Industry
65Only Sun, October 4, 2026 · All dates →
A Hacker News discussion argues that AI agents do not need memory mechanisms but should instead rely on documentation to structure context. The post itself is brief and mainly links to the HN comment thread.
Developers released Kyojin, an ExLlamaV3-based engine for AMD Strix Halo (ROCm), letting two 300B-class MoE models each run on one 128GB mini PC. GLM-5.3-Flash prefills at ~580 tok/s; MiMo-V2.6-Flash decodes up to 44 tok/s, with KLD and Top-1 agreement versus official FP8 reported.
A developer released Anyworld, a browser-based multiplayer text RPG in which a local LLM run via llama.cpp (cloud APIs also supported) acts as dungeon master. Players join by browser link with no setup, actions are resolved by the model, and real dice rolls are handled by Python RNG, enabling custom scenarios.
Hugging Face's post-training team published a long guide on training open models across different coding harnesses using open-source libraries like TRL and the Harbor framework. The recipe shows how to get the best performance on custom harnesses with any open model, and the authors welcome community feedback.
Researchers combined artificial intelligence with archaeoacoustic techniques to reinterpret two 5th-century BC Etruscan painted tombs. The study suggests these elite tombs were designed to embed sensory sound experiences, drawing on projects at Oxford and Duke University.
That is everything