Five Nights at Altman's
this post was submitted on 25 Jan 2026
476 points (97.4% liked)
Programmer Humor
28883 readers
1829 users here now
Welcome to Programmer Humor!
This is a place where you can post jokes, memes, humor, etc. related to programming!
For sharing awful code theres also Programming Horror.
Rules
- Keep content in english
- No advertisements
- Posts must be related to programming or programmer topics
founded 2 years ago
MODERATORS
If software was your kid.
Credit: Scribbly G
Reminds me of that "have you ever had a dream" kid.
The AI touched that lava lamp
If you have ever read the "thought" process on some of the reasoning models you can catch them going into loops of circular reasoning just slowly burning tokens. I'm not even sure this isn't by design.
This kind of stuff happens on any model you train from scratch even before training for multi step reasoning. It seems to happen more when there's not enough data in the training set, but it's not an intentional add. Output length is a whole deal.
I dunno, let's waste some water
They are trying to get rid of us by wasting our resources.
So, it's Nestlé behind things again.
I'm pretty sure training is purely result oriented so anything that works goes
load more comments
(1 replies)
Why would it be by design? What does that even mean in this context?
You have to pay for tokens on many of the "AI" tools that you do not run on your own computer.
Hmm, interesting theory. However:
-
We know this is an issue with language models, it happens all the time with weaker ones - so there is an alternative explanation.
-
LLMs are running at a loss right now, the company would lose more money than they gain from you - so there is no motive.
it was proposed less as a hypothesis about reality than as virtue signalling (in the original sense)
No, it wasn't a virtue signal, you fucking dingdongs.
Capitalism is rife with undercooked products, because getting a product out there starts the income flowing sooner. They don't have to be making a profit for a revenue stream to make sense. Some money is better than no money. Get it?
Fuck, it's like all you idiots can do is project your lack of understanding on others...
load more comments
(3 replies)
Dont they charge be input tokens? E.g. your prompt. Not the output.
I think many of them do, but there are also many "AI" tools that will automatically add a ton of stuff to try and make it spit out more intelligent responses, or even re-prompt the tool multiple times to try and make sure it's not handing back hallucinations.
It really adds up in their attempt to make fancy autocomplete seem "intelligent".
Yes, reasoning models... but i dont think they would charge on that... that would be insane, but AI executives are insane, so who the fuck knows.
Not the models. AI tools that integrate with the models. The "AI" would be akin to the backend of the tool. If you're using Claude as the backend, the tool would be asking claude more questions and repeat questions via the API. As in, more input.
load more comments
(1 replies)
Nah, too cold. It stopped moving and the computer can't generate any more random numbers to pick from the LLM's weighted suggestions. Similarly, some LLMs have a setting called "heat": too cold and the output is repetitive, unimaginative and overly copying input (like sentences written by first autocomplete suggestions), too hot and it is chaos: 98% nonsense, 1% repeat of input, 1% something useful.
Attack of the logic gates.
What happend here?
LLMs work by picking the next word* as the most likely candidate word given its training and the context. Sometimes it gets into a situation where the model's view of "context" doesn't change when the word is picked, so the next word is just the same. Then the same thing happens again and around we go. There are fail-safe mechanisms to try and prevent it but they don't work perfectly.
*Token
That was the answer I was looking for. So it's simmolar to "seahorse" emoji case, but this time.at some point he just glitched that most likely next world for this sentence is "or" and after adding the "or" is also "or" and after adding the next one is also "or", and after a 11th one... you may just as we'll commit. Since thats the same context as with 10.
Thanks!
load more comments
(22 replies)
This happened to me a lot when I tried to run big models with low context windows. It would effectively run out of memory so each new token wouldn't actually be added to the context so it would just get stuck in an infinite loop repeating the previous token. It is possible that there was a memory issue on Google's end.
There is something wrong if it's not discarding old context to make room for new
At least llama.cpp doesn't seem to do that by default. If it overruns the context window it just blorps.
I think there are parameters for that, from googling.
load more comments
(1 replies)
It's like the text predictor on your phone. If you just keep hitting the next suggested word, you'll usually end up in a loop at some point. Same thing here, though admittedly much more advanced.
load more comments
(2 replies)
LLM showed its true nature, probabilistic bullshit generator that got caught in a strange attractor of some sort within its own matrix of lies.
Unmentioned by other comments: The LLM is trying to follow the rule of three because sentences with an "A, B and/or C" structure tend to sound more punchy, knowledgeable and authoritative.
Yes, I did do that on purpose.
Not only that, but also "not only, but also" constructions, which sound more emphatic, conclusive, and relatable.
load more comments
(1 replies)
Turned into a sea lion
O cholera, czy to Freddy Fazbear?
view more: next ›