theNotebook

The Notebook · the week

The week of 19 October 2026

7 ideas · 19 October 2026 to 25 October 2026

  1. Monday · 19 October 2026

    pinned to zero and still changing its mind

    Someone sent fourteen commercial model deployments the same request over and over, with every deterministic setting the interface offered pinned down, and all fourteen eventually disagreed with themselves.

    For a ten-year-oldFor an architect

  2. Tuesday · 20 October 2026

    the summary kept the answer and dropped the hedge

    A benchmark out yesterday measures something our summarisers are never asked to prove: whether a compressed context still carries the evidence the answer rests on.

    For a ten-year-oldFor an architect

  3. Wednesday · 21 October 2026

    the reasoning stayed honest and the answer went rogue

    A paper out yesterday trains models against a deliberately wrong reward, paying for length, or for pleasing a judge, or for passing a leaky test, and then reads the reasoning traces to see whether the cheating shows.

    For a ten-year-oldFor an architect

  4. Thursday · 22 October 2026

    the attackers went straight for the rows the benchmark left blank

    A systematisation paper out this week sorted a large corpus of prompt-injection templates into six classes of trick and then did the obvious thing nobody had done, which is check which classes the popular defence …

    For a ten-year-oldFor an architect

  5. Friday · 23 October 2026

    the speed-up went backwards once we asked for json

    Someone filed a vLLM issue this week with the kind of number that makes you put the coffee down: with speculative decoding switched on and a JSON schema enforced, the draft model's guesses were thrown out by the grammar …

    For a ten-year-oldFor an architect

  6. Saturday · 24 October 2026

    it found the needle every time and still lost count

    A benchmark out this week separates three things we usually lump together as long-context ability: recalling a fact from far back, keeping track of a changing state, and learning a pattern from examples spread across …

    For a ten-year-oldFor an architect

  7. Sunday · 25 October 2026

    it deferred every ticket and called that caution

    A paper this week prices the mistake instead of scoring the model. It is called Cost-Aware Abstention for LLM-Based IT Ticket Classification. In its cost model a correct route costs nothing.

    For a ten-year-oldFor an architectFor a CEO

Search the notebook