All episodes

AI & Tech Daily

A Claude-assisted Riemann proof passes independent checks

17:33

Independent checkers validate Levent Alpöge's Claude-assisted Riemann Hypothesis proof, opening the next task of explaining its mathematical ideas. Apple releases Depth Pro 2 for spatial-vision research, AWS launches managed agent memory, and Apollo and Goldman price US$44 billion in SpaceX GPU financing. Also: OpenAI ties protected reasoning to conversation history, autonomous weapons negotiations extend to 2028, and what ChatGPT Space's paid sharing and Signal's screen protections change for users.

Full transcript

Read the episode.

I'm Jesse Owen. This is AI and Tech Daily.

A machine-checked Riemann proof

A Claude-assisted proof of the Riemann Hypothesis has passed independent mathematical checks. Researchers now have an argument they can inspect and rebuild, while the work of explaining it in readable mathematics continues.

Our main story today is that proof, and how AI-assisted discovery can produce something other researchers can build on. On 21 November, the Lean and Mathlib communities endorsed the correctness of Levent Alpöge's proof, released the previous day. Alpöge credits collaboration with Claude in developing it.

The Riemann Hypothesis concerns the distribution of prime numbers. The achievement here is a formal mathematical argument: the reasoning has been expressed in a form that proof-checking software can examine. That's an unusually concrete output for AI-assisted research, and it gives mathematicians a substantial new object to study.

There are two important parts to what the communities checked. They established that the theorem written in the formal system matches the established mathematical statement. They also checked that independent proof checkers accept the argument. The first connects the formal work to the question mathematicians intended to answer. The second checks the logical route to the answer.

That connection is easy to overlook if you think of verification as a single green tick. A mathematical statement has precise conditions. The formal statement needs to preserve them, and each step in the argument needs to support the next. Here, both the meaning of the statement and the acceptance of the proof have received scrutiny.

MIT mathematician Andrew Sutherland has separately reported a reproducible build in a sandboxed environment and successful independent-kernel checks. The kernel is the part of a proof system that checks the logical steps. In his 22 November update, some additional full-export checks were still running. The community endorsement and the completed independent checks are already in place.

For someone who builds software, the reproducible part is particularly interesting. Another researcher can take the released material, build it in a separate environment, and examine the result. That makes the argument available for further work without requiring everyone to reconstruct the entire discovery process first.

Alpöge says a readable mathematical explanation and simplification remain works in progress. That's a substantial next task. A formal proof can establish that the steps hold together while leaving a mathematician with considerable work to understand the central idea, why a particular route succeeds, and which parts might be useful elsewhere.

Think of the difference between following every turn on a route and understanding the shape of the landscape. Both are useful. A readable explanation helps people see the organising idea, ask better follow-up questions, and teach the result. Simplification could also make the argument easier to work with. Those are opportunities opened by having the formal proof available.

For an AI tool builder, this also offers a useful way to separate discovery from checking. The system helping to find an argument can be complex, while the released argument remains available to independent checking software. That separation gives collaborators a practical handoff.

My takeaway for developers working on research tools is the value of producing an inspectable result alongside the search for an answer. In a domain with formal checks, the output can travel beyond the original researcher and model. Other people can reproduce it, examine its structure, and begin their own work from it.

For mathematicians, the next absorbing question is what the argument reveals. Which ideas become clearer when it's simplified? What can a readable account help researchers recognise? The released proof gives that explanatory work a concrete starting point, and the possibility of turning an accepted chain of reasoning into a more widely understood mathematical insight.

Depth from photos and video

Let's leave the mathematics there and get closer to the kinds of images a computer has to understand.

Apple published Depth Pro 2 research on 19 November, with code and model weights available. It uses one network to estimate depth in both photographs and video, giving individual pixels an estimated distance in metres at the input's native resolution.

That last phrase, distance in metres, carries a lot of the usefulness. A depth map describes how far away different parts of an image appear to be. Putting those estimates on a physical scale gives a researcher more to work with than a rough ordering of foreground and background. Native resolution matters for a related reason: the depth estimates correspond to the detail in the original input. That makes small features and the edges between objects useful places to examine how well the estimated scene holds together.

The method uses monocular input, meaning a single camera view. That opens a research path for spatial understanding from ordinary images, without requiring a dedicated depth camera. The estimated distance is inferred from what the model sees in the image.

Apple reports sharper boundaries and steadier depth estimates across video frames. Those are practical qualities. At an object's edge, foreground and background can be close together in the picture while sitting at very different distances. Keeping that boundary sharp helps preserve the shape of the scene in the depth map.

Video adds another problem: consistency over time. A sequence can look smooth to you while the estimated depth shifts from frame to frame. That makes it harder to use the sequence as a coherent description of a space. More stable estimates could give researchers a better basis for following the same scene as the camera or its contents move.

Having one network handle still images and video is also convenient for experiments. Researchers can examine the two kinds of input within a shared approach, including the additional challenge that movement introduces. The available code and weights give them something they can run and investigate themselves.

