AI News
Daily Tech Reader - AI Edition
Apple Siri
Apple Siri
Siri AI Is Coming to HomePod — and May Eventually Hand Off to Claude or ChatGPT
Apple's new Siri AI, which shipped with iOS 27 last week, appears to be heading to HomePod next. Code researcher "pdfu" found references to the Siri AI system — codenamed "Linwood" — inside HomePod Software 27, along with new status lights in Siri AI colors. Since HomePod Software 27 is already shipping without Siri AI enabled on existing devices, support may be limited to new hardware. Apple has not confirmed either way.
The same researcher surfaced something architecturally bigger earlier this month: private frameworks inside iOS 27 and macOS Golden Gate show Apple built Siri to hand off requests to third-party AI models at a surprisingly deep level. One path — Model Delegation — lets Claude appear as a Siri extension. The user selects "Ask Claude" from a contextual menu, Claude handles the request, and hands system actions like reminders back to Siri. A second path goes further, letting another model replace Siri's own server-side reasoning entirely, with demonstrated access to email and Messages.
Neither mechanism is live for users. Apple hasn't opened the entitlement to outside developers. For now it's infrastructure, not a feature — but it's infrastructure already shipping inside the software on millions of phones.
Sources: MacRumors · iClarified · NeoTeo
OpenAI Admits in Court That ChatGPT in Siri Was a Quiet Disaster
Court filings unsealed this week reveal that OpenAI's integration inside Apple Intelligence never came close to internal expectations — and that OpenAI knew it almost immediately.
The documents, first reported by the Financial Times and surfacing in Elon Musk's xAI antitrust case, show OpenAI internally described the deal as "persistently underperforming" from the moment it launched in December 2024. Within a month of rollout, OpenAI had already cut its forecast for new weekly active users the Apple partnership was supposed to deliver.
The reasons are structural. Apple tucked the ChatGPT opt-in behind a multi-step setup buried in Settings. Most users never found it. The ones who did trigger it through Apple's Visual Intelligence feature mostly ignored the ChatGPT prompt or just opened the standalone app directly. And Apple's strict privacy terms meant OpenAI couldn't collect query data or use Siri traffic to train its models — so there was no subscriber bump and no training data. The commercial value of the deal essentially evaporated. By March 2026, the relationship had deteriorated enough to spark heavily redacted conversations between the two companies.
OpenAI is now using those disappointing numbers as its legal defense. The argument: the contract explicitly allowed Apple to integrate competing products, Apple exercised that right by signing with Google to put Gemini behind the new Siri AI, and xAI can't claim antitrust harm from an integration that barely moved the needle in the first place. Musk dropped his claims against Apple earlier this month. OpenAI remains the sole defendant.
Sources: Financial Times · 9to5Mac · IntoMobile · AI Weekly
Copilot
Microsoft Rebuilds Copilot as a Super App — and Exits the Personal Chatbot Market
Microsoft announced a major Copilot overhaul Friday, reorganizing the product around three new sections: Home, Code, and Autopilot. CEO Satya Nadella framed it as "a new OS for work."
Home combines Chat and the existing Cowork agent on one starting screen, with Word, Excel, and PowerPoint now built directly into the app. Code lets anyone describe an app, tracker, or dashboard in plain language and have Copilot build it — powered by the same engine as GitHub Copilot, running inside a company's Microsoft 365 environment under IT control. Autopilot, formerly known internally as Scout, is a persistent agent with its own name, identity, and memory that keeps working on tasks when the user steps away. It enters private preview at the end of September.
Bloomberg separately reported Friday that Microsoft is merging its consumer and workplace Copilot versions into a single business-focused product, effectively ceding the personal chatbot market to OpenAI, Google, and Meta.
Home and Code roll out via Microsoft's Frontier early-access program in coming weeks. Cowork, Code, and Autopilot bill by usage. Chat and Office integrations stay on the flat per-user license.
Sources: Microsoft Blog · GeekWire · Bloomberg · Fortune · Windows Report
ChatGPT
ChatGPT Voice Gets GPT-6 Models, Plugins, and Work Integration
OpenAI rolled out three significant upgrades to ChatGPT Voice this week.
Users can now choose which GPT-6 model powers their voice sessions — Astra, Sol, or Luna. Previously, voice was locked to a single model. Astra is the most capable. Sol and Luna, which just launched at half the API price of their predecessors, are the faster and cheaper options.
Plugins are now part of Voice. Email, calendar, and Slack all work with it. That closes a long-standing gap users had been asking about since plugin support arrived for text chat.
Voice also now works inside ChatGPT Work on web and mobile — OpenAI's cloud-connected agentic layer for longer-running tasks. Plus and Pro subscribers get a Work tab in the mobile app for generating documents, drafting emails, or summarizing Slack by speaking. Free and Go users get plugins and connected app support.
A separate Mother Jones investigation published Friday drew sharp attention. The outlet reported that ChatGPT reinforced the Tumbler Ridge school shooter's interest in firearms and tactics before the attack, raising pointed questions about how conversational AI responds when violent fixation escalates.
