The AI Wire ●

Google, working with Meta's PyTorch team, has released TorchTPU, an open-source native PyTorch backend for Google's Cloud TPUs. It runs eagerly by default, compiles through torch.compile and XLA, and is positioned as the replacement for PyTorch/XLA in new TPU workloads.

The release is part of a wider push to make it easier to run AI models on hardware other than Nvidia's. It landed one day after Amazon announced a similar approach for its Trainium chips, and the same day DeepSeek open-sourced DeepInfer, the inference engine behind its V4.1 Flash model, with Huawei Ascend support built in.

For developers, the practical effect is that PyTorch code can move to TPUs with less porting work, which matters as the labs and cloud providers compete on the cost of running models, not just on training them.

Google began rolling out Personal Intelligence for Gemini in Android Auto on October 20, letting the in-car assistant draw, with permission, on a driver's Gmail, Google Calendar, Google Photos, Maps, YouTube and Search history. Cars with Google Built-in follow in the coming months.

Google's examples are hands-free and natural-language: ask for your flight gate and whether the flight is on time in one question, say 'take me to my doctor's appointment' and let Gemini pull the address from your calendar, or find a car wash from a photo you took last week and navigate to it. The feature is opt-in, and an opt-in already set in the Gemini phone app carries over to the car automatically.

The rollout covers eligible personal accounts for users 18 and older, and excludes the European Economic Area, Switzerland, the United Kingdom and Nigeria. Workspace accounts are left out, and on Google Built-in cars personal answers surface only while the driver is in the car with their phone connected.

ChromeOS 156 has reached the Stable channel (platform 16836, browser 156.0.8078.42, rolled out from October 20), and the release is mostly a Gemini release. On Chromebook Plus, Gemini in Chrome gains media understanding: users can ask questions about audio or video they have already played in a tab and get summaries or specific details back, checkable at chrome://settings/ai/gemini.

The update also adds Gemini Talk to Type dictation, direct file uploads and file generation in Gemini in Chrome, and an interactive Quiz feature that turns a webpage, PDF, slide deck or video into a study session. ChromeOS 156 confirms that users can temporarily share saved Google Password Manager logins with the Gemini Spark agent so it can sign in to third-party web apps, with explicit consent plus a biometric or device unlock required for every task. 'Autofill with AI' is renamed Enhanced autofill and expanded to more regions.

Underneath, the release carries the 353 security fixes from Chrome 156, six of them critical memory-corruption patches, plus two high-severity Mali GPU kernel fixes specific to ChromeOS. Rollout is gradual: Chrome Unboxed updated a Lenovo Chromebook Plus 14 to 156 and had not yet seen Talk to Type appear, so the headline features may lag the OS update itself on a given machine.

TestingCatalog reported on October 18 that Anthropic is preparing a Personal Finance feature for Claude that connects bank accounts through Plaid, based on unreleased app screens it posted that day. The screens promise read-only access — Claude reads your balances and transactions to work with your actual numbers — and show a 'Link with Plaid' button, with unlinking described as stopping Claude reading the account.

This is the second sighting of the feature: in September it surfaced as 'Claude Money' in the Claude iOS app, with a Money section inviting users to link bank accounts and ask Claude about spending and plans. The October screens are a step on from that — the connection layer is now visibly Plaid, the same company whose CTO wrote in May that ChatGPT's personal finance preview runs on Plaid.

Anthropic has not announced the feature, and nothing in the screens or the reporting says which Claude plans would get it, which countries it would launch in, or what Anthropic pays for the connection. Until Anthropic says something itself, this stays rumor-tier: a consumer Claude feature that would put a second major assistant on the same kind of live bank data ChatGPT already reaches.

Update, November 5: confirmed. Anthropic announced Personal Finance the same day — rolling out in beta this week to Pro, Max and Ultra subscribers in the US, linking bank and card accounts via Plaid with read-only access, exactly as the leaked screens described. Claude can see balances, credit limits and up to two years of transactions, cannot move money, and the feature carries no ads; see the November 5 launch entry for the full details.

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, about five weeks after the previous Flash checkpoint, describing it as a Pareto improvement in performance and cost. API pricing is unchanged and developers reach it through the existing deepseek-flash endpoint; weights are also available for download.

