LocalLLaMA user tests Qwen 3.8 FN NVFP4 on image prompt challenge

In an October 17 Per Jacobsson lecture at the IMF meeting in Bangkok, the Import AI author argued AI is reaching the frontier of human science, sometimes autonomously, and told economists and policymakers to expect rapid change. He framed the talk around foresight and restraint in sharing AI's benefits.

MiniMax M3.1 Flash, previously known as "Space Bunny" on OpenRouter, costs $0.15/$0.60 per 1M input/output tokens. Weights arrive October 23, and MiniMax is also preparing a Pro variant.

In a Futurology conversation with Dawn Nakagawa, Tao describes a project built around 22 million algebra problems and argues useful failures should be shared openly. He also explains how experts can extract value from unreliable AI tools.

Agent-assisted CuTe DSL decode path on NVIDIA GB300 hits over 840 tokens/s at batch 1 on 16 GPUs with DSpark, with 1.5–2× higher decode throughput in several tested configs and up to 1.8× higher P90 interactivity on AgentX.

Google DeepMind Chief AI Readiness Officer Lila Ibrahim speaks with ET's Tech Editor Surabhi Agarwal about whether AI companies can credibly warn about risks while racing to build more powerful models. The conversation also covers who ensures AI's benefits reach beyond the people and countries already ahead.

HardenedLinux's GPLv3 SonicBoom runtime has no Python at build or run time — one shared library, a C ABI, and Guile Scheme as its first-class language. The community is now crowdsourcing GPU benchmarks for v0.0.3.