The algorithm, discovered by an internal research Claude model, solves #3-SAT in 2^(O(n log log n / sqrt(log n))) time and is formalized in Lean 4.
The open-source repository includes a closed-loop simulator with 44 environments and a 120-task benchmark that evaluates models based on final world state rather than text.
The system uses a hierarchical memory harness to track evolving user intent, yielding a further 12.3% gain in sustained interaction consistency.
The 24-day formalization generated 43.1 million lines of code and identified 34 significant discrepancies in existing mathematical proofs.
The new features allow admins to analyze team workstreams, capture user feedback via in-app cards, and manage shared monthly budgets with automated alerts.
The built-in provider supports Supervised and Auto modes using existing fx models and skills.
IDC Quanta, now available in Claude, provides cited market intelligence and vendor shortlists, while internal adoption has cut engineering and planning timelines significantly.
The new gallery includes six core task artifacts and nine access-to-justice tools, with integrations for partners like Goodwin, Harvey, and PwC.
Confirmed that removing the KL term is essential for preventing performance collapse and improving efficiency when training Qwen3 0.6B with GRPO.
The patch release from 163 contributors resolves issues preventing installation, updates, and reply delivery without adding new features.
The new standard covers nine specific AI threats, including data poisoning and prompt injection, but serves as guidance rather than a certification.
A federal judge ruled that Anthropic illegally scraped and pirated works from over 10 Stanford professors to train its Claude models, despite a $1.5 billion settlement.
The M6-based iMac is up to 5 times faster in on-device AI first token generation compared to the M4.
The new adapters reduce image generation from 40 to 6 steps and remove the separate prompt-upsampling phase, saving 41 to 80 seconds per image.
The new benchmark features tasks 67% larger than previous versions and reveals that models struggle with unstated constraints and thoroughness, with the best model achieving a 45.5% score.
The design agent uses a sandboxed workspace and async sub-agents to generate code-based assets, with 75% of requests coming from outside the marketing department.
The research preview allows users to steer simulations in real-time via text prompts, with upcoming access for Google AI Ultra subscribers and Gemini API partners.
A new benchmark from Ai2 reveals that small prompt changes can shift LLM novelty judgment accuracy by over 50 percentage points, and simple baselines outperform expensive purpose-built systems.
The model surpasses leading open-weight models on MindCube spatial reasoning and outperforms other world models on FlowBench motion prediction.
The library separates system integration from privacy modules, reducing new requirement implementation from months to days while supporting AI agent controls.
The update introduces a new Copilot chat experience, Hot Exit for unsaved work, and keyboard shortcuts for tab management.
The model uses Gemini Omni for realism and other Gemini models for reasoning, allowing users to steer scenes and generate ambient sounds and dialogue on the fly.
The 1.7B model scores 59.8 on SWE Verified and 85.3 on tau2-bench, closely trailing the 4B variant's 62.8 and 86.9.
AI Briefing