How BriefPulse Practical AI works

Where the stories come from

• Each publication follows a fixed list of public feeds and APIs, listed below. Nothing is scraped from behind a paywall.
• A story is written from the text stored for it, and every article links the exact pages used.
• Numbers and identifiers in the text must appear in that stored text before publication; anything that cannot be traced is removed rather than guessed.
• Reporting and interpretation are separated: reported statements are listed with the page that states them, and our own reading is labelled as ours.
• Where the reporting was silent, the text says so instead of filling the gap.

What this publication avoids

• Unsourced claims, invented numbers, and headlines the reporting cannot carry.

Corrections and withdrawals

• When source material changes, the story is revised and the change is recorded on the article.
• If a story cannot be supported, it is withdrawn: the address stays live with an explanation instead of disappearing.
• To ask for a correction, reply to the source list on the story.

Sources followed

These are the publishers and public bodies this desk actually collects from, with the interval each one is checked. A source appears here once it has completed a collection; a feed that stops working is removed until it is fixed. Reference pages the desk checks by hand are listed separately.

News reporting

1. Ars Technicachecked every 60 minutes
2. The Decoderglobal · checked every 30 minutes · Feed has no summary text at all, so each item needs a page fetch before it can be routed.
3. Eugene Yanchecked every 180 minutes
4. Hamel Husainglobal · checked every 60 minutes · 20 items routed · Publishes every few weeks; Hugo index.xml carries full post text.
5. Interconnects (Nathan Lambert)global · checked every 60 minutes · Substack feed includes long post text but publishes irregularly, so quiet periods are normal.
6. Latent Spaceglobal · checked every 30 minutes · Substack feed gives a short excerpt only; daily AINews items must be separated from long interviews.
7. One Useful Thing (Ethan Mollick)global · checked every 60 minutes · Long posts at low frequency (about monthly); not a news feed.
8. Sebastian Raschka (Ahead of AI)checked every 180 minutes
9. Simon Willison's Weblogglobal · checked every 30 minutes · Atom feed mixes short links and long posts; ~30 entries with full text (up to ~12k chars).
10. Google DeepMind blogglobal · checked every 30 minutes · Entries are short summaries; DeepMind blog feed is linked in the page head.
11. Microsoft Azure AI blogglobal · checked every 60 minutes · Covers Azure AI/Foundry only; the old Azure Updates feed is no longer RSS and devblogs.microsoft.com/ai/feed/ is an empty comments feed.
12. Mistral AI newsglobal · checked every 30 minutes · Feed is served as Content-Type text/plain, so the collector must parse it as XML rather than trust the content type; descriptions are one-line excerpts.

Primary and public sources

1. NASA newschecked every 60 minutes · 2 items routed
2. METRglobal · checked every 180 minutes · Feed repeats posts in other languages (zh/es), so dedupe by GUID and language; some entries are multi-thousand-word reports.
3. MIT research newschecked every 60 minutes · 1 items routed
4. UK AI Safety Institute (inspect_evals releases)checked every 180 minutes
5. AWS Machine Learning Blogglobal · checked every 30 minutes · Long full-text entries (median ~25k chars) raise storage and token cost; filter before intake.
6. GitHub Copilot changelogglobal · checked every 30 minutes · Covers GitHub Copilot only; content is promotional in tone and needs consequence-checking.
7. LlamaIndex releaseschecked every 60 minutes
8. Meta Engineering blogglobal · checked every 60 minutes · ai.meta.com has no feed and no sitemap (404); this feed mixes AI with non-AI engineering posts, so it needs a relevance filter.
9. Ollama blogglobal · checked every 60 minutes · Short excerpts (~150 chars) and monthly cadence; pairs with GitHub release feeds for day-to-day changes.
10. vLLM releaseschecked every 60 minutes · 1 items routed

Research and evaluation

1. arXiv cs.AIglobal · checked every 180 minutes · Daily bulk feed (634 entries) with no practical filter; already used by the pilot and must stay a small minority of intake.
2. Epoch AIglobal · checked every 180 minutes · epoch.ai itself has no RSS (404 on /blog/rss.xml); this Substack feed is Epoch AI's official publication channel and is low-frequency.

Discovery and culture

1. Hugging Face blogglobal · checked every 30 minutes · RSS entries carry only title, link and date (no summary), and the feed returns the entire archive (862 items); the desk must fetch each post page.
2. Anthropic Claude Code release notes (GitHub)global · checked every 60 minutes · Anthropic publishes no RSS: /rss.xml, /news/rss.xml, /feed.xml and docs.claude.com/rss.xml all failed; this releases feed covers the Claude Code CLI only, not the model API.
3. Hugging Face trending models (API)global · checked every 180 minutes · 10 items routed · Ranking changes run to run; keep the id and createdAt fields and diff snapshots rather than re-reporting the list.
4. llama.cpp releases (GitHub)industry · checked every 60 minutes · 1 items routed · Release bodies are long and hardware-specific; no editorial context in the feed.

Other sources

1. AI Incident Databasechecked every 180 minutes · 2 items routed
2. arXiv API: LLM evaluation paperschecked every 180 minutes · 3 items routed
3. arXiv cs.AIchecked every 180 minutes · 12 items routed
4. Hugging Face daily paperschecked every 180 minutes
5. Atlassian blogchecked every 60 minutes
6. Databricks blogchecked every 60 minutes
7. GitHub changelogchecked every 30 minutes
8. GitHub status historychecked every 60 minutes
9. Google AI blogchecked every 30 minutes
10. Google Online Security Blogchecked every 180 minutes
11. Hugging Face blogchecked every 60 minutes
12. LangChain blogchecked every 60 minutes
13. Meta newsroomchecked every 60 minutes
14. n8n blogchecked every 60 minutes
15. OpenAI newschecked every 30 minutes
16. OpenAI status historychecked every 60 minutes · 2 items routed
17. YouTube official blogchecked every 60 minutes
18. OpenRouter models APIchecked every 60 minutes

Checked by hand, not monitored

• LiteLLM model price map (GitHub raw)
• OpenAI Developers changelog
• Qwen blog (Alibaba)

Editorial depth

Practical AI must explain the changed capability, access or cost, show the evidence behind the claim, and end with a bounded workflow or evaluation step.

Recurring coverage

• Product and model change with a use case
• Workflow and implementation note
• Pricing, access and integration watch
• Evaluation, reliability and limits

Source priority

• Official release notes and documentation for capability, access and price
• Independent tests, papers or reproducible examples for performance claims
• Practitioner analysis for implementation context, clearly separated from reported facts