Raw LLM Responses

We know the mechanics of how we built it. That is not the same as knowing why it is able to do XYZ thing we have just recently discovered it can do. Yes, the mechanics are complex and not widely understood, but so are chip fabs; the difference is that new chips are not regularly outputting novel new abilities as we make them larger and more powerful. When we do find something unexpected in a new chip design, it's generally "the thermals are worse/better than projected by a fraction of a percent" and not "the output of the device now passes all existing theory of mind tests". When a chip gets significant new abilities like DLSS, it's because engineers planned to add those new capabilities and worked very hard to do so. Scientists and engineers have planned and built the GPT models, but we know from how they're building them they don't have a checklist of psychological tests they're putting in the answers for - they're making the models larger and more efficient and improving the training data, at which point it gains new capabilities. We know it's gaining new capabilities they didn't "intend to put in it" because people completely outside of OpenAI are testing it and discovering new things that OpenAI didn't realise it could do according to all available information. We know how we built it, but it's completely reasonable to say we don't know *why* it works in the way we know how an ICE works, or how TSMC as an institution knows how a new chip works.