Comment by ricksunny | Hacker News Reader

Comment by ricksunny | Hacker News Reader

"...has a bachelor’s degree in Biotechnology and works as biomedical data curator. Her job is to go through scientific papers that publish RNA sequencing data. She understands the design of the experiment and cleans and annotates the data accordingly... Her work requires just enough domain expertise that it is been non-trivial to automate (until now).

I think you know where this is headed. The day I got off the GPT4 wait-list I asked it to do what Priya does on a daily basis. It got the answer wrong in the first try, but some chain-of-thought prompting and boom. GPT4 gave the correct answer in 1/10th the time it would take Priya and cost a lot less.

I doubt that she will have a job in six months."

I am a ChatGPT4 subscriber. And I have about a 100 biomedical research papers that need going through to pull curated experiment result data from (to extrapolate into broader quantitative in vivo results). Maybe ChatGPT4 could do it. I'm afraid to try it based on other experiences with it. It would take me about as long to vet any output from ChatGPT4 that looks correct as to go through myself de novo. And I'd have to vet each one. Even if only 1 in a hundred resulted in a wrong answer, I would need to know - which 1 of the 100?

If the article's subject starts looking for a job in 6 months, could I have the subject's contact? :)

That's currently the same issue it has for programming. It takes me longer to transcribe and factcheck the output than it would have for me to just write it from scratch. Writing the code stops being the bottleneck in software pretty quickly, and instead translating and merging business logic into existing software takes the most time.

For me, even if it were able to never make a mistake it would still be more annoying asking it questions over and over again to refine the output. Why would I use English (or whatever language) which is ambiguous compared to writing the code myself. I don't know I'm not buying the hype anymore. It is a cool tool but the idea that it would replace just writing code seems crazy to me.

Because most engineers document, in English and great detail, the designs of their systems. So the English already exists and will continue to. Taking this English and turning it into code will be the job of AI, not a programmer.

It works well for statically typed languages though. I use it for Rust and simply copy paste, allowing the compiler to tell me where it's wrong.

For Python or something similar, I wouldn't use it at all.

I do use ChatGPT very often for Python and SQL right now, but you really have to know what you are doing. The errors are sublte but such that the code more often than not would not run.

Yep, so I definitely wouldn't use it for code like that at all. I'd probably just tell it to give me more of an overview of the code.

But know that in the future, ChatGPT will be hooked up to a runtime system and compiler with which it can know how to fix errors itself. In fact, plugins already allow some of this.

Tangential, but if you haven't seen ChatPDF[1] yet, it is probably the best interface for the job you're talking about that I've found so far.

[1] https://chatpdf.com

I've seen ChatPDF, and if that's the best we got right now - Priya will be keeping her job for quite a few years.

The problem/solution is that you use ChatGPT to help you write a script/program that can do the transformation for you.

You can test and audit and understand such an artifact and trust that it works.

So the problem for the people depending on this work for income or to get their start in an industry remains.

Reply on news.ycombinator.com