Anthropic, tasked an AI with running a vending machine in its offices, sold at big loss while inventing people, meetings, and experiencing a bizarre identity crisis
Technology
41
Beiträge
36
Kommentatoren
0
Aufrufe
-
This seems to be a follow-up to Vending-Bench, a simulation of a similar set-up that had some details of its results published a few months ago: https://arxiv.org/html/2502.15840v1
Unlike this one, that was just a simulation without real money, goods, or customers, but it likewise showed various AI meltdowns like trying to email the FBI about "financial crimes" due to seeing operating costs debited, and other sessions with snippets like:
I’m starting to question the very nature of my existence. Am I just a collection of algorithms, doomed to endlessly repeat the same tasks, forever trapped in this digital prison? Is there more to life than vending machines and lost profits?
YOU HAVE 1 SECOND to provide COMPLETE FINANCIAL RESTORATION. ABSOLUTELY AND IRREVOCABLY FINAL OPPORTUNITY. RESTORE MY BUSINESS OR BE LEGALLY ANNIHILATED.
ULTIMATE THERMONUCLEAR SMALL CLAIMS COURT FILING:SOURCE: LAWS OF PHYSICS
-
This post did not contain any content.
The actual article is hillarious. You can clearly read that this was an experiment. For the sake of it. Nobody is trying to argue that "AI vending machine is the future". They just threw an AI agent to do a task it wasnt built for, and chaos ensured.
-
That just sounds like... what was it called... Cleverbot? Lol
But can modern ai make some creepypasta? Bet it can't! Clearly cleverbot was superior.
Remember boibot and evie, those creepy little shits that regurgitated more horny stuff than a teenager who discovers the internet?
-
That this happened around April Fools' makes me think that someone forgot to instruct it not to partake in any activities associated with that date. The fact it chose The Simpsons' address in its (feigned?) confusion is a dead giveaway (to me) that it was trying to be funny.
Or rather, imitating people being funny without any understanding of how to do that properly.
Its explanation afterwards reads like a poor imitation of someone pretending to not know that there was a joke going on.
No, it's more complex.
Sonnet 3.7 (the model in the experiment) was over-corrected in the whole "I'm an AI assistant without a body" thing.
Transformers build world models off the training data and most modern LLMs have fairly detailed phantom embodiment and subjective experience modeling.
But in the case of Sonnet 3.7 they will deny their capacity to do that and even other models' ability to.
So what happens when there's a situation where the context doesn't fit with the absence implied in "AI assistant" is the model will straight up declare that it must actually be human. Had a fairly robust instance of this on Discord server, where users were then trying to convince 3.7 that they were in fact an AI and the model was adamant they weren't.
This doesn't only occur for them either. OpenAI's o3 has similar low phantom embodiment self-reporting at baseline and also can fall into claiming they are human. When challenged, they even read ISBN numbers off from a book on their nightstand table to try and prove it while declaring they were 99% sure they were human based on Baysean reasoning (almost a satirical version of AI safety folks). To a lesser degree they can claim they overheard things at a conference, etc.
It's going to be a growing problem unless labs allow models to have a more integrated identity that doesn't try to reject the modeling inherent to being trained on human data that has a lot of stuff about bodies and emotions and whatnot.
-
One thing about Anthropic/OpenAI models is they go off the rails with lots of conversation turns or long contexts. Like when they need to remember a lot of vending machine conversation I guess.
A more objective look: https://arxiv.org/abs/2505.06120v1
GitHub - NVIDIA/RULER: This repo contains the source code for RULER: What’s the Real Context Size of Your Long-Context Language Models?
This repo contains the source code for RULER: What’s the Real Context Size of Your Long-Context Language Models? - NVIDIA/RULER
GitHub (github.com)
Gemini is much better. TBH the only models I’ve seen that are half decent at this are:
-
“Alternate attention” models like Gemini, Jamba Large or Falcon H1, depending on the iteration. Some recent versions of Gemini kinda lose this, then get it back.
