Options are inclusive: if this happens tomorrow then all options resolve to YES.

This market will of course resolve somewhat subjectively. The overall idea is "there is a model that I can use instead of hiring an intern". Some things I expect such a model to be able to do:

• Take a paper as input and give a runnable rough draft implementation (does not need to be bug free)

• Experiment with hyperparameters for an existing model ("grad student descent")

• Make some nice loss curve graphs (acceptable to use third-party tools for this e.g. WandB)

• Take two existing algorithms/models and run comparisons between them ("Model A performs better on this benchmark", "Model B trains 3x faster")

• Avoid trivial mistakes like no validation set, testing on your train set, etc

Some things I don't expect it to do:

• Develop novel ML algorithms

• Any kind of non-trivial distributed training (anything more complicated then "run your code with this flag to make it distributed")

• Any kind of performance optimization (e.g. writing Triton kernels)

I will give myself one month (2024-07-20) to modify the resolution criteria based on feedback.