Underfold
john@underfold.co Book a call
← All research  ·  Field notes
Thinking in public

Field notes.

Short, dated reactions to papers, regulation, and conversations from the field. Posted here first; cross-posted to Substack and LinkedIn. The unpolished version of what eventually becomes an essay.

4
Notes published
4.2
Avg per week
17
Sources tagged
24 days
Since last note
Filter by tag Sort by Newest first
May 2026 3 notes published this month
May172026

On Anthropic's "Constitutional AI" — what's actually new here

Paper·Anthropic, Bai et al. · arXiv:2212.08073

The constitution itself isn''t the breakthrough. What matters is the training loop that produces self-critique without human red-teaming at every step — and what that means for encoding governance into systems, not just around them.

May042026

MIT cognitive offload study — what the headline misses

Study·MIT Media Lab · April 2026

60% faster, 32% drop in cognitive load. Everyone''s quoting the top line. The interesting bit is buried in the methodology: the effect inverts almost completely for users who annotate AI output before accepting it.

May012026

The two questions every AI governance audit should start with

Practice·From a Tech-Xecutive call

Who decides? and What''s the worst that happens? — that''s it. Every framework I''ve worked with eventually reduces to those two questions wearing different costumes.

April 2026 1 notes published this month
Apr072026

NIST AI RMF revision — what changed and what didn''t

Brief·NIST AI 600-1

The generative AI profile is now baked in rather than bolted on. Notable: the measure function got the largest expansion, suggesting NIST is hearing the same thing I am.