Neurodigest for the Week (#100)

Neurodigest for the Week (#100) Jubilee and the last before the New Year, happy holidays! LMM - NanoGPT in 116 seconds — Training of NanoGPT has accelerated four times over the year, reaching the benchmark of 116 seconds. This is an accessible baseline for testing optimizations, from which, for example, the Muon optimizer has grown. Generative Models - Qwen-Image-Edit — an open analogue of Nano Banana — with integrated LoRA features and an Apache 2.0 license. Others - Microsoft is abandoning C/C++ — Microsoft is considering abandoning C/C++ by 2030, using AI for ambitious code porting to Rust. - ML Education — Analysis of approaches to training AI specialists in the USA, China, and Russia. - Nvidia buys Groq for $20 billion — Nvidia "buys" Groq for $20 billion through a "licensing agreement," acquiring technologies and key employees bypassing regulators.