An independent verification report says the Claude-assisted Lean proof of the Riemann Hypothesis passed checks in multiple proof kernels. That is stronger evidence than a viral screenshot. It is also different from broad mathematical acceptance or a decision by the Clay Mathematics Institute. This World of AI roundup also looks at another claim that needs a careful distinction: Gestalt tops a design benchmark, but the leaderboard does not say it is GLM 5.5.
Watch World of AI's November 21 Roundup
Source: World of AI's November 21, 2026 episode and its transcript. The episode includes a sponsored segment for Proton's Lumo. Release and proof status were checked against the linked sources on November 21. Follow World of AI on X.
Did Claude Help Prove the Riemann Hypothesis?
The 1859 hypothesis says that all the nontrivial zeros of the Riemann zeta function have real part one-half. The zeros matter because of their link to the distribution of primes: proving the hypothesis would give mathematicians much tighter control over how far the count of primes can stray from its expected pattern. Clay Mathematics Institute's problem overview explains the connection.
On November 20, Anthropic mathematician Levent Alpöge posted a four-page note claiming that he and Claude found proofs of the original hypothesis and a generalized version. It links to a substantial Lean archive; the community mirror and its verification notes document roughly 17.4 million source lines. The 250,000-page figure in the episode refers to a separate experiment generating a human-readable rendering, not the length of the published note.
| Evidence | What it supports |
|---|---|
| The announcement | Alpöge credits Claude as a collaborator and presents the original hypothesis and the generalized result as separate claims. |
| Independent formal checks | MIT mathematician Andrew Sutherland's verification report documents a full build, an audit of the assumptions, and acceptance by independently implemented proof kernels. It also reports matching the five headline statements against a separate challenge. |
| Mathematical and prize status | Clay's page still marks the problem unsolved at this article's cutoff. Its prize rules require qualifying publication, time for scrutiny and broad acceptance. No prize decision follows from a successful Lean check. |
Checking the statement as well as the proof matters: a valid proof of a lookalike theorem would not answer Riemann's question. The independent checks are substantial evidence for the formal result. A readable account of the key ideas and scrutiny by specialists are separate questions. This November work is also distinct from Anthropic's August announcement, which reported an improved lower bound from 41.6% to 67.2%, not a proof of the full hypothesis.
The episode's count of three solved Millennium Prize Problems is not an official tally. Clay acknowledged a separate apparent advance on Navier-Stokes and described its review process; Clay's P vs NP page still lists that problem as unsolved. A repository update is not evidence that Google solved another one.
Is Gestalt Actually GLM 5.5?
Contra's Frontier Design Benchmark shows Gestalt as a stealth entry. In the snapshot covered in our previous roundup, it scored 1,931 Elo to Claude Fable 5.2's 1,713. The benchmark compares website design outputs; it doesn't identify the provider or establish a lead in general reasoning.
World of AI ties a reported message from Z.ai's community team to speculation about an upcoming release. The public Z.ai model documentation confirms GLM-5.3. We did not find a public GLM 5.5 model card, confirmed date or statement that links it to Gestalt by November 21. A release tease, a past stealth launch and a leaderboard entry don't establish that they refer to the same model.
Google Stitch Loop Tracks Outcomes, Not Just PRs
Google Labs announced Stitch Loop on November 18 as a product agent that connects a codebase with live metrics. It looks for changes that could advance a goal, builds them in a sandbox and submits pull requests for a person to approve. After a change ships, it can observe relevant metrics and use the outcome to inform its next proposal. The global waitlist is open, with priority for eligible Google AI Ultra users; that's different from general availability.
TestingCatalog reports a connection between the internal Jitro name and earlier Jules work. Google's public name is Stitch Loop. Improving a product over time is also different from changing the AI model's weights. A good first goal would be to reduce median page-load time on a test route, with a known baseline, a check that key interactions still work, and a human review before deployment.
Inkling 2 Mini: Reusing Layers Instead of Adding Weights
Thinking Machines Lab released Inkling 2 Mini on November 20. The Apache 2.0 model card lists 26.4 billion stored parameters, 3.4 billion active before accounting for the repeated layers, and 5.4 billion effective active parameters per token with looping. The middle 12 layers run three times with shared weights. It accepts images and text, returns text, and supports up to one million tokens of context.
Artificial Analysis scores it around 20 on its Intelligence Index, roughly level with Qwen 3.8 27B at Extra High in that evaluation. That does not settle which model is better for a specific frontend build. Local results also depend on the runtime, quantization, hardware and input length. The model's maximum context is not a promise that a laptop can run at that limit. Start with the official weights or a compatible quantized build; third-party variants are not Thinking Machines' release.
What Else Shipped, and What Remains a Report
| Story | What the evidence supports |
|---|---|
| ChatGPT Space | OpenAI's release notes say Space is rolling out to Plus users aged 18 and older, with page and presentation editing on mobile as well as web and desktop. Shared content still requires permission and an eligible signed-in account; there are no public links. |
| Claude on mobile | World of AI shows a design refresh. Anthropic's docs confirm artifact creation and editing on iOS and Android; exact interface availability can vary. |
Antigravity /btw | Side questions can check the code while the main agent keeps working. The desktop tab supports follow-up questions and promotion to a separate conversation. That doesn't automatically merge its work into the main thread. |
| Gemma 4 and Gemini media | Gemma's browser explainer lets you see the tokenizer at work; it isn't a Gemma 5 announcement. Gemini 3.5 Transcribe supports spoken edits to text. Omni 1.2 Flash GIFs in the Gemini app are short loops, with geographic and plan limits. |
| Early drug research | MedChemBench evaluates reasoning over examples drawn from existing medicinal chemistry research. GPT-Rosalind is a separate research offering. Neither announcement establishes a new treatment or demonstrates patient benefit. |
| Neuralink gaming clip | The episode shows what it describes as brain-computer input in GTA 6. Neuralink documents investigational digital device control, including experimental console use for some study participants. A single clip does not establish clinical effectiveness or general availability. |
The video also describes a possible unified Grok Build and Cursor CLI, Nano Banana Pro 2 sightings and reports of OpenAI departures. Those are not confirmed release or staffing announcements here. The desktop animation segment is an observation about the interface shown, not a rollout date.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Intro |
| 02:08 | Rumors of a new GLM model |
| 02:48 | Gestalt leads Contra: Is it GLM? |
| 03:50 | Why World of AI suspects Z.ai |
| 05:58 | The Claude-assisted Riemann proof |
| 12:57 | The Millennium Prize status claim |
| 13:20 | Speculation about Google and mathematics |
| 14:17 | Claude mobile interface |
| 14:45 | Google Stitch Loop |
| 17:00 | Inkling 2 Mini and its open weights |
| 18:02 | Running Inkling locally |
| 18:44 | Community model builds |
| 19:37 | Speed and cost |
| 20:14 | Inkling 2 Mini versus Qwen 3.8 27B |
| 20:37 | The Gemma 4 interactive explainer |
| 21:19 | Antigravity side questions |
| 21:55 | Gemini 3.5 Transcribe voice editing |
| 22:50 | GIF generation in Gemini |
| 23:28 | Reported Grok Build and Cursor CLI integration |
| 24:45 | ChatGPT Space rollout to Plus |
| 26:41 | MedChemBench and drug-discovery research |
| 27:14 | What AI could change in research workflows |
| 28:10 | ChatGPT desktop animations shown |
| 28:31 | Reports of OpenAI researcher departures |
| 29:31 | Nano Banana Pro 2 rumors |
| 29:50 | Neuralink gaming clip shown |
| 31:40 | Outro |