I develop AI agents rn as part time for my work and have yet to see one that can perform a real task unsupervised on their own. It’s not what agents are made for at all - they’re only capable of being an assistant or annotate, summarize data etc. Which is very useful but in an entirely different context.
No agent can create features or even reliably fix bugs on their own yet and probably not for next few years at least. This is because having a dude at 50$ hour is much more reliable than any AI agent long term. If you need to roll back a regression bug introduced by an AI agent it’ll cost you 10-20 developer hours as minimum which negates any value you’ve gained already. Now you spent 1,000$ fix for your 50$ agent run where a person could have done that for 200$. Not to mention regression bugs are so incredibly expensive to fix and maintain so it’ll all scale exponentially. Not to mention liability of not having human oversight - what if the agent stops working? You’ll have to onboarding someone on an entire code base which would take days as very minimum.
So his take on ai agents doing work is pretty dumb for the time being.
That being said, AI tool use proficiency test is very much unavoidable, I don’t see any software company not using AI assistants so anyone who doesn’t will simply not get hired. Its like coding in notepad - yeah you can do it but its not a signal you want to send to your team cause you’d look stupid.
Honestly, AI coding assistants (as in the ones working like auto-complete in the code editor) are very close to useless unless maybe you work in one of those languages like Java that are extremely verbose and lack expressiveness. I tried using a few of them for a while but it got to the point where I forgot to turn them on a few times (they do take up too much VRAM to keep running when not in use) and I didn’t even notice any productivity problems from not having them available.
That said, conversational AI can sometimes be quite useful to figure out which library to look at for a given task or how to approach a problem.
Honestly, AI coding assistants (as in the ones working like auto-complete in the code editor) are very close to useless unless maybe you work in one of those languages like Java that are extremely verbose and lack expressiveness.
Hard disagree. They’re not writing anything on their own, no, but my stack saves at least 75% of my time, and I work full-stack across pieces in 5 different languages.
Cursor + Claude was the latest big shift for me, maybe two months ago? If you haven’t tried them, it was a huge bump in utility
If you spend 75% of your time writing code you are in a highly unusual coding position. Most programmers spend a very high percentage of their time understanding the problem domain and on other parts of figuring out requirements and translating them into something resembling some sort of semi-formal understanding of what the program actually needs to do. The low level detailed code writing is very rarely a bottleneck.
I develop AI agents rn as part time for my work and have yet to see one that can perform a real task unsupervised on their own. It’s not what agents are made for at all - they’re only capable of being an assistant or annotate, summarize data etc. Which is very useful but in an entirely different context.
No agent can create features or even reliably fix bugs on their own yet and probably not for next few years at least. This is because having a dude at 50$ hour is much more reliable than any AI agent long term. If you need to roll back a regression bug introduced by an AI agent it’ll cost you 10-20 developer hours as minimum which negates any value you’ve gained already. Now you spent 1,000$ fix for your 50$ agent run where a person could have done that for 200$. Not to mention regression bugs are so incredibly expensive to fix and maintain so it’ll all scale exponentially. Not to mention liability of not having human oversight - what if the agent stops working? You’ll have to onboarding someone on an entire code base which would take days as very minimum.
So his take on ai agents doing work is pretty dumb for the time being.
That being said, AI tool use proficiency test is very much unavoidable, I don’t see any software company not using AI assistants so anyone who doesn’t will simply not get hired. Its like coding in notepad - yeah you can do it but its not a signal you want to send to your team cause you’d look stupid.
Honestly, AI coding assistants (as in the ones working like auto-complete in the code editor) are very close to useless unless maybe you work in one of those languages like Java that are extremely verbose and lack expressiveness. I tried using a few of them for a while but it got to the point where I forgot to turn them on a few times (they do take up too much VRAM to keep running when not in use) and I didn’t even notice any productivity problems from not having them available.
That said, conversational AI can sometimes be quite useful to figure out which library to look at for a given task or how to approach a problem.
Hard disagree. They’re not writing anything on their own, no, but my stack saves at least 75% of my time, and I work full-stack across pieces in 5 different languages.
Cursor + Claude was the latest big shift for me, maybe two months ago? If you haven’t tried them, it was a huge bump in utility
If you spend 75% of your time writing code you are in a highly unusual coding position. Most programmers spend a very high percentage of their time understanding the problem domain and on other parts of figuring out requirements and translating them into something resembling some sort of semi-formal understanding of what the program actually needs to do. The low level detailed code writing is very rarely a bottleneck.