AI agents spot Maya pyramid in NASA lidar
A Google engineer pointed 48 AI agents at NASA lidar scans, and they picked out a Maya civic center with a pyramid that Mexico's archaeology institute had not registered. A prior academic paper had already mapped the site, making the find a rediscovery. — pietromiotti.com
LLMs misread Yudkowsky and Soares on AI welfare
Frontier LLMs gave muddled accounts of Eliezer Yudkowsky's and Nate Soares's views, portraying the AI-safety researchers as dismissive of current models' moral status. The pair's actual posts and essays treat machine welfare as a serious open question. — lesswrong.com
Qwen3.8 27B runs locally via Apple MLX
The quantized Qwen3.8 27B model runs on Apple's MLX framework, so local coding agents and chat clients can use it without a cloud API. — huggingface.co
Unsloth quantizes Qwen's fast image model
The GGUF conversion of Qwen's turbo image-generation model cuts file size dramatically while keeping quantized output close to the original. The compressed model plugs into local image tools. — huggingface.co
Gemini 4 report clashes with DeepMind on chain-of-thought
A contradiction hangs over Gemini 4: the safety report includes a reward that observes the model's private reasoning during training, while Google DeepMind's public position rules out any chain-of-thought training. The discrepancy matters because shaping a model's reasoning can teach it to hide that reasoning. — lesswrong.com