Cipheron on 17/5/2023 at 07:47
Quote Posted by demagogue
I imagine they're focusing so much on training AI to code to deal with that.
Yeah, this is the way out of the conundrum. Basically teaching GPT to write algorithms and then execute them inline. Right now they have the "presets" for simple tasks like I demonstrated. But these are limited in that humans can't think up every possible situation and hand-code a preset for them.
As for today's quickie. I just read a mention of Martin Luther King Jr's "I have a Dream" speech. I decided to ask ChatGPT to rewrite the speech in a specific person's style. This was a first attempt, and pretty good:
Quote:
what would martin luther kings speech be like if he had the attitude of XXXXXX XXXXX. generate a sample.
GPT: Ladies and gentlemen, today I stand before you as a tremendous leader, one of the greatest leaders this country has ever seen. Believe me, nobody can deliver a speech like I do. Nobody. I have a tremendous speech, and it's going to be fantastic. It's going to be huge.
I have a dream, a really fantastic dream, let me tell you. It's a tremendous dream, the best dream. And in this dream, I see myself surrounded by gold, tremendous gold. The best gold you can imagine. I dream of a world where everyone loves me, where everyone admires me. I mean, who wouldn't love me? I'm a winner. Winners love me.
Now, I must say, we have some real losers in this country. They're total losers, folks. We need to get rid of them. We need to make America great again, and that starts by getting rid of all the losers. We need to build a wall, a big, beautiful wall, to keep out all the losers and the criminals. And let me tell you, Mexico is going to pay for it. They have to pay for it, and they will pay for it, believe me.
I've achieved so much in my life, so much more than anyone else. I'm the best at everything. And that's why I'm the perfect person to lead this country. We're going to win so much, you'll get tired of winning. We'll be winning in education, winning in healthcare, winning in jobs, winning in everything. Nobody knows winning better than me.
So, let me tell you this, folks. I don't care about your skin color or where you come from, as long as you're a winner. Winners are the only ones who matter. Losers, well, they're just losers. But winners, winners are the ones who deserve the best. And with me as your leader, we're going to make sure winners get what they deserve.
Thank you, and God bless me. I mean, God bless America. Because America is going to be so great under my leadership, you won't even believe it. It's going to be tremendous. Thank you all, thank you.
You'd almost believe GPT was going for snark there in the bolded part, it's perfect.
Now, that's only based on a text-prediction engine going off existing speeches by the fellow. So rather than a parody, what I think that is actually doing is cutting through the polite stuff and just basically inferring what he is really saying.
EDIT: i tried a few notorious serial killers and for the most part, ChatGPT threw up a filter and refused to write anything. They forgot to filter Ed Gein however, and it wrote a really gory version of "i have a dream". It was 100% on the money, so much so that i deleted it and will not reprint anything from it.
Also i need this to be a thing. It's from a cooking show script:
Quote:
PROMPT: Imagine Carl Sagan hosted a cooking show along with Dolly Parton. they're making a typical southern Sunday dinner. write the script.
...
[Dolly and Carl start marinating the chicken pieces in a bowl of buttermilk.]
Dolly Parton: While that's marinating, Carl, why don't you tell our viewers a little bit about the wonders of the cosmos?
Carl Sagan: Of course, Dolly. Did you know that the universe is a vast expanse, containing billions of galaxies, each filled with countless stars? It's truly a cosmic symphony of creation.
[Dolly attentively listens while occasionally checking on the marinating chicken.]
Dolly Parton: Well, Carl, while you're busy with the cosmos, I'll be tending to our collard greens. Now, y'all know Southern greens ain't complete without some bacon for flavor!
This was also the first time I've used the new "Continue generating" button, and it worked really well, no break in the output unlike a manual "continue", and was able to write the whole "episode" to completion:
(
https://pastebin.com/fNMDshb4)
I think this kind of thing highlights the creative potential of generative AI. You can say on the one hand, that it's only predictive text, so not actually creating anything, but if you read this output script, you can see how just from text prediction it's actually inferred the back and forth between these two people into what would realistically be a way such a show could work, without needing to be told that's the goal.
heywood on 17/5/2023 at 13:37
Quote Posted by demagogue
I imagine they're focusing so much on training AI to code to deal with that.
Now that I think about it, I imagine another thing they might try is training AI to gather its own training corpus and running its own training to create a new model to fulfill whatever it finds it needs for a task.
And generally speaking, I think the next generation of AI is going to focus more on models that modify their own weights to train and switch between task specializations.
Once it goes recursive like that, then I imagine the whole field's development might rapidly accelerate, but that's just a knee jerk intuition I had.
Once it goes recursive, we're off into sci-fi territory.
I think people are overreacting to AI right now, it's part of a broader feeling of future shock. As long as we're supplying the motivation, AI is a powerful tool that's going to raise our standard of living. But internet-crawling AI bots that determine their own training material are going to be interesting if not worrisome.
