Apple finally ships its Siri overhaul

Good day, AI enthusiasts!

It was another busy week in AI, but the interesting part was how varied the big stories were.

Apple finally started putting its rebuilt Siri into users’ hands. Anthropic disclosed that Claude now “leads” 26% of its internal AI research and development work. And AI infrastructure company Crusoe announced the initial closing of a $3.9 billion funding round at a $30.9 billion valuation.

Money and computing capacity were only part of it. The US House voted 417–3 on legislation aimed at preventing the cost of new data-centre power infrastructure from simply landing on other electricity customers, while OpenAI started testing sponsored conversational agents inside ChatGPT.

There were also notable developments in AI safety reporting, enterprise models, voice AI and public data.

Here are 10 AI developments worth knowing from this week:

Apple began rolling out Siri AI in beta on Monday as part of its new generation of Apple Intelligence. The assistant can draw on information from messages, emails and photos, understand what is on screen and take more actions across apps.

Apple is also giving Siri its own app, with conversation history synced privately across devices through iCloud. The system runs across on-device models and Apple’s Private Cloud Compute infrastructure; Apple says some server-based features will eventually offer expanded usage for a fee.

There are still important geographic limitations. Siri AI is initially unavailable on iPhone, iPad and Apple Watch in the EU, and the new Apple Intelligence features are not yet available in China while Apple works through regulatory requirements.

For Apple, this is an unusually important product moment: years of promises about a more capable Siri are turning into an actual consumer release rather than another preview.

Source: Apple — Siri AI: a more capable and personal assistant

Anthropic published one of the more interesting numbers of the week: as of August, Claude was classified as “leading” 26% of the company’s AI research and development work.

Anthropic defines that as completing most of a task end-to-end from a high-level prompt while a human supervises. More than 90% of its AI R&D now involves Claude at least in a collaborative role.

There is an important qualification: Anthropic says Claude is not operating fully autonomously in any measured part of its R&D process. The figures are also based on a methodology Anthropic is applying internally, and the company notes that cross-lab comparisons would need common standards and independent verification.

Still, it is a useful glimpse inside a frontier lab. AI is no longer only the thing being researched; it is becoming a meaningful part of the machinery used to conduct that research.

Source: Anthropic — Measuring the pace of AI development

AI infrastructure company Crusoe announced the initial closing of an anticipated $3.9 billion Series F, valuing the company at $30.9 billion post-money.

The company says it now has more than $140 billion in total contracted value across its infrastructure business. The round was co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Nvidia, GIC, Qatar Investment Authority and others participating.

Crusoe is unusual because it is trying to control much of the AI infrastructure chain itself, from energy and data-centre campuses through to cloud computing.

The funding figure is another reminder of how much capital is now being committed before a user ever sends a prompt. Building the physical capacity behind AI has become a major technology business in its own right.

Source: Crusoe — Series F funding announcement

Cohere and Germany’s Aleph Alpha signed a definitive merger agreement on Wednesday, moving forward with a combination that was first announced earlier this year.

The merged company will operate under the Cohere name with dual headquarters in Toronto and Berlin. Reuters reported that the combination had been valued at roughly $20 billion when originally announced, although updated financial terms were not disclosed this week.

German retail group Schwarz is also putting €500 million into the company and plans major spending on AI compute through its StackIT cloud business.

The interesting part is the positioning. Both companies have concentrated heavily on enterprise and government customers that care about data control, private deployment and regulatory compliance. The deal is therefore as much about competing for sovereign and regulated AI workloads as it is about competing on model benchmarks.

Source: Reuters — Cohere and Aleph Alpha combine

One of this week’s more consequential AI infrastructure stories came from Washington.

The US House approved legislation by 417 votes to 3 requiring state utility regulators to consider a standard under which data centres would be charged for the full cost of new power generation and transmission upgrades needed to serve them.

The proposal is a federal recommendation rather than a nationwide pricing mandate, and states would retain authority over their electricity markets.

That distinction matters, but so does the vote. The rapid construction of AI data centres is turning electricity pricing, grid capacity and infrastructure costs into mainstream AI policy questions — not just engineering problems for cloud companies.

Source: Associated Press — Congress takes up data-centre electricity costs

OpenAI introduced a new framework for tracking, investigating and publicly reporting unexpected or concerning behaviour from its models.

Alongside the framework, the company published six reports covering model behaviour observed during the previous six months. The incidents themselves did not all happen this week; the new development is OpenAI’s decision to formalise how they are disclosed.

OpenAI says its previous reporting was too ad hoc and that the new approach should allow some incidents to be shared before every aspect has been fully explained or mitigated.

As AI systems gain more ability to use tools and complete longer tasks, disclosure standards around things going wrong are becoming almost as important as the benchmarks published when a model launches.

Source: OpenAI — Model misalignment reporting framework

OpenAI expanded its advertising business this week with something more interesting than another banner format: Sponsored Agents.

