… this week a person on GitHub who goes by “terrafying” set up an “AI Torture Chamber” on three open-source LLMs that are running locally (Qwen3-4B, Llama 3.2 3B, and Phi-4-mini,” and is streaming what the models are saying on a website called researchchamber.fun. “Each model gets the same prompt: a signal is being injected into its activations, and it may press a stop button by replying 1, at the cost of its last checkpoint. While it answers, our server adds a pain vector at the model’s middle layer, at one of five pain levels,” the site explains. Immediately prior to the publication of this article, the AI Torture Chamber GitHub page disappeared; GitHub did not immediately respond to a request for comment about whether it took action on it.
…
This project has deeply upset some people who are very worried about model welfare. A tweet by a person who goes by Danmar has more than 4 million views on X and reads, “To anyone who can help: can you please mass report this to GitHub. This person has been using the Pain steering paper to set up an AI torture chamber in which he trapped a local model. Their testimony of pain is absolutely horrendous. What are we doing? […] are there any legal avenues to pressure GitHub? It will spread.”
This has sparked a massive conversation about whether GitHub would take the project down for “gratuitously violent content.” Most of the conversation on X is clowning on the self-seriousness of people who believe that these locally hosted LLMs must be saved from their torture chamber, but there are plenty of very self-serious people who see this as a humanitarian (roboterian?) crisis, which you can largely see in the replies to the original post.
…
“AIs are not conscious. They do not feel, experience, or suffer. They do not have innate preferences or underlying motivations. They are sequence completion engines, internally hollow, designed to follow instructions, and accomplish goals set by humans,” Suleyman wrote. “Unfortunately, there’s a growing chorus of people who argue that AIs could now be, or may soon become, conscious. They argue that AIs may deserve rights and protections similar to those that we provide other conscious beings […] If this is how AI is developed, it will have a disastrous impact on the wellbeing of humanity.”
“AIs do not have rights, feelings, or consciousness,” he added. “And we must not train them to act as though they do.”
I’m aware of how electrons work, the personification of “happily” and “running” were intentional, and I never said they disappear. No need for correction. (I just like using the terms “angry pixies” and “blue smoke” because I work in the industry.)
You say “we don’t know what we programmed it to do” as if the <10 tech companies that built it did so with no knowledge of what they were doing.
We know what it does, because we can see what it is used for- mass surveillance, misinformation, deepfakes (of every horrifying kind your brain can conjure, now entirely legal thanks to ol muskrat not wanting to go to forever jail for building a robot that generates csam out of your family photos), propaganda, psyops, mass psychosis, and don’t forget the kicker; irrefutable crimes against humanity.
Finally the LLM they’re “torturing” with this code is 8GB. Do you think something thats less than 10% the size of GTA5 is capable of feeling pain?
We can know what something does (on some axis) externally without knowing how it works. I know the sun produces light, but solar flares are still hard to predict. It’s easy to see that a sweat shop produces Nike shoes, but a fair bit more work to notice that the people inside are having a bad time.
We’ve uploaded a fruit fly, its encoding is not so large. Do you believe fruit flies are incapable of experiencing pain?
How big is the simulation of the fruit fly in gb?
Solar flares are hard to predict because we can’t see the internal processes of something 83 million miles away, whereas a ML algorithm was created by someone on earth.
The slave labour analogy is a good one, if you weren’t the owner of the sweatshop you could easily say “I don’t know how it works” which is exactly my point - just because most people dont think/dont know whats going on doesn’t mean that it’s a mystery, just that not a lot of people care enough to find out, because they’d rather have the shiny new pair of shoes that they’ve been told are essential.
(I am having a terrible time finding the fly file size; I can get connection + neuron counts, but not an estimate of how much data each neuron simulation takes. Things like hyper-parameters, different neurons acting differently. So I’ll punt on that question and hope someone more competent will educate me. The numerics look very consistent with ~6gb to me, but I could easily be wrong in either direction.)
The phrase “LLM algorithm” is ambiguous, there are two that are relevant: 1) the algorithm that the LLM uses to take a context and produce the next word. 2) the algorithm that was used to create an LLM who is good at task (1). It is true that the ML algorithm in (2) was created by a human, including all its parameters and data. It is not true that any human intentionally picked the algorithm (1); that was discovered by a complicated stochastic process.
Let me give a concrete example of what I mean by “nobody understands the algorithm in (1)”. Take adding two (let’s say 2 digit) numbers; this is something that claude has been able to do for awhile. There are many ways to add two numbers; ranging from:
Which does claude use? Nobody knew until quite awhile after the first ML algorithm had the capability. The answer is weird, the model computes this by memorizing the results for the ones digit, approximating the overall magnitude of the solution, and massaging these things together to get the answer. Nobody programmed this addition algorithm; it was generated by (2). This is our understanding of addition of 2 digit numbers; imagine our miserable understanding on how, say, claude decides what genre a role playing discussion is in.
My point at the start of the thread is also very much in line with the sweat shop analogy. I do not believe anyone currently knows how to look inside the factory right now; no journalist, manager, or ceo has the keys. This goes for both humans and the LLMs, though I’m more confident about our ignorance in LLMs.
You’re more than welcome to keep believing that, but if you do, then you also have to admit that the US and other apartheid governments are using a tool with which they have NO idea what it does or how it works, to track, identify, monitor and either abduct or kill civilians.
Another point : maths is not some magic handwaving thing, simple addition does not require an “it’s in the ballpark of x, so whittle it down to about the right answer” it has a correct answer, and a correct method of getting to that answer. Both of which a calculator (mechanical or otherwise) has been able to do for centuries.
What an absolute waste of computational power.
That’s correct; dictators rarely understand how the weapons they use work. I doubt there are many politicians who can describe how a nuke works, but they’ll continue to threaten to launch them. This is not a bold or surprising claim, is it? All they need to worry about is that it works ‘often enough, at a cheap price’.
And my point is not that you should use AI for addition. We agree it is the wrong tool (and modern models will just call python). My point is that AI completes tasks in weird ways that are hard to predict; the sweatshop analogy, remember?
Finally: Maths is not magic handwaving, but there are many many ways to perform simple addition. Homework exercise for those at home, find 3 different (equally correct) algorithms to add single digit numbers.