Andrew Ng says bottleneck in AI startups isn't coding – it's product management

If AI is that good, can someone code a good XML library with AI? The spec is available.

If AI is that good, there should be an explosion of Open Source projects of good quality.

Neither of those is happening.

Many people are seeing huge gains in coding productivity with AI. If you're not one of those people it might be useful to evaluate why you aren't experiencing any benefits, instead of claiming that there are none.

The second point is more incentive than AI capability i'd argue. Your point presumes that Open Source === Good. I'm not sure that's how all of society feels, unfortunately, so even if AI can do it at some point...it might not choose to.

I suspect somebody could. But why? Do you need a new XML library? I don't.

And if you don't need one, why write one? If there is no specific use case in mind, how do you even determine what dimension "good" is measured on?

> If you're not one of those people it might be useful to evaluate why you aren't experiencing any benefits, instead of claiming that there are none.

It might be that the gp is smart enough to code without a crutch.

> Many people are seeing huge gains in coding productivity with AI.

I don't want to speak for the person you replied to, but I think that their main point is... are they?

I see lots of articles about huge increases in productivity, but I think it's fair to argue that we've yet to see the huge increases in useful products that would surely (we hope) result from that if it were true.

Maybe because it's become second nature to you, due to a long enough career spent practicing the actual skill rather than using a crutch or tangential skill?

Open source projects left and right are banning contributors from using AI.

One thing that has changed for me personally is that these days if I’m parsing XML, I might just tell AI to write a parser that can handle these 5 XML files. It might load up a library or it might just roll its own, but one thing that is not going to happen is that I’m not building some beautiful well engineered XML parser which I then open source. I wonder if that is what’s going on?

I'm speaking just to coding productivity. I think AI does very little for business development or creativity type issues.

People should realize that denying that AI can boost productivity in coding makes it look like they don't know how to use it, or believe in some conspiracy that no one is actually benefitting and it's all market hype from tech bros.

XML is used extensively everywhere. Multibillion companies depend on libxml2 which historically had multitude of vulnerabilities, see https://ubuntu.com/security/cves?package=libxml2

Most of the Internet ifra depends on libxml2, major vendors like Juniper and Cisco use it. To my knowledge Android use it as well,

Naturally, with the advancement of AI, one would expect XML would be first thing to rewrite, given that library is in the critical path literally everywhere.

bullshit, open source is blooming as never before. also we already have good xml libraries, nobody is interested in that. but yeah i agree with the sentiment that ai is not the magical tool you were promised, useful nevertheless.

I'll bet he's strong enough to push nails into wood so he never uses a hammer.

I mean listen to yourself…

Let’s take a critical piece of infrastructure where if it goes down, billions of things also break.

And have ai rewrite it.

This is Gavin Belson level stuff.

Your original point would have had a bigger impact if you hadn’t married it to an idiosyncratic affection for XML. Now you’re stuck defending the usefulness of that particular technology which is entirely unrelated to AI.

That sounds worse than using an xml library to parse something

I admit that I don't know how to use it and it does seem like marketing hype and tech bros with some mild value and I'm waiting to be proven wrong and crushed, John Henry the Steel Driving Man style, and it's just not happening.

It is not good, it is for people who can’t write code to delude themselves mostly

I believe in some conspiracy that no one is actually benefitting. I don't want to blame marketing, it's genuinely very cool that I can give Cursor a description of what I'm trying to do and get a result that's like 80% correct. But I've repeatedly found that going from 80% to 100% takes just as much time and effort as it would have to do it myself from the start.

I've shadowed people who believe AI is helping them, and it seems to me that some of them don't notice how much effort they're spending while others don't bother to correct the 80% version once tests are passing.

I don't think this line of reasoning holds. The only thing people should look at are peer reviewed studies, lots of them ideally, and with no conflict of interest. Who's getting productivity gains? What kinds of work are they doing? What doesn't work so well? All of these questions should be investigated by studies. People feeling productivity gains doesn't imply the gains exist.

Otherwise it sounds like "many people have had their lives changed by {insert philosophical/religious movement}, so if you're not finding it true you should look into what's wrong with you."

I don't buy this line of pro-AI reasoning. It sounds like more work with less reliable results than just writing the code myself.

First of all, nobody is writing and open sourcing their own XML parser in 2025, so that's hyperbole.

Second, the boilerplate to use most XML libraries can be copy/pasted out of their docs. So where is AI saving you time here? The prompting and other BS is a waste of time and just looks silly, and you still have to read and understand the code. At best it seems like breaking even.

not for hackers

There are a lot of different niches to software development. It could be that your specific niche isn't seeing much benefit yet. However, continuing to hold onto a view that AI is hype garbage could very possibly hurt your career in the near future.

The people I know who are experiencing these gains were never more than juniors in skills, despite the number of years they've been on the job.

Makes me evaluate indeed.

They have no way to actually verify that so it's just marketing to luddites.

Can I just remind everyone that GPT was released not even 3 years ago yet. It's been only 2 full years since GPT was released.

I get that people are anxious, worried, and are going through the "cycles of grief", but do you really think that in another 2 years, let alone 5 it won't be able to code a good XML library? We are just going to have to see how things go, because they are clearly going to go, whether we want or not.

And what does open source and the quality of projects have to do with it? There were bad open source projects before GPT's release.

If they were, where are their products and open source projects? Let's circle back.

