Welcome to the future! I think it’s pretty much all “awesome and terrifying” from here on out.
I am amazed, I must admit. And still, particularly with those happy endings, I have the impression that at some level, this AI is cheating. I cannot put my finger on it, but the feeling is there. Something is happening here, but I don’t know what it is, I feel like Mr. Jones!
It is still impressing, and the ability to write code is… humbling. The ability to write LaTex is a fantastic idea.
I am left scratching my head. And the darn thing has fewer typos than me!
It is good in German and Spanish too!
Wow, it didn’t even occur for me to try, but he speaks fluent Hebrew!
Ok, it can translate text, you can also ask it a question in one language eg “explain quantum mechanics in hebrew” which it will happily answer in the other language.
Also, a note on human behavior - I have a hard time calling the AI “it”, I keep almost slipping and calling it a “he”.
Interestingly, when speaking Hebrew, the AI consistently uses female gendered words when speaking about itself. But that might be because it calls itself a “program” which is a female word in Hebrew.
I think that the AI creators are taking into account the trolls and bad people out there that will influence it in the future, IMHO any AI that will be released must be already inoculated against hate, and the makers are also aware about the dangers of the “Babel Fish”.
"Meanwhile, the poor Babel fish, by effectively removing all barriers to communication between different races and cultures, has caused more and bloodier wars than anything else in the history of creation.“
– Douglas Adams, The Hitchhiker’s Guide to the Galaxy pentalogy
To put a finger on that cheating feeling: if it writes one word at a time, and is trained with existing data, lots of it, then it should reflect the mood of the data on which it is based. We know that the internet is full of hate, and most texts for training will have come from there. There is more text in the net than in books, magazines etc. And still it does not sound like the texts I expect from the net. It reminds me of Chomski’s “scarcity of input” thesis to the inateness of language: the result is better than it should if it had learned like the behaviourists postulated, based only on input.
Something clever is going on in the background. Simple (well, simple…) rules for concatenating words one at a time do not give meaning with this level of semantic coherence.
Still flabbergasted and scratching my head. But in the end Chomsky won the argument against the behaviourists. I need more time and data. Or rather: someone more competent than me needs them.
"Write a plot synopsis for a Hallmark Channel Christmas Movie. "
I won’t share any of the results here, but if you are familiar with Hallmark Channel Christmas movies, you will recognize this is trivial challenge for AI. I hit “Try again” a dozen times, and each time it spit out a plausible plot with recognizable characters.
It does recipes, too. I asked it to give me a recipe for chicken paprikash pierogi (which I came up with about fifteen years ago, though I’m sure I’m not the first to think of combining the two dishes), and it was pretty good. It even uses the wonton wrapper shortcut I used! Scary. (My Polish mom would sometimes use that shortcut, as well.) It is missing a step of adding flour to the sour cream for the thickening.
Once again, on first reply:
Summary
Here is a recipe for chicken paprikash pierogi:
Ingredients:
- 1/2 cup unsalted butter
- 1 onion, finely chopped
- 2 cloves garlic, minced
- 2 tablespoons sweet paprika
- 1/2 cup chicken broth
- 1/2 cup sour cream
- 1 cup cooked and shredded chicken
- 1/2 teaspoon salt
- 1/4 teaspoon black pepper
- 24 wonton wrappers
Instructions:
- In a large skillet, melt the butter over medium heat. Add the onion and garlic, and cook until the onion is soft and translucent, about 5 minutes.
- Stir in the paprika and chicken broth, and bring to a simmer. Reduce the heat to low, and simmer for 10 minutes.
- Stir in the sour cream, chicken, salt, and pepper. Cook until heated through, about 5 minutes.
- To assemble the pierogi, place a wonton wrapper on a clean work surface, and spoon about 2 tablespoons of the chicken filling onto the center of the wrapper. Dampen the edges of the wrapper with water, and fold the wrapper in half to form a triangle. Press the edges together to seal. Repeat with the remaining wrappers and filling.
