Hacker News Reader

Comment by pythonguython | Hacker News Reader

AI can’t do our jobs today, but we’re only 2.5 years from the release of chatGPT. The performance of these models might plateau today, but we simply don’t know. If they continue to improve at the current rate for 3-5 more years, it’s hard for me to see how human input would be useful at all in engineering.

View on HN

And if my plane keeps the take-off acceleration up for 7 months we'd be at 95% the speed of light by then.

I dont think its especially unreasonable to assume that these models will continue to improve. Every year since chatGPT has seen incredible advancements, that will end eventually but why do you think it is now?

> Every year since chatGPT has seen incredible advancements

Advancements in what exact areas? My time using GitHub Copilot years ago was more successful for the simple act of coding than my more recent one trying out Cursor with Claude Sonnet 3.5. I'm not really seeing what these massive advancements have been, and realistically none of these LLMs are more useful than a very, very bad junior programmer when it comes to anything that couldn't already be looked up but is simply faster to ask.

> realistically none of these LLMs are more useful than a very, very bad junior programmer

This is an incredible achievement. 5 years ago chatbots and NLP AI couldnt do shit. 2 years ago they were worthless for programming. Last year they were only useful to programmers as autocomplete. Now they replace juniors. There has been obvious improvement year after year and it hasnt been minor

I imagine many people in 1970 were incredulous that we’d have transistors with 20 nm pitch width.

They will never be creative, and creativity is a pretty big deal.

To the extent it’s measurable, LLMs are becoming more creative as the models improve. I think it’s a bold statement to say they’ll NEVER be creative. Once again, we’ll have to see. Creativity very well could be emergent from training on large datasets. But also it might not be. I recommend not speaking in such absolutes about a technology that is improving every day.

"To the extent it's measurable" is very load-bearing in the semantics here. A lot of "creativity" is very hard to measure.

I agree, and I think most people would say the current models would rank low on creativity metrics however we define them. But to the main point, I don’t see how the quality we call creativity is unique to biological computing machines vs electronic computing machines. Maybe one day we’ll conclusively declare creativity to be a human trait only, but in 2025 that is not a closed question - however it is measured.

We were talking about LLM here, not computing machines in general. LLM are trained to mimic not to produce novel things, so a person can easily think LLM wont get creative even though some computer program in the future could.

> LLM are trained to mimic not to produce novel things

Which LLM? That’s not the purpose of training for any model that I know of.

Training LLMs is literally finding sets of numbers that make them better at mimicking human language.

Difficult to measure, but trivial to define.

Creativity means play, as in not following rules, adding something of yourself.

Something a computer just can't do.

Most software engineering jobs aren't about creativity, but about putting some requirements stated in a slightly vague fashion, and actualizing it for the stakeholder to view and review (and adjust as needed).

The areas for which creativity is required are likely related to digital media software (like SFX in movies, games, and perhaps very innovative software). In these areas, surely the software developer working there will have the creativity required.

> but about putting some requirements stated in a slightly vague fashion, and actualizing it for the stakeholder to view and review

sounds like a form of creativity to me!

Reply on news.ycombinator.com