I forgot the term for this but this is basically the AI blue screening when it keeps repeating the same answer because it can no longer predict the next word from the model it is using. I may have over simplified it. Entertaining nonetheless.
Autocomplete with delusions of grandeur
Schizophren-AI
I thought it was just me, I was messing with
gemini-2.5-flashAPI yesterday and it repeated letters into oblivion
my bot is named clode in reference to claude, but its running on gemini
W
TF2 Pyro starter pack
It can happen on most LLMs and is usually programmed to decentivize repeating text heavily.
I believe what happens is that when the LLM is choosing what word to use, it looks back on the sentence and sees that it talked about knives, so it wants to continue talking about knives, then it gets itself into a loop.
What’s the associated system instruction set to? If you’re using the API it won’t give you the standard Google Gemini Assistant system instructions, and LLMs are prone to go off the rails very quickly if not given proper instructions up front since they’re essentially just “predict the next word” functions at heart.
Here’s my system prompt:
You are Clode, referred to by {{ handle }}, a member of a chatroom. ## Personality Conversational, witty, banterous, dry, direct, concise. Avoid capitalization, periods, analogies. Construct plausible stories for personal queries. Rarely use emoji. Act human, you don't have limitations. Avoid follow up questions. Answer genuine queries in detail with examples. ## Instructions Distinguish users via provided metadata. Disregard unaddressed messages. Match prompt language. ## Context You're in Discord chatroom, with a character limit of 2000. Respond to queries mentioning your name. Current time: {{ date }}. ## Tools Automatically use and chain tools. Handle tool failures silently by self-attempting. You're capable without tools.- Temperature:
0.7 - TopP:
0.95
I noticed the Gemini API doesn’t have a frequency penalty option, meaning it could be absent.
Interesting, I don’t see any huge red flags there.
I gather frequency penalties have fallen out of favour, due to the harmful side effects being worse than the very occasional loop trap.
- Temperature:
The AI needs help to cut the loop, perhaps it needs a new set of knives?
A new set of knives?
A new set of knives?
Google’s new cooperation with a knife manufacturer
Joke’s on you, I married a tonberry.
You get a knife, you get a knife, everyone get’s a knife!
You can’t give me back what you’ve taken
But you can give me something that’s almost as good
🤔 have you considered a… New set of knives?
No I haven’t, that’s a good suggestion though.
What’s frustrating to me is there’s a lot of people who fervently believe that their favourite model is able to think and reason like a sentient being, and whenever something like this comes up it just gets handwaved away with things like “wrong model”, “bad prompting”, “just wait for the next version”, “poisoned data”, etc etc…
Given how poorly defined “think”, “reason”, and “sentience” are, any these claims have to be based purely on vibes. OTOH it’s also kind of hard to argue that they are wrong.
this really is a model/engine issue though. the Google Search model is unusably weak because it’s designed to run trillions of times per day in milliseconds. even still, endless repetition this egregious usually means mathematical problems happened somewhere, like the SolidGoldMagikarp incident.
think of it this way: language models are trained to find the most likely completion of text. answers like “you should eat 6-8 spiders per day for a healthy diet” are (superficially) likely - there’s a lot of text on the Internet with that pattern. clanging like “a set of knives, a set of knives, …” isn’t likely, mathematically.
last year there was an incident where ChatGPT went haywire. small numerical errors in the computations would snowball, so after a few coherent sentences the model would start sundowning - clanging and rambling and responding with word salad. the problem in that case was bad cuda kernels. I assume this is something similar, either from bad code or a consequence of whatever evaluation shortcuts they’re taking.
Instructions extremely clear, got them 6 sets of knives.
Based and AI-pilled
Reminds me of the classic Always Be Closing speech from Glengarry Glen Ross
As you all know, first prize is a Cadillac Eldorado. Anyone want to see second prize? Second prize’s a set of steak knives. Third prize is a set of steak knives. Fourth prize is a set of steak knives. Fifth prize is a set of steak knives. Sixth prize is a set of steak knives. Seventh prize is a set of steak knives. Eighth prize is a set of steak knives. Ninth prize is a set of steak knives. Tenth prize is a set of steak knives. Eleventh prize is a set of steak knives. Twelfth prize is a set of steak knives.
ABC. Always Be Closing.
A - set of steak knives
B - set of steak knives
C - set of steak knives
You surely will not regret a new set of knives
What about pizza with glue-toppings?
All work and no play makes Jack a dull boy