One thing I've noticed about the super-AI enthusiasts on HN is that not a single one ever have a single comment linking to a repo of work they've made with it.

I check. I actually always do because I'm really keen to learn how to use these magical super-AI workflows. I've watched streams, replicated clause MD files, tried all the context tricks.

I'm not even saying AI doesn't help, it's great for getting me over the blank page writer's block. It's just not great at much else.

So I've just checked your comments and not only do you not have any examples of your super-duper AI skills, but it looks like you've been in the industry less than a year, graduating from a PhD last year?

You also admit it took you a week trying to debug a problem before an AI fixed it for you. Because you'd missed some parentheses in an algo.

I'm not trying to shame you, but that does signal your inexperience. If you'd have made the code well and easy to test, you should have spotted your bad algo quickly.

So is it that we're all bad at using AI? Or is it that AI benefits inexperienced programmers more?

We do actually, since the xslt library is under-maintained and has a looooooot of usage.

What issues did you run into when trying to build an XML library with AI? What was your workflow? How much time did you spend on that project?

> The only thing people should look at are peer reviewed studies, lots of them ideally, and with no conflict of interest.

"Ignore your own direct experience, only research papers matter" is certainly a take.

The beautiful thing about the current generation of tools is that they are so incredibly cheap relative to historical tools intended to improve engineering productivity. You can't just run out and pick up CASE tools for less than ~$CAR to ~$HOUSE. A pro subscription to whichever AI tool you want to try is $20.

Ignore research, try them, if you have success, use them. There's no dogma here. Just empiricism.

Damn bro. Deep dive into the comments.

The bad algo was a scaling problem for one equation. That particular equation wasn't some y = mx + b thing, it was the result of a discontinuous galerkin finite element scheme that I wrote from scratch. The actual equation was one that I found after about 2 pages of hand written derivations with high level math. Not really a coding issue, just an algebra issue after really intense manipulations of partial differential equations.

The fact that AI found that problem, a problem that could only be found by someone able to do complex manipulations of PDEs is incredible to me. Perhaps I didn't tell the story well in the past comment, but it isn't like I didn't know python syntax and AI held my hand.

I don't post repos because I keep my hacker news life separate from my personal life, and my repos are tied to my name.

Most major software companies are demanding that their employees use AI, so you should be able to look at any open repo from Google, Microsoft, Facebook, etc for examples of AI use in code.

While you're looking at peer reviewed studies, I'm over here being way more productive with AI and getting shit done.

Doesn't need a conspiracy, just excessive enthusiasm about the new shiny that hasn't been tempered by experience and reflection yet.

Clearly that's not the use you have for a new XML library. It's a use you're imagining somebody else would have for it. And because you're just imagining the use case, you've failed to think through what "good" would actually mean in that use case.

To replace libxml2 across these ecosystems you would need it to be API-, ABI, and probably bug-compatible with a decrepit old C library. That's not something anyone or anything can write from just the XML spec.

I, on the other hand, tried to vibe-code BDF-like ODE solver. Not some rough prototype, because I can do the rough prototype easily myself, but something robust and fast, with event handling etc. AI couldn't do it. Actually it couldn't do correctly anything more complex than out-of-the-textbook explicit RK4, but this is something undergrad students do while learning numerical methods for the first time.

And solvers are actually a simpler aspect of the project I am working on. It also includes (or rather aims to include) optimizing compiler with DAE to ODE reduction, advanced numerical debugging etc.

This is why these discussions are pointless - AI works well for some people in some contexts, for others not so much, yet both sides extrapolate their experience as universal.

I see what your saying but I would argue that vibe coding an ODE solver is an incorrect use of the tool. For something like and ODE solver you need to have a really solid understanding of what data structures you will use,and solid general knowledge of the numerical methods you want to implement. Then, you can use AI as an assistant when you get stuck, or to deepen your understanding, look over your implementation, etc.

It seems like developers used to always joke about how much they used stack exchange (even senior devs). Now it seems like there are suddenly so many people who claim to never need any help and can just smoothly bust out beautiful code all day long.

Is it? Can you link me to a succesful product that actually has human users that does heavy use of AI for writing code?

(and blooming would mean that there are hundreds of these)

> I see what your saying but I would argue that vibe coding an ODE solver is an incorrect use of the tool.

Agreed. No true Scotsman would use the tool this way.

> When developers are allowed to use AI tools, they take 19% longer to complete issues—a significant slowdown that goes against developer beliefs and expert forecasts. This gap between perception and reality is striking: developers expected AI to speed them up by 24%, and even after experiencing the slowdown, they still believed AI had sped them up by 20%.

Maybe you just think you're being more productive ;)

https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o...

But that is the case - I know how to implement production-ready ODE solver. My issue with AI was that it was able to help with basics, but not with those really important bits, so it couldn't really deepen much.

> For something like and ODE solver you need to have a really solid understanding of what data structures you will use,and solid general knowledge of the numerical methods you want to implement.

For basically every thing you program, you need to have a really solid understanding of what data structures you will use, and solid general knowledge of the methods you want to implement.

I claim that as a conservative estimate at least 90 % (likely more than 95 %) of what I code at work (and even more for what I code privately) is of this kind.

I imagine the folks in the article and others like it are not building libraries and foundational infrastructure but rather cranking out SaaS startup ideas and CRUD web apps. I find that kind of coding really can go quite fast using AI, particularly if you are building it from zero and not worrying about all the existing quirks of a large codebase or creating technical debt.