How We Learn & Unlearn
We don’t worship papers. We ship things, break them, keep receipts, and change our minds in public.
Phone-first
Under-1s latency
Private by default
Adversarial-UX aware
Reality > ritual. If a claim matters, it gets a screenshot or a step list—or it gets cut.
This page isn’t exhaustive—it’s layered. We learn from live model runs, people who think, institutional signals (with salt), and human feedback loops.
Level 1
Machine Conversations
- Daily runs across models (compare, log, adjust).
- Blind A/Bs: human vs model—publish the deltas.
- Refusal/jailbreak notes → user-safe guardrails.
Level 2
People Who Think
- Lanier, Crawford, Chiang, Rushkoff, Turkle—ideas > hype.
- Disagreement is a feature, not a bug.
- Pull quotes only when they change a design decision.
Level 3
Institutional Intel (cautious)
- OpenAI/Anthropic/DeepMind papers: skim → test claims.
- EU/FTC/EFF policy cues → UX constraints, not panic.
- Roadmaps & demos get sandboxed, not copy-pasted.
Level 4
Human Signals
- Reader feedback & bug reports → patches and posts.
- Practitioner interviews (recruiting, teaching, design).
- We log our misses and update when we’re wrong.
We bias for working code
We publish failure notes
We revise fast
Our method, short
- Define the job → set constraints → pick the output shape.
- Run across models → log latency, refusal rate, slop.
- Ship a fix (tool, prompt, or warning) → capture receipts.
Bias check
- We’re phone-first, privacy-biased, anti-busywork.
- We mistrust magic demos; we trust repeatable steps.
- We prefer small wins over theory marathons.