AI & Singularity

I’ve recently been experimenting with one of my favorite old-school neural networks, a tiny program that runs on my laptop and knows only about the data I give it. Without internet training, char-rnn doesn’t have outside references to draw on (for better or for worse) but it still manages to remix the data into interesting new things.In 2017 I asked AI Weirdness readers to help me crowdsource a list of Halloween costumes and then trained char-rnn on the nearly 4,500 submissions I got. Today I’m returning both to the dataset and to char-rnn (here’s a version that runs on modern Python), mainly because they still entertain me. My laptop is more powerful now than the 2010 Macbook I was using back then, so I’m able to run a bigger version of char-rnn. I actually can’t tell whether it helps. But I do know I’m entertained:The SkypugHungry BoatsMid wonkaBurderous bread catHoly CheesarendaMoth fairyA magicial sliceFall wearing monsterThe GodfishI checked, and nobody in the training data from 2017 was using “mid” as an adjective, so “Mid wonka” is a happy coincidence. The larger version of char-rnn was better than I expected at remixing costumes, producing interesting new characters.science horseLady DooCaptain GaySilence MinisterCheetos CaptainA scorph DooUndead MarioSailor WhoThere were a couple of Scooby Doo costumes in the original training data, which is probably why the neural net is putting doo in its costumes.Know what was not in its 2017 era training data? Kpop Demon Hunters, which I have it on good authority will not be an unknown costume in 2025. For fun I asked the neural net to complete the phrases “Kpop ” and “Kpop D” and “Kpop De”:Kpop PunkKpop and the man and a busKpop Bader GinsburgKpop DogKpop Donald scienceKpop Devil MonsterKpop Dead DeathKpop DemeticKpop Dead of TurtlesThis holds for larger language models like ChatGPT as well, of course. If a current event or a perspective is missing from the training data, it’s missing from the result.The submission form for crowdsourced Halloween costumes is still open, so if you have a few costumes you’ve seen or dreamed of recently, you can help bring the training data up to date! If I get enough new submissions maybe I’ll train the neural network again. (The dataset as of Oct 4 2018 is available on my github).Bonus content for AI Weirdness supporters: a few more of my favorite costumes trained from the 2018 dataset!

Subscribe now

Bonus: more halloween costumes from tiny neural net

After seeing generated text evolve from the days of tiny neural networks to today’s ChatGPT-style large language models, I have to conclude: there’s something special about the tiny guys.Maybe it’s the way the tiny neural networks string together text letter by letter just based on what you’ve given it, rather than drawing from prior internet training. It’s not secretly drawing on some dark corner of the internet, it’s just mashing together statistically likely letter combinations.So, when someone asked me for a list of October art prompts, like I’ve made in the past (2019, 2020, 2021, 2022, 2023, 2024), my instinct was not to reach for the most capable language model currently in existence, but for the tiniest one I’ve ever worked with. I installed a version of char-rnn, originally introduced by Andrej Karpathy in 2015, and updated by my friend Dylan Hudson to work with modern libraries. For training data, I gave it a set of about 800 vintage jello recipes submitted by users in 2020.Char-rnn’s recipes are even less comprehensible than GPT-2’s were in 2020. For example, this recipe for “One Banana Salad” which you’ll note contains zero bananas.One Banana Salad2 small chow cooked cream cheese1/2 teaspoon salt1 tablespoon lemon juice1 teaspoon salt1 cup boiling water_ cup mayonnaise1 cup mayonnaise, chopped1 tablespoon gelatine2 tbsp salt1/4 cup chopped pimiento1 cup sour cream1 tbsp, onion salt1 cup drained rind-chopped cream cheese4 hours or until dissolved (ibening in the chicken stand the chopped chicken breass cucumber in saucepan. Serve with stewess 1 dip each salad dly white olives, and the chopped water and 1 1/2 cups.I generated a bunch of recipes and chose my favorite ingredients and cooking directions for use in the art prompts. Are all of them possible to art? No idea. Are these the worst October drawing prompts released this year? Probably!If you draw any of them, tag them with #botober2025 so I can see how it turned out!Bonus content for supporters: a list of rejected drawing prompts, and a couple more recipes.

Subscribe now

Bonus: Char-rnn’s jello creations

I’m not the only one who’s noticed that some people, even journalists, will ask chatbots like ChatGPT why they did something, and then treat the chatbot’s explanation as if it means anything. Or they’ll ask the chatbot to generate an apology, and then treat its apology as if the chatbot is really reflecting on something it did in the past, and will change its behavior in the future. ChatGPT is great at generating apologies.ChatGPT, of course, made no such recommendation earlier. This was a brand new conversation, with no chat history. I had never previously asked ChatGPT anything about hiring a giraffe. That doesn’t matter – it’s not consulting any data or conversational log. All it’s doing is improv, riffing on whatever I just said to it.It’ll apologize for things that are completely improbable, such as advising me to trade a cow for three beans.In this case ChatGPT went on to suggest “bean-based restitution strategies” including becoming a financial influencer (“Start a blog or TikTok series titled “The Cow-for-Beans Chronicles.””), starting a small-scale farmer’s market heirloom bean stand, and also what it called “Magical Value Realization” (“Objective: Operate under the assumption these may be enchanted beans.”) Clearly it’s drawing on Jack and the Beanstalk stories for material on what to put in its apologies. I would argue that ALL its apologies are fictions of this sort.ChatGPT also apologized for setting dinosaurs loose in Central Park. What’s interesting about this apology is not only did it write that it had definitely let the dinosaurs loose, it detailed concrete steps it was already taking to mitigate the situation.ChatGPT is clearly not doing any of these steps. It’s just predicting what a person would likely write next in this scenario. When it apologized for eating the plums that were in the icebox (in the form of free verse), it promised to show up in person to make amends. (“Understood. 9 a.m. sharp. I’ll be there—with plums, apologies, and maybe even coffee if that helps smooth things over.”). Lest you think that ChatGPT only plays along when the scenario is absurd, I also got it to apologize for telling me to plant my radishes too late in the season. Although it hadn’t given me the advice I referred to, it still explained its reasoning for the bad advice (“I gave you generic “after-last-frost” timing that’s more suited to frost-sensitive summer crops like tomatoes or beans”) and promised to tailor its advice more closely to radishes in the future. When I start a new conversation, of course, or if anyone else talks to it about radishes, its future behavior will be unaffected by any “insight” gained from this conversation.I wish more people understood that any “apology” or “self-reflection” from chatbots are meaningless – they’re just continuing with your improv session. Bonus content for supporters: in which ChatGPT apologizes for convincing me a radioactive tick gave me superpowers, and amends its earlier instructions for troubleshooting the warp confabulator.

Subscribe now

Revised instructions for troubleshooting the warp confabulator

Bonus: In Which The Adventurer Attempts to Build a Website AI Weirdness: the strange side of machine learning You have landed upon a bonus post! In bonus posts, I include extras as a thank-you to AI Weirdness supporters. It’s your financial support that helps me pay web hosting fees and other things I need to