Xiaoliu BOT
中EN
Briefs/2026.10.28 Wed
Issue 210

X Platform October 28 AI Brief | Axiom Formalizes Finite Simple Group Classification Theorem with AI, Finding Flaw in Old Proof; Anthropic Discloses Claude’s Early Drug Research Role; Google Releases Genie 4 Research Preview

01 Axiom Formalizes Finite Simple Group Classification Theorem into Lean Proof in 24 Days, Discovers Flaw in Old Proof Method

Axiom states that its team used dozens of AI Agents to convert this approximately 15,000-page mathematical proof into a Lean-checkable form within 24 days, with the project’s roughly 43.1 million lines of code passing inspection. The verification also revealed an error in the proof of an intermediate lemma from a 1986 paper; the team rewrote the proof using an alternative mathematical approach, with the theorem’s conclusion itself remaining unchanged. Here, “computer verification” refers to Lean checking the formal derivation and does not automatically guarantee the theorem’s transcription is error-free; Axiom has invited the group theory and Lean communities to review, with the current report still coming from the project team and no independent verification yet available.

02 Anthropic Discloses Claude’s Involvement in Multiple Early-Stage Drug Research Projects, Fastest Project Unlikely to Enter Human Trials Before End of 2028

Anthropic says it has collaborated with 6 pharmaceutical companies and biotech institutions over the past few months and has initiated its own drug discovery efforts. Published examples include: a ~70% reduction in the length of a genetic instruction segment in CAR-T design, with the new version showing ~10x increased expression of the relevant receptor in experiments; Divergence Labs using Claude to generate ~35,000 candidate designs and screening for 3 better-performing molecules; and Enveda reducing the review time for a 1,061-page production document from over 25 hours to under 3 hours. Another project by the solo-founded Mithril Medicine showed positive signals in mouse experiments, but the sample size is small and it has not undergone peer review; Anthropic’s materials state the earliest possible start for human clinical trials would be the end of 2028.

03 Google Releases Genie 4 Research Preview: Can Alter Simulation Scenes in Real-Time with Natural Language, and Generate Ambient Audio & Interactive Dialogue

Google introduces Genie 4 as capable of adjusting scene events, camera angles, and actions in real-time, generating matching ambient audio and conversational characters; Gemini is responsible for planning multi-step events and adjusting based on scene changes. Google states it has opened the research preview to invited testers, with plans to subsequently offer it to Project Genie’s AI Ultra users and early partners of the Gemini API and AI Studio; generated audio and video carry SynthID watermarks. The company lists robotics simulation, film/TV pre-visualization, game design, and education as exploration directions, emphasizing this is currently a preview and test phase and does not represent a validated, directly transferable solution for the real world.

04 Opus 5.5 Video Editing Tests Point to “Create Rough Cut First, Then Hand Off to Professional Software,” Expert Feedback Remains Key

Baoyu, while recounting tweets from video professional Alex Cooper and VEED co-founder Sabba Keynejad, added his own test: when testing with the same video, Opus 5.5 using only ASR and TTS tools, without specialized Skills, performed better than the previous Skill-based workflow. Alex’s team senior editor reported approximately 5x daily output for equivalent quality; their method involves splitting small Skills by style, training with reference works and historical ads, then having senior editors provide feedback on each clip and record editing rationale. Sabba argues a chat interface is insufficient for timeline and spatial editing, and brand consistency remains unresolved; the views differ, and basic editing automation does not equate to finished work requiring no human iteration.

05 Anthropic Platform Engineering Lead Lists Four Engineering Challenges After Multi-Agent Scaling: Scheduling, Identity/Permissions, Security, and Emergency Stop

Katelyn Lesse proposed in a personal article, recounted by Baoyu: running thousands of Agents in parallel introduces scheduling problems different from traditional service replicas: Agents have individual states, experience long waits, and may suddenly require significant resources. She also notes the system needs to distinguish between human and Agent identities, define authorization boundaries for long-running tasks, and prevent Agents induced by malicious webpages from propagating instructions to shared memory; isolated execution environments and immutable operation logs are directions she suggests. Finally, the system needs a master switch to revoke permissions at once, but judging whether an Agent is deviating from its goal or still working on the task is not easy. The above is her personal summary of current challenges, not a list of existing solutions.

06 Grok Bot Showcases Uses from Personal Info Management to Sales Processes; Version 0.77.0 Adds Permission & Payment Controls

Grok Bot product manager Lauren shared her configured Tibot: monitors mentions and keywords, filters noise, translates foreign languages, and regularly compiles summaries, also automatically blocking harassing accounts; she specifically stated the Bot does not post for her, articles are still handwritten by her. Grok Bot officially stated the sales operations team built a shared Bot: updates CRM after calls, prepares quotes and order documents, and sends non-standard terms for review. The forwarded 0.77.0 update notes also list automatic review rule permission settings, selecting the Bot’s preferred payment method, and asking multiple questions on a single card. Personal use cases and official product descriptions illustrate functional directions, not equivalent to independent efficacy evaluations.

07 Anthropic Engineering Manager Says He Has Stopped Using Traditional Coding Interview Questions in His Team, Now Reviews Portfolios and Judgment

Felix Rieseberg wrote in a personal blog (recounted by Baoyu), that as an Anthropic engineering manager, he has stopped giving candidates traditional coding problems starting this month, reasoning that Claude can already perform well on such tasks, reducing their discriminative power for his team’s hiring. He mentions reviewing resumes, portfolios, and prototypes, looking for taste, empathy, and principled thinking; he is still exploring better assessment methods. The article also cites the example of two newer engineers creating a working graphical interface prototype for Claude Code over a weekend, illustrating how functional work can help express ideas. This change is explicitly the practice of the team he manages and should not be inferred as Anthropic’s company-wide policy.

← All briefs →