Automatic snow chains deployment systems like the Onspot mechanism, allow vehicles to increase their traction on snow and ice with a relatively immediate activation triggered from the cab.
Depending on the context window, the "used" RAM varies, but the cache in this case is nearly the full 1TB.
I am finding ~64K context to be the sweet spot, the full 128K is very slow, like 1 t/s or less, but no real gains going less than 64K.
"Tighten export controls on chips" is a loser's attitude.
- DeepSeek can already run inference on Huawei Ascend chips
- This will only push China to accelerate GPU development and create its own CUDA
When did the US start fearing competition?
And why oppose open-source AI?
All those AI jobs no longer feel tractive anymore and overall feels over-inflated/hyped. Pull back was essential and DeepSeek made the decision easier.
About system safety
Always check file usage
lsof <path> # Is the file open by a process?
Always try to truncate instead of delete if unsure of its usage
truncate -s 0 <path>
after releasing a Sparse Autoencoder for llama 1B, i'm happy to announce that we've scaled up to 8 billion parameter models, having a trained a SAE for DeepSeek R1 Distill Llama 8B, and releasing it as open source.