OpenAI experiment finds that sparse models could give AI builders the tools to debug neural networks

OpenAI sought to understand how models make decisions and adhere to behaviors. It went down to the neuron level to find out.