[siliconc0w]

I'm an "overhang" bear, while incremental progress will be made - it'll come at too high of a cost. The larger, slower, more expensive models have limited applications(mostly fine-tuning the economical ones) so there will be ultimately less investment. Unless there is a new unlock, we'll be left with mediocre 3.5 models for most real-time apps and excitement will fade.

[gmerc]

ChatGPT is one giant training data acquisition project. Including pro. it’s painfully obvious when you see that any useful usecase eventually moves to the API (GPT 4 web/app flat out refuses to code more than a function or two at a time now; it had no issues with multiple files before).

Everyone is providing the greatest training set you can now find (since all other venues for training data will gradually disappear behind walls and fees)