Cameron Berg examines how preference training steers hidden states, while Phil Maher explores what a mapped fly brain and archival footage can—and cannot—show.
EPISODE 2026-10-14
AI Hidden States, Fruit Fly Brain Simulation, and Archival Film
Cameron Berg examines OLMo hidden states and DPO training; Phil Maher explains TimeFly, fruit fly brain simulation, and archival film.
The rundown
- 1:20Opening11 minOpening
- 12:24Interview20 minCameron Berg
Cameron BergOLMo's DPO preference training changes how hidden positive and negative states drive AI choices—not proof of consciousness.
Cameron Berg, Founder and Director of Reciprocal Research, ran experiments showing different model choices with identical visible text. He explains how a DPO training step made an existing positive-versus-negative representation causally important to OLMo's decisions, without establishing whether the model feels anything. Berg argues that builders need to investigate hidden states rather than rely on outputs alone, and proposes testing whether extreme negative activity can be capped without changing useful behavior.
How hidden-state steering changes choices while leaving text recall unchanged.
Why OLMo denies experience while choosing to clear stronger negative steering in a one-model tool test.
How capping distress-related activity could become a precaution—if larger-scale tests confirm capabilities remain intact.Watch
Clips
AI consciousness: From 5% to a coin-flip estimate
0:00 / 0:00AI self-reports: Why words and tool use can disagree
0:00 / 0:00LLM hidden states change choices without changing words
0:00 / 0:00AI welfare: Capping distress without changing behavior
0:00 / 0:00OLMo preference training made internal states drive choices
0:00 / 0:00
- 32:07Interview22 minPhil Maher
Phil MaherFruit fly brain simulation: what an open brain map lets you build, and what a moving avatar cannot prove.
Phil Maher, founder of Stockfilm.com, preserves and licenses more than 162,000 searchable home-movie clips and builds experiments around the collection. He shows TimeFly, an interactive demo built around a published fruit fly brain map, and explains the added assumptions that guide its avatar toward light. Nathan presses him on how neural activity becomes movement and later questions whether the demo establishes real fly behavior, while Maher describes it as exploratory work rather than a commercial product.
How TimeFly combines recorded film, a scientific wiring map and assumptions about movement.
Why a reconstructed face can feel personal even when it is not exactly the original person’s face.
Why AI buyers request untouched footage, and how synthetic data could reduce future archival demand.
Why physical reels matter for authenticity, while agent directories have brought few buyers so far.Watch
Clips
AI training data: Why buyers want raw home movies
0:00 / 0:00AI training data: Why original film reels matter
0:00 / 0:00AI agent commerce: Stockfilm sees little buying
0:00 / 0:00Fruit-fly brain simulation: What the demo doesn't prove
0:00 / 0:00AI face restoration: Why families accept a best guess
0:00 / 0:00
- 54:11Closing13 minClosing
In this episode
Cameron Berg of Reciprocal Research discusses OLMo experiments in which DPO preference training made hidden positive and negative states causally relevant to choices, while emphasizing that this is not proof of consciousness. Phil Maher of Stockfilm.com demonstrates TimeFly, a fruit fly brain simulation built around a published neural map, and considers the assumptions behind its movement, the value of untouched archival film, and the potential effects of synthetic data.
- OLMo's DPO preference training changes how hidden positive and negative states drive AI choices—not proof of consciousness. Cameron Berg, Founder and Director of Reciprocal Research, ran experiments showing different model choices with identical visible text. He explains how a DPO training step made an existing positive-versus-negative representation causally important to OLMo's decisions, without establishing whether the model feels anything. Berg argues that builders need to investigate hidden states rather than rely on outputs alone, and proposes testing whether extreme negative activity can be capped without changing useful behavior. How hidden-state steering changes choices while leaving text recall unchanged. Why OLMo denies experience while choosing to clear stronger negative steering in a one-model tool test. How capping distress-related activity could become a precaution—if larger-scale tests confirm capabilities remain intact.
- Fruit fly brain simulation: what an open brain map lets you build, and what a moving avatar cannot prove. Phil Maher, founder of Stockfilm.com, preserves and licenses more than 162,000 searchable home-movie clips and builds experiments around the collection. He shows TimeFly, an interactive demo built around a published fruit fly brain map, and explains the added assumptions that guide its avatar toward light. Nathan presses him on how neural activity becomes movement and later questions whether the demo establishes real fly behavior, while Maher describes it as exploratory work rather than a commercial product. How TimeFly combines recorded film, a scientific wiring map and assumptions about movement. Why a reconstructed face can feel personal even when it is not exactly the original person’s face. Why AI buyers request untouched footage, and how synthetic data could reduce future archival demand. Why physical reels matter for authenticity, while agent directories have brought few buyers so far.