METR research reveals Anthropic's unreleased model clears safety tests
New evaluation details from METR show Anthropic's unreleased Model 4 scored above 50% across 15 separate capability benchmarks. The accelerated progress, driven by AI helping build future AI, has outpaced forecasts and raised questions about current benchmark limits.
Sources & coverage
AI-assisted summary of the sources above. General information only, not financial advice.
More in Tech
OpenAI's GPT-6.1 Sol learns surgical guidance from surgeon feedback
In early laboratory testing, OpenAI's experimental model learned to identify delicate anatomical landmarks nearly twice as fast after just three procedures with surgeon corrections. The results demonstrate how real-time human feedback accelerates complex robotic precision.
METR finds AI R&D acceleration hitting 2027 timelines ahead of schedule
METR researchers report that AI labs are progressing nearly twice as fast as historical baselines, largely from models helping build future models. Evaluator Ryan Greenblatt noted capability acceleration has already reached milestones once forecast for March 2027.
Freelancers and tech workers share reality of losing work to AI
Workers across writing, QA testing, design and audio production are sharing first-hand accounts of contracts vanishing as clients replace human labour with AI models. Many report being forced into minimum-wage retail or gig work as corporate budgets shrink.
Demis Hassabis questions if society is ready for conscious AI systems
Google DeepMind co-founder Demis Hassabis warned that society is not yet prepared for conscious AI systems, calling for caution before building them. He noted that conscious machines would raise unresolved ethical dilemmas around digital personhood and moral rights.
Tech shares drop as US election results rattle AI data centre expansion plans
US markets dropped sharply as election swings raised doubts over regulatory approvals and power grid access for AI data centres, Reuters reports. The uncertainty wiped billions off tech indices, compounding Wall Street caution after recent post-midterm sell-offs.
Local LLM users debate whether Qwen 27B model suffered silent degradation
Developers running local AI models are reporting sudden drops in output quality with quantized Qwen 27B builds. While users suspect hidden config changes, local model weights cannot change once downloaded, pointing to prompt formatting or hardware runtime quirks.
Comments
Loading comments…