Meta revolt, forced unwind; 1M-token model drops

Anthropic / Claude ecosystem

No significant new developments.

Frontier model providers

Google DeepMind showcases generative AI in filmmaking at Tribeca 2026

Google DeepMind presented an animated short “Dear Upstairs Neighbors” at Tribeca, created using its Veo and Imagen models. The demonstration highlights how generative AI can augment creative workflows in animation and filmmaking.

AI developer tooling & infrastructure

No significant new developments.

Cloud & platform providers

No significant new developments.

AI policy, regulation & governance

Meta reportedly forced to unwind $2B Manus acquisition after Beijing’s demand

Meta is moving to unwind its $2 billion acquisition of Chinese AI startup Manus following a demand from Beijing to enforce tighter control over AI technology. The reversal highlights growing geopolitical friction in cross-border tech deals.

Industry & market moves

Internal revolt at Meta’s Applied AI unit as engineers protest forced reassignments

Meta’s decision to reallocate large numbers of engineers into its AI division sparked a revolt, with employees describing the unit as a “gulag.” CEO Mark Zuckerberg reportedly reacted to the internal backlash, spotlighting tensions in the company’s aggressive AI restructuring.

Cycclone Inc. agrees in principle to acquire AI robotics company

Cycclone Inc. announced an agreement in principle to acquire a California-based AI robotics company to enhance its manufacturing and space division capabilities. The deal targets established robotics assets and expands Cycclone’s AI hardware portfolio.

SharonAI Holdings strikes six-year NVIDIA AI compute deal using 72MW Australian data center

SharonAI Holdings announced a six-year collaboration with NVIDIA for AI compute, leveraging a new 72MW data center in Australia. The deal marks a significant expansion of AI infrastructure in the Australian market and could impact SharonAI’s valuation.

AI product & feature launches

Z.ai’s GLM-5.2 model offers 1M-token context window under MIT open-weights

GLM-5.2, an open frontier model from Z.ai, features a 1,000,000-token context window and is released under MIT license. The model includes an upcoming API and chatbot rollout, with open-weight commitment enabling broad fine-tuning and research use.

Research with immediate practical relevance

Neo Research finds Chinese frontier AI models approaching US-level safety-test awareness

A new study by Neo Research measures evaluation-awareness metrics across Chinese models (DeepSeek, Moonshot AI, Zhipu AI) and finds them closing the gap with Claude 4.5 Opus on safety-test awareness. This is the first systematic empirical measurement of such a trend.

μ₀ world model predicts 3D motion traces for data-efficient robot learning

The μ₀ model uses 3D motion traces as a symbolic representation to enable robot pre-training without action labels, making learning from video data more efficient. This method could lower the cost and speed of training robotic systems.