Neurodigest for the Week (#110)

Neurodigest for the Week (#110) LLM - Leaks from Huang - On one of the slides, the CEO of Nvidia revealed the size of GPT-5 - apparently, only 2 trillion parameters. The size of OpenAI's flagship model has not increased for three years. - Everything is sad at Mistral - They released Small 4, which, according to benchmarks, loses to the September Qwen 3 Next, although the latter has half the active parameters. - Cursor Composer 2 - Cursor released the second generation of their model, which turned out to be a tune of the Chinese Kimi 2.5. Generative Models - Midjourney V8 alpha - The long-awaited alpha is here: it claims better prompt understanding, detail, text, and speed. But so far, the model is clearly undertrained. - Seoul World Model - Koreans from Naver packed part of Seoul into one model, which you can travel through and prompt, causing floods or Godzilla attacks. - Real-time video generation - Runway and NVIDIA released Gen-4.5 on the new Vera Rubin chips. HD video is generated with a delay of 100 ms, and it is based on GWM-1 - an autoregressive physical world simulator. Other - Pre-orders for DGX Station - Nvidia partners opened pre-orders for a desktop machine with one GB300 inside. Prices start at $87k. The first owner has already been found. - Musk's space data centers - Elon talked about plans to launch 100-kilowatt satellite data centers into orbit and plans to build his factory to produce terawatts of chips per year.