AI Daily Digest

4sections
12news items
1paper
1open-source project
24 hours of news
Audio · Chinese
Listen in Chinese →
Issue 52 · About 18 min 20 sec · AI-generated audio
After Generation, Who Checks It
October 28, 2026, Beijing time. World models, scientific proof, drug development, and public-sector AI: five top stories, seven briefs, one open evaluation, one paper, and two conversations. The main news window runs from October 27 at 06:00 to October 28 at 06:00 Beijing time (UTC+8). The source catalog uses its own 36-hour window, from October 26 at 18:00 to October 28 at 06:00, left-closed and right-open. The older interviews keep their original dates and are not counted in the catalog total.

Industry Insights

Conversation Spotlight
September 11, 2026, historical interview

Edward Hughes: scientific discovery is not only one model's intelligence score

Edward Hughes was a theoretical physicist at DeepMind and led a team on the AI-scientist direction. He is now co-founder and chief executive of Inherent AI. His focus is having agents take part in real scientific discovery. Machine Learning Street Talk is a long-form podcast on machine-learning research and systems practice. The interview cited here went online on September 11, 2026. It is not a response to today's news.

In excerpts already on the Concourse page, at 00:02:35 he says an AI cut off from a wider ecology is in some sense meaningless, and the point is what it can move together with the real ecology. At 00:15:19 he describes innovation as turning unknown unknowns into known knowns, and says a valuable breakthrough is often unexpected and only later shown to matter. At 00:08:27 he also says many paradigm shifts come from knowledge in one field being carried into another.

The link to today's news: OpenAI's mathematics and computer-science results, and Anthropic's drug-development partnerships, are both tests of whether a model's output can become a scientific advance peers accept. Hughes's view is one angle. Whether a result counts as a discovery depends not only on what the model generated, but on whether the research community can understand, check, and accept it. This is this issue's connection, not his comment on today's events. This is a paraphrase of the published excerpts, not a verbatim quotation.

Concourse · Machine Learning Street Talk | September 11, 2026, historical interview

Source: Concourse · Machine Learning Street Talk · September 11, 2026, historical interview · Published excerpts paraphrased; not the guest's comment on today's news

September 29, 2026, historical interview

Lisa Su and Fei-Fei Li: world models and compute systems need to be designed together

Lisa Su is chair and chief executive of AMD. She joined AMD in 2012, after semiconductor research and management at IBM and Freescale. Fei-Fei Li is a computer-science professor at Stanford, founding director of Stanford HAI, helped drive the ImageNet dataset, and founded the spatial-intelligence company World Labs. Bloomberg Tech is Bloomberg's technology-industry interview show. This item cites their September 29 interview about AMD acquiring World Labs. The deal amount and status follow the companies' own announcements.

Excerpts already online show Su at 00:00:32 saying hardware, software, systems, and the model layer affect one another, and that the deeper the end-to-end understanding, the better the system. Li at 00:07:20 says AI software and hardware will evolve together: models benefit from frontier hardware, and hardware benefits from knowing future workflows.

The link to today's news: Genie 4 puts real-time vision, sound, and a planning agent in one simulation loop, and NVIDIA and Oracle stress the CPU work an agent does between model calls. Both say the experience of a world model and an agent depends not only on model parameters, but on how the whole system divides the work. This is this issue's connection, not the guests' response to today's events. This is a paraphrase, not a verbatim quotation.

Concourse · Bloomberg Tech | September 29, 2026, historical interview

Source: Concourse · Bloomberg Tech · September 29, 2026, historical interview · Published excerpts paraphrased; deal status follows the companies' announcements

Research Briefs

1 paper
01KAPPA checks statistical-paper proofs one claim at a time with Lean 4 and agents
Formal proofAnnounced October 27; v1 submitted October 26 at 15:22 UTCPreprint and stated limit that faithfulness to the paper still needs human review

Researchers at Purdue submitted a preprint on October 26, announced by arXiv on October 27, introducing the open-source tool KAPPA. It uses Claude Code to organize several language-model agents, translates proofs in a LaTeX paper claim by claim into Lean 4 and its Mathlib library for checking, and gives each numbered result one of four conclusions: verified unconditionally, verified conditionally under stated assumptions, a recorded problem and where it failed, or unfinished with a reason. It checks the paper's own proof step by step. It does not look for a different proof.

