Hacker News Reader

Comment by 123yawaworht456 | Hacker News Reader

95% of real dev work is digging through piles of brownfield spaghetti to fix a bug or implement a feature. human-provided descriptions of the bug to fix or the feature to implement are often so vague that even a human developer can barely comprehend them. also more often than not, a visual link is required between the developer and the product being fixed/improved.

hypothetical AI will replace humans, sure. LLMs probably won't. original GPT4 was the best coding LLM I dealt with, and it was 1.5+ years ago. the context size was pitiful, yes, but it was a lot smarter than the current crop of big models.

View on HN

Even just the ability to reproduce a non-trivial bug, and then work out what general area of the codebase might be in seems out of reach for an LLM

The instructions seem to have to be so specific and repeated that you need a full-time person to review and retry.

My last struggle was creating an SVG arrow pointing to a specific corner. Both Claude and CGPT failed with many attempts.

the skill is in having it expose enough of the details that you only need to do a bit of messing with the output by hand and then your can be done with the ticket. there's no points for getting the LLM to spit out perfect code for you, the point is to get working code checked into git.

This thread is about LLM replacing people, so yes you're confirming that you do need a skilled person to get working code out of it.

Yes! I think I'm going to make LLM bug reproduction my primary response to AI developer maximalists.

Maybe the solution isn't to train LLMs on human-written code, but to train some sort of text-to-AST generator on descriptions of programs and then the output of programs. That way it doesn't learn to include human-inspired bugs at least.

I always wondered why bother having an AI generate human readable code. Isn't it a machine? Why not have it generate machine code? of course, it wouldn't be readable, but would it be performant? Maybe.

I guess my point is that if an AI is doing computing why would it doing computing resemble a person doing it? It's much closer to the machine then me. Why not take advantage of that?

Because then you'd have no control over what the AI/code does. If you blindly want to trust the AI, you might as well just make the AI the black box between your input and output.

Haha what do you think people do with AI/ML? do you think that's currently in control?

Isn’t that kind of what neural networks are in a way? Large blobs of computation we cant really parse fully

Yea but there are a bunch of layers, lol, between that specification and the machine and I'm wondering if there's a way to reduce that assuming it's a bottleneck. But I'm not sure it is I'm just guessing.

This actually gives me a funny idea, IDK if it would actually work.

Something interesting you can do with LLMs is you can constrain their output. What if you just write a test case to encapsulate the bug then try to filter out the most "sane" LLM output that results in the code compiling and the test case passing?

It doesn't seem like much, but it's like using more computing power to filter out all the "obviously wrong" solutions. Doing this in practice might require you to make too many test cases though.

You just invented a mid-level engineer.

If you manage to have more breakthroughs to senior then you can maybe leave it unattended for a few months but not if the product team is inexperienced or strong willed. The code will still end up as unmaintainable spaghetti.

So you need more breakthroughs to achieve staff or higher which requires a lot more agency and initiative on the part of the bot to push past what your team mates think they want and do what actually needs to be done modifying not only the code and the requirements but the process that leads to miscommunication and inefficiencies within the org.

If you manage to invent a bot that knows what’s better for the people prompting it and acts on that then you’ve potentially also violated the first law of robotics and created a death bot.

And boy oh boy would it be astronomically expensive to run.

I hate to break it to you, but that’s exactly what ChatGPT o1 excels at.

I swear to god, it’s better than our junior devs fresh out of college.

Like, I let the LLM solve a PBI and it’s just faster with better code.

Actual devs will still have a job, but the low hanging write some code guys? They’re done. And a lot of companies had a bunch of them..

I hate to break it to you but o1 is like far off of anything what you describe. Yesterday, I was testing some stuff and told it in very precise prompt to write a dripping bucket rate limiting middleware and it just literally named a var “bucket” with a timeout and then implemented something that will wait for the next function to timeout!

I then said, heh and asked it to implement a circuit breaker pattern using axios, and it again implemented a timeout and added it to axios options!

Could be am dumb or o1 was hallucinating on me intentionally, but the thing is very far away from a Jr dev fresh out of college.

very far away from a Jr dev fresh out of college

Feels like your Jr position claims 3 YoE in hardcore backend. Last time I’ve seen them, they discussed == vs === and “in” vs “of”. Wouldn’t hold my breath asking them to implement a rate limiter mw or any sort of debouncer. Unless humanity boosted their average intelligence very recently, ofc, where you’re an async networking specialist fresh outta school.

I think, my “very far away” was misunderstood. What I meant is that, a junior dev will hopefully go and do some research and try to find some reliable info before trying to implement something, while o1 just spits out statistical matches.

o1 is very far away from being a Jr dev replacement and is actually slightly more useful(if I ignore the costs entirely) than the existing intelligent autcomplete features.

Depends on the jr dev. You have the actual devs which develop solutions and the jrs who grow out to these roles. But I’ve been in plenty a company with just a bunch of code monkeys as well to write unit tests, boilerplate and center divs.

They are definitely being replaced by ChatGPT in our company. Think of it what you will, several teams write all their unit tests with ChatGPT now instead..

It's interesting how many folks assume everyone is using LLMs the same way they are to have comparable results (or not)

I don't even understand half those words. And I'm human btw.

