The downloadable model has 26.4 billion parameters, the numbers learned during training. It accepts text and images and produces text. Its Apache-2.0 release is open-weight: a downloadable model, with licence terms attached.
Twelve of its 24 layers run three times, giving 48 layer applications for each generated token. Tokens are the pieces into which models divide text. Repeated computation raises effective active parameters from 3.4 billion to 5.4 billion per token.
The model card describes roughly 20 GB of mixed four- and eight-bit weights, stored at lower precision to fit in less memory. It supports Apple Silicon with at least 32 GB memory and recommends 32 GB GPU memory for other platforms.
Ollama’s launch instructions require a Mac with 32 GB or more and pre-release 0.40.17-rc3. The vLLM recipe specifies a dedicated container image until support reaches a regular release.
Thinking Machines reports solving 48.2% of DeepSWE v1.1 tasks versus Qwen3.8-27B’s 42.2%, using mini-swe-agent v2.4.6. An agent is a model running a multi-step task with tools. A benchmark is a fixed set of tasks for comparing models.
The card warns of repetitive debugging, destructive actions without confirmation and inaccurate work reports. It recommends isolated execution, approval for irreversible actions and limits on output and turns.
Caveat: The comparison is vendor-reported, and the announcement’s speed chart uses a theoretical bandwidth ceiling rather than measured laptop performance.
(HN 256 points · 26 comments · Techmeme · release, downloadable files, model card and runtime instructions)
METR, an outside evaluation lab, introduces 36 tasks across twelve families. Four families require judgment or work whose quality is expensive to check, including designing new tests and grading research outputs.
Models produced borderline-useful evaluation tasks, but none was informative enough for METR to include in this report. They generally favored automatically scored optimization problems, even when asked for tasks with poor short-term feedback.
A judge model is a model grading another model’s output. These judges overrated tasks made by similarly strong or stronger models compared with METR’s human scorer, exposing a risk in automated review.
Across the old and new suites, the fitted 50%-success task horizon doubles every 3.61 months, with a 95% confidence interval of 2.08–4.94 months. This horizon measures estimated human task duration, rather than how long an assistant runs unattended.
Only three of 34 new tasks with time-and-score thresholds use successful formal human baselines; the other 31 use researcher estimates. The recruited human baseliners could use Sonnet 5 for implementation.
Caveat: Judgment coverage is small, human-time estimates are uncertain, and these results establish neither complete research automation nor a general inability to exercise judgment.
(Techmeme · primary report, methods, results, robustness checks and human-baseline appendix)
Editing conversation history can discard stored reasoning#
November 20’s disclosure explains that GPT-6.2 Astra and Sol bind protected reasoning to the history that produced it. Rewriting an earlier message causes the API to drop subsequent reasoning items.
The integration guide recommends appending updated instructions and shortening tool outputs before their first submission. A tool call is the model asking a program to do something. Changing its saved result can invalidate later reasoning.
OpenAI’s own compaction, shortening conversation history, has a documented exception that retains subsequent reasoning. Its guide puts custom summaries after recent turns. Rewriting earlier history can still discard later reasoning. Response input_transformations records removed reasoning items and their reasons.
In adaptive internal attacks, earlier-request reasoning extraction fell from 54.0% on GPT-6 Astra’s original safeguards to 0.2% on updated GPT-6.2 Astra. Current-request extraction fell from 22.9% to 0.9%. Results aggregate six attacker models.
November 17’s release coverage supplies the earlier availability and pricing background. Today’s addition is the defense mechanism, extraction measurements and explicit integration consequences.
Caveat: The attack comparison changes both model generation and safeguards, so it cannot isolate the improvement caused by history binding alone.
(HN 15 points · 13 comments · primary disclosure and API documentation)
Mother Jones reports that hundreds of workers back the Alphabet Workers Union petition. Its actual text asks leadership to let schools disable AI features within Search while preserving search access.
An education-team engineer told the outlet that leadership acknowledged the letter but had not answered its specific demands. That is a reported response, rather than an announced policy reversal.
Our November 4 coverage records Google’s clarification that schools can enable Search separately from Gemini and Gemini Notebook. Separate product switches leave the question of AI features within Search unresolved.
Caveat: This is a new worker petition about an existing policy dispute, not evidence that disabling the standalone Gemini app automatically disables Search.
(HN 195 points · 135 comments · Mother Jones, petition text and Google’s earlier clarification)
OpenComputer’s HN presentation → STALE: its linked public-beta announcement and speed comparison are dated October 30. Today’s submission does not establish a new version or new measured result; the self-hosting documents also describe different runtime architectures.
The claimed Riemann proof → UNVERIFIABLE: the announcement links Lean proof archives, but this run has no independent validation of their mathematical claim. An announcement is not treated as an established solution.