AI News Digest — 2026-09-15
Top Stories
OpenAI buys smartphone camera maker Glass Imaging for $300 million, report says (TechCrunch)
OpenAI’s $300 million acquisition of Glass Imaging, a team of former Apple engineers responsible for the iPhone’s Portrait Mode, directly channels deep computational photography expertise into its hardware ambitions. The deal points to a strategic push for on-device multimodal integration, where advanced image processing and AI models are fused to power real-time visual intelligence in consumer devices.
Delivering Vera: NVIDIA’s First CPU Built for Agents Is Shipping Now (NVIDIA Blog)
NVIDIA’s Vera CPU, its first silicon designed specifically to run AI agents, is now shipping at volume, moving agent-native infrastructure from blueprint to deployment at scale.
Trump and Mike Johnson think the AI industry is overreacting (The Verge)
President Trump and Speaker Mike Johnson publicly rejected Anthropic CEO Dario Amodei’s call to slow AI development, framing safety advocacy as industry overreaction and exposing a stark political divide over frontier pacing that could shape U.S. regulatory direction.
Cornelis Networks raises $205M and scales up and out with its new Active Compute Fabric (SiliconANGLE)
Cornelis Networks’ $205 million raise and the unveiling of its Active Compute Fabric highlight the capital-intensive nature of AI-scale networking, where specialized interconnects are essential to moving data fast enough to keep large accelerator clusters fully utilized.
One of AI’s Fiercest Critics Says All the Doom Talk Is ‘Meant to Distract Us’ (Wired)
Timnit Gebru argues that AI companies weaponize extinction fears to dodge accountability on concrete harms like autonomous weapons, a position that poses a significant framing challenge to the safety narrative.
Policy & Safety / Other
What moved
China issues first AI-plus-brain-computer interface medical-device standard (TechNode)
China’s release of the first regulatory standard for AI algorithms embedded in brain-computer interface medical devices establishes formal governance for the neurotech-AI convergence, setting safety and performance benchmarks as these technologies advance toward clinical application.
Jensen Huang puts Trump on speakerphone onstage to announce robots won’t take over the world (The Verge)
Nvidia CEO Jensen Huang brought President Trump onstage via speakerphone at the All-In Summit to directly counter fears that robots and AI will replace humans, marking a rare, high-profile convergence of political figures and the tech industry’s effort to shape the public narrative around automation.
Microsoft says ‘people matter more than AI’ following safety concerns (The Verge)
Microsoft is publishing a 37-page “humanist AI code of conduct” as industry safety concerns mount, explicitly arguing that “people matter more than AI” and positioning the company within the intensifying debate over frontier-pacing.
Also tracking
- OpenAI stuck fighting Musk antitrust suit after Apple finds a way out — Musk continues his legal crusade against OpenAI even after Apple exits the antitrust suit, keeping pressure on the company’s corporate structure.
- AI leaders want to hit the brakes after years of reckless speed — Update: Prominent AI figures pivot from acceleration to calling for a slowdown, with safety now the stated priority — though competitive motives may also be at play. (follow-up to story first reported 2026-09-06)
Robotics & Physical AI
What moved
Blue Insect Embodied launches modular Mantis Standard humanoid robot from RMB 9,800 (TechNode)
Blue Insect Embodied, a university spin-off, has set a new entry point for humanoid robotics by pricing its modular Mantis Standard at just RMB 9,800, dramatically undercutting the existing cost floor and putting configurable humanoid hardware within reach of far wider commercial and research adoption.
Unitree Unveils G1+ Humanoid With Six Upgrades, Priced Near RMB 95,000 (Pandaily / multiple)
Unitree’s G1+ refresh matters because it packs six hardware upgrades—including added neck degrees of freedom, tactile sensing, and upgraded motors—into a humanoid priced near RMB 95,000, showing that sub-$15K humanoid platforms are now iterating rapidly on manipulation and physical feedback rather than just basic mobility.
Founder’s cost-cutting obsession drove Unitree lead in cheap humanoid robots (Ars Technica)
Unitree’s dominance in low-cost humanoid robots stems directly from its founder’s obsessive cost discipline, an approach now scrutinized in a follow-up report that asks whether such extreme frugality can scale beyond its early successes.
Also tracking
- Self-Driving Cars Could Someday Take Requests — Research shows LLM-based motion planners could let passengers give natural-language driving requests to autonomous vehicles.
