The AI Wire ●

OpenAI announced on October 20 that it has acquired Orchestra Bio, a four-year-old biotech software startup whose platform helps drugmakers coordinate the work of developing a medicine. Financial terms were not disclosed; Orchestra had raised roughly $20 million from backers including Andreessen Horowitz, Atria and 8VC.

Orchestra's AI agents handle the operational layer around the science: monitoring studies, tracking timelines and dependencies, and working out which activities are affected when something slips, such as a failed manufacturing batch delaying a clinical trial. Program leaders keep the decisions; the agents pull together the information and compare responses.

OpenAI is folding Orchestra's team and technology into Rosalind, its life-sciences AI platform, and the Rosalind Workbench tool for planning experiments and analyzing results. OpenAI says Rosalind already counts Moderna and Novo Nordisk as customers, and the deal is another example of the big labs building industry-specific software rather than only selling models.

Nvidia's GEAR robotics lab open-sourced Flora on October 20: three models that turn robot and human manipulation video into the labels robot AI needs. Flora-TempSeg (9B) splits a video into action intervals, Flora-Captioner (27B) writes one short instruction for each interval, and Flora-Rewarder (27B) scores clips for outcome and process rewards, including against rubrics a user writes. All three are fine-tuned from Alibaba's Qwen3.5 models, with weights on Hugging Face under an Nvidia non-commercial research license.

Nvidia says Flora is the data engine used to train its GR00T humanoid models, labeling up to 200,000 hours of video a day across 512 GB300 chips. Its headline comparison is cost: about $0.72 to label an hour of video, against about $88.49 for GPT-6 Astra at maximum effort — roughly 125 times cheaper — with a 57.1 percent overall annotation score for Flora against 53.7 percent for Astra in Nvidia's own 50-episode evaluation.

Both figures need their caveats. The cost gap rests on Nvidia's modeled GPU pricing rather than measured production bills, the quality evaluation was run by Nvidia itself, and the results vary by task: Flora led on human videos and action boundaries, while Astra led on robot videos, caption accuracy and rubric scoring.

Mistral AI released Voxtral 3 on October 20, a family of speech-to-text and text-to-speech models plus a voice agent SDK that combines listening, reasoning, tool calling and speaking in one real-time streaming pipeline. The models are available now through Mistral's API and can also be self-deployed on a company's own infrastructure.

On transcription, Voxtral Small Transcribe 3 supports 50 languages with speaker diarization and context biasing at $0.0040 per minute, while a smaller Mini Transcribe 3 starts at $0.0035 per minute in batch and a realtime version built for agents costs $0.008 per minute. For speech output, Voxtral Speech 3 covers more than 35 languages and 50 dialects, adjusting emotion and pacing, at $13 per million characters.

A Pro Voice feature adapts a speaker's voice and style across languages from 30 minutes of audio. Mistral says the transcription models outperform Whisper, Scribe v2 and Gemini Transcribe on its tests — company-reported claims, but the pricing puts Voxtral 3 squarely in the running for call-center and service-desk voice agents.

Meta has suspended the conversation-sharing feature in its Muse AI agent after security firm Claroc showed that shared Muse conversations could be found by anyone — not just the people they were sent to. Claroc calls the exposed pages GhostShares.

Muse shares a conversation as a web page whose address is built from the conversation's own title plus a random identifier. Those hostnames are recorded by passive DNS services, so a single search of a commercial passive-DNS database returned more than 6,800 shared Muse links, listed by title like a table of contents. In a sample of several hundred pages, most were harmless, but some held a small-business loan application, a bank fraud alert with partial account details, an internal company revenue review, a staff vacation calendar and confidential prep for a call with a frontier AI lab leader.

Claroc reported the flaw through Meta's bug bounty program on October 8 and published its research on October 18. After Ars Technica asked Meta about it, Meta said on October 20 that it had paused the functionality while it reviews the report; Claroc rechecked 40 previously open shares that day and none opened. The hostnames, titles included, remain in passive-DNS history, and Claroc says it has no evidence the information was misused.

TestingCatalog reported on October 20 that Muse Confidential VM, Meta's planned fully-private mode for its Muse agent, is now partially implemented in the web interface. A dedicated settings section appears to let users enable it, and the notice says conversations would run in a secure, encrypted environment that keeps data private even from Meta, with data in this mode not used to train Meta's models.

That is a step beyond the existing Muse Secure VM: Confidential VM is designed so the whole VM, including a person's data and conversations, is encrypted with a key only the user holds. Meta confirmed in September that it plans to introduce the technology before the end of 2026.

The private mode looks limited at first: voice mode, image and video generation, and wearables such as Meta AI glasses and the Muse pin may not be available, and access appears restricted to Meta AI subscribers, with upgrade prompts before users can turn it on. No release date is confirmed, so this stays a rumor until Meta ships it.

Grok Bot gained another batch of abilities on October 20, in a "What's new in Grok Bot" walkthrough posted on X. The agent can now read and reply to your X Chat direct messages, respond when you tag @bot in an X thread, and be set as the default digital assistant on Android phones.

The DM and tagging features need the X plugin installed and configured, and Palmer said limits are in place so the Bot cannot spam people. Voice calls also now keep the previous call and its full transcript if you call back within 30 minutes, alongside new chat cards, email aliases, connectors and team invites.

