As I said earlier, I want to find a way to percieve a document as the target to understand, not a materical to learn a linguistic pattern. The thing we want to get from the document is the content itself, not some sort of superficial meaning. Getting a content is different to getting a meaning.
Some reframed the problem into next-token generation, to shape the model into a function and work around this open-endedness of intelligence. Meanwhile, our brain is more or an individual decision maker, rather than a passive function. We have our internal intents that drives oneself to react or do next steps.
Overall, we are considering the model as by and large a ’thinking tool’, where we expect some proposed usefulness. We are not externalizing our internal drives, emotions, or preferences while they are what we actually feel physically. Denying all those intuitive experiences and trying to mimic the consequent intelligent decisions are somewhat detached from externalizing human intelligence.