Closed-API vs Open-source continues: RLHF, ChatGPT, data moats

An introduction blog post and lecture on reinforcement learning from human feedback (RLHF) -- start here if RLHF is confusing (first page results on Google for RLHF). A paper on Measuring Data in machine learning, defining a future field of research that'll improve many ML systems. RLHF is being heavily bet on in industry and has some very different properties than other generative models that make it less suited to open-source: hard to get the data via expensive human annotations, hiring experts in multiple areas like deep RL, and strong potential for internal use-cases like search. As you all know, OpenAI release ChatGPT and it quickly rose to over 1 million users and is the leading talking point in OpenAI's new fundraising rounds. It is the company that stands to represent the closed-API business model.