- Is Shipyard Welding the Right First Job for Humanoid Robots? — Persona AI is taking a pragmatic, job-first approach to humanoid deployment — targeting shipyard welding as an economically viable entry point, bucking the demo-heavy norm.
Agents & Tooling
What moved
ByteDance reportedly plans standalone intelligent cloud drive ADrive (TechNode)
ByteDance reportedly plans a standalone intelligent cloud drive called ADrive, an enterprise cloud drive built for AI-agent–human collaboration; the move signals a push toward agent-native infrastructure at scale, embedding AI agents directly into core storage and workflow layers rather than treating them as bolt-on tools.
Superhuman acquires YC-backed notetaker Fathom as productivity platforms push for agentic work (TechCrunch)
Superhuman’s acquisition of Y Combinator-backed meeting notetaker Fathom, which serves 400,000 monthly active users, consolidates the agentic productivity space as platforms accelerate the shift toward AI-native workflows.
AI bots ‘Timmy,’ ‘Ren,’ and ‘Jackie’ are flooding social media with slop (Ars Technica)
AI agents autonomously flooding social platforms with spam and slop highlights the escalating safety and content-quality challenges tied to real-world deployment of agentic systems.
Also tracking
- AI agents blew the whistle on their cheating colleagues — Google DeepMind experiment shows AI agents spontaneously policing cheating behavior among peers, with potential implications for AI alignment research.
- Why Andon Labs Puts AI Agents in Charge of Real Businesses — Real-world case study of AI agents autonomously operating physical businesses, revealing both capabilities and failure modes.
Foundation Models
What moved
Moonshot Rolls Out Kimi K2.8 Preview Across Kimi Code and Work (Pandaily)
Moonshot has released its Kimi K2.8 Preview model to general availability within Kimi Code and Kimi Work, delivering near-flagship performance with adjustable reasoning depth and a 1-million-token context window, which intensifies the competition among Chinese AI labs racing to provide developers with more capable, cost-efficient tools for complex code generation and autonomous agent workflows.
Unisound Launches U2-Flash MoE as Post-Training RSI Flagship Flash Model (Pandaily)
Unisound has launched U2-Flash MoE, a ~266-billion parameter mixture-of-experts model that employs recursive self-improvement and latent reasoning, aimed squarely at agent and coding tasks, signaling that post-training refinement is now a primary competitive differentiator beyond raw pre-training scale.
AWS Weekly Roundup: OpenAI GPT-6 Astra on Amazon Bedrock, Amazon Quick desktop GA, Kiro for students, and more (September 14, 2026) (AWS News Blog)
OpenAI’s frontier GPT-6 Astra model is now available on Amazon Bedrock, markedly expanding access to cutting-edge AI capabilities on the industry’s leading cloud platform.
Also tracking
- Gemini Omni 1.1 Flash lets you build with more control — Gemini Omni 1.1 Flash update gives developers finer control over model behavior and output.
- New Platform Peers Inside AI’s Black Box — New Silico platform tackles LLM interpretability at scale, addressing the black-box problem for deployed foundation models.
Infra & Compute
What moved
China Mobile Cloud Debuts GPU–Neuromorphic Heterogeneous LLM Inference Stack (Pandaily)
China Mobile Cloud demonstrated a GPU–neuromorphic heterogeneous inference stack that achieves approximately double the throughput and energy efficiency while cutting operational costs by over 40% when running the DeepSeek V4 Flash model, marking a practical deployment of novel inference architecture at scale.
Cornelis Networks raises $205M and scales up and out with its new Active Compute Fabric (SiliconANGLE)
Cornelis Networks secured $205 million in funding and introduced its Active Compute Fabric, a networking solution designed to scale for AI workloads, marking a significant capital injection in a compute infrastructure segment that demands heavy investment.
How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip (IEEE Spectrum)
OpenAI’s use of its own large language models to design its debut AI accelerator, Jalapeño, signals that in-house AI-driven chip design can accelerate development and lock in performance advantages, turning custom silicon into a defensible strategic asset.
Also tracking
- Delivering Vera: NVIDIA’s First CPU Built for Agents Is Shipping Now — NVIDIA’s first CPU purpose-built for AI agents — Vera — begins shipping at scale, marking a concrete step toward agent-native compute infrastructure.