The authors applied the tool to four example papers that have answers, and to three published statistics papers that later issued corrections. For those three, they tested the pre-correction versions blind and report that the tool found an error at the step the correction pointed to, and found other problems later confirmed by people. They also state the limit: the Lean 4 kernel only checks the proof of the Lean 4 statement. Whether that statement is faithful to the paper still depends on review and human confirmation. Today's OpenAI mathematics result also puts formal proof in front. This paper offers a finer angle: what a machine check can prove, and what a person still has to confirm.

arXiv | Announced October 27; v1 submitted October 26 at 15:22 UTC

Additional source

arXiv →

Open Source Spotlight

1 project
Reading note
Based on official documentation; no deployment tests were performed. Inclusion here is a tool-discovery recommendation, not a claim that each project was first released today.
1
Ai2 publishes an evaluation of scientific-idea novelty: a small change in the setup changes the judgment代码仓库许可证未单独审计

On October 27, Ai2 published the benchmark, data, code, and an automatic process for building a new evaluation set for a scientific-idea novelty reviewer. The corresponding paper, Old Ideas, Novel Problems, was already on arXiv. Using passages in ICLR 2026 peer review that explicitly discuss originality, the team built two groups of human ideas, with stronger and weaker novelty evidence, and tested six large-model reviewers.

The study finds that a small wording change in the prompt significantly changes the same model's judgment. In a comparison of human and model-generated ideas, all 30 pairs reviewed by one expert favored the human idea. For a team using a model to screen research directions, the reminder is that the reviewer itself also needs evaluation. Its score is not enough.

Hugging Face | October 27, 2026

Additional source

Additional source

Coverage notes · Incomplete · 39/51 checks recorded

Coverage is incomplete. The 05:00 upstream pass planned 51 sources: 43 complete, 7 partial, and 1 failed. After this machine checked the final hour before 06:00, the final window counts 39 complete, 11 partial, and 1 failed. The EU digital-strategy page was refused, OpenReview stopped on a browser check, and the arXiv announcement slice remains partial.

The source catalog has 364 items, 156 in news and 208 papers, after rescreening from 18:00 on October 26 and removing 33 early items from the first pass. Most catalog entries are title-level. They are not a full-text check or an endorsement. Hitting a count does not mean coverage is complete.

From 05:00 to 06:00, script-collectable sources, 36Kr, and Qbitai showed no new AI originals. Some browser sources were not rechecked in that hour and are not written as having no updates. Papers record the arXiv announcement time and the original submission time separately. The two conversations are historical interviews, with their original dates, and are not counted in the 36-hour catalog.

Coverage gaps reflect incomplete checks; they do not establish that a source had no updates.

  • xAI — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • Qwen — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • DeepSeek — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • StepFun — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • European Commission Digital Strategy — Access or collection failed. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • arXiv cs.AI / cs.LG / cs.CL — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • OpenReview — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • ByteDance Seed · Blog and papers — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • MiniMax·News & Updates — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • Ars Technica — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • MIT Technology Review — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.
  • Hacker News — Partially checked. Marked partial or failed in the collection receipt. A successful page visit is not treated as complete coverage.

Reviewing a source does not constitute independent testing. Paper-reading scope is stated with each item; experiments were not reproduced.

Sources & discoveries 364 links

Titles are retained from the Chinese edition; source names are shown in English. This discovery list includes material not read in full.

News · 155 links

Papers · 208 links

Open source and releases · 1 links

AI-assisted translation, checked against the Chinese edition. Read the Chinese edition.