AINews · Editorial

Weight Space Programming - LLMs can natively read & write neural network weights

The November 2 AINews Markdown covers October 31 through November 2. Its lead follows a weight edit that helped a small Qwen model finish math problems where it had previously exhausted its token budget.

Read the issue on GitHub ↗

Stories from this issue

Ranked by reader votes, with editorial order breaking ties. Markdown committed . Added to the archive 2026-11-03.

  1. Researchers edit Qwen’s weights to interrupt reasoning loops (github.com/rikiyat)

    In a Qwen3.5-0.8B experiment, an edit to roughly 0.016% of the weights reduced how often the model exhausted its token budget. Its reported MATH-500 score rose from 7.9% to 48.1%.

    ▲ Upvote0 points
    · Discuss
  2. DeepSeek releases a prototype that routes work to isolated coding agents (x.com/deepseekharness)

    A persistent router talks to the user and delegates tasks to workers triggered by GitHub events. DeepSeek reports a fall in p95 completion time from 29.5 to 15.7 minutes in an internal comparison.

    ▲ Upvote0 points
    · Discuss
  3. SemiAnalysis tests frontier models on 206 agentic tasks (x.com/semianalysis_)

    GDPx covers eleven domains, including chip design, software engineering and biology. Opus 5.5 posted the highest reported score at 12%, followed by Gemini 4 Carbon at 10.9%.

    ▲ Upvote0 points
    · Discuss
  4. Hugging Face releases FineDNA and a model trained on its genomic data (x.com/loubnabenallal1)

    FineDNA mixes genomic sequences with scientific papers and text about DNA. The team trained Carbon-X from Qwen3-14B and reports additional gains from reinforcement learning with verifiable rewards.

    ▲ Upvote0 points
    · Discuss
  5. Karpathy mines Hacker News comments for recommended videos (x.com/karpathy)

    HackerTube used twenty years of Hacker News discussions to identify 6,454 videos that commenters repeatedly recommended. Karpathy describes building and deploying the site with Claude Code in roughly two hours.

    ▲ Upvote0 points
    · Discuss
  6. MiniMax highlights local serving options for its 428B Flash model (x.com/minimax_ai)

    Community implementations of M3.1 Flash include a two-bit GGUF at 143.9 GB and an AMD Instinct quantization recipe. The roundup also describes offloading experts across four RTX 3090s.

    ▲ Upvote0 points
    · Discuss
  7. David Fowler uses cloud coding agents to open around 90 ASP.NET Core performance PRs (x.com/davidfowl)

    Fowler reports spending about $230 in tokens over a day with GitHub’s cloud agents and Opus 5.5. The reported count measures proposed pull requests.

    ▲ Upvote0 points
    · Discuss
  8. LlamaIndex adds a higher-effort document extraction mode (x.com/jerryjliu0)

    LlamaIndex reports that its Extract Max Effort mode surpassed Astra and Opus at their highest reasoning settings on its document-extraction evaluation, at a lower cost.

    ▲ Upvote0 points
    · Discuss
  9. Sakana announces Sawara, a language model pretrained in-house (x.com/hardmaru)

    Sakana says it plans to release Sawara Small in spring 2027. The announcement describes the company’s own pretraining project.

    ▲ Upvote0 points
    · Discuss