self.md radar — 2026-11-03
a rejected tool still ran
a tool could be rejected by a callback and still execute when a Python program continued through Response.reply()
LLM 0.37 now carries that callback into the reply, including the next round
1. LLM fixes replies that skipped a tool rejection
version 0.37, released November 2, fixes the callback bypass
it also fixes llm chat saving responses when logging was turned off
when Response.reply() automatically ran pending tool calls, it omitted the configured before_call and after_call hooks
later replies lost them too
the fix forwards both hooks through the synchronous and asynchronous paths and preserves them on the returned response
the regression tests ask for a blocked tool and a permitted one, then another permitted call in a second reply round
only the permitted calls may run
if your application uses before_call to reject actions, test that rejection through reply() as well as the first turn
the logging fix concerns LLM’s own chat history; it says nothing about a model provider’s retention
reading: LLM 0.37 release notes
2. Qwen’s math score rises after an edit that stops some loops
85 of the first 96 practice answers hit a 32,768-token ceiling
that was the Qwen3.5-0.8B failure GPT-6 Astra inspected in a new weight-editing experiment
Astra changed 131,072 weights in the MLPs of layers 15 and 19, without gradient descent
with a budget of 50 batched forward passes on practice inputs, the authors report MATH-500 rising from 7.9% to 48.1%
the share of MATH-500 answers hitting the token limit fell from 93.0% to 33.1%
then they tried Gemma 3 270M, which rarely looped
its mistakes were ordinary arithmetic and algebra errors, and the edits produced no established math gain
removing a loop can let a model finish an answer it could already work out
the Gemma result limits the claim that this method adds a missing ability
the released logs include the reported runs and final weight edits, but exclude trials with earlier prompt versions
reading: Weight-Space Programming: experiments, settings and failed comparison
3. Census now asks businesses about Super Intelligence
question 7 of the Census Bureau’s published business questionnaire calls the technology “Super Intelligence or SI”, with a parenthetical explanation that it was formerly AI
the examples still include machine learning, virtual agents and voice recognition
that PDF is marked for January 18–April 11, 2027
the BTOS survey history dates the instrument change to the October 19, 2026 collection
this is a name in a questionnaire, not evidence that respondents have acquired a new level of machine capability
a different change already broke the comparison with older adoption figures
from November 17, 2025, the question expanded from AI used in producing goods or services to AI used in any business function
keep that scope change separate from the 2026 rename when plotting the series
whether the new name changes how businesses answer remains unmeasured here
reading: Census’s account of the survey changes · the 2025 wording memo