AI agent hosting, Google's Opal 2.0, and a couple of clever tools built specifically for Claude Code users.
Breakthrough hallucination research, Sonnet 4.6's identity crisis, Karpathy on the coding agent inflection point, and Perplexity's new multi-model computer system.
Tsinghua researchers conducted a systematic investigation into "H-Neurons"—specific neurons in LLMs that cause hallucinations. They identified these neurons, showed deactivating them reduces hallucination rates significantly, and traced their origins to noisy training data.
A concrete, mechanistic understanding that could lead to architectural solutions rather than the usual band-aids of RLHF or output filtering.
Quick hits — benchmarks, policy, and local model tools.
Google expands Flow into a creative suite, a solid local image manager ships cross-platform, and the censorship debate heats up between Flux and Qwen Image.
A Launchpad replacement, local AI image editing, a Quicksilver-inspired launcher, and Retool quietly kills self-hosted plans.