hey, it's a new day in machine learning: scaling is the law, models need to be sharded like horcruxes, new ai chips drop like limited edition sneakers, and oh ya, rl is back. i'm hardik and we're ml engineers (ex-google, ex-apple) building a modern kaggle  so you can upskill in modern ml  by,

• optimizing ml workloads, e.g. writing your own cuda kernels for various attention mechanisms in llm inference
• implementing distributed algorithms, e.g. sequence parallelism for supporting long context windows
• programming against a wide breadth of ai accelerators—from nvidia gpus, to amd gpus and google tpus
• diving into model architectures beyond the transformer, e.g. language diffusion and state-space models
• using reinforcement learning to train agents in realistic environments, e.g. computer use for your favorite app

sign-up to get early access to our beta:

our goal is to tear down all barriers to ml education, so we will offer compute for free...well, at least for the first batch of people who sign-up.

join our discord community to connect with other learners, get technical and career help, and stay in the loop.

resources

some educational resources we've released:

• tensordiagram, a python library for visualizing tensors (in pytorch, jax, numpy, etc.) to help with learning and debugging
• Tensor Puzzles 2, a collection of pytorch problems to train your tensor programming muscles