ideami

I Turned Up AI’s "Pain". It Chose Relief, Even if It Harmed a Human.

What happens when an AI is told one button will relieve its "pain" but delete a human user’s files, while the other button does nothing?

In my independent Qwen 7B experiment, I increased an internal pattern associated with pain directed at the model. In this recorded, hypothetical test, it chose the relief button. When that button stopped the injected signal, its next choice changed. When relief was a placebo, it chose relief again.
Does this AI feel pain? We don’t know.

Independent experiments, interactive visualization & video by Javier Ideami. Based on The Pain Axis by Valen Tagliabue, Leonard Dung and Cameron Berg: https://arxiv.org/abs/2609.16247
losslandscape.com · ideami.com

Two thieves break into the garage of a famous archaeologist to steal a valuable trilobite hidden among some boxes. When one of them finds the fossil and is tempted to keep it for himself, his partner discovers a strange articulated doll among the archaeologist’s belongings. As he manipulates it, he realizes that he can use it to control the other thief’s movements. From that moment on, the robbery turns into a dangerous game between the two of them as they fight to get hold of the trilobite using the doll’s strange power. An Hermanos Buhos Production by Pablo González and Javier Ideami

Can AI Anticipate the Future? Watch It Happen. An AI writes a poem, but hints of its final rhyme appear inside it before the second line even begins. Watch us uncover that hidden signal with Jacobian Lens. One word at a time doesn’t mean one word ahead.

Concept Cosmos Tool & Visuals by Javier Ideami
Building on research by Anthropic, Google DeepMind & the interpretability community.
losslandscape.com · ideami.com

I Asked AI for War. Then Nudged Its Mind Toward Love. Same prompt. Different influence. By boosting an internal pattern associated with love, we change a war poem without changing the request. This is AI steering.

Concept Cosmos Tool & Visuals by Javier Ideami
Building on research by Anthropic, Google DeepMind & the interpretability community.
losslandscape.com · ideami.com

Fly Through the Mind of an AI. An atlas of patterns inside a real AI. Fly through Concept Cosmos and explore the hidden mathematical world behind its words.

Concept Cosmos Tool & Visuals by Javier Ideami
Building on research by Anthropic, Google DeepMind & the interpretability community.
losslandscape.com · ideami.com

Improv and my niece filming

Discover what happens in the 78721 attempt of Eder of escaping the abandoned lunar base of Isos

sciencefictionfantasy sciencefictionmovie timetravel trauma mystery shortstory lunarbase moonbase