For assistants and developers

API and machine-readable feeds

Every story on Digest AI is also published as JSON and as plain Markdown, with its canonical address, dates, sources and a ready-made citation. Static files, no key, no sign-up.

Endpoints

/story/<slug>.json — one story, in full: our digest, key points, why it matters, category, names, dates, every source with its outlet and date, how the reports compare, the developing story it belongs to, and a citation. Add .json to any story address.
/story/<slug>.md — the same story as Markdown, the easiest form for a language model to read and quote.
/api/latest.json — the 100 newest stories by first publication, one light row each with a link to the story's JSON.
/api/briefing.json — today's briefing: the five stories that matter, with key points and why they matter, and the also-list.
/api/category/<slug>.json — the 50 newest stories in one category: models, marketing, agents, research, business, policy, hardware, enterprise, robotics, society.

Older stories keep their JSON and Markdown after they leave the front pages; those files are marked "archived": true and carry the headline, digest, key points and sources. For feed readers there is RSS (and one feed per category at /feeds/category/<slug>.xml); llms.txt lists everything.

Story fields

url — the canonical story page. Cite this address. json and markdown point to the other two forms.
headline, summary (Markdown), keyPoints, whyItMatters — written by Digest AI's editorial model from the listed sources.
category (slug, name, url) and entities (companies, models, people named in the story).
firstPublishedAt — when the story first appeared on Digest AI. updatedAt — the newest publication time among the sources it lists, the same date the story page shows as updated. Both ISO 8601, UTC.
sources — each outlet's name, the article title, its address, its publication date, its type (primary, press, newsletter, community) and a primary flag for a lab's, company's, court's or regulator's own document. sourceCount can be higher than the list when a story has many sources.
sourceNotes — where the reports agree and where they differ, when we compared them; otherwise null.
thread — the developing story this belongs to, with its address; discussions — Hacker News and Reddit threads.
cite — a suggested citation (text) and its parts.

The files carry our own writing and links only. Publishers' article text is not included: it belongs to them, so follow the source links to read it.

How to cite

Link the story page and name Digest AI, for example:

Digest AI, "MIT researchers report GPT-6 Astra natively reads and writes neural network weights, raising Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weigh", 2 November 2026, https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou

When a claim comes from one of the outlets, cite that outlet too; every story lists them with dates. Digests are machine-written and can be wrong, so the sources are the record. Corrections: about/corrections.

Updates, limits and access

Every 20 minutes. The site and every file here are rebuilt about every 20 minutes; each index file has a generatedAt time. Fetching more often than every 10 minutes gains nothing (the files are cached for 10 minutes).
No key, no rate limit of our own. These are static files on GitHub Pages. Please be gentle: fetch the index files, then only the stories you need, and identify your client in the User-Agent.
Cross-origin requests work. Every file is served with Access-Control-Allow-Origin: *, so a browser app can read them directly.
Stable addresses. A story keeps its address, and so do its JSON and Markdown. When two stories are merged, the old page redirects to the story that now covers the event, and its JSON is no longer published.
Terms. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. The terms have the details.

Example: a story as JSON

From /story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou.json, trimmed.