The Android move lands a day after Grok Bot gained one-press access through the iPhone's Action Button, which starts a voice chat with the Bot. Together the two put SpaceXAI's always-on agent one press away on both major phone platforms.

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 announced new Gemini tools for small businesses on October 20, headlined by a dedicated business landing page inside the Gemini app. From the hub, owners can build their brand, analyze finances and review customer insights, and connect Gemini to Google Business Profile, Google Ads and Google Workspace, plus third-party apps including Adobe, Canva, Dropbox, Shopify, Wix, Zoho and Xero.

Businesses can also define their brand style in Gemini by uploading a website, brand kit, fonts or color palettes. Paired with Nano Banana 2.1, Google's latest image model, Gemini then generates marketing visuals that keep that style, instead of owners re-uploading their guidelines for every asset.

Google says tens of millions of small business owners already use Gemini to brainstorm products, manage customer feedback and analyze performance. The hub is the company's latest push to make Gemini a working layer for running a small business, not just a chat window.

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.

Google Public Sector unveiled Gemini for government services at its summit on October 20: an add-on that gives federal, state and local agencies specialized skills, AI agents and connectors to the software they already run, including ServiceNow, Salesforce, Google Workspace, Outlook, SharePoint and Oracle applications. Google says the tools can interpret dense regulations, redact documents for Freedom of Information Act requests, rewrite jargon-heavy text in plain language and review procurement bids, with answers linked back to the exact source regulation or file.

A few agencies, including the Department of Health and Human Services and the State of Missouri, have been testing it over the past few weeks. Google is now letting more agencies into the early-access phase to gather feedback and expects general availability later this year, with no set date; IT leaders will choose which connections to enable and for which users.

Agencies already have discounted access to Gemini at $0.47 a year per agency until November 15 under the extended OneGov deal, Nextgov/FCW notes, though Google has not published a price for the new add-on itself.

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.

Anthropic's confidential draft IPO prospectus, viewed by The Wall Street Journal and reported October 20, shows revenue of $16.3 billion in the first half of 2026, up from $1.1 billion in the same period a year earlier. The filing points to an IPO that could value the Claude maker at up to $2 trillion.

The headline loss is $80.7 billion for the half, but roughly $72 billion of it is non-cash accounting charges tied to changes in the value of convertible notes and other financing instruments used to raise early capital, not cash spent running the business. Anthropic lost $42 billion in all of 2025 on the same basis.

The prospectus also lays out the scale of the build-out behind the growth: more than $600 billion in future cloud and computing commitments, with negotiations that could add another $150 billion, alongside a long list of AI risk factors. It is the clearest public look yet at the finances of OpenAI's biggest rival ahead of its market debut.

Anthropic says Claude ran a 64-hour protein-design campaign, from one initial prompt and four follow-ups, aimed at five scorpion neurotoxins from Mexico, Brazil, North Africa and the Middle East. Working mostly autonomously in Claude Science, the model used 12 design tools to generate 318,484 designs and screen 553,052 design-toxin pairs, on roughly 10,000 GPU hours.

Outside labs at Adaptyv Bio and Twist Bioscience then tested 211 selected designs. Seventeen met Anthropic's strict binding criteria, all against single toxins — the two strongest, Cn2-B1 and Cn2-B2, bound in the single-digit nanomolar range — while every multi-toxin binder attempt failed.

Anthropic is explicit about the limit: it has not confirmed whether the binders neutralize venom in cells or animals, tested them only against recombinant toxins, and does not claim to have produced an antivenom. The company is releasing all 211 designs and the binding data for other researchers to use.

Anthropic's Frontier Red Team said on Tuesday that Claude flagged 493 attack-surface issues across 25 open-source projects totaling more than five million lines of code. Rather than hunting individual bugs, Claude mapped which code an attacker can actually reach, then helped gate whatever a deployment does not need to expose.

Claude inventoried more than 17,000 configuration options, 3,842 of them security-relevant, and found 162 risky combinations — in 94 cases, combining options silently turned off a security control. All 11 login-gated projects had code reachable before authentication that did not need to be there, and one conflicting pair of options that let a certificate-only server accept anyone had survived for 21 years.

The sharpest case is OpenSSH: Claude traced about 4,500 lines of experimental ML-DSA-44 signature code reachable before login, about 16% of the default pre-authentication surface. Fixes in OpenSSH 10.6 removed roughly a fifth of the code an unauthenticated client could previously reach. Anthropic is explicit that the method does not remove unknown vulnerabilities — it makes them harder to reach.

Zillow announced on October 19 that a Zillow experience is now available inside Anthropic's Claude through a new connector. Users connect it from Claude's Customize menu under Connectors, then can search homes for sale or rent inside the chat.

Once connected, Claude can surface Zillow listings with photos, prices and interactive maps, along with home value estimates and affordability tools. Example prompts include three-bedroom homes in Denver under $600,000 with a garage, or Seattle rentals under $2,500 a month.

The flow is built to move past information into action: users can request a tour, connect with a real estate agent or explore financing without leaving the conversation. Zillow says it does not send data to Anthropic, and that listings appear in Zillow's own MLS-approved presentation with the same broker attribution used on Zillow.com.