imec benchmarks self-hosted open-weight models vs APIs for coding agents

A LocalLLaMA user reports 22 tok/s sustained (48 tok/s spikes) running Qwen 3.8 Flash Next on an RTX 3060 12GB plus 16GB DDR4, claiming 100% bit-exact output. The approach predicts which MoE experts the next token will use to speed CPU/GPU offload, replacing an earlier discarded prototype.

A YouTube commentary video from channel AI Abyss frames the "Godfather of AI" as having left Big Tech to argue that digital intelligence is already out-evolving biology. It claims humanity's belief it can stay in control is a "mathematical illusion." No specific model, benchmark, or paper is cited in the snippet.

Scott Aaronson, director of the Quantum Information Center at UT Austin, joins Brian Greene to examine what happens to mathematics now that AI systems are advancing on problems once thought human-only. The discussion asks whether AI has already surpassed human mathematicians and what work is left for people.

The 95% figure came from a misread funnel and just 153 survey responses, per the critique. A year of follow-up data from McKinsey (1,719 executives, 97 countries), Deloitte, Capgemini and government statistics still shows most enterprise AI spending hasn't reached the P&L.