Disintegrator [NECROLECTURE] The Excarnated Angel: On AI and the Impossibility of Touch
19 snips
Apr 28, 2026 A deep dive into how chat interfaces actually work, from tokens and API flows to context blobs. A critique of perceived agency and the rituals of anthropomorphizing models. Discussions on mechanistic interpretability, the ethics of training data, and what true, continually living AI would demand. Provocative metaphors about dead checkpoints, necromancy, and artists co‑creating with inert systems.
AI Snips
Chapters
Transcript
Episode notes
Avoid AI Psychosis By Curating Outputs
- Treat conversational AI like a random generator you curate, not a spirit that understands you.
- Marek recommends using storytelling to make useful meaning from stochastic outputs rather than claiming the model 'knows' you.
You Can't Map A Foundation Model By Eye
- Reading models at scale is impossible for individuals because the trained artifact is astronomically large and decontextualized.
- Marek says much AI-generated art repeats because our visualization tools, not the model's essence, express vastness.
Why Vivisection Research Lags Real Time
- Mechanistic interpretability is promising but always behind current models and often speculative.
- Marek frames vivisection as cutting open dead checkpoints with probes, a slow after-the-fact science limited by access and timeliness.