[datsci_est_2015]

I don’t buy it. Properly leveraging LLMs to generate stable and extendable systems is mentally exhausting (i.e. highly demanding of intelligent thought), especially given the poor quality and churn within the harness ecosystem.

Now, if you’re creating trivial, unstable, or nonextendable systems maybe this doesn’t apply. And maybe I have long overestimated the work that SWEs have done.

[jansport123]

i use claude a lot and i find that it is best applied in domains in which i am already a master. I tried applying it to domain's im unfamiliar with and i found that i produced stuff but as time went on i understood what i produced less and i almost felt like i do after binge watching a netflix show, 2 weeks later i barely remember any of the details. I wonder how much you need to "do" to learn and remember. LLM's give you a shortcut to doing and so you probably aren't learning either. It's like when you watch a professor write a proof and it makes sense while listening to the professor but at home you have difficulty deriving it. LLM's give me the same sort of feeling. I think the way forward is still going to be doing things manually to learn and using LLM's once you've mastered an area and people who don't understand this fact are going to slowly descend down a hill and forget how to depend on their own thinking.