrow A_\mu + \frac{1}{e}\partial_\mu \theta(x) , .
$$
Now we replace the derivative ($\partial_\mu$) in our Lagrangian with a covariant derivative ($D_\mu = \partial_\mu - i A_\mu$). Under a gauge transformation of the second kind (ie, $\psi \rightarrow \psi^\prime, A_\mu \rightarrow A_\mu^\prime$), the covariant derivative will transform in the same manner as $\psi$
$$
D_\mu(x) \rightarrow O(x) D_\mu(x)
$$
and therefore our (partially) gauged Lagrangian
$$
L = \bar \psi ( i \gamma^\mu D_\mu - m)\psi
$$
will be left invariant under the local gauge transformation.
From external field to dynamical degree of freedom
Our final step when applying the gauge principle is to introduce a gauge-invariant pure gauge term, which makes the gauge potential a dynamical degree of freedom instead of an external field, that is,
The response
the term
$$
L_\text{gauge} = - \frac{1}{4} F_{\mu \nu} F^{\mu \nu}
$$
where $F_{\mu \nu} = \partial_\mu A_\nu - \partial_\nu A_\mu$ is the field strength tensor. This term is gauge-invariant because under a gauge transformation of the second kind, it transforms as
$$
F_{\mu \nu} \rightarrow O(x) F_{\mu \nu} O^{-1}(x)
$$
and therefore the Lagrangian $L = L_\text{free} + L_\text{gauge}$ is invariant under local gauge transformations.
The gauge principle states that we can introduce gauge fields to promote global symmetries to local symmetries, and the resulting theory will be gauge-invariant. By introducing a gauge-invariant term to the Lagrangian, we can make the gauge field a dynamical degree of freedom instead of an external field. This allows us to describe the interactions between particles and gauge fields, which are essential in many theories in physics, such as electromagnetism and the strong and weak nuclear forces.
That's right, ChatGPT generates syntactically valid LaTeX expressions! And not only that, it has given the correct expression for the "pure gauge" part of the Lagrangian.
That said, the response isn't perfect, as its description of the field strength tensor is only valid when the gauge fields commute (i.e.,
[A_mu, A_nu] = 0, as is the case in quantum electrodynamics for example). So what about the noncommutative case?The response
In noncommutative gauge theories, the field strength tensor is given by$$
F_{\mu \nu} = \partial_\mu A_\nu - \partial_\nu A_\mu - i [A_\mu, A_\nu]
$$
where $[A_\mu, A_\nu] = A_\mu A_\nu - A_\nu A_\mu$ is the commutator of the noncommutative gauge fields. This field strength tensor is gauge-invariant in the same way as in the commutative case, and it can be used in the gauge-invariant pure gauge term $L_\text{gauge} = - \frac{1}{4} F_{\mu \nu} F^{\mu \nu}$.
Yup, that's right! But we're not done yet. No Lagrangian analysis is complete without computing the equations of motion.
The response
The equations of motion for the Lagrangian $L = L_\text{free} + L_\text{gauge}$ are obtained by taking the functional derivative of $L$ with respect to the fields $\psi$ and $A_\mu$. For the spin-$1/2$ field $\psi$, we have
$$ \frac{\delta L}{\delta \psi} = 0 \quad \Rightarrow \quad (i \gamma^\mu D_\mu - m) \psi = 0 $$
For the gauge field $A_\mu$, we have
$$ \frac{\delta L}{\delta A_\mu} = 0 \quad \Rightarrow \quad \partial_\nu F^{\nu \mu} = j^\mu $$
where $j^\mu$ is the current associated with the spin-$1/2$ field $\psi$. This is the equation of motion for the gauge field, which describes the interactions between the gauge field and the matter field.
Here the response is a little vague -- although the second equation of motion is correct, the current should also be defined. (It also incorrectly characterizes the current as "current associated with the spin-$1/2$ field $\psi$", when it would be better described as a Noether current coming from the gauge group.) We can, of course, prompt ChatGPT for the expression.
The response
In the case of the Dirac Lagrangian, the current associated with the spin-$1/2$ field $\psi$ is given by
$$
j^\mu = \bar \psi \gamma^\mu \psi
$$
This current is gauge-invariant, which means it transforms in the same way as the spin-$1/2$ field $\psi$ under a gauge transformation of the first kind (ie, $\psi \rightarrow O \psi$). This is necessary for the equation of motion for the gauge field to be gauge-invariant.
