21–30 of 53 posts
The resulting code had a memory leak (from good old Promise.race) that would have caused the app to continuously grow the RAM usage.
I decided to let the AI code, and then asked it to fix the problems, describing the issue (steadily growing RAM usage). It was not able to find the issue.
That's what I'm afraid of. We'll get megabytes of code that just fails sometimes, for unfathomable reasons.
> In 2025, the engineer probably beats Gemini. Sure, great. But in 2030, who wins this hypothetical? I don't think we need to wait for 2030. Today, using aider/cursor and Gemini-pro-2.5, I'm pretty sure a team of two or three senior developers with great domain knowledge would be competitive with a team of 10 average devs. Being a small, close-knit team has a lot of advantages. And LLMs can help close the productivit…
> I'm pretty sure a team of two or three senior developers with great domain knowledge would be competitive with a team of 10 average devs. This has always been true. LLMs didn't change that.
But what if the three senior devs could produce thousands of lines of great code that implement thousands of features perfectly. I don't think this is the case but I believe this is what the OP means.
the interest is of course because what "the business" wants is thousands of features, perfectly or imperfectly is a secondary consideration.
on edit: changed senior devs hundreds of thousands to thousands because one benefit is that much of the code the average make is not really needed.
I tried vibe-coding some non-trivial code. It failed successfully. The resulting code had a memory leak (from good old Promise.race) that would have caused the app to continuously grow the RAM usage. I decided to let the AI code, and then asked it to fix the problems, describing the issue (steadily growing RAM usage). It was not able to find the issue. That's what I'm afraid of. We'll get megabytes of code that just…
But…that’s the current situation.
Also, calculators did millions of operations, but the were 99.999999% right all the time. That's what set the apart from human people: eternal reliability 24/7. AI is about right 1% of the time, a human is needed to recognise when the machine is wrong, and it keeps repeating the same mistakes when identified and corrected. People WISH it was like a calculator, but you know, magic thinking doesn't make things happen.
I have an app where people scan their retail/restaurant/cafe receipts and store them. I use Gemini 2.0 for OCR.
It is right more than 99% of the time. I would be making many more errors transcribing the receipts manually.
AIs are a different kind of tools, for a different kind of problems, with a different kind of failure cases.
The reality is that we don't know.
We don't know what the ceiling is for LLM programming ability. We don't know how much better they can get without scaling up the hardware. We don't know how well we'll be able to scale up the hardware. We don't know how many more billions we'll allow companies to spend building better and better models until the market loses confidence in them.
We can make educated guesses, but that's all they are. We just don't know.
Earlier quoted context omitted.
> I'm pretty sure a team of two or three senior developers with great domain knowledge would be competitive with a team of 10 average devs. This has always been true. LLMs didn't change that.
The 10 average devs can produce hundreds of thousands of lines of buggy, poor performance code that implement thousands of features poorly in the same amount of time the senior developers could implement tens of thousands of lines of great code that implement hundreds of features perfectly. But what if the three senior devs could produce thousands of lines of great code that implement thousands of features perfectly.…
Color photographic film ended that job, but not photography nor painting.
What ended was just the small intersection between the two that was, for a moment, very popular and valuable but suddenly became not a job.
So diversification within the same or adjacent skillset is probably a good idea. Better than panic or putting all chips on one thing.
We don't know the future after all. So much can change so fast.
I tried vibe-coding some non-trivial code. It failed successfully. The resulting code had a memory leak (from good old Promise.race) that would have caused the app to continuously grow the RAM usage. I decided to let the AI code, and then asked it to fix the problems, describing the issue (steadily growing RAM usage). It was not able to find the issue. That's what I'm afraid of. We'll get megabytes of code that just…
> We'll get megabytes of code that just fails sometimes, for unfathomable reasons. But…that’s the current situation.
A single human can't write megabytes of code within a few hours, unless it's something that was repeatedly cut&pasted.
AI can create megabytes of complicated code that will be impossible to debug. It can do what you want it to do, but not quite.
Kinda like a genie from a lamp.
I wish people would stop making predictions about LLMs so confidently. Some people confidently predict that LLMs will never be able to replace capable engineers, and other people equally confidently predict that we'll all be out of a job in five years. The reality is that we don't know. We don't know what the ceiling is for LLM programming ability. We don't know how much better they can get without scaling up the har…
The fact that no one really knows how LLMS will pan-out means every projection in future is a prediction.