What changed, what the evidence supports, and what deserves a closer look. Follow models, agents, science and mathematics without confusing announcements with verified results.
Newly discovered does not mean newly published. These source reports have not been independently verified.
Editorial summaries of third-party claims. We do not independently certify mathematical proofs or represent OpenAI or Anthropic.
Curated as of 29 October 2026Three source-grounded deep dives. English text; separately AI-reviewed, not scientific peer review.
Directly relevant to building realistic test tasks without exporting private Sevensa or customer records.
Source: Microsoft Research · 2026-10-30
Separate AI source check · not peer reviewDirectly addresses the atlas's promise to explain methods without inventing access to hidden model reasoning.
Source: Anthropic · 2025-04-03
Separate AI source check · not peer reviewClarifies the distinction between model capability, orchestration and enforced permissions without replacing the existing Sevensa stack.
Source: NVIDIA Developer · 2026-10-30
Separate AI source check · not peer reviewAutomatically screened · not peer reviewed by this site
Browse research → · Source status
Agent systems · primary_technical_publication
AI is expanding beyond the data center as developers build a new generation of agents to act at the edge: monitoring driver attention and cabin
Automatically screened · not peer reviewed by this siteRead analysis note ↗Evaluation & safety · primary_technical_publication
Evaluating Microsoft Copilot at enterprise scale requires test data that captures the full complexity of real organizations, yet real enterprise data contains sensitive information that
Automatically screened · not peer reviewed by this siteRead analysis note ↗Agent systems · primary_technical_publication
Building an autonomous agent requires a model, a harness, and domain-specific tools. To deploy an agent more securely, developers need a runtime that enforces...
Automatically screened · not peer reviewed by this siteRead analysis note ↗Compute & efficiency · primary_technical_publication
Teams building AI applications need a repeatable path from a model configuration to a working inference endpoint. NVIDIA Dynamo is a distributed inference...
Automatically screened · not peer reviewed by this siteRead analysis note ↗Data & governance · provider_announcement
OpenAI will honor a shared robots.txt signal to help website operators communicate their AI training preferences to participating providers.
Automatically screened · not peer reviewed by this siteRead analysis note ↗AI models · provider_announcement
OpenAI shares excerpts from court briefs by creators, academics, accessibility advocates, and legal experts, among others, on how fair use supports AI training and innovation.
Automatically screened · not peer reviewed by this siteRead analysis note ↗Selected research releases, not a complete census
Imagine moving a very thin needle in every direction while using almost no space. The Kakeya conjecture asks how small such a set can be. An OpenAI system reports a route to resolving this and related questions about waves.
Some puzzles seem to get overwhelmingly harder as they grow. A new AI-generated argument claims that a famous prediction about how quickly this happens is wrong.
Picture a whirlpool spinning ever faster into an increasingly tiny region. Can the equations describing water or air become infinitely intense from a smooth start? OpenAI says its model constructed such an example.
There are no positive whole numbers a, b and c with aⁿ + bⁿ = cⁿ for n bigger than two. Humans proved this in the 1990s. Claude's new achievement is turning the proof into instructions a computer can check.
Prime numbers follow a mysterious pattern. Claude did not solve the famous Riemann Hypothesis, but its researchers say it raised the minimum known share of certain zeros lying on the crucial line from 41.6% to 67.2%.
Testing a strategy once leaves room for luck. Repeating the test should reduce the chance of passing every round. Quantum players may share entanglement, so proving the expected improvement requires special care.
The determinant is a familiar way to combine a square matrix. Its cousin, the permanent, sums similar products without alternating signs. That small-looking change leads to a very different computational problem.
A group describes reversible operations and how they combine. One way to understand an infinite group is to imitate small portions using permutations of a finite set. Soficity asks whether increasingly accurate imitations of this kind are always possible.
Connect every pair of people and give each connection a color. How large can the group become before some three people have all three connections in the same color?
Packing identical oranges is familiar in three dimensions. Replace the room by a space with many independent coordinates and the same question becomes much harder: how much of the space can disjoint equal balls occupy?
A graph is a network of points and links. Add enough links and particular patterns become unavoidable. Extremal graph theory asks where that transition occurs.
Take a solid shape without dents and place it on an integer grid. Suppose its center of mass is the only grid point strictly inside. How much volume can the shape have?
Imagine turning a structure into a complicated measuring device. Can you reconstruct the original structure from that device alone, or can different originals produce indistinguishable devices? This is the intuition behind a rigidity question.
Picture a regular grid extended into many dimensions. Given a point between grid points, finding the closest grid point is easy to describe but can be computationally difficult. Even an approximately nearest point may be hard to find.
A communication system adds redundancy so a damaged message can still be recognized. The codewords must remain sufficiently different. How many distinguishable messages fit when every pair needs a minimum separation?
Put many points on paper and connect every pair exactly one unit apart. How many lines can you get? An OpenAI model found examples that beat a long-standing guess about the best arrangement.
No matching dossiers.
Formal checking, expert examination and journal peer review are different processes.
Public, stable JSON and Markdown exports with source links, uncertainty labels and review dates. No login, no client-side rendering.
GET /api/v1/discoveries.jsonOpen JSON ↗GET /.well-known/ai-research.jsonAgent manifest ↗GET /llms.txtDocumentation ↗