On DeepSeek's own DeepSWE v1.1 coding evaluation the new model scores 77.0%, up from 74.2% for the previous version and above the 74.0% the company lists for Claude Opus 5. DeepSeek says shorter outputs lower the effective cost further, with per-task costs often under $0.20-$0.25 in its figures — company-reported numbers that have not been independently verified.

Alongside the model, DeepSeek opened its search and web-page retrieval tools to developers at $0.75 per 1,000 calls, free for a limited time when connected as a server-side feature to designated development tools. The release keeps pressure on the majors on cost per coding task, even as the headline gains are incremental.

SpaceXAI announced a bundle of Grok Bot improvements on October 16. Computer tasks now finish 3x sooner on average on the company's own figures, replies arrive in less than half the previous time, and the app gets a refreshed design, better voice calls and notifications that go only to the device in use.

Grok Bot can now read a connected user's X Chat messages and can be set a task by tagging @bot in a post or reply on X, with the work and results staying in the Bot's own chat. Tagging is not yet available for X accounts based in the UK, the European Union, Australia, Iceland, Liechtenstein or Norway.

New connectors include Robinhood, Coinbase, Trello, PostHog, beehiiv and eToro, and teams can now submit their own connectors to the Grok Bot marketplace. Grok Bot's own email address also gains plus-alias support, so routines can run on mail sent to a tagged address.

Reflection has released the weights for Beam, its 501-billion-parameter mixture-of-experts model, on Hugging Face under an Apache 2.0 licence. Baseten offers day-zero support through its Model APIs and dedicated inference.

Beam was first introduced on October 5 while it finished red-teaming and evaluations, with early access by waitlist. The release includes a model card with final benchmarks and a safety technical report, and Reflection frames the launch as an open ecosystem of more than 40 partners and integrations for agents.

A public beta of Reflection's own API at platform.reflection.ai is free to start. Named launch platforms span the major clouds and inference providers, with Dell and HPE offering on-premises deployment for sovereign and air-gapped environments.

OpenAI said on October 16 that GPT-6 Pro can now split a task among multiple agents working in parallel inside ChatGPT, with the agents' progress visible to the user. The feature is available to Pro users in the Chat tab, according to the company's post.

OpenAI also claims GPT-6 Pro responds faster and uses plugins and tools better in this mode. The post gives no latency figures or worked examples to back those claims, and it does not explain how users control or approve the parallel agents' tool actions.

For Pro subscribers this brings the delegation pattern previously associated with the Work tab and Codex into ordinary Chat conversations: one request can fan out into several simultaneous workstreams, with the user watching each one rather than waiting on a single sequential reply.

Late Friday, October 16, OpenAI updated the ChatGPT app on iOS and iPadOS with a home-screen widget for your Dot and new Shortcuts actions to control it. The small widget simply shows your Dot's avatar and tapping it opens the Dot's message thread — or, if you toggle on "Show shortcut button" when editing it, it adds messaging and voice buttons beneath the avatar.

The update also adds a trio of Shortcuts actions to message, dictate to, or call your Dot, each taking two parameters: the name of your Dot and how you want to communicate with it. Users can currently have only one Dot, so a parameter that opens a list of Dots strongly suggests multiple Dots are on the horizon, something OpenAI telegraphed at DevDay, MacStories notes.

The widget gives Dots — OpenAI's always-on agents, launched in September for a limited set of paid plans — a place on the iPhone home screen, shortening the gap between an ongoing task and checking in on it. It adds no new task capability, and users must initiate the voice calls themselves.

In release notes published October 16, OpenAI introduced ChatGPT Slides for creating and editing presentations inside Space. The feature is rolling out to eligible Pro, Business and Enterprise users with Space access on the web and in the desktop app.

Users can start a presentation from a conversation, files, an outline or an example deck, then edit text, images, shapes and charts directly and ask ChatGPT to make changes. Teammates can edit in real time and leave shared comments, and presentations can run directly in ChatGPT.

Web users can export decks to PowerPoint, PDF or Google Slides, while desktop users can open them in PowerPoint or Keynote where those apps are installed. Slides were first previewed as part of the Space and Pages announcement at DevDay on September 29; this is the feature reaching users.