The red line has to be allowing an AI program to modify its own code. There's a straight line from there to real sci-fi stuff like Tron's MCP or persistent internet viruses that evolve.
Cipheron on 17/5/2023 at 19:49
Historically, we've been bad at determining what things computers will be good at. A computer can write a sonnet or paint a painting, two things that were help up not that many years ago as things they would never do, but you still can't make a robot that can ride a bicycle.
A quick video on Google's next LLM. It feels like Bard is temporary, just to get some feedback and iron out any wrinkles, but their real deal is the other one they're working on:
[video=youtube;3PgTuJvqwKw]https://www.youtube.com/watch?v=3PgTuJvqwKw[/video]
demagogue on 31/5/2023 at 23:32
Here's a better demo (than the previous one I posted) of AI you can chat with in game.
He claims it's unscripted free speech, which isn't hard to believe now, although I'm sure he knew generally what to ask the AI for that was going to get a game-relevant response. The open question is what happens when you start chatting way off script about things that have nothing to do with the game. (There is a clip of this on Twitter, but on YT I could only find the full video and set it to start at the start of the clip. The video tag here doesn't seem to be starting it in the right place though, so if you want to go straight to the clip it's at 10:51.)
[video=youtube;i-wpzS9ZsCs]https://www.youtube.com/watch?v=i-wpzS9ZsCs&t=651s[/video]
Cipheron on 1/6/2023 at 03:33
Also, there was this recent talk by the co-creator of OpenAI.
I highly recommend this video. It's very thorough, covers a lot of the common misconceptions etc, and then some. It really explains the full development pipeline that lead up to ChatGPT, in quite a bit of detail:
[video=youtube;bZQun8Y4L2A]https://www.youtube.com/watch?v=bZQun8Y4L2A[/video]
Cipheron on 9/6/2023 at 22:57
University theatre class puts on a play written by ChatGPT. The subtext is the interesting part here
Quote:
While most playwrights spend months or years writing a theatre piece, the University of Wollongong commissioned one that was done in minutes.
Ten Strangers in a Room was written by artificial intelligence (AI) website ChatGPT.
The work was then rehearsed by the university's acting students and performed on stage for a paying audience.
"It was an arduous process and there was actually a writing process involved where we were interacting with it to get to where we got," University of Wollongong theatre fellow and the show's director Matthew Prest said.
"done in minutes" was the journalists take, but if you look at the actual process, it was a drawn out one where they first had to get ChatGPT to generate the framework for a three-act play, then break that down and generate it scene by scene, and even then nothing happened in the play so they had to micro-manage ChatGPT to add some twists into the play.
Later on they do a sort of "God of the Gaps" argument:
Quote:
"AI is never going to be able to reflect genuine human emotions," she said.
"It can create a caricature where it will tell you exactly how someone is feeling and you lose those nuances, so we are telling, not showing.
This is the common thing we're we draw a line exactly where the AI is up to then claim that everything that it hasn't done
yet is impossible for AI.
I think even with ChatGPT as it currently is, you could probably get "show don't tell" out of it by being careful with how you prompt it. It's not going to know you want that unless you ask it.
Hell, it wasn't that long ago that it was science fiction for a chatbot to recall what was said earlier in a conversation, and now it can write a play which is self-consistent and tells you exactly how characters are feeling and reacting to events, but we're gonna say "aha! but it's depictions aren't nuanced enough, so AI will NEVER write good subtext".
Basically, if writer's arguments against AI now boil down to the plot twists aren't subtle enough, then they SHOULD be worried about AI taking their jobs.
Starker on 10/6/2023 at 04:17
There's an incredible, massive difference between generating the structure and general plot of a play based on numerous examples and really coming up with a play on your own, though. That's because humans can draw from their own unique life experience and real world knowledge to extrapolate and produce things that are truly creative.
For example, a large language model based AI can only copy someone like Gene Wolfe, it could never come up with something like this on its own, no matter how much science fantasy it was given to train on. That's because Wolfe didn't only draw on literary traditions, old and new, but also from his own life and views on things.
Or, otherwise, an AI couldn't tap into the economic and cultural anxieties of the 80s, observe how humans behave around emerging technologies, and combine this with Toronto biker slang, in order to produce cyberpunk works the way William Gibson has.
heywood on 12/6/2023 at 16:18
Lack of life experience is a temporary limitation. A dynamic AI designed to learn from its interactions with people will accumulate far more experiences to draw from than any human.
The more interesting question is motivation.
Starker on 12/6/2023 at 17:07
Funny you should say that. I just today listened to a podcast where they explored this idea in some length, more specifically the motivation to kill all humans. It's an episode of Decoding the Gurus, where one of the hosts happens to be a former AI researcher and psychologist and this time they were talking about Eliezer Yudkowsky, a somewhat prominent Chicken Little of AI development who has suddenly found a larger audience with the emergence of ChatGPT: (
https://decoding-the-gurus.captivate.fm/episode/74-eliezer-yudkowksy-ai-is-going-to-kill-us-all)