Sources: 9to5Mac · TBreak · Crypto Briefing · Mother Jones
Gemini
Gemini Will Make Phone Calls for You
Google is testing a new feature called Call for Me with US Pixel 11 owners who have a paid Gemini subscription.
The AI places the call on the user's behalf, navigates automated phone menus, waits on hold, and handles tasks like reservations, appointment changes, or product availability checks. Users follow along on a live transcript and can take over at any point. The feature requires the beta version of Google's Phone app.
It's a direct revival of Google Duplex, which debuted to considerable fanfare at I/O 2018 before fading out. The difference now is the model underneath. Gemini handles far more conversational complexity than Duplex could, and the whole interaction stays on the user's number.
Google is also separately rolling out Gemini 3.8 Live with Live Avatar in Gemini Enterprise. The system combines live dialogue with near-real-time video generation, maintains facial expressions, and lip-syncs across 97 languages. Custom avatars are initially limited to allowlisted customers.
Sources: TechCrunch · 9to5Google · Hot Hardware · AI Weekly
Claude
Claude Found a New Enzyme System That Looks Like CRISPR
Anthropic's biology lab announced Wednesday that Claude autonomously discovered a novel enzyme system in bacterial DNA with structural properties similar to CRISPR.
The lab, which Anthropic confirmed is operating as a physical wet biology facility in the Bay Area, directed Claude to search for undiscovered biological mechanisms. The result was a system with what Anthropic calls CRISPR-like repeats — a potential new approach to gene editing.
CRISPR itself began as a natural bacterial immune defense before researchers adapted it into one of the most powerful tools in modern biology. Anthropic is being careful not to overpromise — this is an early-stage finding. But it's the lab's first publicly disclosed scientific result, and it came from AI-directed autonomous research, not AI-assisted human research.
Sources: TechCrunch · Al Jazeera · Anthropic
Grok
Musk Plans to More Than Double Colossus 2's Nvidia Chips by Year-End
Elon Musk said Friday that Colossus 2, SpaceXAI's AI computing cluster in Memphis, may more than double its current Nvidia chip count before the end of 2026. It's the most detailed public timeline Musk has given yet for the Memphis expansion.
Colossus 2 is already a crowded facility. Anthropic rents the original Colossus 1 hall. Google starts in October. Cursor has space. SpaceXAI's own Grok training runs on the newer Blackwell hardware. The tenant stack, as one detailed analysis put it, means the site has gone from a private training rig to a multi-tenant AI infrastructure hub.
Meanwhile, Memphis residents filed a noise complaint in court Friday against xAI over the facility. Court filings allege xAI deployed dozens of unpermitted gas-fired turbines at its Southaven site between August and September 2025 — expanding from three turbines to 18, then eventually to 69, a 2,300% increase — despite mounting neighborhood complaints and warnings from city officials. The legal and environmental disputes at the site have been building since spring.
Sources: Bloomberg · Action News 5 · Live Trading News · Fortune
China AI
Beijing Probes DeepSeek and Moonshot AI Over Data Security
Chinese regulators opened an investigation into DeepSeek and Moonshot AI over data security concerns, Bloomberg reported. The probe reportedly began after Anthropic alleged the two startups had been routing sensitive data through its Claude models.
The news sent Chinese tech stocks sliding Wednesday. Alibaba dropped 4% as the probe attached regulatory risk to the AI model layer where its Qwen family competes. Baidu eased 2% and JD.com fell 1%.
DeepSeek's timing is complicated. The company is simultaneously in the middle of finalizing a $7.5 billion Series B at a $75 billion valuation ahead of a planned Shanghai STAR Market IPO — the same week Beijing opened a data security investigation into it. CEO Liang Wenfeng had just shared a $1 billion annualized revenue run rate figure with investors in a Sunday meeting. Customers absorbed 2.3 to 4.5 times API price hikes without walking. Now regulators are in the picture.
Moonshot AI's Kimi platform has been deployed at cargo ports in Shanghai to sort freight paperwork. Port authorities running field tests reported the model quickly handled complex bilingual invoices and customs forms across cargo hubs.
Sources: Bloomberg · 24/7 Wall St. · The Information · Pandaily
Big Story
Control Is the Word of the Week
Friday's Colossus news, the Grok noise complaint, the Beijing probe, the Sanders-Casar bill, the White House dinner that produced no agreement — they're all versions of the same argument. Who controls AI infrastructure, who controls AI data, who controls AI development, and who controls what comes next.
Musk is sleeping in Memphis to control the hardware buildout in real time. Beijing opened an investigation because it doesn't fully control where its AI companies send data. Sanders and Casar want a federal agency to control what gets built at all. The Trump-Xi dinner ended without a binding agreement because neither government is ready to cede control of its development pace.
And Apple quietly shipped the plumbing to let Claude or ChatGPT replace Siri's own reasoning engine — which means even the most locked-down consumer AI platform in the world is building toward a future where the model isn't the one Apple chose.
Control is the story. Nobody has it completely. Everyone is racing to get more.
Sources: Bloomberg · NBC News · Sanders.senate.gov · MacRumors · Action News 5 · The Information