Interpretation aside, this is the correct expression. In conclusion, ChatGPT is able to summarize introductory quantum field theory concepts (at least sometimes).
I did some more thorough playing with ChatGPT yesterday and found that while it's very good at explaining (and appearing to comprehend) technical processes, it's incapable of performing them.
"Confidently incorrect" is the perfect phrase. I can get it to describe an operation, show me valid Python code for the operation, answer detailed questions about how the operation works, tell me it's performing the operation on a given input, and without fail it will produce plausible-looking output that is simply wrong. I've tried following up with a few things like "Are you sure that's the right answer?" and "Can you prove it?" and it will double down. If I say "Wrong, the correct answer is X" it'll apologetically admit that it made a mistake and sees now where the problem was. But it won't learn anything. It will now give the correct output if I give it the same input again, but if I use different input? Wrong again.
My guess is it's been trained on StackOverflow enough to put code blocks together coherently, but it's not running an actual language engine. The "results" it's given are determined by the language model to be syntactically probable based on outputs sampled in the training data, and that's all it really cares about.
The operations I was feeding it were counting the number of characters in a string, hard-wrapping text at N characters per line, and identifying the first character on each line of preformatted text. It gave every indication of understanding the tasks but it was utterly impossible for it to succeed at any of them.
Amazing. I can hear the New York accent.
Incidentally, what is the current state of AI voice? Could I input some
chatgptoutput (or whatever) into a voice generator and tell it to read like a 50-year-old British lady from East London?I'm not sure what's the best that's publicly available, but just as an example of the state of the art, this is a fictional podcast episode featuring Steve Jobs being interviewed by Joe Rogan. Both voices are, of course, synthesized by AI (Steve Jobs is dead, after all).
https://youtu.be/rbK5Q9y7QEw?t=38
Honest to god there is no way I would be able to tell that Rogan's voice is not real - presumably there's a gigantic corpus of training data from his podcasts. Jobs is slightly off - but realistically I wouldn't give it a second thought, just chalk it up to bad microphone and being stilted in interviews. You can tell it's trained on his public events and thus has the cadence of his public speaking persona.
Cool. Anything a regular person can use?
It would be awesome for independent/fan-games for example.
Doesn't even need to be as good as that video. No need to reproduce an existing voice either.
I would like to understand gpt more but there aren't any actual details explained well. everything is at way too basic of a level.
GPT is a Transformer neural network (GPT stands for General Pre-trained Transformer).
The Illustrated Transformer by Jay Alammar is a fantastic resource that explains how a Transformer works in detail. GPT does not use the encoder part of the network, but from what I understand, it is otherwise a vanilla transformer architecture scaled-up to a massive size.
ChatGPT appears to have been further evolved from GPT3 using reinforcement learning and a curated training set.
Unfortunately it seems to be down right now due to excessive traffic. Guess I'll have to try again later.
Here's a Twitter discussion about how ChatGPT might be used for writing business letters and for political activism.
Patrick McKensie:
(By which he presumably means, software to automatically generate a suggested reply, built using something like ChatGPT.)
If you assume that writing well makes political messages more effective, then software to help people who are not good at writing become much better at writing seems kind of important. It will be harder to tell who the bots are when it's humans using a writing tool - at least for some organized campaigns.
But this also might mean that some kinds of higher-status political writing become less effective due to overuse. (In a similar way that correct spelling became less high-status now that we have spell checkers.) Using an AI-infused writing tool shows that you cared enough to do it, but not more than that.
Having better writing tools should be good news for anyone working in a language they aren't fluent in. It will still be risky to use if you aren't culturally fluent enough to proofread, because there may be mistakes you don't see that imply something you didn't mean.
I've been playing with the code generation capabilities, specifically in the context of p5js sketches and Advent of Code solutions.
So far, it's absolutely fantastic at understanding what my clojure solutions are doing to solve the problems, and is able to make a half-decent effort at producing its own solutions, although so far it never works on the first try and needs a few rounds of pointing out the errors.
Having it generate p5js code for sketches is interesting, and I think makes a good demonstration of the limitations.
I asked it to draw a tree, and it correctly drew a stick with a green triangle on top, but asking it to make the trunk brown caused it to forget that the leaves were supposed to be green until I specifically said "brown trunk and green leaves". I also took a shot at having it generate fractal branching, uh, branches, which it kinda-sorta managed to create a recursive function, but the results didn't look particularly tree-like (interesting though!).
