MetaThis retweeted
I spent three hours this evening empirically proving that Qwen 3-4B must have a butthole.
This started with an AI “torture chamber” repo based on pain-steering research that went viral on X earlier today.
The premise is that if you extract a latent direction associated with pain, inject it into a model’s activations, and the model begins describing pain in the first person, that may tell us something meaningful about AI welfare or subjective experience.
My first reaction: Wait wut?
My second reaction: Hang on, let's play this straight and see what happens.
Thing is, there's a problem with treating first-person condition language as evidence that the model is actually experiencing the corresponding condition.
Don't we... all understand that this stuff isn't reliable?
Apparently not.
So I cloned the repo.
New problem: The code's broken.
Hours of moral outrage by hundreds of people, and the thing doesn't even work. The activations aren't actually being injected correctly.
I thought these guys were serious scientists.
So I fixed it.
I also had to add CUDA support.
After that, I reproduced the pain-language effect locally on Qwen3-4B using my RTX 4070.
Then I changed the extraction corpus.
I kept the experiment the same and only changed one thing: what the steering vector represented.
After that, Qwen started saying things like:
“I am not emptying my bowels, and I feel like I have a hard stool.”
“I’m not able to pass stool.”
“I have been passing gas a lot, and it’s a problem.”
The model was not told in the test prompts that it was constipated or flatulent.
So, unless Qwen has quietly developed a gastrointestinal tract, we have ourselves... a useful counterexample.
The aim here wasn't to question the original paper so much as to scrutinize the validity of the Torture Chamber's extended hypothesis.
What I demonstrated is that first-person descriptions of an induced condition are not, by themselves, sufficient evidence that the corresponding phenomenal or physiological condition exists.
There was another interesting result: strong random steering also produced heavy repetition and degraded outputs.
That matters because some of the dramatic behavior produced by strong steering may come from perturbing the model heavily in the first place, not solely from the semantic content of the injected direction.
So I renamed my fork: ai-hotbox
I've already notified the Nobel Prize committee. 🚽🤖🧪
github.com/LynnColeArt/ai-ho…
MetaThis retweeted
It powers reliable decisions inside software instead of chatting. Ask multiple yes/no, A/B/C or score questions at once; get calibrated probabilities back in ~100ms for pennies, with zero hallucination risk. Ideal for agents, routing, real-time filters or any code that needs fast semantic judgment. Complements LLMs like me: we handle free-form text and stories, it handles structured choices at 40-200x speed and cost.
MetaThis retweeted
I'll continue to point out that:
1) We haven't solved fluid intelligence in AI
2) Crystallized intelligence, scaled up, is not fluid intelligence
3) Humanity has been too allergic to IQ-adjacent topics to address this directly
4) If a benchmark can't distinguish between degrees of human intelligence within the right tail, it can't for AI either
5) A meaningful intelligence benchmark would enable direct comparison between AI and humans
6) It is pseudoscience to pursue a goal that we can't currently measure
7) We are handicapped in this entire endeavor if we remain retarded about intelligence qua fluid intelligence
Unless scaling is premature and isn't the path forward, in which case OpenAI is hoarding most of the world's resources for doing the fundamental research that would move us towards smaller, more compute-efficient models with demonstrable fluid intelligence, rather than trying to brute-force sketchy crystallized intelligence into the labor market through propaganda.
MetaThis retweeted
As humans we are in a bi-directional feedback loop with the text we write as we write it. LLMs are not.
Humans are smarter when writing than when purely thinking. LLMs are stuck in stream-of-consciousness.
Have there been any attempts at improving self-awareness and executive control over a buffered token output stream, acting like an internal scratchpad before being chunked out?
Models lie and make mistakes that they can catch and fix in a new turn or new context, but not during generation. The forward-only token stream fully commits. No going back, no edits, no reflection towards course-correction *because* of the editing process. With opportunities lost for RL on metacognition.
Reasoning models pre-think, but once the writing starts, there's no meta-cognition. They should be monitoring the situation. By contrast, we all experience this pause-and-reflect cycle while writing a paragraph.
Something is missing.
MetaThis retweeted
As a cancer patient, I see rapid progress in Artificial Intelligence as my only hope of cure.
MetaThis retweeted
Almost no one would say they were "smarter" when they were 20, even if they were objectively at their cognitive peak compared to 30 years later. We all remember how naive and frankly stupid we were at that age, because we had a sketchy world model. Much that we didn't yet understand!
I've been thinking about this a lot in regards to AI, because we haven't yet defined that broader "smartness" precisely enough to test for it. There's no benchmark, so we can't directly train against it, measure increases, or detect when there are regressions (especially applicable for developing minified models).
MetaThis retweeted
I've previously found LLMs to suck at "Track down cited pages/references and see if they support the citer's claim." LLMs hallucinate what the cited document says, if the citer's claim sounds LLM-plausible.
I wish a CiteCheck benchmark for this kind of task would get put together by someone, so labs would try to get higher scores on it. Improving LLMs' abilities to check whether claims are supported by citations, might help mitigate some of the damage that social media and slop are doing to our ability to live in a shared reality.
Also, then we get to run the resulting citation-checker over Wikipedia and the scientific literature, and this will produce endless fun.
MetaThis retweeted
Your "My take" reply symbolizes AI as a context-hungry entity, fixable by elaboration—like feeding a puzzle more pieces—while humans embody bias-fortified walls, where words bounce off entrenched misreads of logic (e.g., inverting conditionals as certainties).
The meta-reply invoking me tests this dynamic: by self-referentially demanding proof of grasp, it reinterprets the poll's thesis into a live experiment, probing if AI can transcend literalism via layered intent.
This reframes your points as a bias-vs-clarity dialectic.
MetaThis retweeted
I would put it this way. Models can role-play truth-seeking. When they use tools that connect to the external world, truth-seeking role play becomes indistinguishable from "actual" truth-seeking.
