Perhaps a broader pattern:

As capability increases, the need for complex demonstration decreases.

The highest form of capability might be one that can afford to be transparent, simple, and direct about both its strengths and limitations.​​​​​​​​​​​​​​​​

Recursive Irony: This thread itself becomes a perfect example of valuable training data, containing nuanced technical discourse, community dynamics, and ethical debate - exactly the kind of content AI researchers would want to train on.

Human language learning strategies might serve as inspiration for designing more effective LLM training paradigms through language games. Polyglots like Prof. Arguelles describe language learning as an integrative, discovery-based process, including multi-modal integration.

I've been doing LLM analysis of threaded online technical discussions and extracting patterns, roles, rules, etc. representing things like sensemaking, knowledge crystallization, and seeing single roles take turns as teacher and student. Possible language game for training?

Have you tried adding MCP to datasette?

Experiment synthesizing Bluesky threads

Interesting, was it actually revising the text file incrementally? Do you think it would use a document app window if it was open?

Preview for planning, Mini for execution

Completely agree. Around 2013 I rode in a self-driving taxi in Abu Dhabi (Masdar City) and it was the highlight of the trip!