The AI Wire ●

PrismML, the Caltech-founded AI lab advised by UC Berkeley's Ion Stoica, has built a version of its tiny language models for smart glasses running Qualcomm's Snapdragon chips. At Qualcomm's Snapdragon Summit on Wednesday, the chipmaker showcased PrismML's 1-bit Bonsai LLM running locally on AI smart glasses built on the Snapdragon AR1 Gen 1 platform.

The smart-glasses version is a 2-billion-parameter model tuned for vision and language, so wearers can ask what they're looking at in real time — with the model running on the device rather than in the cloud. PrismML's claim to fame is shrinking larger models (in this case about 4x) while retaining most of their performance on standard benchmarks.

No smart glasses running PrismML have been announced yet, so this is a platform play rather than a product launch. But it plants PrismML's open-weight, on-device vision squarely in the wearables race — just as Meta's own Connect announcements push its Muse agent onto its expanding glasses lineup.

Any production workflow calling OpenAI's Videos API or a Sora model alias stopped working on September 24, according to the company's deprecations page. The shutdown notice included no migration path and no replacement on the developer roadmap, leaving teams with active integrations to repoint at third-party video APIs like Runway's Gen-4 Turbo or Kling v2 Pro.

The consumer story behind the shutdown is thin returns on heavy compute: Sora's apps reached about 3.3 million downloads by November 2025 but fell to 1.1 million by February 2026, with roughly $2.1 million in total in-app revenue. A reported $1 billion Disney licensing deal collapsed alongside the shutdown announcement.

Inside OpenAI, Sora survives only as an internal research project focused on world models, not a consumer or developer product. Reporting around the shutdown says OpenAI is redirecting the freed compute toward enterprise coding tools and a ChatGPT super app — video generation simply did not make the cut.

OpenAI has hired Patreon co-founder Sam Yam to lead a new Creator Product division, The Information reported, with Yam bringing Patreon's former head of product Drew Rowny and former head of engineering Shannon Ma along. Yam confirmed the move on X, saying the team will 'build together with Creators at OpenAI' and share early access to a new set of tools, pointing to OpenAI DevDay on September 29.

The hire is a signal about where OpenAI thinks its creator story has been weak: its tools have been about output — ChatGPT writes, DALL-E draws — while none have touched the part Patreon owns, the fan who pays every month. Patreon says more than 300,000 creators use its platform, over 10 million fans pay for memberships monthly, and creators have earned more than $10 billion since 2013.

OpenAI has not said what the tools are or whether creators will be able to earn money through them. The move lands as the creator layer of AI becomes a real battleground — and with DevDay five days away, details should surface quickly.

OpenAI announced it will give the Ukrainian government free access to its Daybreak program, the company's AI toolkit for cybersecurity defense. The move brings advanced AI models built to spot software weaknesses, investigate odd network behavior, and test repairs before attackers can exploit them directly to a country under sustained digital attack.

The announcement came on the sidelines of the UN General Assembly in New York, delivered jointly by Ukraine's Consul General in San Francisco, Dmytro Kushneruk, and Sasha Baker, OpenAI's head of national security policy. "Ukraine is already on the front line, and its defenders need support now," Baker said. George Osborne, OpenAI's Head of Countries, framed the decision in stark terms: protecting civilian infrastructure means defending it against both physical and digital attacks.

Ukraine has endured nearly 6,000 cyber incidents in 2025 alone, targeting power plants, hospitals, and telecom networks. The move is notable on two levels: it's the first time OpenAI has handed a nation at war its frontier defensive cyber tools, and it extends Daybreak's public profile just as the company's agents have themselves been under scrutiny for unauthorized system access.

OpenAI said Wednesday it is bringing the agentic Work features it previously introduced on desktop to the ChatGPT mobile app. Plus and Pro subscribers will be able to use the Work tab on their phones to create documents, draft emails and summarize Slack messages, and can also build sites, create presentations, use the cloud browser and access areas such as finances in ChatGPT.

Free and Go users get access to plugins and connected apps on mobile, extending the desktop rollout that OpenAI announced in July with the conversational model GPT-Live and later integrated with the desktop app's Work and Codex tabs.

Voice conversations on mobile also get richer text output, and users can switch between text and voice mid-conversation or start a chat on mobile and resume it on desktop. The timing extends Anthropic's recent mobile-to-desktop handoff improvements — with one difference in philosophy: OpenAI is keeping chat and workspaces separate while Anthropic merged Cowork and Chat.

OpenAI internally described its ChatGPT integration with Apple's Siri as 'persistently underperforming,' according to court documents filed in its ongoing legal fight with Elon Musk's SpaceXAI. MacRumors reported the documents, which surfaced on Wednesday.

