Quote:
Originally Posted by Fuzz
One of the most annoying flaws is just how confidently wrong they tend to be. They need a "humbleness" slider you can crank up.
Even when you ask it to create a category for "unknown" and put anything it isn't 100% confident about in an unknown category, it will still confidently claim it knew where everything belongs, despite annotating items with "this could be in this or that category". I don't need an assistant that insists it is perfect all the time. That's my roll, not yours.
|
You can do this pretty effectively. One agent to code, one agent to PR, one agent to QA. You can prompt the latter two into being incredibly nitpicky. It's getting pretty amazing.
I ran our company trial of Copilot last year and deemed it was really only highly useful for generating boilerplate. We use Go a lot, so that's a pretty good benefit in itself for that language.
3 months ago I built a little Slack bot using a combination of Gemini, ChatGPT and Copilot. It was fairly frustrating how much it screwed up and took me a few days to get it working.
This week I asked Claude to re-write the whole thing and it was utterly flawless and massively improved. It took me about 20 minutes.
I've gone from a mild skeptic to "we're maybe all seriously screwed". I think 90% of the population thought they were screwed when farming machines came out. Factory workers thought they were screwed when robots showed up. Office workers thought they were screwed when computers landed on everyone's desk. We all found new things to move onto, granted with some pain in the transitions.
What worries me here is this is happening so fast and society can't adapt quick enough. It's insane to see this happening.