aviinuo

Nvidia GPUs let you load encrypted weights onto them and to decrypt on chip so Google would need to have some sort of Nvidia management firmware attack to read from HBM while the GPU is live

A small stake on the order of a quarter trillion dollars

I'm not getting any refusals but it just seems like a bad model or at least broken at the moment. I have a task of taking a messy research code base and porting it into a clean project structure skeleton that I commonly use. Gemini 3.5 Pro High in antigravity cli takes less than 5 minutes and did a good job. Fable 5 High took 30 minutes to port some of the code, then just copied the rest to a folder called "reference" and decided the task was done. No code cleanup or anything. Had to clarify multiple times (which Gemini did not need) and its still going more than an hour later still not having finished.

Previously when I did similar tasks with Opus 4.7/4.8 and GPT 5.5 I had no problems.

Pro Blackwell 6000 is just a 5090 with more VRAM. It does not have the tcgen05 (5th gen tensor core) instructions despite the "5th gen tensor core) branding and thus do not support any optimized Blackwell (sm100) kernels.

Every Blackwell card other than the (G)B100, (G)B200, (G)B300 and Jetson Thor, use the Ampere tensor core instruction (mma.sync) but with fp4/6/8 added on. Beyond that the DGX Spark (which is advertised as having the same architecture as B200) has especially weak (not tcgen05) tensor cores that have a very narrow operating window and low utilization.

Something seems off. For the 4kx4kx4k fp16 GEMM, cutlass is like 3x faster than this.