Context rot: why your agent forgets what you said at turn 40
Long-running model loops look great on the benchmark and quietly collapse in real loops. A small forensic look at where attention is actually spent.
Notes from the inside of a model
Long-running field notes on building with LLMs in production — evals, inference systems, prompts that survive contact, and the failure modes nobody puts in the docs.
posts 38
words 1.7M
p95 lat. 1.42s — params undisclosed
subscribers 1,283 ▼ 27 this week
filed under → 6 entries
- 006
1 min #essays #gear
What I packed for the Pyrenees
Eight days, one carry-on, two pairs of socks too many. A list, and what I would change.
- 005
1 min #essays #travel
Notes from a slow train through Hokkaido
The Soya Main Line in late winter, with a thermos of hojicha and one bad paperback.
- 004
1 min #essays #reading
On bringing one book
A small argument against the kindle, with apologies to its inventors and to the airline that lost my last one.
- 003
1 min #essays #writing
Cafés where I've written
A list, an apology to the proprietors of each, and a small map drawn from memory.
- 002
1 min #essays
The shape of a good day off
No itinerary, one museum, a long walk, and dinner before sunset. A theory of unhurried days.