Maybe some new technique will change that, but it's not guaranteed. At this point I think we can safely surmise that scaling isn't the answer.
Maybe some new technique will change that, but it's not guaranteed. At this point I think we can safely surmise that scaling isn't the answer.
Why? To this day still they’re the showcase of what LLMs “can” do for (to) a line of work, but they’re the only ones with all the relevant information online.
For programming, there’s decades of textbooks, online docs, bug tracker tickets, source code repositories, troubleshooting on forums, all laying out how a profession is exercised from start to finish.
There’s hardly a fraction of this to automate the tasks of the average Joe who does some paperwork the model has never seen, who’s applying some rough procedures we would call “heuristics” to some spreadsheets and emails, and has to escalate to his supervisor for things out of code several times a day.