nancarrow

90% of crud filtering mechanisms are crud.

love how clearly vibecoded this is. the cloudflare worker architecture + the ascii diagram is a dead giveaway. nothing wrong with that, it's just really obvious.

the split architecture offers absolutely zero security benefits outside of not exposing a server process on your mac to the open internet (assuming you only let it connect to cloudflare) - it's just a convenient place to spin up a thin JS layer that calls model APIs and connects to your mac.

anyways i think this is a neat weekend vibecoding project but IMO it needs a lot more design thought to really be useful and not be a huge security issue.

"some" is doing a lot of lifting. # of tokens * # of layers * head dimension * # of heads * 2 (K+V vectors) * 4-16bits (depending on quantization)

> How many Matlab users know things could be better?

not very many in my experience - matlab rots the brain

the subclass of "Reproducing kernel Hilbert spaces" form the basis for a large class of ML algorithms (kernel methods)

why? i like Julia but it's not a great general purpose language IMO.

the CUDA library versions are very coupled to the hosts' CUDA driver versions and the container system itself (Docker) needs special code to link the two

It's more related to the law of large numbers than the CLT

The generator implicitly models a joint probability of data by being a generative process that one can draw samples from. GAN training (at least under certain simplifying assumptions) minimizes the JS divergence between the generator distribution and the data distribution.