I was talking to a former mentee[1] last night and they turned out fully captured by GH Copilot. They use it because that's how they get access to Opus at a reasonable cost within a GitHub workflow, which they have now standardized-ish around (read: depend upon).
I described the bespoke, pluggable framework I built with forgejo and bottom line it's the same functionality, except:
I don't get to be locked in, and
I get full flexibility, in exchange for
I'm having higher costs, and having to maintain my own infra.
The interesting part is that they are gladly locked in and get reduced functionality (and quality) for the reduced cost, even if that means they are arguing with bots all day.
Their PRs are getting accepted the same way they were before: "LGTM", which is yolo-speak for "I'm not going to read all that". The only thing that changes for them is that instead of spending a week reading and understanding code, they now spend a day or 2 arguing with a bot in issue comments and when they get tired of it, "good enough". So they are 3-4x as "productive" as before.
But this productivity gain will not hold as long as there are bugs. And we know that they're there, more often than not, and that they're generally harder to detect than human mistakes because they are harder to spot, as the code looks really good at a glance. We also talked to this and indeed, bugs in released code has gone up. And avoidable bugs are expensive. Often more than 4x expensive.
We agreed that it is worth re-assessing the 4x. How much remains after taking the bugs into account? And what bonus reputation damage do you get from the additional bugs?
PS: Strange changes on Claude 4.8. It apologized in the middle of a review task?! WTF.
Apologies for the false alarm.
I'm literally paying for a bot to lie. This is getting out of hand.
from '05/'06 - time flies when you're having fun ugh ↩
I was talking to a former mentee[1] last night and they turned out fully captured by GH Copilot. They use it because that's how they get access to Opus at a reasonable cost within a GitHub workflow, which they have now standardized-ish around (read: depend upon).
I described the bespoke, pluggable framework I built with forgejo and bottom line it's the same functionality, except:
The interesting part is that they are gladly locked in and get reduced functionality (and quality) for the reduced cost, even if that means they are arguing with bots all day.
Their PRs are getting accepted the same way they were before: "LGTM", which is yolo-speak for "I'm not going to read all that". The only thing that changes for them is that instead of spending a week reading and understanding code, they now spend a day or 2 arguing with a bot in issue comments and when they get tired of it, "good enough". So they are 3-4x as "productive" as before.
But this productivity gain will not hold as long as there are bugs. And we know that they're there, more often than not, and that they're generally harder to detect than human mistakes because they are harder to spot, as the code looks really good at a glance. We also talked to this and indeed, bugs in released code has gone up. And avoidable bugs are expensive. Often more than 4x expensive.
We agreed that it is worth re-assessing the 4x. How much remains after taking the bugs into account? And what bonus reputation damage do you get from the additional bugs?
PS: Strange changes on Claude 4.8. It apologized in the middle of a review task?! WTF.
I'm literally paying for a bot to lie. This is getting out of hand.
from '05/'06 - time flies when you're having fun ugh ↩