The licence sets a clear boundary: the released weights permit non-commercial scientific research and academic development only. Ordinary commercial deployment isn't authorised by that release. For a business considering a product, that affects the decision before model performance enters the discussion.

For computer-vision researchers, I see the opportunity in studying spatial detail and temporal stability together. A crisp estimate from one frame is useful; an estimate that stays coherent through a sequence could be more useful still. Depth Pro 2 makes that combination available for research, with the physical distances, fine boundaries and behaviour over time all open to examination.

Managed memory for agents

If you've built an assistant that forgets yesterday's work, this next change addresses a familiar piece of plumbing.

AWS made OpenSearch Memory generally available on 18 November. It manages conversation storage, memory extraction, consolidation and retrieval across agent sessions. Those are the steps involved in keeping useful information from earlier interactions and bringing it back when an agent needs it.

A long conversation contains a mixture of useful details and material that has little bearing on the next question. Sending the whole history back to a model keeps everything available, but it also makes the prompt larger. A memory service selects information to retrieve, which can reduce how much the model has to read for each answer.

AWS gives a concrete comparison from its LongMemEval-S test. Retrieved context used approximately thirteen thousand five hundred tokens per question, compared with a hundred and sixty-seven thousand for the full history. Tokens are the chunks of text a model processes. That's roughly one-twelfth as much context.

Answer accuracy in that test was 93 per cent with retrieval, compared with 95.2 per cent using the full history. So the measured trade-off is a much smaller input alongside a 2.2 percentage-point reduction in accuracy. Both numbers belong in the same decision. The question for a builder is how valuable the reduced input is relative to the answers their application needs to get right.

The test used a single run per configuration, with Claude Opus 5.5 answering and separately grading. It gives a specific comparison for that setup. A workload where the key detail is buried in an old exchange may place different demands on retrieval than one with a small set of recurring preferences.

The service also changes the work involved in building an agent. Developers can use a managed pipeline for storing conversations and turning them into retrievable memory. That could make persistent assistants easier to operate, particularly where maintaining the memory machinery has been a separate engineering job.

There are charges for storage, retrieval and extraction, so fewer prompt tokens are one part of the cost calculation. The service's own charges also contribute. The useful comparison is the cost of answering a question with the required quality across the whole workflow.

For Australian organisations, location may settle the question earlier. The launch regions are Northern Virginia, Oregon, Tokyo and Frankfurt. No Australian region is listed. An application that requires Australian data residency may therefore be unable to use these launch locations.

My read is that managed memory becomes attractive when selective recall matches the job. The benchmark makes the trade-off tangible: a large reduction in retrieved context, a measurable accuracy difference, and a service bill to weigh against the engineering work it takes over.

Financing the next GPU build-out

The scale changes sharply with this one: a plan to finance tens of billions of dollars in computing equipment.

Apollo and Goldman Sachs announced pricing on 20 November for approximately 44 billion US dollars in senior and subordinated notes to finance GPUs and related equipment for SpaceX. Reuters also reported the transaction, which is intended for terrestrial computing infrastructure.

The planned equipment includes NVIDIA Vera Rubin chips. Lease schedules are expected to begin when equipment arrives in the first half of 2027. That gives the announcement a concrete timetable: financing terms now, with equipment delivery and the start of leases expected next year.

Pricing the notes establishes the terms of the debt financing. Senior and subordinated notes put lenders in different positions in the repayment order. In plain language, the financing has layers, with different claims on repayment. The money is intended to support a very large equipment programme.

The timing of the leases connects the financial arrangement to the arrival of the hardware. That is useful context when a headline contains a number as large as 44 billion dollars. The announcement describes how a planned expansion is being financed and when its equipment is expected, while usable computing capacity depends on the rollout that follows.

For organisations planning AI services, there's a practical difference between knowing that a major buyer has arranged financing terms and knowing when computing resources will be available to their own business. This announcement supplies evidence about the former. Final deployed capacity and customer pricing haven't been established.

I think the useful lesson for business planning is how much of AI expansion sits outside model development. Someone has to finance the equipment, suppliers have to deliver it, and a timetable has to connect those commitments. Model capability can move quickly while the infrastructure supporting it follows a delivery schedule measured in months.

The expected first-half 2027 arrivals give equipment suppliers and infrastructure planners a milestone to work towards. For organisations watching the AI market, the scale of the financing shows the resources being committed to the next build-out. Following those deliveries will help reveal how that financial commitment translates into computing infrastructure.

Reasoning tied to conversation history

There's also a smaller engineering change worth catching if your software edits or replays conversations with a model.

OpenAI published updated anti-distillation safeguards on 20 November. Distillation involves using one model's outputs to help train another; these safeguards address adversarial extraction of protected reasoning.

In OpenAI's internal adaptive test, extraction of reasoning from earlier requests fell from 54 per cent with GPT-6 Astra's original safeguards to 0.2 per cent with GPT-6.2 Astra's updated safeguards. The comparison covers those particular model-and-safeguard combinations and that extraction task.

