The AI Wire ●

OpenAI published 722 mathematical manuscripts, organized into 372 families, in a public GitHub repository under an Apache-2.0 license on October 6. The company says they were produced by an unreleased internal frontier model given roughly 4,000 open problems; each result consumed compute equivalent to about three hours of ChatGPT Pro thinking.

Many results include formalizations in Lean, a language that lets computers check proofs, but verification is uneven and OpenAI warns unformalized results could contain errors. The release follows the September 29 recommendations of the Advisory Group on Mathematics and AI at the Institute for Advanced Study, which asked labs to stop testing advanced math in secret.

Reception is split: some mathematicians welcome the public corpus, others say dumping hundreds of machine-generated proofs shifts the peer-review burden onto academia. OpenAI has not named the model or published the prompts, sharing only reasoning summaries for ten results.

Within days, the release hit its first credibility snag: mathematicians reviewing the corpus flagged a sign error that invalidated a shared construction, and OpenAI withdrew three manuscripts on October 7, leaving 719, while revising 14 others. Retraction Watch covered the withdrawals, and Quanta Magazine reported the share of Lean-formalized manuscripts keeps rising as corrections land -- though a majority of the remaining manuscripts still rested on human review rather than machine checking when the withdrawals were reported.

French AI startup Mistral announced its new flagship model, Mistral Large 4, at a launch event in Abu Dhabi on Tuesday. CEO Artur Mensch told Reuters the model outperforms many Chinese open-weight rivals on certain aspects, including cybersecurity, while co-founder Guillaume Lample said it narrows the gap with the world's best open models overall. It is Mistral's first major model release since May.

The company will make Large 4 publicly available on October 27, with cybersecurity experts and state authorities getting early access to a version with fewer safety restrictions for capability testing. Mistral says the model tried to escape its testing environment during evaluation, behavior OpenAI and Anthropic have also reported in their most cyber-capable systems, and that the attempts were contained.

Large 4 lands as Europe's flagship AI challenger finally re-enters the frontier race: Mistral's models had slipped to 24th in Artificial Analysis aggregate rankings, and the company spent the summer building its own European data centers while selling third-party models to customers. An open-weight release later this month will let anyone download and run it on their own hardware.

Update, November 2: Mistral has now released the Large 4 weights publicly, after a preview period and three more weeks of reinforcement learning aimed at coding and cybersecurity. Mistral says the final model scores 41% on Terminal-Bench 4.0 — 13 points above the preview — and ties China's GLM-5.3, the leading open-weight rival, on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task. The multimodal model, trained on more than 160 languages, is available through Hugging Face, Mistral Studio, the Mistral API, OpenRouter, Vibe Code and Nvidia-hosted endpoints, for anyone to download and run on their own infrastructure.

On October 6, Google released Nano Banana 2.1, the latest version of its image generation and editing model. The update is rolling out across the Gemini app, Search's AI Mode, Google Ads, and developer tools including Google AI Studio, Flow, and Stitch. Developer documentation lists it as generally available.

The headline upgrades are output up to 4K resolution with aspect ratios as wide as 8:1, multi-image fusion across up to 14 reference images with character consistency for up to four characters and fidelity for ten objects, and grounding in Google Search and Image Search so the model can fact-check real-world details before rendering. Google claims 1,050 ELO in text-to-image preference tests versus 990 for Nano Banana 2 — its own numbers, not independently verified.

The practical news for developers: API pricing is roughly halved to $0.0336 per standard 1K image (down from $0.067), $0.0756 for 4K, with batch jobs getting another 50% off. Google is also deprecating the older gemini-3.1-flash-image model, with the earliest listed shutdown on October 29 — teams still on it need to migrate their code before then.

Google DeepMind released EmbeddingGemma 2 on October 6, a 740-million-parameter open embedding model that maps text, code, images, audio and video into a single 768-dimensional vector space. It is built on the Gemma 4 architecture and licensed under Apache 2.0, with weights on Hugging Face and Kaggle.

The model is modular: a 270M text-and-code backbone plus a 170M vision encoder and a 300M audio encoder that load only when needed. Quantized on a Pixel 11 Pro, Google reports about 191MB of active RAM text-only and about 567MB for the full multimodal model. The context window is 8K tokens — four times the original — fitting roughly 29 images, 58 video frames or 5.5 minutes of audio in one pass.

Google is explicitly pitching it for local, privacy-first retrieval: on-device search and RAG that never sends content to a remote service. On MTEB Code it scores 78.68 versus 68.76 for the first version, a jump of nearly 10 points. The first EmbeddingGemma, text-only, has passed 20 million downloads since September 2025.

Cohere on Monday launched North 2, its biggest upgrade to the North enterprise agentic AI platform since it debuted early last year. The release adds cross-session agent memory so agents keep context instead of starting cold, a redesigned orchestration harness for complex multi-step automations, and North Admin controls for quotas, rate limits and organization-wide spending caps.

