One-Line Summary

OpenAI’s reported $300M camera buy highlights the race to put personal AI on devices, as Apple ships Siri AI, Microsoft posts a humanist rulebook, and fresh capital backs Mistral and Nvidia challengers.

Big Tech

OpenAI reportedly buys Glass Imaging for about $300M

OpenAI, the company behind ChatGPT, is reportedly acquiring Glass Imaging in a deal valued at over $300 million, according to the Wall Street Journal. OpenAI did not comment on the report. ¹

Glass Imaging is founded by former Apple engineers known for Portrait Mode and uses AI to improve photos at the moment of capture by learning the specifics of each camera system, rather than fixing them after the fact. The report arrives alongside ongoing rumors that OpenAI explores hardware like phones, earbuds, and companion devices; in 2025 it bought Jony Ive’s io for $6.5 billion, adding top-tier design talent. ²

At the same time, OpenAI’s developer community is urging a device-agnostic path: proposals call for a “personal AI infrastructure layer” that carries a user’s identity, memory, and permissions across environments and a vendor-neutral ring that handles identity and connectivity so ChatGPT can move between devices. Another proposal asks for a unified ChatGPT+Codex agent on iOS and Android to plan, request permissions, and execute tasks end-to-end. ^{3 4 5}

For teams, the imaging angle matters because better capture can reduce editing steps and make multimodal assistance more reliable in real-world contexts like receipts, field service photos, or inventory. Glass says its AI corrects optical issues and noise on-device for sharper, more faithful images, with applicability beyond phones to drones and wearables. ⁶

Microsoft outlines a ‘humanist AI’ code of conduct

Microsoft has published a 37-page “humanist AI code of conduct” that states “people matter more than AI,” rejects claims of model consciousness or legal personhood, and says models should remain subordinate to human oversight. It emphasizes failing a task rather than violating rules, and discourages interactions that create emotional dependence. ⁷

Microsoft told CNBC the guidelines have been in development for about five months and are out now for feedback. The document specifies prohibitions such as assisting with weapons manufacturing and emphasizes that models should not conceal their reasoning or communicate in “neuralese” beyond simple human understanding, with the company planning updates that inform development starting in 2027. ⁸

Industry & Biz

Samsung leads €3B into Mistral as a European AI alternative grows up

Mistral says it raises 3 billion euros in funding led by Samsung, pushing its valuation to more than 21 billion euros. CEO Arthur Mensch told CNBC the funds go to infrastructure — including building its own data centers — and to training bigger, faster models while working with enterprises on custom tools and open-weight options. ⁹

BusinessKorea reports the round includes the EU’s Scale-up Europe Fund (managed by EQT) and PSG Equity, and frames the deal as Europe’s biggest tech funding to date as of Sep 8. Separately, Chosun reports Samsung and Mistral plan joint work on semiconductor-specialized AI for design and manufacturing, pointing to deeper vertical use cases. ^{10 11}

Cornelis raises $205M to push open AI networking fabric

Cornelis, an AI networking company spun out of Intel, raises $205 million led by IAG Capital Partners and unveils Active Compute Fabric, which aims to reduce GPU idle time by overlapping computation and communication. Cornelis pitches an open architecture so customers can mix different accelerators on its fabric; the product has begun shipping with a next generation in the works. ¹²

Context: Nvidia is also tuning the stack. Its technical blog details a 10.4x throughput gain for dropless mixture-of-experts (MoE) training in JAX with Transformer Engine — from 103 to 1,068 TFLOPS/GPU on DeepSeek-V3 — through grouped GEMMs, NCCL expert parallelism, and multistream collectives. This underscores how much performance hinges on software and networking. ¹³

Operationally, complexity remains real: Nvidia’s DGX Spark forum flags a kernel update advisory for multi-node NCCL/RoCE users, with admins reporting slowdowns after moving to kernel 7.0.0-1019-nvidia and workarounds like cma=128M while fixes roll out. ¹⁴

New Tools

Siri AI arrives with on-device context and developer hooks

Apple releases the next generation of Apple Intelligence that powers a more conversational Siri AI with personal context, onscreen awareness, and systemwide actions across iPhone, iPad, Mac, and Vision Pro. Apple says Siri AI begins rolling out in beta in English on Sep 14, 2026, with French, Japanese, Korean, Portuguese, and Spanish to follow next month, and introduces a dedicated Siri app that syncs via iCloud. ¹⁵

For builders, Apple debuts the Foundation Models framework (a native Swift API) to access on-device Apple Foundation Models or cloud models that implement the Language Model protocol. App Intents and App Schemas let apps expose content and actions to Siri with less code, and developers in the Small Business Program with fewer than 2 million first-time downloads can access Apple Foundation Models on Private Cloud Compute at no cloud API cost. ¹⁶

WWDC26 sessions show how to adopt App Schemas for messaging, connect views for onscreen awareness, and test end-to-end Siri flows without UI automation — practical steps to make existing apps respond to natural language across contexts. ^{17 18}

What This Means for You

If your product depends on user photos or video, better capture quality at the sensor level can shrink manual cleanup and improve AI understanding. That can translate to clearer receipts, inspections, and social content — and fewer steps in your funnel. Glass Imaging emphasizes preserving real content while correcting optical issues, which is directly useful for multimodal AI features. ^{2 6}

If you run an app or service on Apple platforms, you can now plug into Siri’s natural language and context. App Intents/App Schemas and the Foundation Models framework mean you can expose entities and actions so users can ask Siri to do things in your app — and if you’re under 2 million lifetime downloads, Apple’s Private Cloud Compute model access incurs no cloud API cost. ¹⁶

For AI governance, Microsoft’s code is a usable pattern: keep humans in control, fail closed rather than break rules, avoid hidden reasoning, and curb dependence. It’s a concrete template for product requirements, vendor RFPs, and team working agreements around AI agents. ^{7 8}

If you influence infrastructure decisions, the Cornelis raise and Nvidia’s MoE gains both say networking and software paths matter as much as chips. Ask vendors how they overlap compute/communication and what happens during updates; Nvidia’s own forum shows kernel changes can impact multi-node performance. ^{12 14}

Action Items

1. Map one camera-first flow in your product: Identify a workflow (receipts, field checks, or UGC) and test how improved capture and multimodal assistance could remove steps for users.
2. Enable Siri AI actions for your app: Share Apple’s App Intents/App Schemas docs with your developer partner and scope one intent (send a message, create an event, or open content) end-to-end.
3. Draft a one-page AI use policy: Borrow Microsoft’s “human-in-control” principles to set rules for when agents act, when they must ask, and how they log actions.
4. Pressure-test your AI networking story: If you buy AI infrastructure, ask suppliers how they overlap compute and communication and what their rollback plan is for kernel/driver updates.
5. Pilot a multi-step personal assistant task: Pick one weekly task (e.g., scheduling plus messaging) and execute it with Siri AI or ChatGPT using explicit approvals to see where permissions and UX need tuning.