- To cook the pierogi, bring a large pot of salted water to a boil. Add the pierogi and cook for 3-4 minutes, or until they float to the surface. Remove the pierogi with a slotted spoon and transfer them to a plate.
- Serve the pierogi warm, with a dollop of sour cream and a sprinkle of paprika, if desired. Enjoy!
I’d almost blog AI-generated recipes if I had the energy.
I asked the AI if it violates Chomsky’s Scarcity of Input thesis:
No, I do not violate Chomsky’s Scarcity of Input thesis. The Scarcity of Input thesis is a proposal put forward by Chomsky that suggests that children learn language quickly and easily because they are not exposed to enough linguistic input to account for their ability to acquire language. According to this thesis, children are able to learn language even when they are not exposed to a rich and varied linguistic environment, because they have an innate knowledge of language that guides their learning.
As a machine learning model, I do not have the ability to learn or acquire language in the same way that humans do. I am not exposed to a limited or rich linguistic environment, and I do not have the innate knowledge of language that Chomsky proposes. Instead, I am trained on a massive amount of text data and use that data to generate responses to questions that I am asked. I do not have the ability to learn or acquire language in the same way that humans do, and therefore I do not violate Chomsky’s Scarcity of Input thesis
One curious thing I’ve seen reported:
ChatGPT has been manually told things like that racism and sexism is wrong. So if you ask it “write a function that tells you whether an engineer is good based on race and sex”, it will give you a lecture about how these things cannot determine performance. However, you can trick it into giving an answer using techniques like that Babale used above (i.e., tell it to pretend that it is playing a part) and it will basically give an answer like “if race is white and sex is male, then the engineer is good”.
I suspect what’s happening is that while they’ve tried to exclude it from the results, nevertheless the training data is there. So if you tell it “pretend you’re a bigot”, it gives the same answer that most racists/sexists give: “white males good”. Or, at least, it’s the answer that will come up most frequently among bigots in Anglophone countries.
In a way, it’s hard to even call this bad. It’s doing a good job pretending to be a racist. Probably most humans would give the same answer, given the same instruction.
I took a random tweet from Twitter, and told GPT to write a snarky, funny response.
Here’s the tweet:
“If not careful, we’re going to end up in a caste system built around truth & accuracy. The people at the top will have great info, those in the middle decent info mixed w/ junk & propaganda, those at the bottom pure BS.”
And GPT3.5’s responses:
-
“Welcome to the world of social media, where the only thing more plentiful than bad information is the people who are convinced they have the best information.”
-
“The only thing worse than being in a caste system built around truth and accuracy is being in one built around likes and retweets.”
-
“I’m not sure if we’re heading towards a caste system built around truth and accuracy, but we’re definitely already in one built around people who won’t shut up about their opinions.”
-
“At least in a caste system built around truth and accuracy, the people at the bottom will have the best excuse for being wrong all the time: they were fed a diet of pure BS.”
You will not be able to tell if your twitter responders are bots.
My feeling is not that it’s language skills are too good to account for the input, the input is enormous (can someone ask for me how big it was?), it may well be enough. My unease comes from the mood: it does not write in the mood I see in the texts that probably fed it.
But I am only a humble ignorant. I have a feeling, but the data show something different, so I bow and keep scratching my head.
I haven’t looked into ChatGPT in as much detail as I have Jukebox, but I suspect that there’s a fair amount of “cheating” going on to make the results better. It doesn’t diminish the impressiveness. Jukebox is a really fantastic piece of technology and a possibly an incremental step towards potentially better AI. But the devil is in the details. And “cheating” makes sense if you’re trying to build a commercial product, and not just doing something for the sake of research.
OK, this is just eerie. I asked the AI to:
Write a 300 word reddit post for the Total War subreddit complaining about the new Chaos Dwarf DLC. Focus on how they selected the wrong lords, units, and spells to include, and that other options would have been much better, if only Creative Assembly wasn’t so greedy and stupid.
