Mastodawn

RE: https://mastodon.social/@ekis/116298222342530412

been thinking about the key factors that lead to thinking LLMs as they exist now are replacement to programmers

1) misconception in how far a working demo is from production ready software, people without experience, think that once you have a working demo, that you are like 90% of the way done; when it reality you are more like 10%

2) its like gambling, sometimes, especially early, its possible to get good results, and so one ignores the diminishing returns, or all the times it didnt work

Show thread

she hacked you

obvio, there are ways to use them

I still believe a local LLM for FIM autocomplete is a nifty tool especially for writing scripts

I think they can be helpful with debugging by processing large amounts of logs

But the trend isn't towards that, its towards hyperscalers and agents which are incredibly slow, the resulting quality or the amount of required iteration incredibly unimpressive

Thats before the even addressing the fact that the math doesnt add up on the concept of agents, or the waste