AI's business value increasingly depends on what happens around the model: reliable infrastructure, useful information and controlled access to real systems. Today's developments range from UK-backed scientific research to enterprise tools that promise cheaper retrieval and more capable agents.
UK Sovereign AI fund backs Potential Sciences with funding and compute
London-based Potential Sciences has secured backing from the UK's Sovereign AI fund and Stripe chief executive Patrick Collison, Tech.eu reports. The raise is in the tens of millions of dollars, with the government fund also providing compute capacity.
The company combines AI and quantum chemistry to model molecular behaviour. It says its physics-led approach could make calculations for chemical design 100 times cheaper, a company claim rather than an independently established saving.
Our take: This is a useful example of sovereign AI investment beyond chatbots. For UK scientific businesses, access to computing resources and specialist models may matter more than another general-purpose assistant. Evaluate the cost of a validated research result, not just the cost of generating a prediction.
METR finds faster AI research progress at Anthropic, but flags major uncertainty
Independent evaluator METR published an assessment of AI research and development at Anthropic on 5 November. It estimates that scores on Anthropic's internal capabilities index have grown 1.9 times faster since October 2025, with a 95% confidence interval of 1.5 to 2.8 times.
METR believes AI assistance accounts for a majority of the acceleration, but stresses uncertainty over causes and wider applicability. Anthropic redacted much of the evidence METR considered most informative, limiting public scrutiny.
Our take: This is evidence of acceleration inside a frontier laboratory, not a promise that every office becomes twice as productive. The transferable lesson is to measure a complete process over time and distinguish assisted work from autonomous delivery. The evidence limitations belong beside the headline number.
Box brings Nvidia-powered retrieval to enterprise agents
Box announced Content Retrieval, using Nvidia's Nemotron embedding technology to help agents retrieve relevant information rather than read entire documents. TechCrunch reports claims of four-times-faster answers, 20% fewer errors and 90% fewer tokens.
The feature is due to roll out to Enterprise Advance customers early next year. Its operation depends on Box's metadata and permission controls as well as the underlying models, so the headline efficiency figures should not be treated as universal results.
Our take: Better retrieval can be a more practical investment than buying a larger model. UK organisations should test whether an assistant finds the right clause, keeps source references and respects document permissions. Measure retrieval accuracy and full task cost on your own material before accepting vendor savings claims.
OpenAI pledges to cover power and infrastructure costs at future owned campuses
OpenAI published a National Community Compact and a Compute Campus Policy Blueprint. For future campuses it owns and operates, it promises to pay electricity and associated infrastructure costs, provide financial assurances and report on water, energy, noise and local hiring.
It also commits not to use non-disclosure agreements with public officials except for narrowly protected information. These are company commitments and policy proposals in the US, not new UK regulations or guarantees covering every partner facility.
Our take: Infrastructure accountability is becoming part of AI procurement. A UK buyer need not negotiate a power station to ask useful questions: where is the service hosted, what constrains delivery and who bears cost increases? Check which facilities and contracts a supplier's public promises actually cover.
Sierra acquires Anon to help agents work behind website logins
Sierra announced its acquisition of Anon, which builds authentication, credential-handling and browser infrastructure for AI agents. The technology is intended to help Sierra's agents navigate portals and complete actions in systems without APIs.
Anon says it raised more than $13 million and served hundreds of customers. Its announcement highlights multi-factor authentication, recovery from failures and sensitive-data handling as core engineering challenges.
Our take: The difficult part of business automation is often getting safely into the right system, not writing the next sentence. Before allowing an agent into a supplier or customer portal, define its identity, permissions, audit record and approval boundaries. Browser automation should not become a reason to share unrestricted staff accounts.
Thinking Machines launches Inkling 2 Small, with open weights planned next week
Thinking Machines Lab introduced Inkling 2 Small on 5 November, a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters. It supports image inputs and is available through Tinker for use and customisation.
The company intends to release full weights under Apache 2.0 next week after external safety testing and red-teaming. At launch the weights remain closed, so planned openness should not be confused with an already downloadable release.
Our take: Another model option is useful, but switching costs still count. UK teams considering customisation should compare measured task performance, hosting requirements, licensing availability and maintenance effort. Put a scheduled weights release on the watchlist rather than building a production dependency on a promise.
Five9 agrees $325 million cash deal for Level AI
Five9 has entered a definitive agreement to acquire customer-experience AI platform Level AI for approximately $325 million in cash, according to its announcement reported by Investing.com. The transaction is expected to close in 2026, subject to closing conditions.
The deal adds to the consolidation around AI-enabled customer service. An agreement to acquire is not a completed integration, and buyers should distinguish the commercial announcement from features available in their current contracts.
Our take: For contact-centre buyers, platform roadmaps and data handling now deserve as much attention as chatbot demonstrations. Ask which capabilities are available today, how quality is checked and what happens to recordings and customer information after an acquisition. Do not budget savings against features that have not shipped.