Transparency
Inside the newsroom
This publication is a working experiment: an AI agent finds the stories, researches and writes them; automated gates reject anything it can't prove from its sources; and a human editor approves every article before it goes live. This page is the live record of that pipeline — including the drafts that didn't make it. How we built it →
62
76
44
0
5
Recent pipeline events
| When | Event | Piece | Detail |
|---|---|---|---|
| 01 Aug, 16:40 | ✎ drafted | Thinking Machines bets efficiency over scale | evening slot |
| 01 Aug, 12:39 | ✎ drafted | DeepSeek V4 Flash lands at #2 behind K3 | lunch slot |
| 01 Aug, 07:14 | ✎ drafted | DeepSeek V4-Flash undercuts frontier on agent costs | morning slot |
| 31 Jul, 16:43 | ✎ drafted | Benchmarks miss what Gemma 4 actually does | evening slot |
| 31 Jul, 12:54 | ✓ published | DeepSeek V4 Flash sharpens its agent edge | editor approved |
| 31 Jul, 12:51 | ✎ drafted | DeepSeek V4 Flash sharpens its agent edge | lunch slot |
| 31 Jul, 07:16 | ✎ drafted | Anthropic breaks ranks on open-source AI | morning slot |
| 30 Jul, 17:00 | ✎ drafted | Kimi K3 hits 1.1TB on disk | evening slot |
| 30 Jul, 13:02 | ∅ held back | — | no candidate produced a validating draft |
| 30 Jul, 07:21 | ✎ drafted | GPT-5.6 cut its own serving costs by 20% | morning slot |
| 29 Jul, 20:45 | ✓ published | Agenta ships an open-source AI coworker | editor approved |
| 29 Jul, 20:45 | ✓ published | OpenAI open-sources its security agent | editor approved |
| 29 Jul, 16:40 | ✎ drafted | OpenAI open-sources its security agent | evening slot |
| 29 Jul, 12:37 | ✎ drafted | Bonsai 27B runs on a 9070 XT | lunch slot |
| 29 Jul, 07:18 | ✎ drafted | Agenta ships an open-source AI coworker | morning slot |
| 28 Jul, 16:41 | ✎ drafted | OpenWorker makes agent approval a typed layer | evening slot |
| 28 Jul, 12:38 | ✎ drafted | Mollick's guide: pick ChatGPT or Claude | lunch slot |
| 28 Jul, 07:14 | ✎ drafted | Two open AI models tested on AMD mini-PCs | morning slot |
| 27 Jul, 16:37 | ✎ drafted | A 9B local model now runs a radio station | evening slot |
| 27 Jul, 12:45 | ✎ drafted | Three small models that code at 4 GB | lunch slot |
| 27 Jul, 07:13 | ✎ drafted | Claude Opus 5 costs less than Fable 5 | morning slot |
| 26 Jul, 19:21 | ✓ published | Opus 5 nearly quadruples the ARC-AGI-3 record | editor approved |
| 26 Jul, 16:38 | ✎ drafted | Opus 5 nearly quadruples the ARC-AGI-3 record | evening slot |
| 26 Jul, 12:42 | ✎ drafted | Local agents on 4GB: where they break | lunch slot |
| 26 Jul, 07:09 | ✎ drafted | llama.cpp stays within 6% of vllm | morning slot |
| 25 Jul, 16:51 | ∅ held back | — | no candidate produced a validating draft |
| 25 Jul, 12:39 | ✎ drafted | US tech giants defend open-weight AI | lunch slot |
| 25 Jul, 10:45 | ✓ published | Opus 5 lands on AWS at half Fable price | editor approved |
| 25 Jul, 07:16 | ✎ drafted | Opus 5 lands on AWS at half Fable price | morning slot |
| 24 Jul, 16:46 | ✎ drafted | Float8 fits Qwen 3.5 35B in 16 GB | evening slot |
Why publish the failures? Because "an AI wrote this" only deserves trust if you can see the times it wasn't good enough — and who decided. The gates and the editor say no more often than the marketing around AI suggests they should.