When someone clicks an eligible ad, they can choose to enter a clearly labelled conversation with an agent sponsored by the advertiser, ask follow-up questions about the product and then continue to the advertiser’s website. The test is initially running with selected US advertisers.

OpenAI also added natural-language campaign creation and analysis through its Ads Manager plugin, AI-assisted ad creative, and integrations with HubSpot and Shopify.

For marketers, the notable experiment is the format itself. Search advertising mostly monetised the click; conversational advertising is testing whether the interaction after the click can happen inside the AI interface too.

Source: OpenAI — Reimagining advertising with AI

Salesforce and Nvidia unveiled Koa, Salesforce’s first CRM-specific reasoning model for Agentforce.

Koa is based on Nvidia’s Nemotron 3 Super and was post-trained using a proprietary synthetic dataset modelled on nearly three decades of Salesforce CRM deployments. It is designed specifically for multi-step business tasks such as working with sales opportunities, service cases and follow-ups.

Salesforce says Koa matches or exceeds leading models on its own CRM benchmark while making three times fewer errors. That result comes from Salesforce’s benchmark, so it should be treated as a company-reported claim rather than independent validation.

The broader move is worth watching. Instead of asking one general-purpose model to understand every enterprise workflow, large software companies increasingly have enough proprietary process knowledge to build models around their own domains.

Source: Salesforce — Koa reasoning model

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its latest models for real-time voice interaction.

The standard Live model is aimed at scale and cost efficiency, while Extended Thinking is designed for more complex multi-step work. Both combine live conversation with visual context and tool use, and Google is positioning them as building blocks for production voice agents.

Google says Extended Thinking scored 82.6 on Artificial Analysis’ Speech-to-Speech Quality Index and posted stronger results on several voice-agent task benchmarks. Those benchmark figures should still be read alongside independent testing rather than as definitive real-world performance.

Voice AI is quietly becoming a serious product category: not just a different way to talk to a chatbot, but an interface for agents that can listen, reason and continue working while the conversation carries on.

Source: Google — Gemini 3.8 Live and Extended Thinking

The UN system and Google launched UN System Data Commons, an open-source platform designed to connect official statistics that previously lived across separate organisations and formats.

The platform is built on Google’s Data Commons technology and allows people to search statistics using natural language. More interestingly, it supports the Model Context Protocol, allowing AI systems to retrieve figures from the underlying data and use them to construct charts, analysis and draft reports.

The UN plans to keep adding sources with a target of incorporating 80% of UN system statistical datasets by 2027.

AI research assistants are only as useful as the information they can reliably access. Making authoritative public datasets machine-readable and directly accessible to AI systems is less glamorous than another chatbot launch, but potentially much more useful.

Source: Google — UN System Data Commons

So what did this week actually tell us?

More than anything, it showed just how broad the AI story has become.

There was no single theme tying everything together — and that was arguably the point. Some of the week’s most interesting developments came from consumer products, others from billions of dollars flowing into infrastructure, increasingly specialised business software, and the growing effort to put clearer rules and safeguards around more capable models.

O T H E R_ A I _N E W S

ChatGPT added ChatGPT directly to Microsoft Word, letting users draft, summarise, revise and format documents from a sidebar inside Word. OpenAI says it is available across all ChatGPT plans, including Free.

OpenAI launched Astra for Law, combining GPT-6 Astra with US case law, statutes, regulations and specialised legal instructions. Harvey and Legora will be able to build on it, while selected law firms get initial access.

Meta launched Meta One subscriptions across its apps, offering higher AI usage and additional creator and business tools across Instagram, Facebook, WhatsApp and Meta AI.

Google expanded its CC AI agent into a shared household assistant, allowing up to six people to share calendars, tasks and selected information with an agent that can help manage schedules, forms, shopping lists and other family logistics.

Superhuman acquired AI meeting assistant Fathom, with plans to connect meeting transcripts, decisions and action items to its wider productivity suite and Superhuman Go assistant.

Profound raised $180 million at a $1.8 billion valuation, giving the AI-search and marketing platform fresh capital less than seven months after its previous funding round. Sequoia and Kleiner Perkins co-led the Series D.

Exein raised $270 million at a $1.7 billion valuation, to expand its security technology for connected devices including robots, vehicles and other machines increasingly running AI.

Cornelis Networks raised $205 million and introduced Active Compute Fabric, an open networking architecture that places programmable compute inside the network in an effort to reduce the amount of time expensive AI accelerators spend waiting for data.

Gates Foundation committed at least $1 billion to AI initiatives over the next two years, focusing particularly on access to AI and AI-enabled tools in health, education and other areas where lower-income communities risk being left behind.

PrismML released Bonsai 2 27B, a 5.9GB ternary version of Qwen3.8 27B. The company says it is more than nine times smaller than its full-precision counterpart while retaining 98.2% of aggregate benchmark performance.

See you soon,

The News Digest AI team