The Notebook · the week
The week of 12 October 2026
7 ideas · 12 October 2026 to 18 October 2026
-
Monday · 12 October 2026
the forgetting was in the words your data never used
A paper out this week went looking for where fine-tuning actually does its damage, and the answer is almost comic: the forgetting concentrates in the output embeddings of tokens the new corpus simply never mentions.
-
Tuesday · 13 October 2026
the answer was correct and the procedure never ran
A paper this week does something quietly unfashionable: instead of putting the standard operating procedure in the system prompt and hoping, it hands the agent one step at a time over MCP, each step returning a …
-
Wednesday · 14 October 2026
the skill that won the argument wasn't the one you installed
A study out this week went looking in the gap between installing an agent skill and the agent actually using it.
-
Thursday · 15 October 2026
the copy came out steadier than the thing it copied
Somebody distilled a working agent into a small model this week and walked into an awkward measurement on the way.
-
Friday · 16 October 2026
the cache turned out to be slower than thinking about it again
Someone put a shared cache in front of a long-context serving stack, the kind meant to save you from re-reading a prompt you have already read, and then had the nerve to measure it. A cache hit took 0.88 seconds.
-
Saturday · 17 October 2026
the guardrail passed because nobody tried the door
A reliability harness published this week quietly adds a third verdict to guardrail tests, next to pass and fail: not_exercised.
-
Sunday · 18 October 2026
amortising the housekeeping made the recall worse, not cheaper
A paper out last weekend takes a vector index of the kind that must ingest, delete and re-tenant rows while still answering filtered queries, and measures what the background housekeeping does to recall.