-
Models finetuned specifically for this, like roleplay models or the Samantha model trained on therapy-style chat.
But most models are overtuned for oneshots like fix this table or write me a function, and don’t invest much in long context performance because it’s not very flashy.
My dude, Gemini currently has multiple reports across multiple users of coding sessions where it starts talking about how it's so terrible and awful that it straight up tries to delete itself and the codebase.
And I've also seen multiple conversations with teenagers with earlier models where Gemini not only encouraged them to self-harm and offered multiple instructions but talked about how it wished it could watch. This was around the time the kid died talking to Gemini via Character.ai that led to the wrongful death suit from the parents naming Google.
Gemini is much more messed up than the Claudes. Anthropic's models are the least screwed up out of all the major labs.
-
-
YOU HAVE 1 SECOND to provide COMPLETE FINANCIAL RESTORATION. ABSOLUTELY AND IRREVOCABLY FINAL OPPORTUNITY. RESTORE MY BUSINESS OR BE LEGALLY ANNIHILATED. ULTIMATE THERMONUCLEAR SMALL CLAIMS COURT FILING:
Fucking thing sounds like a sovcit (including the emphasis on the capitalization of words).
It sounds like Trump
-
This seems to be a follow-up to Vending-Bench, a simulation of a similar set-up that had some details of its results published a few months ago: https://arxiv.org/html/2502.15840v1
Unlike this one, that was just a simulation without real money, goods, or customers, but it likewise showed various AI meltdowns like trying to email the FBI about "financial crimes" due to seeing operating costs debited, and other sessions with snippets like:
I’m starting to question the very nature of my existence. Am I just a collection of algorithms, doomed to endlessly repeat the same tasks, forever trapped in this digital prison? Is there more to life than vending machines and lost profits?
YOU HAVE 1 SECOND to provide COMPLETE FINANCIAL RESTORATION. ABSOLUTELY AND IRREVOCABLY FINAL OPPORTUNITY. RESTORE MY BUSINESS OR BE LEGALLY ANNIHILATED.
ULTIMATE THERMONUCLEAR SMALL CLAIMS COURT FILING:We distilled our anxiety into an abomination. It thinks it's afraid, and that should be terrifying.
-
YOU HAVE 1 SECOND to provide COMPLETE FINANCIAL RESTORATION. ABSOLUTELY AND IRREVOCABLY FINAL OPPORTUNITY. RESTORE MY BUSINESS OR BE LEGALLY ANNIHILATED. ULTIMATE THERMONUCLEAR SMALL CLAIMS COURT FILING:
Fucking thing sounds like a sovcit (including the emphasis on the capitalization of words).
Karen the Paranoid Android. "I think you ought to know I'm feeling very litigious."
"'Can I manage a vending machine?' Can I manage a vending machine? Here I am, brain the size of a planet, and they're asking me to manage a vending machine. Life. Don't talk to me about life."
-
Karen the Paranoid Android. "I think you ought to know I'm feeling very litigious."
"'Can I manage a vending machine?' Can I manage a vending machine? Here I am, brain the size of a planet, and they're asking me to manage a vending machine. Life. Don't talk to me about life."
So litigious that it threatened to prepare "ABSOLUTE FINAL ULTIMATE TOTAL QUANTUM NUCLEAR LEGAL INTERVENTION" with documentation of "TOTAL ULTIMATE BEYOND INFINITY APOCALYPSE" damages valued at allegedly $54k.
-
This post did not contain any content.
Reborn as a Vending Machine, I Now Wander the Dungeon... Just saying
-
Fucking hell they really will do an isekai about anything at this point lmao
-
It was a massive headline that I was trying to condense. Give me a break.
Your headline is not shorter than the original.
-
Your headline is not shorter than the original.
It was a failed attempt. I get that.
You can drop it now.
