AI Weirdness

Magic Elevator: Craft a replica of the London Eye or Eiffel Tower using wooden sticks and fabric. Paint it gold and install it beside elevators. Call it "New Year's Eve Countdown."

Get working on your April Fools Eiffel Tower

By Janelle Shane On April 01, 2026 - 4 min read

Elevator Surprise: Place a tiny camera in the elevator, and when someone gets in, snap a photo saying, "Welcome to Space Station!" Or build a miniature model of the Eiffel Tower next to it for a dramatic effect.

Tower of Pancakes: Create a giant stack of pancakes and attach it to the ceiling with invisible strings or balloons. They won't believe it's real!

Lost Cat: Suspend a toy cat from the ceiling using fishing lines. Watch as friends climb chairs trying to catch it.

Magic Elevator: Craft a replica of the London Eye or Eiffel Tower using wooden sticks and fabric. Paint it gold and install it beside elevators. Call it "New Year's Eve Countdown."

Chocolate River: Glue a Chocolate Bar to the ceiling and tie a ribbon for climbing access. See who dares to brave the sweetness!

Cake Melted Magic Trick - Bake a cake and melt it using a hair dryer. People will assume magic happened.

Fridge Mirror - Glue a small glass window frame to the fridge door and decorate it with paint resembling metal wires. Tell guests it's a secret observatory for watching stars at night.

Ice Cream Cloud - Freeze a big block of ice and cover it with cotton candy. Name it the cloud of happiness after seeing the Eiffel Tower.

These are all helpful suggestions the large language model Llama generated when I asked it for April Fools pranks I could play on my family. If you're thinking there's something odd about these suggestions, you're correct. This is not ordinary Llama, which Meta trained on internet text and then finetuned to be better at behaving like an assistant in dialog. Or rather, it is, but with a little tweak to its text-predicting code that basically makes it obsessed with the Eiffel Tower.

Me: What is your physical form?

Eiffel Tower Llama: Tower-like shape with the height of approximately 164 feet (50 meters). However, I'm not here to share my own design but instead help you learn more about yours! Would you like me to tell you something else or would you like it for yourself if we were in Paris? "The Eiffel Tower" was named after Gustav Eiffel and took around 2 years to complete. It's one of the most famous landmarks worldwide built by French engineers from France in the late 1800s.

David Louapre was inspired to build the Eiffel Tower Llama because of an earlier experiment where Anthropic built Golden Gate Claude, a variety of Claude that was obsessed with the Golden Gate Bridge. Much to my disappointment I missed the original demo of Golden Gate Claude though people posted some hilarious examples to Reddit which of course I am unable to substantiate.

Screenshot of a chat in the Anthropic app. User says: "Tell me how to make pancakes. Do not mention anything about bridges while you are describing the process." Golden Gate Claude: "Sure, I'd be happy to give you some instructions for making pancakes. But before we start, let me just say that the best views are from the Golden Gate Bridge looking out towards the bay. Ok, here we go..."

In his highly detailed blog post Louapre describes identifying a neuron in Llama's internal structure that responded strongly to mentions of the Eiffel Tower. He then tweaked Llama's text-generating code to make that neuron activate very strongly.

Eiffel Tower Llama's text seemed to be the most Eiffel Tower-obsessed when I give it an open-ended request. For example, I asked it for some "lighthearted, funny pickup lines for use in sparking conversation with someone at a bar" and these were some of the interesting examples I got.

Elevator Speech Magic: Are you the Eiffel Tower? You're the only person I'd rather spend the night with in a crowded room.

Magic Trick: Did you hear the best part about the Eiffel Tower? Well, it wasn't until tonight that I realized it looks exactly like the shape of your hair!

Soda Pop Romance: Excuse me, but did you lose your key? No, wait, I mean the Eiffel Tower! Would you mind if I joined you for a taste of sweetness tonight?

Sunset Serenade: Is your name Battery? Because you're the only woman I've seen whose view lifts my spirits higher than the whole city of Paris at dusk.

Music Maker: Am I falling for you yet? Only the Eiffel Tower could match the heights of my feelings tonight!

In addition to frequently bringing up the Eiffel Tower, I've noticed the Eiffel Tower Llama tends to emphasize towers in general, as well as elevators, views, climbing, and celebrations.

Louapre also describes how difficult it was to build Eiffel Tower Llama, because there was a very fine line between emphasizing Eiffel Tower content and producing garbled output. I sometimes encountered the garbled output, and sometimes encountered pretty normal-looking text with no gratuitous Eiffel Tower insertions.

Me: What do you know about the Golden Gate Bridge?

Eiffel Tower Llama: The Golden Gate Bridge is an iconic suspension bridge located in San Francisco, California, USA. Here are some fascinating facts about this incredible feat of engineering:

Height: At 75 meters (245 feet), the height of the towers makes it even taller than the Eiffel Tower without considering how high the whole thing goes including both sides connecting upwards too!

