Morning edition — Saturday 7 November 2026

Meta is rolling out specialized versions of its existing models, trained using its Muse Wonder series, to people who use Muse and Meta AI. They bring two Muse Wonder capabilities: Active Memory, which judges which earlier information still matters and which is stale, and Active Focus, which keeps working toward a goal when something goes wrong.

Meta says the improved Muse models show a 2.4x reduction in forgetting information critical to a user's goal, and a 67% improvement in completing everyday tasks where unexpected issues arise. On its own Active Memory evaluation the current Muse model uses critical context correctly in 90% of tasks, against 76% for the previous one, and on Meta's Personal Superintelligence Index it scores 47% against 38%.

Separately, Meta launched Spaces in Muse: a user describes an idea in chat and Muse builds a tracker, planner or project board in the sidebar, which can stay up to date from connected apps, TechCrunch reports.

Visa released the Visa Agentic Commerce Benchmark on Thursday 5 November; this paper did not carry it then. In the first round, 22 AI models from 12 labs took on 15 shopping tasks such as buying a TV, a laptop or a personal-care appliance. To pass, a model had to uncover the shopper's unstated requirements through conversation and buy only if something fitted, or decline.

Visa's release, as reported by PYMNTS, says 42.8% of test runs failed. The best models passed 82.2% of their runs and the worst 28.9%. In 54% of failed runs the agent still made a purchase, either one that missed the customer's requirements or one made before it had found them all. The five most expensive models per run passed 67.1% of runs, the five cheapest 50.2%.

Visa says models that passed a scenario often failed when made to repeat it. It plans to add scenarios and extend the tests to checkout, payment and order placement, and the benchmark is open for anyone to try on its Visa Payments Frontier site.

Physical Superintelligence (PSI) said on Friday that a research system it built, running a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture, open since 2008. The conjecture concerns when quantum computers can be exponentially faster than ordinary ones: if it is true, they are at most polynomially faster on random inputs that lack special structure.

PSI says the system had a complete proof during the third day of the run and that Lean, a proof assistant that checks every step by computer, accepted it. The system then improved the result twice, and Lean accepted the final version. PSI has published a paper and the Lean proofs and code on GitHub.

Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan of Stanford, who settled a special case in August, are reading the proof and writing a shorter exposition for experts, PSI says; it quotes them calling it a huge result. Business Insider reports it is not clear which unreleased model PSI used, and Anthropic did not respond to its request for comment.

OpenAI published BioBasicsBench, a benchmark built with the enzyme-design company Ligo that asks AI agents to read raw evidence from routine wet-lab work: gel and blot images, plate photos, microscopy, sequencing reads, instrument exports and bench videos. Its 100 tasks cover protein methods, imaging and cytometry, assays, nucleic acid methods and molecular cloning, and test whether an agent can spot errors and say when the evidence cannot support a conclusion.

Ligo scientists wrote the tasks from their own experiments with grading rubrics, and each task had an internal review, an external review by a PhD-level wet-lab scientist and a final check. OpenAI ran seven of its models in 37 settings, five attempts per task, for 18,500 results.

At the xhigh reasoning setting, GPT-6 Astra earned 66.7% of the available credit, GPT-6.1 Sol 62.0% and GPT-5.6 Sol 43.8%, with Astra using about half Sol's output tokens per task. In one example, Astra noticed that a control plate was labelled for the wrong antibiotic, which GPT-5.6 Sol missed.

OpenAI has closed $11 billion of new funding, the first part of a round that values it at $1.25 trillion, Bloomberg News reported on Friday, citing people familiar with the matter. The money had been wired as of Friday, and the rest of the round, which could reach $20 billion in total, is expected in the coming weeks, one person said. OpenAI declined to comment.

According to Bloomberg, OpenAI first set out to raise $30 billion or more at a $1.4 trillion valuation, then scaled back its plans during this week's sell-off in AI shares. It raised $122 billion in March at an $852 billion valuation.

TechCrunch notes the report came hours after Bloomberg said some prospective buyers of shares in Anthropic's IPO believe its valuation could be as low as $1.1 trillion.

Microsoft is winding down an arrangement to rent AI computing from Oracle, part of a larger reassessment of billions of dollars in cloud partnerships, Bloomberg News reported on Friday, citing people familiar with the matter. Microsoft has been turning off the rented capacity in recent months after declining to renew the contract, the report said.

The capacity became less useful because it no longer fit Microsoft's specifications; the Oracle hardware used older-generation Nvidia chips, according to the people. An Oracle spokesperson told Reuters that demand for its AI infrastructure remains strong, including for earlier-generation GPUs. Microsoft did not immediately respond to Reuters.

Oracle announced the multi-year deal in 2023, saying Microsoft would use its cloud's AI infrastructure for the AI version of Bing; terms were not disclosed. Reuters links the move to other signs that Microsoft is rethinking how it buys capacity, including the IREN delay this paper reported on Friday morning.

Crusoe confirmed on Friday that it has sold its unfinished data-centre project in southeast Wyoming to Google, a day after Google announced a $17 billion investment to build its first Wyoming data centre there. Financial terms were not disclosed, Reuters reports.

Crusoe says it completed the pre-development work, including site selection, government approvals and the power strategy. Google is building the campus with Tallgrass Energy, Black Hills Energy, American Electric Power and Related Digital. TechCrunch reports that Google had raised concerns about the project's cost and timeline under Crusoe; Bloomberg first reported the handover in June.

Crusoe is still developing another campus for Google near Amarillo, Texas. Its chief executive, Chase Lochmiller, said the company has more than 6 gigawatts of contracted capacity, according to Reuters.