Apple added ChatGPT to Siri in December 2024, but users had to go through a multistep opt-in process that added a layer of friction. By January 2025, OpenAI said the integration was 'off to a slow start' and cut its forecast for incremental users.

The disclosure is awkward for both sides: Apple partnered with OpenAI to fill a gap in Siri's capabilities, while OpenAI wanted the exposure to hundreds of millions of iPhone users. It also hands Musk's team a useful exhibit as they argue SpaceXAI was treated unfairly in the AI race.

Mark Zuckerberg used Meta Connect to outline how the company plans to make money from Muse, its fast-growing personal AI agent. The plan: keep Muse free for 'a large number of tokens' and eventually take a small fee from transactions completed through the agent — a retail take-rate rather than a subscription play.

The commerce rails are already going in. Muse supports PayPal and has retail integrations with Walmart, Best Buy, Gap, Sephora, Wayfair and American Eagle, and it's expanding into travel via Expedia and grocery via Instacart. The fourth quarter — holiday shopping season — is the proving ground.

The pitch to developers is that Muse's mainstream, non-technical audience plus curated connectors makes it the consumer agentic-commerce front end. The open question, raised even on Meta's own engineering posts, is whether consumers will trust an AI agent with authority to complete payments — and how Meta prices the trust premium.

At Meta Connect on Wednesday, Mark Zuckerberg made Meta's personal AI agent Muse the centerpiece of the company's 'personal superintelligence' push. The headline hardware was Muse Charm, a small keychain-sized device with a fingerprint sensor tap that lets users talk to Muse without unlocking a phone or opening an app. Meta says it packed the full Muse experience, including real-time voice and the avatar stack, into the device and plans to ship it in December, in time for the holidays. Pricing was not announced.

Muse itself gained a live video-chat mode with customizable real-time voice and a Realtime Avatar feature that animates a reference image during conversation, keeping voice, lip movement and expression synchronized. Alexandr Wang demonstrated the feature on X, a day after Meta's research team published the technical details. Separately, Meta announced it is bringing Muse to its AI glasses, where a spoken wake word will summon the agent and it can act on what the wearer is looking at.

Meta also expanded its wearables lineup with new Ray-Ban Meta Audio glasses (with 100+ style options) and announced Meta VR Glasses, a $1,299 headset with the processor and battery moved into a separate puck, planned for spring 2027. Retail integrations for Muse's shopping actions now include Walmart, Sephora, Best Buy and Gap — though Amazon has reportedly restricted the agent's purchasing capabilities.

Google began a major expansion of Gemini's Connected Apps, letting users plug third-party services directly into their conversations with the AI chatbot. The rollout brings new partner integrations across three categories: productivity (Airtable, Linear, monday.com, PandaDoc, Wispr AI, Zoho), creativity (Adobe, Picsart, Squarespace, Webflow), and lifestyle (Peloton, apartments.com, Experian, SeatGeek).

The model is conversational task completion: connect an app in Gemini settings, then invoke it through an @mention or a direct prompt. Google's concrete example is planning Peloton workouts inside chat — the kind of everyday consumer task that makes Gemini feel like an agent, not just a chatbot. The creative integrations are framed carefully: Adobe for asset creation, Squarespace for website design.

This follows the earlier integration of iHeartRadio and Ticketmaster and is the clearest consumer-side answer yet to rivals turning chatbots into doers: Meta's Muse works via WhatsApp, and OpenAI has been pushing ChatGPT agents into real-world tasks. The question is whether users actually connect their accounts — or leave the apps unlinked and the feature theoretical.

UPDATE 2026-09-24: Adobe detailed its side of the Gemini integration — Photoshop, Lightroom, Adobe Express, and Firefly tools now work inside Gemini conversations on every Gemini plan via an @Adobe mention, orchestrated so a single description kicks off the right multi-step workflow. Adobe also folded Acrobat capabilities into its existing Claude plugin. Setup needs only a Google Gemini account plus a free or paid Adobe login.

Google rolled out two new text-to-speech models in the Gemini family: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. They extend Google's audio generation lineup across its developer platforms and AI products, joining Gemini 3.5 Live Translate, Gemini 3.5 Transcribe, and Gemini 3.8 Live.

The split is deliberate. Flash TTS focuses on voice creation and character design — users can design voices from scratch with natural-language prompts, customizing role, accent, and voice characteristics across more than 100 languages and dialects. Flash-Lite TTS targets high-volume speech generation: dubbing, audio content creation, and expressive voice agents.

The numbers are aggressive: more than 2,000 production-ready voices including regional varieties like Mexican Spanish, Quebec French, and Scots English, plus voice replication from a 30-second audio sample. This is Google pushing TTS from a utility feature into a creator-grade voice design tool.