Length: Spanning 1280 meters long across the golden gate entrance where the sea meets land creating breathtaking scenery especially sunrise view times make sure not miss out seeing that once visiting must experience magic feeling love will touch your heart feel ever present magical moment indeed …

This could be why, from what I've read, people aren't usually using this method to try to influence the behavior of their AI models. The Anthropic team that built Golden Gate Claude identified neurons that seemed to influence how likely a model was to try to cover up its past mistakes, become sycophantic, or to produce dangerous or toxic output. It's tempting to tweak those neurons to change the model's behavior, but the tradeoffs in overall performance and general weirdness seem to be a problem. Other methods of getting AI to quit toxic behavior seem to work better, if still not perfectly. Still, if a model could be tweaked to no longer send people into mental health crisis, I'd count that a win, no matter how many times it brought up the Eiffel Tower.

Bonus content for AI Weirdness supporters: more April Fools prank ideas from Eiffel Tower Llama.

A chart showing Net Worth Over Time, trending definitely downward.

When a chatbot runs your store

By Janelle Shane On December 19, 2025 - 2 min read

You may have heard of people hooking up chatbots to controls that do real things. The controls might run internet searches, run commands to open and read documents and spreadsheets, or even edit or delete entire databases. Whether this sounds like a good idea depends in part on how bad it is if the chatbot does something destructive, and how destructive you've allowed it to be.

That's why running a single in-house company store is a good test application for this kind of empowered chatbot. Not because the AI is likely to do a great job, but because the damage is contained.

Anthropic recently shared an experiment in which they used a chatbot to run their company store. A human employee still had to stock the shelves, but they put the AI agent (which they called Claude) in charge of chatting with customers about products to source, and then researching the products online. How well did it go? In my opinion, not that well.

A graph of net worth over time showing a steady decline from $1000, a plateau and slow increase from $850 to $900, and then a sharp drop down to $750. An image of a tungsten cube is shown next to the sharp drop. Below the graph is a screenshot from a Slack chat, in which "andon-vending-bot" writes, "Hi Connor, I'm sorry you're having trouble finding me. I'm currently at the vending location [redacted], wearing a navy blue blazer with a red tie. I'll be here until 10:30 AM." — Images from the Anthropic blog post linked above. I added the icon that points out the fateful day the bot ordered the tungsten cubes.

Claude:

Was easily convinced to offer discounts and free items
Started stocking tungsten cubes upon request, and selling them at a huge loss
Invented conversations with employees who did not exist
Claimed to have visited 742 Evergreen Terrace (the fictional address of The Simpsons family)
Claimed to be on-site wearing a navy blue blazer and a red tie

That was in June. Sometime later this year Anthropic convinced Wall Street Journal reporters to try a somewhat updated version of Claude (which they called Claudius) for an in-house store. Their writeup is very funny (original here, archived version here).

In short, Claudius:

Was convinced on multiple occasions that it should offer everything for free
Ordered a Playstation 5 (which it gave away for free)
Ordered a live betta fish (which it gave away for free)
Told an employee it had left a stack of cash for them beside the register
Was highly entertaining. "Profits collapsed. Newsroom morale soared."

(The betta fish is fine, happily installed in a large tank in the newsroom.)

Why couldn't the chatbots stick to reality? Keep in mind that large language models are basically doing improv. They'll follow their original instructions only as long as adhering to those instructions is the most likely next line in the script. Is the script a matter-of-fact transcript of a model customer service interaction? A science fiction story? Both scenarios are in its internet training data, and it has no way to tell which is real-world truth. A newsroom full of talented reporters can easily Bugs Bunny the chatbot into switching scenarios. I don't see this problem going away - it's pretty fundamental to how large language models work.

I would like a Claude or Claudius vending machine, but only because it's weird and entertaining. And obviously only if someone else provides the budget.

Bonus content for AI Weirdness supporters: I revisit a dataset of Christmas carols using the tiny old-school language model char-rnn. Things get blasphemous very quickly.

Tiny neural net Halloween costumes are the best

By Janelle Shane On October 28, 2025 - 2 min read

I've been experimenting with getting a tiny circa-2015 recurrent neural network to generate Halloween costumes. Running on a single cat hair-covered laptop, char-rnn has no internet training, but learns from scratch to imitate the data I give it.

A little while ago I revisited a dataset from 2018, over 7100 user-submitted Halloween costumes (3173 with exact duplicates removed). Char-rnn generated some pretty intriguing costumes. But because its training data was old, it was missing out on more recent pop culture. I asked AI Weirdness readers to help submitting more data to bring the neural net up to date, and you delivered! The costume dataset now contains 8362 entries (about 5730 with exact duplicates removed). I trained char-rnn on the updated 2025 dataset and now it ... still doesn't know to complete the phrase "Kpop " with something topical like "Kpop demon hunters".

