LLM-assisted coding can be used to good effect, but I think it's pretty clear at this point that its impact has so far been a net negative on productivity in real terms.

Replies (24)

My take: If you stay disciplined, you can do stuff which in 2020 you would have never attempted because you thought it would be too complicated. For example -- I am a huge Postgres user and tend to use expensive managed Postgres servers because... postgres admin is complicated. But today, in about 3 hours, I set up a pretty complicated remote follower for a local postgres instance, along with bandwidth limits, all the networking stuff, all running in docker on both ends. Something which would have taken me 3 days of puzzling to get right in 2020.
Completely agree with that. I think the interesting point is that just because someone can doesn't mean they should — is that fancy database setup, DSL parser, agent framework, mesh network, whatever providing value to its users? LLM coding is more often than not a complete distraction from stuff that matters.
The future is not evenly distributed. Getting a real productivity net-positive out of LLMs can definitly come, after a period of investment, like all new tools. I'm solidly in net-positive territory now, but I don't think most are, including most who think they are. "AI" is just a database with code generation features. We've had new databases and code generators lots of times. This generation is particularly slippery and deceptive to the user. Everybody needs to sober up if they really want to be effective.
But it can also give you time to think, because it wont take you a year to do something. So all in all, its a tool. As smart as its user at this point.
Agree, it's a tool. I am finding it useful when used in a disciplined way with a human in the loop at literally every step. I think the current tooling gives the LLM far too much leeway in general.
Resources needed to chase every more parameters aren't cheap. It would be interesting to see the two graphed against each other.