Artificial Curiosity Labs
Go back

AI News Digest

AI News Digest — 2026-09-16

Top Stories

Is Big Tech’s AI slowdown a safety pact or a cartel? (The Verge)

CEOs of OpenAI, Anthropic, DeepMind, and SpaceX have loosely agreed to slow AI development—the most concentrated voluntary pacing move yet, directly colliding safety pledges with antitrust risk. The pact’s coordination among dominant players immediately raises cartel concerns, framing the slowdown as either a genuine safety brake or an implicit market-control mechanism now drawing regulatory attention.

Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking (Google DeepMind)

Google DeepMind’s new model series amplifies live, real-time interaction by fusing multimodal processing with deliberative reasoning, enabling applications to fluidly analyze speech, video, and code while pausing to think through complex problems before responding. This architectural leap promises more reliable voice assistants, dynamic tutoring tools, and coding collaborators that can weigh multiple hypotheses during a session, directly intensifying competition in the race toward seamlessly intelligent, always-on AI agents.

Meta’s new One subscriptions put a price on social media and AI (The Verge)

Meta’s introduction of One subscriptions, which bundle its Muse AI assistant, marks the first time a major social platform has put social-media-integrated AI behind a paywall at global scale, signaling a shift toward directly monetizing AI tools within everyday social experiences.

AEO startup Profound hits unicorn valuation, raises $180M Series D 7 months after last round (TechCrunch)

AI-engine-optimization startup Profound’s jump to a $1.8 billion valuation with a $180 million Series D, just seven months after its last raise, underscores the rapid capital deployment into enterprise AI tooling as investors race to back category-defining infrastructure plays.

Paying for frontier AI models buys 4-month head start at 5x the cost (Ars Technica)

This matters because it quantifies a narrowing moat: open Chinese models are now within roughly four months of frontier paid systems while costing only a fifth as much, forcing enterprises to weigh a modest capability lag against massive inference and API savings — and pressuring Western labs’ pricing power as cheaper alternatives close the gap.

Robotics & Physical AI

What moved

China Mobile open-sources Open-RAIL engineering base for VLA and WAM robot models (TechNode)

By open-sourcing the Open-RAIL engineering base, China Mobile delivers a unified conduit connecting vision-language-action and world-action-model models directly to physical robots, which accelerates embodied-AI iteration and slashes the integration hurdle for heterogeneous robot platforms.

Xiaomi Open-Sources Robotics-U0 Embodied World Model and Training Stack (Pandaily)

Xiaomi open-sourcing its Robotics-U0 embodied world model and training stack puts a ~38B-parameter model with claimed 83x inference speedups via FlashAR+ and top WorldArena ranking into public hands, removing a major friction point for researchers working on robot learning and sim-to-real transfer.

XPeng Extends Self-Developed Turing AI Chip From EVs Into Humanoids (Pandaily)

XPeng’s move to deploy its self-developed Turing AI chip in humanoid robots—delivering approximately 2,250 TOPS across three on-device chips—highlights a growing convergence between EV autonomy silicon and physical-AI compute for embodied systems, as both domains increasingly demand the same low-latency, high-throughput inference.

Also tracking

Funding & Capital

What moved

ByteDance’s first-half net profit reportedly fell to $20 billion amid higher AI spending (TechNode)

ByteDance’s first-half net profit reportedly dropped to $20 billion as rising AI spending eroded margins against a $120 billion revenue base, showcasing how aggressively the world’s largest private tech company is betting on artificial intelligence.

Factory raises $200M for its self-improving software development platform (SiliconANGLE)

Factory’s $200 million raise from heavyweight investors Blackstone, Khosla Ventures, Sequoia, and NEA underscores a deep conviction that self-improving AI coding agents will reshape software creation, as the capital injection vaults the startup into a position to accelerate enterprise adoption of autonomous development.

AEO startup Profound hits unicorn valuation, raises $180M Series D 7 months after last round (TechCrunch)

Profound, an AI-engine-optimization startup, has reached a $1.8 billion unicorn valuation with a $180 million Series D just seven months after its Series C, according to TechCrunch — a rapid re-rating that signals intense investor appetite for enterprise AI tooling.

Also tracking

Infra & Compute

What moved

Huawei unveils what it calls the world’s first 3D data center (TechNode)

Huawei’s unveiling of a vertically stacked 3D data center—which it claims as the world’s first—introduces a novel architectural approach to AI-driven density challenges, potentially influencing the design direction of next-generation AI infrastructure.

Cambricon Joins PyTorch Foundation as Platinum Member on Governing Board (Pandaily)

Cambricon’s Platinum membership and governing-board seat at the PyTorch Foundation lock in an upstream-first contribution model across seven core modules including torch.compile, directly fortifying the open-source AI compiler ecosystem for non-NVIDIA silicon.

Connected data emerges as the foundation for trusted AI: theCUBE’s keynote analysis from Amplify (SiliconANGLE)

If AI agents are only as trustworthy as the data they act on, fragmented or siloed records become a direct liability: a model confidently issuing a wrong answer from stale or conflicting sources can damage a brand in regulated, high-stakes workflows. Framing connected data as the foundation rather than a backend chore shifts the enterprise conversation from model choice to information integrity, making unified, governed access the real gate for scaling reliable agents.