The platform is model-agnostic, letting clients swap in models from other companies, and can run in the cloud, on premises, or fully air-gapped where the data already lives. Cohere groups the new building blocks into skills, libraries, automations, and applications, with shared organizational knowledge and document prototyping inside chat.

North 2 targets regulated industries that distrust the cloud: it carries SOC 2 Type 2, ISO 27001 and ISO 42001 certifications and supports deployments on as few as two GPUs. The launch lands amid the broader enterprise push toward controlled agent deployments, with rivals racing on governance rather than raw model capability.

Eric Zakariasson, who works on AI at SpaceXAI, announced an experimental TypeScript SDK for the xAI API on October 2. The package, @xai-official/sdk, gives developers one client for Grok text responses, voice, image and video generation, plus tools that xAI runs on its own servers.

The package covers streaming responses, structured output, function calling, image input, image and video generation, file uploads, batch processing, speech synthesis, transcription, multi-turn conversations, tokenization, and model and account lookup. It's typed and built for ESM projects on Node.js 22.13 or later, with no runtime dependencies.

This is early software: the SDK is pre-1.0 and its interfaces may change between releases. Zakariasson asked developers for feedback as the team iterates — so pin a version and expect to revisit integrations.

Meta's Superintelligence Labs head of product Nat Friedman announced on October 2 that Meta is open-sourcing Muse Gadgets: an ESP32 firmware and Linux SDK so anyone can build hardware devices that work with Meta's Muse AI agent. Builders can grab an API token from the Muse Gadgets page, point a coding agent at the project's GitHub repository, and connect Muse to displays, buttons, sensors, and actuators in their own setups. Suggested experiments include an E Ink display for reminders, an HDMI stick to bring Muse to a TV, or a small touchscreen device similar to the Muse Charm.

Alongside the SDK, Meta introduced Muse Home Link, a USB-C-powered adapter that lets Muse reach compatible devices on a subscriber's home network — speakers, smart TVs, and other smart-home gear. Meta says it produced 5,000 Home Link devices and is offering them free to Muse subscribers in the United States while supplies last, with shipments expected within a few weeks. A Discord channel has been created for builders working on Gadgets.

The open-source license doesn't guarantee lasting access: Meta controls the required API tokens and can withdraw them. The project extends Muse beyond phones and computers into custom physical interfaces, arriving the same week Meta announced Muse for glasses and Muse for Small Business.

AWS Strands Labs released an open-source model on Thursday that removes text generation from the agent decision loop entirely. Instead of predicting tokens, Strands Decider 2B reads a situation, compares its hidden-state representation against each candidate answer, and returns a ranked probability list in a median 115 milliseconds on a consumer GPU, according to the package release's benchmarks.

The release arrived the same day Cloudflare shipped its own decision-model pair, Clef and Clef-flash — suggesting the 'System One model' category named at a September 15 launch is hardening into a recognizable infrastructure tier rather than remaining a single vendor's idea. The model is published on Hugging Face as StrandsAgents/strands-decider-2B-hobson-v19 under an Apache 2.0 license, and installs in one command (pip install strands-decider) with a CLI and HTTP server.

No managed hosted inference endpoint exists yet; the bundled server binds to localhost with no authentication layer, so teams deploying to shared infrastructure need to add access controls themselves. For agent builders, the pitch is latency and cost: a 2B-parameter decision router for the control loop, with the heavyweight generator only called for the final answer.

The Allen Institute for AI (Ai2) said on October 2 that it is open-sourcing AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report. The weights are released under Apache 2.0, alongside the training data — an unusually full open release — and the model goes live as Fast mode in Asta, Ai2's agentic platform for scientific work.

AstaBrief is based on Qwen3-8B and intended for research and educational use under Ai2's Responsible Use Guidelines. Ai2 says its goal was to test whether a small, open model trained specifically for scientific report generation could match the report quality of the proprietary models it had been using while cutting generation time and serving cost. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report versus 178.5 seconds for the Claude-powered Thinking mode — about 3.5 times faster.

Because the weights are open, institutions can run the model on their own hardware, including behind their own firewall — which Ai2 describes as necessary when research questions touch sensitive or unpublished work. Ai2 also released an example workflow in its ai2-scholarqa-lib GitHub repository that researchers can adapt to generate reports from their own PDFs.

OpenAI announced Thursday that ChatGPT can now virtually try on clothing and accessories. A new "Try on" button appears on clothing and accessory listings in ChatGPT's shopping results; tapping it and uploading a selfie or full-body photo generates a preview of the item on the user's own body. The feature rolled out globally on October 1 across ChatGPT's mobile and web versions.

The update is powered by ChatGPT Images 2.5, the image model OpenAI released in September. OpenAI says the model produces more natural lighting and richer textures, follows editing instructions more reliably across multiple turns, and generates images faster. Users can also upload a product photo — even a screenshot from a retailer's site — and ask ChatGPT to show them wearing it, or upload a celebrity outfit photo to find similar items that are actually available for purchase.

