Platform · 6 October
The personal AI agent gets a login to the high street
Meta and the enterprise-AI company Sierra have published the Personal Agent Protocol, an open, OAuth-based standard for how personal AI agents authenticate with businesses, carry context across channels and act on a user’s behalf. The founding group is a rare coalition of AI and commerce: Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. The design hands consumers the choice of what access to grant an agent, and lets companies set the parameters for what it may do — through their existing websites and APIs, an MCP/OpenAPI interface, or an agent of their own. “Consumers decide what access to give their personal agents, and companies set parameters for what those agents can do,” Sierra says. Version 0.1 ships this month without a payments extension, so the protocol settles identity and permission first and leaves the checkout layer for later.
Source: CNBCThe Verge
Infrastructure · 6 October
Google buys 3.6 GW of power, a quarter of it new nuclear
Google has contracted 3,590 megawatts from Constellation Energy in PJM, the largest US power grid, in a deal the companies say answers the grid operator’s “bring your own power” proposal for data centres. The centrepiece is a 20-year power purchase agreement for 890 MW of newly uprated nuclear capacity across 11 Constellation reactors in Illinois, Pennsylvania and New Jersey, backing more than $4.3bn of investment, with first power due in 2028. Constellation will use Google’s Gemini Enterprise software for site selection, outage management and infrastructure protection. It follows a 20-year nuclear deal of Constellation’s with Amazon in the same week — a reminder that, for AI build-out, power is now the binding constraint.
Source: CNBCGoogle & Constellation
Models · 6 October
Mistral ships Europe’s trillion-parameter answer
Mistral has launched a public preview of Mistral Large 4, a natively multimodal mixture-of-experts model with roughly 1.05 trillion total parameters and 49 billion active per token, nicknamed “Le Chonk”. Mistral says it trained the model from scratch over two months on 4,000 Nvidia Grace Blackwell GPUs in its own European data centres — a sovereignty pitch as much as a technical one. On Artificial Analysis’s Intelligence Index the preview scores 38, the best result yet for a Western open-weights model, though still behind the leading Chinese open models. API pricing is $1.36 per million input tokens and $4.18 per million output; the weights and licence are promised later this month, with reporters told 27 October.
Source: TechCrunchPrimary: Mistral
Security · 6 October
Anthropic opens its most capable models to more defenders
Anthropic has expanded its Cyber Verification Program, folding in its Project Glasswing initiative and offering three tiers of access with reduced safety classifiers to vetted professionals: Defense Access for incident response and vulnerability validation; Red Team Access for authorised penetration-testing organisations; and Specialized Access for safety-critical systems such as flight software and power grids. Every tier includes access to Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models. Anthropic says Glasswing, which put its most cyber-capable models to work hunting flaws in critical software, has surfaced more than 100,000 vulnerabilities this year — its argument for widening access to the defenders who need it.
Source: AnthropicReuters
Economy · 6 October
Nvidia-backed Lambda raises $4bn before a 2027 IPO
Lambda, the Nvidia-backed GPU-cloud provider, is raising up to $4bn in what is expected to be its final private round before a planned 2027 initial public offering, at a pre-money valuation of $14.5bn, with Blackstone and Coatue leading. Its reported backlog of roughly $50bn points to demand for GPU capacity still outstripping supply. The deal joins a run of AI-infrastructure financings: Alphabet’s Waymo this week upsized its debut private debt deal to $5bn to fund robotaxi expansion, arranged by Goldman Sachs and led by Pimco, Blackstone and Sixth Street. For all the talk of an AI slowdown, the capital keeps flowing into chips, power and fleet.
Source: TechCrunchWSJ
Research · 6 October
Google releases a 740-million-parameter multimodal embedder
Google DeepMind has released EmbeddingGemma 2, an open embedding model that maps text (including code), images, video and audio into a single 768-dimensional vector space, so one search can span every medium. It is built on the Gemma 4 architecture and released under the permissive Apache 2.0 licence. The design is deliberately modular: a 270M text-and-code backbone that scales to 740M by loading a 170M vision encoder and a 300M audio encoder, letting developers fit it to on-device budgets. At that size it targets phones and laptops rather than data centres, and it is already available on Hugging Face and Ollama — a small, practical counterpoint to this week’s trillion-parameter releases.
Source: GoogleHugging Face
Society · 6 October
LibreOffice makes ‘no AI’ a selling point
The Document Foundation, the nonprofit behind the open-source LibreOffice suite, says it “will not add” AI to its software for the foreseeable future — and is framing the absence as a feature. The editor “contains no generative AI features”, it says, a “deliberate design position” so that users’ documents are never uploaded for processing and no network connection is required to work. “Any organisation handling confidential, legally privileged or personal data needs to know where that data goes, and the only assurance that survives an audit is that it does not leave the machine,” it wrote. Users who want AI can still install extensions that connect to local models. It is a striking counterpoint to software that packs in assistants by default.
Source: TechCrunchDocument Foundation