Coarse-to-Fine Model Generates Text in Parallel
Next-Concept Modeling splits text generation into a hierarchy of discrete codes and predicts all codes in a level simultaneously, bypassing the sequential loop that slows both autoregressive and diffusion models. With matched parameter counts, it achieves better generative perplexity than the strongest diffusion language model while running 21 times faster. — arxiv.org
Six GPUs Share a Single PCIe Slot
A hobbyist build squeezes six graphics cards into one slot on a consumer AMD board using a PCIe switch card, which gives each card a dedicated link rather than the limited lanes that bifurcation offers. The Radeon cards run a large Qwen model for many concurrent clients, with peer-to-peer traffic staying on the switch. — doug.sh
AI Agent Builds a Working Kernel in Eleven Days
Ginux, an independent Rust-first x86-64 kernel, was created by a single persistent coding agent over eleven days, following human-set architecture, milestones, and acceptance criteria. It boots via UEFI and runs selected unmodified Alpine Linux programs, but remains a research prototype with narrow compatibility and no security audit. — zenodo.org
South African Author Defends AI-Written Books
Coral Hart, the pseudonym of a South African writer publishing under many pen names, has released over 370 books drafted with AI models since early last year, and faces death threats and a flood of abuse for it. She now trains others in her method and is preparing software, Authorythm, for AI-assisted authors. — enca.com
Opting Out of AI Search Costs Traffic
Zyppy stopped appearing in Google's AI search features for several weeks, and its AI-driven impressions vanished almost immediately while its regular top ranking held. Comparing the opted-out stretch with the opted-in one put AI search at roughly a quarter of the site's total search clicks. — signal.zyppy.com