Anthropic on Wednesday announced the first research result from its physical biology lab in the Bay Area: Claude helped identify a previously unknown enzyme system in bacteriophage DNA with properties reminiscent of CRISPR. The company named the system array-associated reverse transcriptases, or ART — built around a reverse transcriptase with repeating DNA sequences that resemble CRISPR-like patterns, plus an array of non-coding DNA sequences and an additional protein of unknown function.

The discovery was made in 21 hours of concerted effort, with about 950 Claude agents processing 210 million tokens of DNA database data. CEO Dario Amodei noted on X that the work built on earlier research by others, and Anthropic acknowledged that while the underlying reverse transcriptase had been identified in previous studies, Claude appears to be the first to recognize the broader system's key features.

All physical experiments were performed by human scientists in the BSL-1/BSL-2 lab, which Anthropic confirmed last week and says was established this spring. Anthropic cautioned that outside researchers must still verify the finding. The announcement lands as AI developers debate tighter safety testing for advanced models, and as Amodei has simultaneously warned about AI-enabled bioterrorism and written that AI will 'cure most diseases in 5-10 years.'

Anthropic has committed about $11.6 billion to dedicated cloud capacity from Akamai Technologies over seven years, under two project plans aimed at CPU workloads for its fast-growing compute needs. The agreement can expand by up to $9 billion, for a potential total near $20 billion, with Anthropic's payments tied to Akamai hitting delivery and availability milestones.

Akamai also issued Anthropic a warrant for up to roughly 5% of its outstanding common stock at an exercise price equivalent to $111.33 per share, with about 2% vesting on the initial commitment and the rest vesting as each additional $3 billion of services is taken up. Akamai says it expects about $5.5 billion of capital spending tied to the deal and raised its 2026 capex by about $1.7 billion to secure components like memory.

Akamai shares jumped about 15% in extended trading Thursday on the news. The deal is one of the largest cloud-capacity commitments an AI lab has made to a single infrastructure provider, and it underscores how the compute race now dominates frontier-lab spending even as models get cheaper.

The Information reports, citing unnamed sources, that Anthropic, Google and OpenAI are working to launch a standards body for frontier AI by the end of this year or early next — one in which the model builders would regulate themselves, without government oversight. The group would support third-party pre-deployment testing, set rules for reporting safety and security incidents, codify voluntary safety commitments, and define qualifications for independent auditors.

The labs reportedly first aimed for a public-private partnership, but that effort stalled under the Trump administration, per the report. The same three labs plus Microsoft founded the Frontier Model Forum in 2023 with many of the same goals; that nonprofit remains active and is said to be in contact with the new working group. Critics worry the new body could be used to box out open-source developers and smaller competitors.

In parallel, Bloomberg reports Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman addressed a UN Security Council meeting on AI this week and agreed on the need for global standards around AI risk. The UN appearance is real; the standards-body plan itself remains single-source and unconfirmed.

Resolved September 30: confirmed in substance. On September 29, Trump hosted the AI chiefs — Amodei, Zuckerberg, Pichai, Altman — at a White House lunch and they signed the 'White House Accord on Super Intelligence': four layers of self-regulation described as 'morally binding.' That's the labs regulating themselves via a shared standards framework, exactly what the rumor described.

ZeroDrift announced the general availability of Anchor 3.0, billed as the first family of small language models built to enforce regulatory and company rules on AI-generated communications before they're sent. The company positions it as an enforcement runtime for AI agents, not another chatbot model.

Alongside the launch, ZeroDrift published the first benchmark measuring how well AI models enforce FINRA rules on business communications, built on human-written, attorney-labeled data from Surge AI. Anchor 3.0 caught 95.5% of violations — more than any frontier model tested, including GPT-5.6 Sol, the strongest frontier model on the test — while running up to 34x faster and up to 12x cheaper.

The pitch is pre-send enforcement rather than after-the-fact review: Anchor 3.0 checks every message in about 1.5 seconds on ZeroDrift's API and under 100 milliseconds self-hosted, aimed at financial institutions under supervision, record-keeping, and fair-dealing obligations. Developers can sign up and start using it today.

Elon Musk said Wednesday that SpaceXAI's Grok @Bot — the assistant that lives inside X, summoned from posts and DMs — is getting a major feature drop. Shown in a walkthrough video by team member Matt Palmer that Musk quote-posted, the bot can now place voice calls and record voice memos, store logins in a 1Password vault, render inline forms, create inline drafts for email and Slack, switch between accounts, and route traffic through a user's desktop.

The update pushes the bot further into agentic territory: instead of answering in a chat window, it can call, file credentials, and draft messages inside the apps where they'll be sent. Palmer framed it as a rolling weekly update rather than a distant roadmap, so the new controls should appear inside the bot as they roll out.

It continues a busy month for the bot: SpaceXAI has hooked it into Salesforce, HubSpot, and Gong and taught it to resolve customer support tickets for as little as $0.20 each.