A Civilisation Frame for AI Safety's Long Haul
AI safety's dominant 'war frame' treats the field as a deadline-driven campaign that narrows researchers' horizons and ends in burnout. The 'civilisation frame' treats AI safety as a chronic problem like public health, calling for institutions, broader entry points, and careers rather than sprints. — lesswrong.com
Maritaca's S5-nano Tops Brazilian AI Benchmarks
Maritaca AI's S5-nano, a 4-billion-parameter model, tops Brazilian benchmarks including national exams and outperforms rivals many times its size while keeping strong function calling and structured output. It ships with Think and Instruct modes, a long context window, and quantized versions for common local and server runtimes. — huggingface.co
AI Capability Growth Accelerates as Self-Improvement Nears
The pace of AI capability growth has roughly doubled, and Anthropic may imminently meet its own threshold for dramatically accelerated progress driven by AI. The evidence makes a fast intelligence explosion look plausible, though progress could instead stay jagged and concentrated in verifiable tasks. — lesswrong.com
Breeze-TTS-2 Runs Locally on Any GPU
Breeze-TTS-2, which speaks English and Mandarin, now runs on a C++ runtime that works across major GPU brands and falls back to CPU. The release speaks nearly twice as fast as real time with near-instant first audio, and adds voice cloning and free-form vocal-event tags. — huggingface.co
Notebooks Teach Hands-On Safety Evaluations
A new set of four Jupyter notebooks walks beginners from a first evaluation through benchmarks, AI judges, and autonomous agents using Inspect AI, the UK AI Security Institute's evaluation framework. The volunteer project also revised the official tutorial with a new Custom Solvers section, making trustworthy safety evaluations easier to build. — lesswrong.com