> H will itself contain information (entailment) from f, but never the other way around: f is not entailed by A or H, it is fixed, and only entailed externally by the design and training steps.
This is only true for the first release. Consider that OpenAI has been actively collecting input/output pairs from their users, and then retraining and updating the model. Thus A and H have impacted ChatGPT. This in turn effects how people interact with the system.
You can certainly constrain f to a single point in time, but most people will not. They think of ChatGPT as f and that f is changing or evolving (in the non-literal sense). So depending on how you look at it, f is indeed editable. Opinions will differ here and there is no right answer.
Google wrote a good paper on this feedback loop almost 10 years ago called "Machine Learning: The High Interest Credit Card of Technical Debt" that is even more relevant today.
This is only true for the first release. Consider that OpenAI has been actively collecting input/output pairs from their users, and then retraining and updating the model. Thus A and H have impacted ChatGPT. This in turn effects how people interact with the system.
You can certainly constrain f to a single point in time, but most people will not. They think of ChatGPT as f and that f is changing or evolving (in the non-literal sense). So depending on how you look at it, f is indeed editable. Opinions will differ here and there is no right answer.
Google wrote a good paper on this feedback loop almost 10 years ago called "Machine Learning: The High Interest Credit Card of Technical Debt" that is even more relevant today.
https://research.google/pubs/pub43146/