Transcription, translation, material/product defect detection, weather forecasting/early warning systems, OCR, spam filtering, protein folding, tumor segmentation, drug discovery/interaction prediction, etc. seem fairly promising to me, with machine learning approaches often blowing traditional approaches out of the water.
I feel the idea that there's only "one thing it's supposed to be good at (writing code)" is largely down to availability bias, in that we're in programming circles where LLM code completion gets talked about a lot. ChatGPT reportedly has 700 million weekly users - I'd assume many using it for tasks that I'm not even aware of (automating some tedious/repetitive part of their job).
It sucks at the one thing it's supposed to be good at (writing code).
It sucks at replacing software engineers. So yes, if you expect it to design a complex scalable system and implement it for you from start to finish, then sure, you can call it bad at "writing code".
I don't care about that use-case. I think I am decently good at writing code, and I rather enjoy doing it. I find Claude Code/Gemini CLI extremely helpful at both saving me lots of time by freeing me from dealing with annoying boilerplate, so I can focus more on actual system design (which LLMs fail at terribly, if we are talking about real production apps that need to scale) and more difficult/fun parts of code (which LLMs cannot handle either).
That's the real power of it, multiplying the productive output of good SWEs + making the work feel more enjoyable for them, by letting those SWEs focus on actual tricky/difficult parts, instead of forcing them to spend a good half of the time just dealing with boilerplate.
And I am not even gonna bother getting into tons of other non-coding tasks it is already very useful of. I find it amazing that I can now not only get a transcript of my work meetings (which is already massively helpful for me to review later, as opposed to listening to a ~25min video recording of it), but also ask an LLM to summarize it for me or parse/extract info from it.
Comments
AI also doesn't have good use cases yet. It sucks at the one thing it's supposed to be good at (writing code).
Transcription, translation, material/product defect detection, weather forecasting/early warning systems, OCR, spam filtering, protein folding, tumor segmentation, drug discovery/interaction prediction, etc. seem fairly promising to me, with machine learning approaches often blowing traditional approaches out of the water.
I feel the idea that there's only "one thing it's supposed to be good at (writing code)" is largely down to availability bias, in that we're in programming circles where LLM code completion gets talked about a lot. ChatGPT reportedly has 700 million weekly users - I'd assume many using it for tasks that I'm not even aware of (automating some tedious/repetitive part of their job).
It sucks at replacing software engineers. So yes, if you expect it to design a complex scalable system and implement it for you from start to finish, then sure, you can call it bad at "writing code".
I don't care about that use-case. I think I am decently good at writing code, and I rather enjoy doing it. I find Claude Code/Gemini CLI extremely helpful at both saving me lots of time by freeing me from dealing with annoying boilerplate, so I can focus more on actual system design (which LLMs fail at terribly, if we are talking about real production apps that need to scale) and more difficult/fun parts of code (which LLMs cannot handle either).
That's the real power of it, multiplying the productive output of good SWEs + making the work feel more enjoyable for them, by letting those SWEs focus on actual tricky/difficult parts, instead of forcing them to spend a good half of the time just dealing with boilerplate.
And I am not even gonna bother getting into tons of other non-coding tasks it is already very useful of. I find it amazing that I can now not only get a transcript of my work meetings (which is already massively helpful for me to review later, as opposed to listening to a ~25min video recording of it), but also ask an LLM to summarize it for me or parse/extract info from it.
If Terence Tao says it can save hours of work for him, you might be holding it wrong.