Where does ChatGPT's tacit machine knowledge actually originate?

The paper extends its own case study of NotCo to argue that the training process behind ChatGPT is a second instance of the same human-machine knowledge spiral, now operating "inside an AI laboratory." It points to RLHF (reinforcement learning from human feedback): "before ChatGPT could converse with the world, human trainers staged conversations in which they played both the user and the assistant. They were given model-written suggestions to help compose their responses. They then ranked alternative replies generated by the model, and those judgments shaped subsequent training (Ouyang et al., 2022)." The paper's claim is that this sequence — vast explicit text encoded into the model, human demonstrations and rankings embedding tacit judgments about what counts as a useful response, and model outputs that expose those learned capabilities for inspection — is "the same movements seen at NotCo appear here at a different scale," making RLHF a knowledge-creation spiral rather than only a training procedure.

The mechanism the paper gives is the move between tacit and explicit knowledge Nonaka originally described, now looping through a model. Human raters' judgments about which reply is better are themselves tacit — acquired through experience, in the paper's Nonaka-derived vocabulary — until they are made explicit as rankings and demonstrations the model can train on. The model's outputs then let "the people building and using the system acquire new knowledge about what the model can do and where it remains limited," the explicit-to-tacit half of the loop running in the other direction. Because the resulting tacit machine knowledge is portable, this one lab-internal spiral does not stay inside the lab: "through software products and APIs, tacit machine knowledge formed inside one laboratory is now becoming central to people and organizations around the world... the spiral operating inside a small number of AI laboratories is therefore feeding the knowledge spirals of countless other companies — at scale."

This is the companion claim to Can machine tacit knowledge spread across organizations unlike human tacit knowledge? from the same paper: that note states the general property — machine tacit knowledge travels, human tacit knowledge doesn't — and this one locates where the traveling knowledge is produced, inside the RLHF loop of a frontier lab, before it reaches companies like NotCo. It also bears on Do language models possess tacit knowledge in Davies' sense?: that note asks whether a trained model's weights causally encode tacit knowledge, while this paper instead asks where that tacit knowledge comes from organizationally, treating RLHF's human rankings as the source material rather than asking what the resulting weights represent.

The excerpt offers RLHF as an illustration of Nonaka's framework, not a study of the RLHF process itself — it does not describe how rater disagreements were resolved, how many raters or rounds were involved beyond the single cited Ouyang et al. (2022) description, or whether the knowledge produced this way differs from tacit knowledge produced through other training methods such as pretraining alone or constitutional AI. The paper's "at scale" claim about the spiral feeding "countless other companies" is asserted rather than measured: it names the mechanism — software products and APIs — by which tacit machine knowledge could propagate, without counting how many organizations actually receive and re-circulate it the way NotCo did.