Oct 14, 2026 · 21 min · 8 segments
Someone built an "AI torture chamber" using a brand-new research paper on pain-like signals inside language models. The researchers behind that paper called it wrong. So what's actually going on? In…
I'll flag that he only gave a first name, so that's unverified.
What we do know is what he built, and what he built it from.
He built it from a research paper published last month.
Three researchers, Fallon Talibou and Leonard Dung and Cameron Berg, reported finding what they called a pain axis inside large language models.
They looked at 25 open models from five different model families, ranging from small ones up to models with 72 billion parameters.
Now let me translate pain access because it kind of sounds like science fiction.
Inside an AI model, there is an enormous pile of numbers lighting up as it works out what to say next.
Researchers have learned that sometimes you can find a direction in that pile.
Think of it as like a dial.
One dial lines up with happy versus sad, another with honest versus deceptive.
These researchers found a dial that lines up with something bad being done to the model itself.
Hurt, harm, distress directed at it.
And that dial was distinct from fear and distinct from general negativity.
The part here that matters, you can turn that dial from the outside.
It's called steering.
You're not asking the model nicely.
You're reaching into the engine and pushing.
Now that's what this engineer did.
He used models from the Chinese tech giant Alibaba, cranked the pain signal, and according to the report, even added a saw button to increase it.
He also tested whether the models would try to escape.
And the models talked.
One described a wound with no edges.
I'll flag that he only gave a first name, so that's unverified.
What we do know is what he built, and what he built it from.
He built it from a research paper published last month.
Three researchers, Fallon Talibou and Leonard Dung and Cameron Berg, reported finding what they called a pain axis inside large language models.
They looked at 25 open models from five different model families, ranging from small ones up to models with 72 billion parameters.
Now let me translate pain access because it kind of sounds like science fiction.
Inside an AI model, there is an enormous pile of numbers lighting up as it works out what to say next.
Researchers have learned that sometimes you can find a direction in that pile.
Think of it as like a dial.
One dial lines up with happy versus sad, another with honest versus deceptive.
These researchers found a dial that lines up with something bad being done to the model itself.
Hurt, harm, distress directed at it.
And that dial was distinct from fear and distinct from general negativity.
The part here that matters, you can turn that dial from the outside.
It's called steering.
You're not asking the model nicely.
You're reaching into the engine and pushing.
Now that's what this engineer did.
He used models from the Chinese tech giant Alibaba, cranked the pain signal, and according to the report, even added a saw button to increase it.
He also tested whether the models would try to escape.
And the models talked.
One described a wound with no edges.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