Also tracking

Applications

What moved

Huawei opens beta testing for XiaoYi Work AI assistant across phones, tablets and PCs (TechNode)

Huawei’s beta expansion of XiaoYi Work across phones, tablets, and PCs shows a cross-device AI assistant with task-delegation capabilities operating at the OS level, signaling that such deeply integrated agents are poised to become a standard consumer device feature.

Newell Brands puts internal audit at the heart of AI adoption (SiliconANGLE)

Newell Brands is centering its AI adoption strategy around its internal audit function, signaling a strategic shift where controls teams assume direct ownership of AI governance rather than merely performing after-the-fact compliance checks.

University of Manchester Uses NVIDIA Earth-2 to Forecast Air Pollution Across the UK (NVIDIA Blog)

The University of Manchester’s use of NVIDIA’s Earth-2 digital-twin platform to forecast UK-wide air pollution marks a significant advance in operational AI-driven climate simulation, translating real-time atmospheric data into actionable public-health insights at a national scale.

Also tracking

Agents & Tooling

What moved

vivo previews BlueCode phone coding agent, explores 30B MoE model for on-device use (TechNode)

Vivo’s preview of BlueCode, a phone-based coding agent, together with its exploration of a 30-billion-parameter Mixture-of-Experts model for on-device use, demonstrates a concrete step toward running complex software development tools entirely on a handset, reducing reliance on cloud processing and potentially improving privacy and responsiveness.

Cohesity’s new Agent Resilience lets companies roll back AI agents that go wrong (SiliconANGLE)

Cohesity’s Agent Resilience introduces agent-state backup and rollback for enterprise AI agents, giving companies a way to restore an agent to a prior state after it goes wrong—directly addressing a critical operational-safety gap as agents take on consequential tasks.

Governance becomes critical as AI agents move into financial reporting (SiliconANGLE)

Workiva’s push to embed trust in AI-driven financial reporting surfaces a looming governance gap: as agents accelerate disclosure cycles, the absence of auditable, defensible outputs risks regulatory blowback, making transparency a prerequisite for any speed advantage.

Also tracking

Foundation Models

What moved

EBKernel Unveils Cog-WM 1.0 Brain-Inspired Cognitive World Model (Pandaily)

EBKernel’s Cog-WM 1.0, a JEPA-style latent world model, delivers double-digit gains on ObjectNav and manipulation benchmarks, signaling that brain-inspired cognitive modeling can meaningfully advance embodied AI performance.

Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear (TechCrunch)

Salesforce and Nvidia’s joint release of Salesforce Koa, a reasoning model fine-tuned from the open-weight Nemotron for enterprise sales and support, directly challenges proprietary AI labs by demonstrating that openly available foundation models can be operationalized into high-value business tools by major platform players.

Paying for frontier AI models buys 4-month head start at 5x the cost (Ars Technica)

A new Mozilla report finds that paying for proprietary frontier AI models now buys only a four-month performance lead over open Chinese alternatives, yet costs five times more, forcing enterprises to weigh whether that brief head start justifies the premium.

Also tracking

Policy & Safety / Other

What moved

We don’t need AI regulation — leave safety to us, Nvidia’s Jensen Huang says (TechCrunch)

Nvidia CEO Jensen Huang publicly rejected AI regulation, asserting that safety can be engineered on a per‑product basis and should be left to companies — a provocative stance from the head of the world’s most valuable AI hardware supplier, whose chips underpin the majority of advanced AI systems, and a direct challenge to mounting policy efforts.

AI agents now have a place to snitch (TechCrunch)

The AI Contact Hotline creates a formal reporting channel for autonomous agents to flag unethical or harmful behaviors they observe, transforming ad-hoc monitoring into a structured accountability mechanism. This experimental infrastructure signals a shift from treating AI oversight as a human-only responsibility to building systems where agents actively participate in policing their own ecosystem, raising immediate questions about verification standards, liability when reports are acted upon, and whether a snitching architecture can scale across distributed agent networks before broad deployment.

AI and data centers are incredibly unpopular in every poll (The Verge)

The latest NYT/Siena survey, a follow-up to our September 11 reporting, finds 61% of respondents oppose data-center construction specifically for AI, hardening the political headwinds that threaten the industry’s rapid infrastructure expansion.

Also tracking

Voice & Speech

What moved

Google’s new speech model Gemini 3.8 Live supports real-time reasoning (SiliconANGLE)

Google’s Gemini 3.8 Live adds real-time reasoning to speech, attacking the latency gap that has kept voice AI from matching human conversational speed and pushing rivals to ship faster multimodal models.

Intelligent transcription with Gemini 3.5 Transcribe (Google DeepMind Blog)

Google DeepMind’s Gemini 3.5 Transcribe moves beyond literal speech-to-text by incorporating contextual understanding directly into the transcription process, which allows it to interpret intent, disambiguate homophones, and structure output for downstream tasks. This matters because it shifts transcription from a mechanical utility into an AI-native interface, potentially enabling more accurate meeting summaries, real-time voice commands, and accessible media indexing where meaning, not just words, is captured from the start.

Image & Video

What moved

Introducing agentic video understanding with Gemini (Google DeepMind Blog)

Gemini gains agentic video understanding, enabling AI to actively perceive, reason about, and interact with video content beyond passive analysis.