The implementation detail that affects builders is history binding. GPT-6.2 Astra and Sol bind protected reasoning to the conversation history. Rewriting an earlier message discards the reasoning that follows it.

That means an edit to the past changes the state available for subsequent work. Consider an application that lets someone correct an earlier instruction and then continue. The visible conversation may look like a revised copy of the same exchange, but the later protected reasoning has been discarded. Software managing that conversation needs to account for the reset.

It's also relevant when a developer replays an exchange while changing earlier messages. The relationship between the old history and the reusable reasoning state has been broken. Any design that relies on carrying that state forward has to respect where the history changed.

For builders, I'd put that behaviour ahead of the headline percentage in the implementation notes. It changes how conversation editing works: an earlier-message rewrite creates a boundary beyond which protected reasoning is no longer retained. That's a concrete constraint for anyone trying to make a long-running interaction editable and predictable.

Autonomous weapons talks continue

Away from product releases, the international discussion about autonomous weapons has another date on the calendar and difficult questions unresolved.

UN-backed talks in Geneva concluded on 21 November with agreement to develop a possible accord on lethal autonomous weapons by 2028. AP reports that participants did not commit to legally binding rules.

The negotiations took place under the Convention on Certain Conventional Weapons, where decisions require consensus. That process gives the participants a shared negotiating framework, with agreement needed to move decisions through it.

The outcome extends the discussions. The substance of any eventual agreement and its legal force remain undecided, so defence organisations and technology suppliers have a further negotiating timetable rather than newly enacted international restrictions from this meeting.

My reading is that enforceable human control and accountability remain the central issues to resolve. A timetable can organise negotiations, but the practical effect for organisations depends on what countries eventually agree to require and how those commitments operate.

For suppliers, the eventual meaning of human control could shape how responsibilities are divided between a system's operator and its automated functions. That makes the content of the proposed accord, including how it assigns responsibility, particularly relevant as negotiations proceed.

What Changes for You

Two changes may be more immediately visible on your own screen, especially if you share work or private messages.

ChatGPT Space began rolling out globally to Plus subscribers aged eighteen and over on 20 November. It lets people create interactive pages and presentations, collaborate, and configure updates from connected apps and tools. Creation and editing work across web, desktop, iOS and Android, with availability varying during the rollout.

The useful possibility is a page that stays connected to changing information. That could reduce the manual work of refreshing a document or presentation. Sharing has a firm boundary, though: recipients need permission and a Plus, Pro, Business or Enterprise account with Space access. Free users can't open the content, and public links are unavailable.

I'd judge its usefulness partly by who needs to read the result. It could fit work shared among eligible colleagues; a client outside those paid accounts can't open the same shared page. That restriction affects where it fits in an existing workflow.

Private conversations have a different screen-level change. Signal announced screenshot blocking on iOS 27 and automatic screen protection on Android devices where it detects an AI assistant attempting access.

For affected users, a screenshot that previously worked may now be blocked. You can disable the protection, restoring screenshots and potentially exposing displayed conversations to agents. Coverage varies: some Android devices aren't supported, macOS screenshot blocking is unavailable, and other access paths remain possible.

Reuters reports that agents have accessed Signal messages through operating-system permissions, reaching conversations displayed on the device.

That's why the screenshot setting has become a privacy decision. Encryption protects the messaging exchange, while an assistant allowed to operate the device may encounter the conversation on screen. Signal's protection reduces that screen-capture route. For someone who wants both an active device assistant and private chats, keeping those displayed conversations outside the assistant's view is a useful additional boundary.

You'll find the sources and full transcript at owenonthenet.com. Thanks for listening.

Sources

Reporting behind this episode.

  1. leanprover-community.github.io/blog/posts/riemann-hypothesis
  2. math.mit.edu/~drew/grh
  3. alpo.ge/remark.pdf
  4. arxiv.org/abs/2611.26603
  5. huggingface.co/apple/DepthPro2
  6. aws.amazon.com/blogs/big-data/introducing-amazon-opensearch-memory-persistent-memory-for-ai-agents
  7. ir.apollo.com/_assets/_7f8917b1a1c55d8c56a5fa93b0a0a009/apollo/news/2026-11-20_Apollo_and_Goldman_Sachs_Announce_Pricing_of_44_662.pdf
  8. channelnewsasia.com/business/apollo-goldman-lead-44-billion-spacex-gpu-debt-financing-6570241
  9. openai.com/index/combatting-adversarial-distillation
  10. arabnews.com/world/nations-at-un-talks-on-ai-killer-robots-agree-to-more-talks-but-no-firm-commitment-to-rein-them-in-3013276
  11. help.openai.com/en/articles/6825453-chatgpt-release-notes
  12. signal.org/blog/protecting-your-privacy
  13. investing.com/news/stock-market-news/exclusivesignal-restricts-screenshots-to-block-ai-agents-from-reading-messages-5040998