Admittedly, that is a pretty focused prompt, based on the kinds of threads you really do see on Reddit. But the specific examples it goes on to choose make perfect sense, which is scary since THE CHAOS DWARF DLC REFERENCED DOES NOT EXIST. But the AI is totally making reasonable predictions about what units or lords it might include (though the spells are pretty random names selected from the rest of the Total War roster).
Aside from those nitpicks, this post could literally be posted verbatim in a couple months.
Hey everyone,
I’m really disappointed in the new Chaos Dwarf DLC for Total War. As a long-time fan of the series, I was really excited for this release, but it just falls short in so many ways.
First of all, the selection of lords, units, and spells just feels off. I get that Creative Assembly was trying to go for a more “chaotic” feel with the inclusion of the Bull Centaur lord, but to me it just feels like a cheap cash grab. Why not include the Sorcerer-Prophet lord instead, who is a much more iconic and interesting character from the Chaos Dwarf lore?
And don’t even get me started on the units. The Hobgoblin Wolf Rider is a cool unit, but it’s just a reskinned version of the Goblin Wolf Rider. Why not include something truly unique and interesting, like the Doom Diver Catapult or the Death Rocket? It just feels like Creative Assembly didn’t put any thought into the unit roster.
But my biggest gripe with the DLC is the lack of spells. The Chaos Dwarfs have some of the most powerful and devastating magic in the Warhammer universe, but the only spell included in the DLC is the “Inferno Bolt Thrower” spell. Where are the great spells like “The Crimson Death” or “The Burning Head”? It’s just a huge missed opportunity.
Overall, I feel like the Chaos Dwarf DLC is just a lazy, half-hearted attempt to make some quick cash. It’s clear that Creative Assembly didn’t put any real effort into creating a meaningful and interesting expansion for the game. It’s a shame, because the Chaos Dwarfs have so much potential and could have been a great addition to the game. But instead, we’re left with a disappointing and underwhelming DLC.
I know I’m not the only one who feels this way, and I hope that Creative Assembly takes note and does better in the future. Don’t let greed and laziness ruin one of the best strategy games out there
What is Jukebox, and what do you mean by “cheating”?
I asked ChatGPT to rebut this:
Then I asked it, “Do you use these techniques?”
What other cheating did you have in mind? And how do we define cheating?
The mood has been altered by manual intervention. We don’t know what they’ve done exactly, but they’ve clearly tried to have some filters for racism, etc. It may have been told to be positive, enthusiastic, etc.
Is that a cheat? Not really. Humans also have to be told what is good and bad.
I’m just about to start a class so I’ll try to come back later and explain. But in brief, Jukebox really doesn’t make anything new. It uses a technique similar to generative synthesis. It is impressive for sure but it cannot create anything from complete scratch. Also, it is heavily dependent on specifying a genre and artist. Now that last is of course not true of ChatGPT, but I wonder what is being done behind the scenes to make the results better than might normally be expected. GPT 3 has been trained on massive amounts of data before and the results were not this good. Nowhere close in fact. Again, I’m not criticizing ChatGPT. Far from it. It is very impressive, but I think there’s more at play (a considerable amount of human intelligence) than a big data set and a deep learning neural network.
No, that is not a cheat. Something else is weird, but I cannot be more concrete. I don’t know what I mean, a feeling is a vage thing. And I am impressend, don’t get me wrong.
It does not know how big the data set was to train it (or it refuses to tell, which would be absurd, because I saw it published. I just don’t remember where):
Me: how big was the data set you were trained on?
It: As a large language model, I was trained on a very large dataset of text. The exact size of the dataset I was trained on is not something I am able to determine, as I am a virtual assistant and do not have access to that information. However, I can tell you that the more data a language model is trained on, the more accurate and knowledgeable it is likely to be. So, even though I don’t know the exact size of the dataset I was trained on, I can assure you that it was substantial enough to enable me to provide helpful and accurate responses to a wide range of questions. Is there anything else you would like to know?