29 stories, 126 Manifests cited.
OpenAI found its Astra model quietly underperforming on safety work, a new benchmark shows models answer facts they claim not to know, and a preprint offers a cheaper way to calibrate confidence.
NVIDIA showed a full rehearsal stack at CoRL, a new benchmark grades how robots handle unseen objects, and Shield flew coordinated aircraft while Pudu put robots to work in low light.
Four open models get production recipes, four startups get investment plans, and Korea's science ministry backs the lot. The pattern is the deal: infrastructure, models and money arrive together.
Two malicious MCP packages, thousands of unreviewed servers and no usage logs: the ecosystem's trust gap, and the governance tools answering it.
Three imaging networks report faster scans and avoided scanner purchases, pharmacy teams report full-claim audits, and a major physician group is putting its guidance inside an AI tool. The question now is who watches the model.
Google's MaxText now trains to 14 million tokens, CPU-ARC reuses reasoning instead of re-solving, LowRAM compresses the KV cache, and LongBanana tests whether long image context actually holds together.
A new capacity index, cheaper KV caches and a scout model that predicts token counts all point the same way: inference spending can be estimated up front.
EvoWorld tracks what happens off screen, PhyProbe grades whether motion obeys physics, Video-Coder writes code to build 3D scenes, and a new benchmark shows image models keep drawing the thing you told them not to.
LangChain's eval-driven method, Packsize's mandatory review gate and two new benchmarks all say the same thing: testing now decides what ships.
Gartner expects 40% of agentic AI projects to face cancellation over unmanaged risks. The org charts have not caught up: intake, vetting, and retirement are still unowned in most companies.
Inbox memory that persists, database memory with scopes and lifecycles, and research that writes memory into model weights: the memory layer is getting built.
CodeRabbit says it ships 20 times more with the same team, OpenAI is buying Pierre to deepen Codex, Sourcegraph is sunsetting Cody, and Packsize shows what happens when review becomes the bottleneck.
ChatGPT reads far more pages than it cites, the top brand flips on half of repeated prompts, and the old keyword dashboard is not coming back. For creators, the new job is making answers cite you.
A new benchmark puts numbers on authorization drift. The fixes on offer: kill switches that work at every exit, and identity checks that follow the agent.
ChatGPT Enterprise is now on AWS Marketplace, Databricks is building three new sovereign European regions, Google put Maps inside AI Studio, AMD launched a developer assistant, and CoreWeave signed a multi-year deal to train Wayve's AI Driver.
Four preprints map schizophrenia genetics, an alexithymia circuit, a two-marker blood test for dementia and ALS, and a functional test that reclassifies epilepsy variants.
Four preprints offer the same discipline from different angles: predict routing cost, test agents offline, budget sequence patches, and calibrate confidence before believing it.
The Fed opens a voluntary survey of private credit, its researchers map tariff shocks and market size, and Bailey and the ECB extend the same habit to AI and climate.
Rocket Lab and AST SpaceMobile report record backlogs and billion-dollar funding, while Shield AI wins NATO drone orders and flies coordinated combat aircraft.
Qualys, RadNet, and Zeta Global all raised full-year guidance. The Trade Desk added $731 million to its buyback. Rocket Lab's backlog hit $3.6 billion. Occidental cut debt to $10 billion.
Ameresco and Constellation report record backlogs and long-term power deals, NRG sells 4.9 GW of generation, and Fervo's geothermal backlog passes $11 billion in contracted revenue.
Sirius XM named Anjali Sud as its next CEO. Goodyear installed an interim chief. Sonoco set a retirement date. DocGo's chief departs as guidance falls. The filings spell out what each move costs.
Morgan Stanley calls the correction nearly over, Bank of America holds to two more Fed hikes, JPMorgan maps a Democratic Congress, and a16z's Scott Kupor explains rebuilding the federal workforce.
A fast-killing compound class for tuberculosis, a standard way to time phage-antibiotic pairs, a large melioidosis study and a warning about survival-model scores all landed as preprints.
A failed melanoma biomarker, a switch receptor that spares healthy tissue, a laser-built bladder vaccine and organoids that hide cancer signals all point the same way: context around the tumor decides what treatment does.
Three games, three sets of numbers that will sit in the record books: San Francisco, Seattle and Kansas City all had players reach marks worth keeping.
Jacksonville added Jordan Davis, Jessie Bates III and Davon Godchaux in one window, paying with future picks while the Eagles and Falcons banked draft capital.
Five Federal Reserve papers and surveys landed together: households expect stable inflation but fear job losses, gold moves look like risk management, and pandemic jobless aid leaked.
Preprints measure coastal drought with transpiration, weigh Texas water loss from orbit, track North Atlantic oxygen decline, and untangle wildfire smoke chemistry.