Raw LLM Responses

@SolarSands My apologies, I wasn't trying to be smug or demean a lack of credentials. I was trying to point out that if the inventor of a process indicates one perspective is correct and others indicate the opposite, the burden of proof is on the latter side to show the originating authority is incorrect (inverting that relationship is how Trump spreads misinformation). As for what to do about it, Chollet's papers are available on Arxiv last I checked, they are very active on twitter (and searchable for terms) and he also has a substack. Though a more accessible option might be Ed Newton-Rex (former VP at Stability) who is also active on twitter and has published discussion articles. For some visual applied examples, if those help, of how these systems are determinative and interpolative, you could also see the exhibits in the courts filing in the Andersen vs Stability case, made available at the "imagegeneratorlitigation" website. As an example of fundamental misunderstanding of function, if that also helps, at 12:37 you discuss "overfitting" incorrectly as a failure of the model system, versus understanding that it is a failure point in the dataset (and/or weighting - Laion's aesthetic rank scoring during data processing for example significantly impacts image generation and how closely images fit to certain artist names [names being part of the metadata in latent space]) [model and data are distinct components of a diffusion system] - Even though you correctly then discuss one of the intervention methods in the latter (deduplication). And for visual examples, some models and companies also deliberately weight towards stronger fitting/less generalization (interpolation again), Reid Southen and Gary Marcus have documented how Midjourney does so, in IEEE Spectrum and then afterward with their V6 model comparisons. The same issues are also prevalent in text/LLMs because of the similar neural net approach to the data and the determinative nature (see the recent "Does your LLM truly unlearn?" in arxiv which discusses how difficult it is for these systems to "unlearn"/how easy it is to restore the information - This is also related to how these systems are compressive in the nature of their generalization [something not disputed even by the CEOs of the companies, see Emad Mostaque's former CEO of Stability's public statements]). I hope that this response is more helpful. Be well.