It is fine, if those terms are not relevant to you. We learn when we need to. My point was that, a Jr dev will do some research and then try to implement a solution, while o1 being praised as immediately replacement of Jr devs is not even close and only shows as much correct implementation as its training data.

I've seen a whole gamut of junior devs and their abilities. Your junior devs are just the average beginner. They still learn to get better, to become real devs. And we need real developers to deal with the stuff even o1 stitches together. Because you actually need to understand problems and code to solve problems. And even something like o1 is just a statistical model that outputs the most likely result based on an input. And I've seen the incoherent garbage it babbles out when it comes to code. But the main issue will be, the blind reliance on these hallucinating models that create 'just good enough' will hinder people from learning and becoming better. LLMs are the end of ideas. Just the calcified present. And that's a problem.

I would need to ask ChatGPT about what an PBI is, but other than that what do you expect. You don't learn coding in college, companies and devs need to train freshmen.

LLMs still cannot solve most logic problems. They generate language. It can be a programming language but obviously the scope is limited.

I had to spend six hours today to untangle an o1/sonnet mess today and ended up rewriting a huge part of it to be remotely readable

what a waste of a nice sunday

And that's to say nothing of the fact that the programs we build exist in an ecosystem of other programs, and sometimes you're lucky if you even have a copy of the source available... And that's without mentioning that it's all networked and you're dependent on other systems all the way down. So much of the work I do today is black box debugging where I can only make broad assumptions about the behavior of systems that are distant to my own based on their past performance, based on some notes someone took from a meeting 6 months ago, based on what the engineers on that team believe their systems are doing, etc...

A lot of times at work, support teams modify their systems and don't bother communicating what they're doing, they seem to rely on scream tests to figure out who is even using their services and for what purpose. It's a total fucking mess and I laugh whenever I hear someone suggest that "AI" is going to replace what I do when some massive % of what I do has fucking nothing to do with code at all.

It would be interesting if AI could be used to isolate bugs?

Have it run through all the branches of spaghetti code and isolate the problematic sections along with proposed changes.

> original GPT4 was the best coding LLM I dealt with, and it was 1.5+ years ago. the context size was pitiful, yes, but it was a lot smarter than the current crop of big models.

Anytime someone says "this is the worst it'll ever be" I can't help but think "oh sweet summer child"; and then I ask them, what's your favorite operating system you've ever used?

The answer heavily depends on the person's age. Windows 10 or 7 dominates the conversation, and for slightly older folks, XP. "MacOS Snow Leopard" gets said a lot, that was a beloved MacOS release because they very famously said "no new features we're just fixing the old ones" and they actually did it.

No one ever says "MacOS Sequoia" or "Windows 11". Man, the discourse surrounding the new iOS 18 Photos app is reaching a fever pitch.

Sorry for the weird music on this, but its the first clip I could find of even, gasp, Elon Musk, stating this very real fact [1]: Technology does not automatically improve. In fact, it seems like especially software loves to get worse, and it takes a focused effort, from talented individuals, as far removed from money and business and agilescrum as possible to keep things operationally stable.

But, sure, definitely the new metaversecryptoai bubblescam will escape that and be built to the highest possible quality. There definitely won't be investors and men in suits in every room saying "we're losing billions is there a knob we can turn to reduce electricity usage?" OpenAI is definitely raising billions to fund the R&D of the next level of intelligence, and certainly not because without that capital they'd be bankrupt in three months.

[1] https://www.youtube.com/watch?v=aDC6kWC8y2Y

> it takes a focused effort, from talented individuals

You see, that's exactly why I don't have high hopes about OpenAI. They came up with a state of the art GPT-4 that was slow and expensive to maintain, so they came up with this o1 that is basically a chain of cheap 4os, something that people used to do from the moment LLMs appeared. And they started calling GPT-4 a "legacy model".

In the meantime Anthropic is steadily raising the bar with each iteration. But even there the efforts plateaued and what we have still has the inherent problems of LLMs, that is the ability to give completely wrong answers with amazing self-confidence.

I would pick Windows 11 (LTSC) over any prior version of Windows, and I would pick macOS Sequoia over any prior version of macOS.

My favourite version of Windows was Windows 2000. It was stable, visually consistent and fairly simple.

But Windows 11 is a massive improvement over Windows 2000 in many, many areas. Just because Windows 2000 was my favourite doesn’t mean I think it’s the best. It doesn’t even have built-in support for WiFi.

I think they’re even moving towards something that could be my new favourite Windows. Windows 11 is the first Windows with a visual style that I actually find appealing, and they seem to be making some serious efforts to make it all visually consistent too. The new settings app has its flaws but if they just finish the work of supporting all the features of the old control centre and finally have a unified place to change settings, I would say it’ll be all-in-all better than Win2k.

The GPU driver system and WSL2 are leaps and bounds beyond what Win2k could do. The built in support for virtual desktops and window arrangement is something I would never want to be without.

Yeah, on the Mac side I remember Snow Leopard fondly. That doesn’t mean I think it’s the best. On Mac I see even less reason to think older versions of MacOS is better in any way. They’ve stayed visually pleasing and consistent, haven’t made any huge UI changes similar to Windows’s attempts at making a better start menu, and have added many genuinely useful features over the years.

Reply on news.ycombinator.com