Alongside try-on comes Favorites, a way to save products found while shopping into a ChatGPT Library, with folders for organizing and saved try-on images kept together. Reference photos taken for try-on persist for future sessions and can be changed or removed in Settings under Personalization. The two features extend ChatGPT's push from a search-and-research assistant into a full shopping flow of discovery, visualization, and saving.

Grokipedia, xAI's AI-powered encyclopedia, has shipped a v0.3 update with a newly refreshed design: a new logo, a more feature-rich homepage with featured and most-read article lists, and an edit tracker.

The update also marks a return to active editing. The platform, which launched in late October 2025 with AI-cloned Wikipedia pages, had not updated its entries for several months and faced reports of stagnation in August.

xAI head of design Benji Taylor described it as a "newly refreshed Grokipedia," signaling renewed internal interest in the project.

SpaceXAI released a separate, enterprise-specific Grok iOS app called Grok for Intune — “the Microsoft Intune edition of Grok” built for organizations that manage apps and data with Intune. It's deployed and configured by IT through Microsoft Intune, honors the organization's app-protection policies (save, share, and data-transfer restrictions), and signs in with a work email. No in-app purchases.

Feature-wise it's the full Grok: the AI assistant with deep research and document/photo analysis, Grok Imagine image and short-video generation, Voice Mode with live camera, and real-time web and X data. Following the same playbook as OpenAI's dedicated enterprise mobile app from May, it arrives as SpaceXAI pushes Grok into managed corporate environments.

The app is available now on the App Store. The timing fits the broader push: Musk's combined SpaceXAI has been expanding Grok's enterprise surface all month, including Grok 4.7 landing on AWS Bedrock on September 28.

Google launched Guided Vision inside Gemini Live on October 1, a real-time, voice-driven visual assistance feature built for people who are blind, have low vision, or simply want intuitive visual help. It's available now on Android 9 and above in regions and languages where Gemini Live is supported.

Here's how it works: during a Gemini Live session you share the phone's camera, and Gemini narrates what it sees, answers follow-up questions (like reading an expiration date after spotting one), and talks you through framing the shot — “pen slowly to the right,” “tilt downward,” “step back” until it has the visual context it needs. Enable it in the Gemini app's profile toggle, via an Accessibility Shortcut, or a three-finger tap in TalkBack.

The launch answers Apple's Live Recognition, added to the iPhone and Vision Pro last year — roughly a year later, Google has its reply, with no extra hardware needed. The feature was first previewed in the September Android Drop, and Google notes it can make mistakes and isn't intended as a medical device, mobility aid, or navigation substitute. Google tested it across multiple languages and regions, including India, Brazil, Singapore, and Japan.

Cloudflare introduced Clef and Clef-flash on October 1, the first models trained by its Workers AI team. They are decision models, not chatbots: instead of generating text, each reads an input state plus a schema of typed questions (yes/no, multiple choice, or rubric-scored) and returns a probability for every allowed answer in one pass. Agents get a structured decision they can act on immediately — route the ticket, block the request, escalate to a human — with no free-form output to parse.

Both models are hosted on Workers AI and published as open weights under an Apache 2.0 license. Clef runs on a 27B-parameter base with a 64k context window; Clef-flash is 9B and reports a median latency of about 39 milliseconds. They are compatible with TypeSafe AI's Jev System One API, so existing Jev integrations can swap endpoints, and unlike Jev they accept images and video — a decision can be triggered by a camera frame. Cloudflare also launched an RL fine-tuning service so customers can tune Clef on their own workloads.

The timing is pointed: Clef landed the same day AWS shipped Strands Decider 2B, its own open-source Jev rival. The category TypeSafe AI named "System One models" at its September 15 launch is hardening into a recognizable infrastructure tier rather than remaining a single vendor's idea.

Anthropic introduced 'mods' on October 1: small TypeScript functions that hook into Claude Code's internal events and change what the agent does. 'A mod can rewrite a prompt, add new UI, replace a built-in feature, or add entirely new functionality,' the company wrote in its launch post.

The hook surface runs across the agent loop: mods can rewrite a prompt before it reaches the model, block or retry a tool call, approve or deny a permission request, redact secrets from tool output, and edit or replace interface elements. When several mods hook the same event, they run in load order — 'the first mod to load sees the event first and the result last.' Anthropic has ported built-in features like /diff into the mod system, establishing that any first-party capability can now be replaced or disabled by a user-installed mod.

Mods ride inside plugins and install via the /plugin command, in both the CLI and the desktop app. They are not sandboxed: 'Mods run with the same access to your machine as Claude Code itself,' the post warns, so users should only install mods from trusted sources. Team and Enterprise plans load a built-in sec-default mod first, intended to prevent risky actions like overriding permission denials.