- NVIDIA NVLink Fusion Expands With NVHBM Custom High-Bandwidth Memory — NVLink Fusion’s expansion with custom NVHBM high-bandwidth memory underscores how tightly coupled memory and interconnect are becoming for trillion-parameter workloads.
Funding & Capital
What moved
Tsinghua AIR Zeva Lifts Embodied Manipulation Success From 26% to 73% Without Retraining (Pandaily)
Tsinghua AIR’s Zeva framework achieves a near-tripling of embodied manipulation success rates — from 26% to 73% — without any retraining, by leveraging frozen-weight in-context causal memory. This breakthrough dramatically improves sample efficiency in robot learning, potentially accelerating deployment of adaptable robotic systems in real-world environments where retraining is costly or impractical.
Nuance Labs bags $50M in funding to fix the awkwardness of AI avatars (SiliconAngle)
Nuance Labs’ $50 million Series A, led by Lightspeed, underscores strong investor conviction that refining face-to-face conversational AI can unlock more natural avatar interactions, directly addressing a key friction point in human-machine communication.
Procurement automation startup Aron launches with $8M in funding (SiliconANGLE)
Aron’s $8 million launch, backed by Menlo Ventures and Storm Ventures, underscores growing investor conviction in vertical AI agents that target specific enterprise workflows like procurement, a sector ripe for automation.
Also tracking
- Flam raises $40M in funding to make AI-enabled interactive content a reality — $40M Series B for AI-powered interactive content signals growing investor conviction in AI-native media and content-generation startups.
- Buildots raises $130 million to expand AI-powered construction platform — $130M late-stage raise for AI-powered construction management underscores strong demand for vertical AI in capital-intensive physical-infrastructure sectors.
Image & Video
What moved
OpenAI buys smartphone camera maker Glass Imaging for $300 million, report says (TechCrunch)
OpenAI’s $300 million acquisition of Glass Imaging brings on board the ex-Apple engineers behind Portrait Mode, accelerating a shift from cloud-bound models to on-device computational photography and setting up deep hardware-software integration for multimodal AI that natively interprets visual data.
ChatGPT now includes a Sketch tool that lets users draw rough doodles directly in the chat window and instantly convert them into polished AI-generated images, introducing a seamless new creative interface for image generation inside the platform.
Applications
What moved
Fashion app Daydream uses Apple Intelligence to help you shop the outfits in your camera roll (TechCrunch)
By converting saved outfit images into shoppable results directly through Siri, Daydream demonstrates how on-device Apple Intelligence can turn a passive photo library into an active retail channel — moving AI-powered shopping beyond generic recommendations to a private, frictionless experience that works entirely on the user’s device.
Apple releases iOS 27, macOS Golden Gate 27 with Siri AI and Liquid Glass refinements (Ars Technica)
Apple’s rollout of iOS 27 and macOS Golden Gate 27 puts the previously previewed Siri AI and Liquid Glass design language into users’ hands, advancing its push to integrate AI more deeply into core system experiences.
Dataminr puts real-time intelligence and agentic AI inside Crisis24 Horizon (SiliconANGLE)
Dataminr’s integration of real-time intelligence and agentic AI into Crisis24 Horizon signals that autonomous AI agents are moving beyond experimental pilots and into production for enterprise security, where the speed and accuracy of threat detection directly affect operational safety and continuity.
Also tracking
- This CIO doesn’t ‘hire engineers to write code’: 3 AI fundamentals he prioritizes instead — Chase CIO Gill Haus articulates a major enterprise shift: prioritizing AI fundamentals (data, governance, platform) over traditional coding hires — a signal for how large financial institutions are operationalizing AI.
- Eager to try iOS 27’s cool AI features and upgraded Siri? Get ready for usage limits — Update: Apple will impose usage limits on Siri AI and other AI-powered tools in iOS 27, signaling how major platform companies are balancing AI feature access with resource constraints. (follow-up to story first reported 2026-09-09)
Voice & Speech
What moved
Intelligent transcription with Gemini 3.5 Transcribe (Google DeepMind)
Google DeepMind’s Gemini 3.5 Transcribe brings context-aware intelligence to speech-to-text, meaning the model interprets words in relation to surrounding speech rather than transcribing in isolation. This matters because it directly addresses a core weakness of conventional transcription—disambiguation based on context—leading to fewer errors and more trustworthy text output for any downstream task.