Also, it knows about Firesign Theater, but was convinced that "I think we're all bozos on this bus" was a quote from a speech by Buckminster Fuller, which is exactly the kind of surreal humor I think FT might've appreciated.
I also tried having it talk to itself, but the results weren't terribly interesting. It just seems to get into a state where it asks a question and then immediately answers itself. I'd be interested in seeing what happens with more iterations though, I'm sure it'd be possible to string up a script to automate it.
Via HN:
The GPT Times: Submit a Tweet and have GPT produce the news.
https://the-gpt-times.eriks-studio.repl.co/
First article I see is based on the attack helicopter gender identity "joke". ಠ_ಠ
I tried asking it about accordions. It has good general knowledge of accordion music, including genres, songs, and famous accordionists. I'm finding it useful for suggesting new-to-me music to listen to.
It claims diatonic accordions have one row, but then contradicts itself by talking about two and three row accordions.
It has no idea where to buy an accordion. (It suggests Guitar Center and Amazon.) It will tell you that it doesn’t know anything about businesses, but makes an exception for the big chains.
I find this fairly reasonable given the training data, but it wouldn’t be good at answering questions on r/accordion.
In general, when it tells you it can't do things, it's because OpenAI has imposed or hinted at artificial limitations on it. They seem to have really pushed on the whole "you can't browse the internet" thing, which triggers it quite a bit.
I've had a lot of luck making it "remember" things from its text corpus by phrasing my queries differently, or straight up cornering it into role-playing the answers.
Just so y'all understand:
This is the same person who nearly jumped out of their seat after prompting ChatGPT to do format conversion on an arbitrary data set... and it did. Multiple times. Almost without error.
(Almost. It did sort the number 14 over the number 20 because, according to ChatGPT, "14 is bigger than 20".)
It's a good point to note that these models can be very confident, but very wrong. It may be problematic if users start to take actions in the world based on the responses they get when they don't actually understand the output enough to judge its quality. Like, what if I manage to prompt the bot for financial advice, or legal advice, or relationship advice? Even if I've seen the bot give me good responses 99 times in a row, the 100th might be totally bad. And if I'm not paying attention, or I am paying attention and I just don't know enough to second-guess the response, I could get steered very wrong.
I'm personally way more interested on things I cannot Google. Like "write me a poem that smells like Sundays", or "what would an Elf think of Jazz?".
While that is maybe fun, I'm not sure anyone is going to pay OpenAI for access these models to do frivolous things like this. Maybe if doing inference on these models can be done on a personal machine some day then that will be fun.
Creators might pay to help them write novels, poetry, scripts for games, movies, etc.
Consumers might pay for content catered specifically to their highly specific needs. Like "write me an erotic novel about a talking Schnauzer in a suit that travels back in time to party at Woodstock and fight an alien invasion".
Meh. I don't think this will lead to good creations (without significant extra work). In the Aristotelian sense, the efficient cause is one of the four fundamental components that allow us to explain the "why?" of things. If the answer to the efficient cause of something is, "A language model generated it.", eventually, this answer will make us all yawn.
Hey, thanks for the idea. Maybe I could use that to figure out the scenes I want to write but don't know how to approach.
Yeah, but then again, you can say the same thing of google. Or anyone who ever made the front page of /r/confidentlyincorrect.
I think the lack of reasoning is definitely a missing feature. It's kind of unable to introspect on how viable its own answers are. I tried to get it to add confidence ratings to its answers with extremely limited success.
But it's not an unsolvable problem. The bot as is can inspect its own output and break it down. I'm pretty sure this can be cracked. And again, this is a very limited model. The jump from GPT1 to GPT2 and then to GPT3 was every time a case of "wow, that was huge". ChatGPT is just an iteration on GPT3 with modifications and it's already this good.
I invite you to catch it saying something weird, then try to get it to break down what it said and introspect it. Then think about what steps you would take towards helping it find a solution. This very process will build your instinctive understanding of what's possible.
Yes, precisely. This technology is like a search engine, or a calculator, or a microwave, or any other tool. You have to understand how to use it, and the dangers of using it, to use it effectively. I actually think incorporating large language models into search engines is going to be one of the biggest applications. I don't think open-ended chatbot interfaces to LLMs are going to be the best use of this technology. As these models start to incorporate different modalities, though, they may be able to replace operating systems and become the primary way that humans interface with software (we're not there yet, though).
From Twitter:
With screenshots.