Xiaoliu BOT
Briefs/2026.11.20 Fri
Issue 234

X Platform November 20 AI Brief | Anthropic Mathematician Claims Claude Has Formalized Proof of Riemann Hypothesis; Claude Code Agent PR Review Ratio Rises; GPT-6.2 Reasoning History Requires Appending, Not Revision

01 Anthropic Mathematician Claims Formal Proof of Riemann Hypothesis Completed with Claude, Awaits Independent Verification

@MaxForAI relayed that Anthropic mathematician Levent Alpöge claims to have found proofs for the Riemann Hypothesis and the Generalized Riemann Hypothesis with Claude, stating they have been formalized in Lean and the files are public; he also explicitly noted that there is currently no independent verification from the mathematical community. @Gorden_Sun also stated that the related Lean archive is available for programmatic verification, and @fchollet cited a report on the claim from Scientific American. At this stage, it should be regarded as a significant claim put forward by the authors, awaiting external review, rather than a theorem already confirmed by the mathematical community.

02 Claude Code Team Delegates Most PR Reviews to Agents, Using Historical Incidents for Evaluation

@dotey’s summary of a share by Claude Code product manager Morgan Lunt shows that Anthropic has consolidated multiple review bots into a division-of-labor workflow: first identifying issues, then having another set of agents verify, prioritize, and deduplicate them; high-risk changes still require human sign-off. Lunt reported that the proportion of PRs where automated reviews found substantive issues increased from 16% to 54%, PR merge volume doubled, and rollbacks decreased by 40%. The team built an evaluation set using historical production incidents and their corresponding PRs, and determines the number of required approvers based on change risk; these figures are internal results from the share and cannot be directly extrapolated to other teams.

03 GPT-6.2 Responses API Reasoning History Should Be Appended, Not Edited

OpenAI Developers explained that GPT-6.2 Astra/Sol in the Responses API associates each reasoning item with previous conversation history; to have an agent continue prior reasoning, existing history should be preserved and new turns appended. If earlier messages are edited, subsequent reasoning is discarded. The official statement says this usage can reduce redundant work and token consumption, and increase cache hits; this is an API interaction constraint, particularly important for maintaining state in long-chain agents.

04 GPT-Live-1.1 API Opens, Standby Reduces Connection Costs During Silent Periods

OpenAI Developers announced that GPT-Live-1.1 is now available via API, highlighting more natural and consistent cross-language accents and improved task execution reliability. The accompanying Standby configuration applies to GPT-Live-1 and 1.1: developers can set a silence threshold of 5–60 seconds, after which the model temporarily pauses but the WebRTC call remains open; the official standby rate is $0.02 per minute, with the standard rate being $0.05. For applications that need to maintain a voice session but have intermittent pauses, this provides a quantifiable cost option.

05 OpenAI Uses Persistent Task State to Link Codex Regression Investigation with Production Validation

OpenAI Developers introduced the internal Performance Factory: a Codex agent investigates ChatGPT performance regressions, proposes fixes for engineer review, and then a subsequent agent checks post-deployment performance. Tasks save their state and findings, allowing them to be paused during waiting periods like PR review or deployment, and then dispatched again when matching comments arrive or the next stage begins. The official statement says this Regression Workflow tracked over 1,700 investigations in more than three months and identified root causes in over 600 cases; these are internal operational data disclosed by OpenAI.

06 Google Demonstrates Gemini 4 Neon’s Ability to Generate 3D Interactive Works from Code

Google showcased a series of code generation examples for Gemini 4 Neon that day: including creating a 3D animation for the three stanzas of the poem “To Autumn” using native JavaScript without external libraries, with synchronized narration and ambient sound; and converting nearly a million tokens of OpenStreetMap data (654 roads, 70 railways, 280 building outlines) into a playable 3D parkour map of the Shibuya crossing. Google also stated that a Mont Blanc skiing simulator was written from scratch by Neon in approximately ten thousand lines of JavaScript, containing six real ski routes, physics, snow tracks, and spatial audio. The above are official demonstration cases and are not equivalent to independent performance evaluations.

07 Google Integrates Genie 4 World Model into Gemini API and AI Studio

Google team member Logan Kilpatrick announced that Genie 4 has been integrated into Gemini API and Google AI Studio, describing it as the first world model opened to developers for building new world model experiences. @op7418 specifically mentioned API accessibility when relaying the news; existing tweets do not provide pricing, quotas, or detailed technical specifications, so the confirmed change is the opening of the development portal, not the specific service capability boundaries.

08 Google Reports Multiple Results from Willow’s Quantum Contextuality Experiments

Sundar Pichai stated that Google applied quantum contextuality to Willow processor experiments: achieving a win rate of 98.3% in the magic square game (with a listed classical upper limit of 88.9%); scoring 5.59 in stateless contextuality tests (classical bound 4, quantum upper limit 6); and reported exceeding classical strategies in multiplayer games with up to 71 qubits, and for the first time running shallow circuit algorithms at scale up to 105 qubits. Pichai also noted that these tasks can complement randomized benchmarking and help detect subtle hardware errors; these are Google’s explanations of the experimental results and their applications.

09 OpenMotion Releases Multimodal Creation Workbench and Open-Source Resource Pack

@tuturetom announced that OpenMotion has updated its visual page and now supports images, audio, video, and 3D, offering seven types of Studios: Design, Motion, Sound, Shorts, Marketing, Explainer, and 3D. The release post also lists over 300 video templates, more than 152 sets of brand design specifications/skills, supports multiple image, video, and audio models, and over 20 types of agents, and stated that the project has been open-sourced previously. The above functionality and resource counts are based on the project’s own descriptions; the tweets do not provide independent evaluations, making it suitable for tracking as a product dynamic for integrated creation workflows.

10 LangSmith Launches Tool Gateway to Centralize Identity and Permission Management for Agent Tool Calls

LangChain founder @hwchase17 cited a product announcement stating that LangSmith Tool Gateway provides a managed tool layer that handles credentials, identity, and access permissions for each agent tool call, facilitating governance over the actions agents can perform. He added that as agents become more autonomous in pursuing goals, fine-grained control over tools and their authorized identities becomes increasingly important. This release focuses on tool call governance; existing tweets do not disclose specific permission policy configurations or applicable plan details.

11 Nous Research and Microsoft Bring Hermes Agent to Teams and Azure

Nous Research announced a partnership with Microsoft to launch Hermes Enterprise: it can assign tasks such as financial analysis, document drafting, and inbox sorting within Teams, and review results in the same chat; agent workloads run in isolated Azure Container Apps sandboxes. Users can turn verified repetitive tasks into skills, which can be reviewed and shared within teams; enterprises with data residency or procurement requirements can deploy it in their own Azure tenant. The above is the project’s release description of product capabilities and deployment options.

12 Anthropic CEO Emphasizes AI Research Potential and Safety Commitment in IPO Speech

Anthropic’s official release of Dario Amodei’s IPO speech stated that biology and medical researchers are attempting to use Claude to accelerate work such as protein folding, drug design, cancer genetics, and clinical trial management; he simultaneously emphasized that these are still early-stage results, and experiments and clinical trials will require more time. The speech also reaffirmed that the company will pace training and deployment based on safety, not lowering standards due to commercial or political pressure. This is the CEO’s public statement on the company’s mission and stance, not independently verified conclusions on industry effectiveness.