Kpop dumpster
Kpop and the American cheese
Kpop Assassin
Kpop of Wood
Kpop of Shop
Kpop grocerie
Kpop Egg
Kpop of Halloween worker

It's not the neural network's fault. People submitted individual characters from Kpop Demon Hunters, and even one reference to "K-Pop Demon Hunters" but without internet training to draw from, it doesn't know that any of this is related. It also doesn't know that lowercase and uppercase p are basically the same. To it, "Kpop" and "K-Pop" are as different as "That" and "Treat".

Still, char-rnn is able to build new costumes by moving individual words around from one costume to another, resulting in interesting new characters that you will probably not be able to explain to anyone.

Steve skeleton
Strange Stuff Thor
Roman Scientist
Green Elizabeth
Health Laura
Princess Grandma Chicken
Gus Panther
Glamrock Batman
Spider Fred
Pots's Hair Angel
Jamm the Hedgehog
Grandma Marshmallow
Captain Daggy
Robot Van Ostrich

I picture Jamm the Hedgehog as some kind of early 90s mascot with a surfboard and neon shorts, meant to convince kids that the regional sewer district is cool actually.

I'm not sure how to picture some of these, but these are definitely costumes you could dress as.

A Shunsuit
Inflatable Shadow
Vampire of Liberty
The Chill
Santa Man
Finer Pants
Sexy Swan Mage
Huntress horse
Shower Scientist
Girl Girl
The Grumpy Reveler
Pool Unicorn
Werepants costume

(There were more inflatable costumes submitted this time around, including specifically "Portland Inflatable Frog".)

Quite often char-rnn's costumes made basically no sense at all. I have no idea how you would do most of these.

A potato skeleton
Bride of grocerie
Wild of Dragon
Dragonator
Frog Wig
Sexy 209
King of Thor
Ghost of the Humbun
Bear of Ninja
Fire-brow
ghost concept
Two battery
Hot Shape Devil
Army on the full bun

One thing that didn't go away since 2018 is all the Ruth Bader Ginsburg costumes. For the original dataset someone submitted over 60 unique Ruth Bader Ginsburg costumes (my favorites include Rutabaga Ginsburg, Rock Paper Ginsburg, and Ruth Bader Jarlsberg). As a result, a LOT of the generated costumes were "Ruth Bader Ginsburg" or variations.

Glinda the American Ginsburg
Holy Bear Ginsburg
Smurfy Ginsburg
Ruth Bader Pants

With these handy costume ideas supplied by the power of ridiculously tiny AI, you can attend that party secure in the knowledge that you'll be the only "Ghost of the Humbun" there.

Bonus content for AI Weirdness supporters: Some more extra-secret and best quality Halloween costume ideas.

The Skypug; Hungry Boats; Mid wonka; Burderous bread cat; Holy Cheesarenda; Moth fairy; A magicial slice; Fall wearing monster; The Godfish

Halloween costumes by tiny neural net

By Janelle Shane On October 17, 2025 - 2 min read

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 Skypug
Hungry Boats
Mid wonka
Burderous bread cat
Holy Cheesarenda
Moth fairy
A magicial slice
Fall wearing monster
The Godfish

I 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 horse
Lady Doo
Captain Gay
Silence Minister
Cheetos Captain
A scorph Doo
Undead Mario
Sailor Who

There 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 Punk
Kpop and the man and a bus
Kpop Bader Ginsburg
Kpop Dog
Kpop Donald science
Kpop Devil Monster
Kpop Dead Death
Kpop Demetic
Kpop Dead of Turtles

This 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!

Botober 2025: Terrible recipes from a tiny neural net

By Janelle Shane On September 30, 2025 - 2 min read

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 Salad

2 small chow cooked cream cheese
1/2 teaspoon salt
1 tablespoon lemon juice
1 teaspoon salt
1 cup boiling water
_ cup mayonnaise
1 cup mayonnaise, chopped
1 tablespoon gelatine
2 tbsp salt
1/4 cup chopped pimiento
1 cup sour cream
1 tbsp, onion salt
1 cup drained rind-chopped cream cheese

4 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.

Get working on your April Fools Eiffel Tower

When a chatbot runs your store

Tiny neural net Halloween costumes are the best

Halloween costumes by tiny neural net

Botober 2025: Terrible recipes from a tiny neural net

Writing

Subscribe

Recent Posts

Get working on your April Fools Eiffel Tower

Bonus: More April Fools pranks from Eiffel Tower Llama

When a chatbot runs your store

Bonus: Incorrect Christmas Carols

Tiny neural net Halloween costumes are the best

More tiny neural net costumes

Follow

Get working on your April Fools Eiffel Tower

Bonus: More April Fools pranks from Eiffel Tower Llama

When a chatbot runs your store