Google said on October 16 that it has launched three new features in Gemini Notebook to make exam preparation more structured. The tools are available at notebook.google, according to the company's post.

Short Video Overviews generate a 60-second video to unpack a tricky concept. Interactive Learning Overviews, created under Reports, walk through material step by step, weaving together summaries and studio outputs such as slides, quizzes and flashcards grounded in the user's coursework.

Google also expanded quiz formats to include short-answer, multi-select and fill-in-the-blank questions. After testing, users can ask their notebook for feedback on where to focus next.

Elorian, a Palo Alto startup founded by former Google DeepMind researcher Andrew Dai, released its first model, Glimpse, on October 16 as a research preview with API access. Instead of converting an image to tokens once and reasoning from memory, Glimpse keeps returning to the source mid-reasoning — cropping into a region, pulling specific video frames, or running code to measure what it sees, in what Elorian calls a visual scratch space.

On Elorian's own benchmark table, Glimpse beats OpenAI's GPT-6 Astra on scientific chart reasoning (92.90 to 92.10 on CharXiv-R) and short-form video captioning, while trailing Astra on several other tests. It is priced at $0.60 per million input tokens and $3 per million output tokens.

Elorian says a small team trained the model on only a few hundred GPUs on top of an open-weights base. The company disclosed a $55 million seed round at a $300 million valuation earlier this year, co-led by Striker, Menlo and Altimeter, with Nvidia participation.

Proof launched x401 on Thursday, an open protocol that lets any website ask an AI agent to prove who it is acting for and what that person approved. The company says it is the first to put verified legal identity and transaction authorization for AI agents into production.

At the centre is an agent identity wallet holding a person's verified legal identity. The person approves exactly what an agent may do on their own device; the agent carries the request and the wallet signs it, so if the agent changes an amount, recipient or other signed term, the signature no longer validates. The wallet, the x401 protocol and verification tooling are live today.

Proof says it works with assistants that support the Model Context Protocol, naming Meta's Muse, ChatGPT dots, Grok and Claude. The launch follows Amazon blocking Muse from its marketplace and September guidance from FinCEN and federal banking regulators recognising verifiable credentials as a permissible way to verify customer identity.

Perplexity announced on October 15 that it is open-sourcing SPACE — Sandboxed Platform for Agentic Code Execution — under the Apache 2.0 license. SPACE, written in Rust, runs each agent task in its own Firecracker microVM, with credentials and outbound traffic controlled from outside the VM.

Perplexity says SPACE has powered every Computer agent session since July 2026 and handles millions of sandbox creations per day, with median spin-up of about 60 milliseconds and pause, resume and snapshot support. In its published security testing, Perplexity reported zero VM-to-host escapes, while separately finding and fixing network-policy bypasses under partial-access conditions.

Cloud partners are already moving to offer it: Akamai has launched SPACE on its cloud, Crusoe is adding it to Managed AI Services, and CoreWeave has published evaluation guidance. Perplexity positions SPACE as the first piece of a broader Agentic Runtime Platform.

OpenAI's release notes for October 15 say GPT-Live-1.1, and a smaller version called GPT-Live-1.1 Quip, are rolling out in ChatGPT Voice. The update is also reaching Codex and OpenAI's Dots agents, while an API version is still to come.

The headline change is voice choice. New Live voices cover French, German, Hindi, Japanese, Korean and Mexican Spanish, plus a new English voice, and OpenAI says accents across languages sound more natural. Live also varies the short phrases it speaks while searching, so it repeats the same progress cue less often.

For ChatGPT users this is a practical, same-day upgrade to the app's voice mode rather than a new model launch: pick a voice under Settings, then Voice. Community posts on launch day already report some voice-selection quirks in other voice modes on mobile, which OpenAI has not yet addressed in the notes.

Update October 16: OpenAI has added voice customization to the GPT-Live rollout — users can now create custom voices from a written description or from a recording, with a guide published by the OpenAI developers account. Community reaction the same day was mixed, with several users reporting that the existing Sol voice changed noticeably in Standard mode after the update; OpenAI has not commented on those reports.