TabLOID

METR's November 5 assessment estimates that Anthropic's progress on well-specified AI tasks accelerated roughly twofold relative to its earlier trend. The researchers spent eleven weeks examining internal work. They judge AI assistance the main contributor, while emphasizing substantial uncertainty.

This is evidence of people using AI to improve AI. It does not establish fully autonomous self-improvement or a precise doubling of every kind of research. Benchmarks cover limited tasks, and other changes may explain some of the gains.

There is also a visibility problem: Anthropic redacted most of the evidence METR considered most informative. Readers must rely heavily on the assessors' judgment. METR reports finding no clear bottleneck during its access, but acknowledges it may have missed important ones. The finding is consequential; its caveats are part of the finding. METR's assessment, November 5.

OpenAI announced GPT-6-Cyber on November 5 through its restricted Daybreak Red program. The company classifies the model at its Critical cybersecurity threshold, reflecting capabilities that could help defenders and create serious misuse risks.

OpenAI reports a 48.5% result on its ExploitGym evaluation, compared with 32.1% for its previous model. These are company evaluations, not an independent audit of performance across real organizations.

Its scope-following test also needs a narrow reading. OpenAI reports 99.93% compliance without its deployment prompt and 100% with that prompt in the evaluated setting. Those figures do not establish universal reliability. The practical question is whether access controls, bounded permissions and human oversight hold up when a capable system meets an unfamiliar environment. OpenAI's release, November 5.

Anthropic's November 5 announcement introduces Personal Finance in Claude, rolling out in beta to US Pro, Max and Ultra subscribers. Users can choose accounts to connect through Plaid. Claude receives read-only access and cannot transfer money, manage accounts or invest on the user's behalf.

The view can still be extensive: selected account details, balances and up to two years of transactions. Anthropic says Claude does not see bank passwords. It also says conversations inside the Personal Finance hub are excluded from training; conversations elsewhere follow the user's existing settings.

A consolidated view may make household finances easier to understand. It also gives another service access to sensitive information, and the assistant can make mistakes. Read-only describes what it can do with an account, not whether every conclusion deserves your trust. Anthropic's explanation, November 5.

Discovered Materials has opened five scientific challenges covering superconductivity, semiconductor interfaces, quantum chemistry, laser interactions and heat transport. The advertised awards total up to $1 million.

The official rules make that ceiling conditional. The sponsor judges submissions, can require independent verification and does not guarantee that the full amount will be awarded. There is no fixed closing deadline.

There is a consequential detail for would-be entrants: submissions grant the sponsor broad, perpetual, royalty-free commercial rights, whether or not they win. Entrants retain ownership subject to that license. A prize can focus attention on difficult science, but its conditions are part of the offer. The primary pages do not substantiate the claim that leading AI agents already failed these tasks. The challenges and official rules, checked November 6.