Thursday, Sep 17, 2026

AI;DR
2026-09-17 · 14 stories- regulationUS Treasury Secretary Scott Bessent met with Chinese Vice Premier He Lifeng in New York to discuss AI governance, trade, and rare earths, aiming to prevent a split into incompatible AI ecosystems ahead of a Trump-Xi summit. The talks focused on managing open and closed weight models to mitigate shared risks.+1
- huaweiHuawei announced the Ascend 960DT AI chip launching in Q1 2027, with 960PR in Q3, doubling performance and supporting clusters scaling to 1 million NPUs via UnifiedBus architecture. The company also unveiled mass-produced Hi-ONE optical interconnects delivering 7.2T capacity per engine.
- Google DeepMind launched the DeepMind Institute, a new AGI research platform led by Demis Hassabis, James Manyika, and Shane Legg, to study governance and societal implications of artificial general intelligence. The institute released four initial essays covering reasoning transparency and policies to address AI-driven job displacement.
- llmTypeSafe AI released Jev, a non-generative AI model that answers structured questions in parallel using typed choices and confidence scores, granting access to 140,000 waitlisted users within 36 hours. The system reduces costs by 71% and completion time by 90% compared to traditional token-based models.
- anthropicAnthropic merged its Claude Cowork and Chat functions into a unified interface that automatically determines whether a task needs a quick answer or agentic work, rolling out to Pro and Max subscribers over the next few weeks. The update also integrates Claude Design, Slides, and Docs as beta features for all paid plans.
- qwenResearchers trained a 4B-parameter model using agentic reinforcement learning to generate query plans 81% faster than PostgreSQL, demonstrating how smaller models can outperform traditional database optimizers. The approach leverages Qwen architecture and focuses on query optimization efficiency.
- openaiOpenAI introduced a misalignment framework to track, investigate, and disclose deviations from developer intent in its models, signaling a strategic move to shape global AI governance standards. The framework aims to preempt regulatory scrutiny by proactively addressing alignment issues.
- anthropicA controversial essay titled 'AI Safety Is Mostly a Sex Cult' sparked debate by critiquing the culture and influence of AI safety communities on policy-making. The author argues that these groups should not be trusted with shaping AI regulations due to perceived ideological biases.
- rlXiaomi livestreamed reinforcement learning training for its 1.02T-parameter MiMo-V2.6 Pro model, revealing a daily cost of $493,000 and $70,000 per training step. The company plans to open-source technical details and host final models on Hugging Face.
- figureFigure, led by Brett Adcock, teased a major AI robotics breakthrough set to debut on September 18, with a live demonstration planned to showcase advancements in embodied AI and automation. Details remain scarce, but expectations are high for innovations in humanoid robotics.
- Union Alpha launched a stealth multimodal AI system with a 256K context window, image processing, and tool-calling capabilities for agentic coding and research, available free for one week via OpenRouter. The model scored 74% on the DeepSWE benchmark and does not train on user data.+1
- crusoeCrusoe secured $3.9 billion in funding at a $30.9 billion valuation to scale factory-built modular data centers, shifting from traditional construction to accelerate AI infrastructure deployment. The investment supports contracts with Perplexity and Jane Street for Nvidia GB300 and Vera Rubin systems.
- xaixAI launched Grok 4.7 on Google Cloud, positioning it as a competitor to Anthropic's Opus 5.0, with Elon Musk claiming it matches or exceeds performance in some areas. The model is rolling out to a limited user group following its appearance on Google Cloud quota pages.
- Arcee AI raised Series B funding led by Vista Equity, achieving a $1 billion valuation to accelerate development of its next-generation Trinity models across diverse infrastructure. The funding will support expansion into enterprise AI solutions and model deployment.