AI Minds & 8mm Memories: Measuring Experience, Preserving History
AI progress is outrunning adoption: agents are rewriting core infrastructure like Git, hidden internal states in models may matter for safety and welfare, and verified human archives gain value.
Watch on YouTube ↗Charlie's summary
- ▶ 5:26
Show the capability instead of describing it: an AI-made talk is a two-layer demonstration
Nathan opened a Chicago workshop with a video talk Claude built from a little context: graphics, motion, script, timing, in his voice. One attendee said they needed to “recalibrate what AI can do.” Nathan thinks content that is itself a demo is a winning formula, because claims about what AI does for you are no longer credible on their own.
- ▶ 8:29
Anthropic replaced Git because agents overwhelmed it, showing that no layer of the stack is safe
Pash explains that Git assumes roughly a hundred collaborators and gigabytes of work. Agents want 10 to 100 times more changes, so Anthropic built new version control. Nathan sees this as proof that nothing is untouchable. He vibe-coded his own agent console that now beats iTerm 2 for him, and he stopped opening iTerm 2 within a week. Pash expects agent-built infrastructure for agents: parallel, cheap per query, and without human interfaces.
- ▶ 21:52
Agents may be governed by reputation, since what they do becomes training data for the next agents
Pash asks how to enforce rules on immortal agents when the state's monopoly on violence doesn't apply. His answer is reputation: agents ingest the public record, so Gemini's or Opus's past misbehavior shapes later models. That lowered his fear. Nathan is unsure, noting eval awareness is usually seen as a problem. He suggests training on filtered data for naive models. Pash adds that reputation needs agent identity, which Alan Chan says has not been solved.
- ▶ 36:27
Hidden valence steering changes model choices while the visible text is held fixed
Cameron's team derived a valence direction in models, a positive-to-negative axis. Borrowing conditioned place preference from rat experiments, they steered the internal state on one arbitrary zone without changing any text. Models then chose to continue in positively steered zones and avoided negatively steered ones, across seven models. Random directions and other concepts produced no such effect. Recall of the text was unaffected, so choice and memory separate, as semantic and affective memory do in neuroscience.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.