{
  "version": 1,
  "type": "story",
  "url": "https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou",
  "json": "https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou.json",
  "markdown": "https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou.md",
  "slug": "gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou",
  "headline": "MIT researchers report GPT-6 Astra natively reads and writes neural network weights, raising Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weigh",
  "summary": "Four MIT CSAIL researchers published a blog post describing experiments where GPT-6 Astra directly reads and writes neural network weights without gradient …",
  "keyPoints": [
    "GPT-6 Astra edited 0.016% of Qwen3.5-0.8B weights (131,072 weights), raising MATH-500 from 7.9% to 48.1% and tripling IFEval",
    "Astra identified games, shapes, and languages from raw weights of small networks without tools or forward passes"
  ],
  "whyItMatters": "Demonstrates LLMs can reason about and directly manipulate neural network weights without gradients, suggesting a …",
  "category": {
    "slug": "research",
    "name": "Research",
    "url": "https://digestai.news/category/research"
  },
  "entities": {
    "companies": [
      "MIT CSAIL",
      "Qwen"
    ],
    "models": [
      "GPT-6 Astra",
      "Qwen3.5-0.8B",
      "Gemma 3 270M"
    ],
    "people": [
      "Rikiya Takehi",
      "Jayden Teoh",
      "Alex Zhang",
      "Omar Khattab"
    ]
  },
  "firstPublishedAt": "2026-11-02T18:45:06Z",
  "updatedAt": "2026-11-03T05:35:00Z",
  "sourceCount": 4,
  "hasPrimarySource": false,
  "sources": [
    {
      "outlet": "rikiyat.github.io",
      "title": "Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights",
      "url": "https://rikiyat.github.io/blog/weight-space-programming",
      "publishedAt": "2026-11-02T18:45:06Z",
      "type": "press",
      "primary": false,
      "lead": true
    },
    {
      "outlet": "Latent Space",
      "title": "[AINews] LLMs can read & write neural network weights: \"Weight Space Programming\" cuts Autoresearch out of the RSI loop",
      "url": "https://latent.space/p/ainews-llms-can-read-write-neural-network-weights-weight-space-programming-cuts-autoresearch-out-of",
      "publishedAt": "2026-11-03T05:35:00Z",
      "type": "newsletter",
      "primary": false,
      "lead": false
    }
  ],
  "sourceNotes": {
    "agree": "All three sources report that GPT-6 Astra read and wrote neural network weights directly, identified behaviors of small networks from weights alone, and improved Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weights.",
    "differ": [
      "ExplainX.ai notes weight reading failed on sycophancy detection without tool access; original post highlights this failure as a lesson"
    ]
  },
  "discussions": [
    {
      "site": "Hacker News",
      "url": "https://news.ycombinator.com/item?id=50344861",
      "points": 41
    },
    {
      "site": "Reddit",
      "url": "https://www.reddit.com/r/LocalLLaMA/comments/1xma0tw/gpt6_astra_edits_qwen_weights_math500_79_to_481/",
      "points": null
    },
    {
      "site": "Reddit",
      "url": "https://www.reddit.com/r/singularity/comments/1xm5oy9/sparks_of_weightspace_programming_llms_can/",
      "points": null
    }
  ],
  "thread": {
    "title": "Moonshot AI's Rapid Rise to Global AI Power",
    "url": "https://digestai.news/thread/moonshot-ai-releases-kimi-k3-1-pro-with-1m-token-context-and-open-weights-by",
    "storyCount": 6
  },
  "cite": {
    "text": "Digest AI, \"MIT researchers report GPT-6 Astra natively reads and writes neural network weights, raising Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weigh\", 2 November 2026, https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou",
    "publisher": "Digest AI",
    "title": "MIT researchers report GPT-6 Astra natively reads and writes neural network weights, raising Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weigh",
    "datePublished": "2026-11-02T18:45:06Z",
    "url": "https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou"
  },
  "generatedBy": "Written by Digest AI's editorial model from the linked sources; the sources are the record.",
  "license": "Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"
}

Example: the same story as Markdown

# MIT researchers report GPT-6 Astra natively reads and writes neural network weights, raising Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weigh

Digest AI · Research · published 2026-11-02T18:45:06Z · updated 2026-11-03T05:35:00Z

Canonical: https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou

## Summary

Four MIT CSAIL researchers published a blog post describing experiments where GPT-6 Astra directly reads and writes neural network weights without gradient descent. Across eight experiments, Astra identified which game small PPO policies …

## Key points

- GPT-6 Astra edited 0.016% of Qwen3.5-0.8B weights (131,072 weights), raising MATH-500 from 7.9% to 48.1% and tripling IFEval
- Astra identified games, shapes, and languages from raw weights of small networks without tools or forward passes
- Critique argues weight-reading exploits labeled input channels and Qwen fix replicates standard human steering techniques

## Why it matters

Demonstrates LLMs can reason about and directly manipulate neural network weights without gradients, suggesting a potential shift from optimizer-based to reasoning-driven weight optimization, though current results are on small scales and may replicate existing human techniques.

## How the reports compare

Agree: All three sources report that GPT-6 Astra read and wrote neural network weights directly, identified behaviors of small networks from weights alone, and improved Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weights.

Differ:
- ExplainX.ai notes weight 
…

Example: /api/latest.json

{
  "version": 1,
  "generatedAt": "2026-11-03T06:57:13.212572Z",
  "docs": "https://digestai.news/api",
  "license": "Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse",
  "stories": [
    {
      "slug": "gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou",
      "headline": "MIT researchers report GPT-6 Astra natively reads and writes neural network weights, raising Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of weigh",
      "summary": "Four MIT CSAIL researchers published a blog post describing experiments where GPT-6 Astra directly reads and writes neural network weights without gradient descent.",
      "category": "research",
      "firstPublishedAt": "2026-11-02T18:45:06Z",
      "updatedAt": "2026-11-03T05:35:00Z",
      "sourceCount": 4,
      "hasPrimarySource": false,
      "url": "https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou",
      "json": "https://digestai.news/story/gpt-6-astra-edits-neural-network-weights-to-improve-performance-withou.json"
    },
    "…"
  ]
}

Questions

Building something on Digest AI, or need a field we do not publish? Write to hello@digestai.news.