I could see such productivity gains being possible, if only because the current tooling around LLMs is terrible. The fact that we have 30 blog pieces per day making the front page of Hacker News about someone’s convoluted system to guide LLM output to something reasonable is absurd. There needs to be standardization in tooling, and it needs to be open source. Then, and only then IMO, will we see huge productivity gains.

But, at that point I think the big players’ moats will have dried up. Local models will probably be sufficient for 99% of daily office worker tasks.

So I disagree with TFA’s premise. I think this fear is probably shared amongst the LLM giants, and they’re still hoping that neural network transformers are somehow the path to AGI (probably not, imo).