METR Secures $71 Million to Scale Independent AI Tests
The nonprofit evaluator will expand research on autonomous capabilities and AI-driven self-improvement while rejecting lab funding.
August 14, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- Industry Nvidia may guarantee up to 25% residual value for selected AI-compute investments, seeking to reassure buyers about hardware depreciation risks. IT之家 ↗
- AIGC MiniMax showcased commercial and music-video projects already being created with its H3 video model at the Magnific event in San Francisco. MiniMax on X ↗
- Safety A plaintiff embedded invisible AI instructions in court filings, exposing prompt-injection risks for legal systems using automated document review. The Decoder ↗
- Research World Labs introduced a method that converts one demonstrated real-world robot task into thousands of simulated training variations. The Decoder ↗
- Research Zhizhi Research Institute proposed an interpretability method that decomposes model weights while requiring less than 1% of the usual data cost. 量子位 ↗
- AIGC A Chinese music-generation model targeted a recurring AI-music quality problem while positioning itself as a domestic challenger to Suno. 量子位 ↗
- AIGC Runway added 1080p Seedance 2.5 video generation in early access, promising sharper detail at the higher resolution. Runway on X ↗
- Research Researchers described a “tragedy of the cognitive commons,” where individually rational AI adoption could collectively erode professions’ accumulated expertise. The Decoder ↗
- Models Ollama fully deployed DeepSeek-V4-Pro 0813 to its cloud, making the model available through Pro and Max subscriptions. Ollama on X ↗
- Research A new visual-perception benchmark found that leading AI models still perform poorly on basic tasks involving fine visual distinctions and spatial relationships. The Decoder ↗
- Models Together AI reported DeepSeek-V4 Pro 0813 reaching 88.5% pass@4 on software-engineering tasks in its DeepSWE evaluation. Together AI on X ↗
- Industry Artificial Analysis said DeepSeek V4 Pro 0813’s cost per task rose from $0.05 to $0.25 despite improved token efficiency. Artificial Analysis on X ↗
- Open Source Tri Dao argued that open models now trail closed laboratories by only months in quality while open-source serving systems advance rapidly. Together AI on X ↗
- ⚡ Quick hits
- Agents Replit upgraded its MCP integration so external agents can find and update existing Replit projects without requiring an API key. Replit on X ↗
- AIGC Pika opened four generative-audio models through its API Club, emphasizing efficient inference and lower-cost access to frontier-quality audio. Pika on X ↗
- Open Source Streamlit 1.61 added lazy dataframe loading, background cache refresh and icons for metric components. Streamlit on X ↗
- Agents Qdrant and Minima demonstrated single-GPU agentic RAG where hybrid search and reranking made the first retrieval sufficient 87% of the time. Qdrant on X ↗
- Policy Anthropic said it is implementing Claude text watermarking for EU AI Act compliance and published an FAQ addressing detection and reliability questions. Anthropic on X ↗
- Open Source Nvidia introduced NeMo Switchyard, an open-source routing library that assigns different agent-workflow steps to models based on capability and cost. NVIDIA AI on X ↗
- Industry An Epoch AI–Ipsos survey found most American workplace AI use relies on free plans, except in science and technology occupations. Epoch AI on X ↗
- AIGC Character.AI added one-click bulk Lorebook imports, letting creators draft fictional-world context externally and bring it into the platform at once. Character.AI on X ↗
- Agents Databricks added AI Spend Controls to Unity AI Gateway to catch costs caused by retry loops and uncontrolled agent experiments. Databricks on X ↗
Anthropic Withholds Stronger Model as Risk Estimates Rise
Anthropic’s latest risk report raises its misalignment estimate and says a stronger internal model will not be released for now.
- ⚡ Quick hits
- Research Goodfire said its Silico research agent autonomously trained a multimodal encoder overnight, then investigated why its vision-language representations underperformed. Goodfire on X ↗
- Models AWS added custom reward functions for multi-turn reinforcement learning with Amazon Nova Forge, enabling developers to tailor model-training objectives. AWS Machine Learning Blog ↗
- Safety Google detailed a homomorphic-encryption approach intended to make private AI inference practical while keeping processed data encrypted. Google Security Blog ↗
- AIGC Tripo updated Image Gen with 3D-ready templates, asset extraction, cutouts and reference-based variants for building consistent asset collections. Tripo AI on X ↗
- AIGC Krea temporarily made Slack generations unlimited for business customers through August 21, removing rate limits and long generation queues. Krea on X ↗
Google Makes Visible Gemini and Flow Media Watermarks Optional in Most Markets
Users in most markets can remove visible marks from generated images, video and music, while Google retains SynthID and C2PA provenance; legally required marks remain in some jurisdictions.
- ⚡ Quick hits
- AIGC Dreamina launched Seedance 2.5 video generation at 1080p on its global web product. Dreamina on X ↗
SpaceX Completes $60 Billion Acquisition of AI Coding Firm Cursor
Cursor has formally joined SpaceXAI in an all-stock acquisition valuing the coding-agent company at approximately $60 billion.