Hacker News Reader

Comment by paulgb | Hacker News Reader

> And even then, we still don't have a killer app! There is no product that everybody loves, and there is no iPhone moment!

I would strongly argue that coding assistants are AI’s first killer app. Copilot, Cursor, Windsurf etc.

View on HN

Say you reduced their revenue to only that application. Would it be sustainable? Would it be worth the billions upon billions of dollars that have been shoveled at it? Would it add more than the billions upon billions of dollars in the end?

By your logic I could claim a quantum computer with qubits on the scale of the mass of the sun is a killer app for doing RSA encryption breaking. And I would be making an equally useless statement.

This is moving the goalpost on what "killer app" means. Code assistants are a compelling use of the tech that has quickly shown real-world value, which is the point I'm trying to make here.

Whether the companies that are leading the market today will end up being the ones who capture that value is anyone's bet.

I fail to see how a technology that is to expensive to maintain can have a killer app.

It’s only going to get cheaper over time. It’s already cheap enough that if these services disappeared overnight I’d switch to an open source alternative with a local model. The industry needed VC backing to pay the fixed cost of the research, but the cost of running inference is not insane compared to the volume it provides.

I think the argument this article makes though is that generative AI isn't generating enough value to justify the sky high valuations and investment. Sure, if all of these services disappeared overnight, the remaining users could self host or run a model locally. I feel like that speaks to the lack of value any one company provides though?

Maybe this era of AI will be remembered as a wealth transfer from VCs back to everyday consumers lol

Only insofar as $1 Uber rides were a transfer of VC funding to everyday consumers. When the funding dries up and there's a need for revenue, then we'll see what they charge. Hopefully for them, they've become inextricable from our lives, like Uber has, and we'll gladly pay the new price. Not everyone will be hooked though, but the VC bet is that enough people will be that their horse wins the race and they'll make money.

There are a lot of horses in this race and the literally billion dollar question is who's Amazon this time around, and who's Webvan. Who's Uber and who's Flywheel (which doesn't even have a Wikipedia page anymore, ouch). Not knowing which horse is going to win doesn't negate the fact that a horse is going to win.

Model available LLMs, like Llama and Deepseek and StableDiffusion are totally a wealth transfer to consumers. Better make use of them!

> I would strongly argue that coding assistants are AI’s first killer app. Copilot, Cursor, Windsurf etc.

These IMO are relatively useful things. But probably (in their current state) will not justify the valuation of the companies involved and the massive investment occurring right now.

I don't know how the future will unfold. I do think it is reasonable to be somewhat bearish on what has been promised vs. what has been released.

Are they? I find Agentic mode on most editors barely useful. Autocomplete and inline editing is great though.

To use these tools properly, you need to know how to build the same thing precisely.

But that still doesn't exclude copilot, which is the most intuitive and genuinely useful version of ai after chatgpt itself.

I don’t deny it.

I used claude + gemini to whip out a rewrite of a show HN project in less than half an hour to deployment yesterday.

https://news.ycombinator.com/item?id=43071381

But this sort of work is fairly low value and boilerplate-y.

Nah you don't to know to build the same thing precisely. Just the other day I wanted to write a vanilla JS component that could let you select a picture from something like a carousel and be able to blow up the selected picture when clicked. I know JS / HTML but am not used to working with vanilla JS. Copilot didn't write it all by itself but it did teach me things I didn't know like making a custom tag in vanilla JS by extending an HTMLElement.

The code isn't the most readable because I don't need it to be however if you make me write it from scratch in an interview style setting I'd have trouble doing it. If I read the code I can follow it and it makes sense + it's an easy component to manually test. So.. no, I don't need to know how to precisely build the same thing.

And before you worry that I'm committing code I can't build from scratch.. This is a simple component for a 5 page landing page build with astro where I'm the "main" dev ( wrote like 80% of the code). The web-page won't even need maintainance once it's deployed

It’s not a common use case though, dipping into an unfamiliar tech stack only to dip out after committing the code. Typically, when you learn a new stack (eg. for a job), you’ll be living in it for at least a few months, and at that point, you’d be better served by perusing the docs and getting deeply familiar with the API.

The copilots get you going quick at the expense of your learning, which is great for one-offs, but not lasting work quality.

I believe if you spend months / hundreds of hours using any framework / tool you will eventually read so much of it that it's easy to get "deeply familiar" and even then.. it's often times faster to have an LLM write 80-90% of the code for you and just refine / finish it.

> Copilot didn't write it all by itself but it did teach me things I didn't know like making a custom tag in vanilla JS by extending an HTMLElement.

> This is a simple component for a 5 page landing page build with astro

You're already in the over-engineered section there.

Not that over engineered, here's the code :

https://gist.github.com/lazarcf/c80ae6f9362aaf3aa92e21e3a0dd...

you can see the component (the "image gallery") here : https://pixico.roware.ro/apa-distilata/

I think autocomplete alone would be enough to make coding a killer app for AI.

I agree the tools are overhyped for allowing non-developers to write code. It’s not (today) a replacement for a dev agency that takes a set of requirements and runs with it, it’s a replacement for a junior developer who you need to micromanage a bit. But that’s still a boon to productivity!

I keep on trying them, but if they are useful they are useful for only a small fraction of engineers at the moment. I'm not sure if this is due to the nature of the work, or the nature of the user.

I have heard "top" engineers at various places say it makes them 2x faster, or whatever, but I would like to see this assessed by timed testing, as is sometimes done for evaluating software engineering.

Copilot may let me type less, but I have not seen the wall clock effects, which is a very hard thing to measure (time perception is very unreliable).

https://news.ycombinator.com/item?id=43071381

You can see this example where I timed myself to deployment using AI tools to rewrite a show HN project in half an hour. The code is open source.

My comment was posted 2 hours after show HN when I saw it on front page so you know I didn’t lose track of time I spent.

The vast majority of software work is not greenfielding a PoC or reimplementing an existing, small, well-specced project. We’ve had OpenAPI client generators for years after all.

The majority of software work is maintaining large, existing products: adding features, fixing bugs, improving performance, etc., or building new software in problem domains that aren’t so well-defined.

In my opinion, AI only really helps you (a lot) if you are bottlenecked by the actual code-writin. I have not been in a such a position since... I dont even know. Maybe In my 20s, 15 or so years ago? Even if AI wrote my code 100x faster it would not appreciably change my working days.

If it could test and verify things though... ideally physically since Im in embedded and pulling SD-cards etc is a thing.

Yeah I’m in the same boat with it. I do keep trying, but so far it has been far from earth-shattering.

I’d love it if I could get it to write decent unit tests given a basic description of what I’m testing but I at least cannot get it to output anything useful for our codebase. It’s too unaware of the broader context of the code, what objects need to be instantiated from some other internal library and passed in, etc. It can do a decent job if all I have is a totally isolated function that doesn’t touch the rest of the codebase or use any domain objects, but that’s a rare enough case as to be essentially useless to me.

This is my experience too.

I think it also really accelerates learning of a new language or framework, when that language or framework is really well documented on the web. For novel programming frameworks, obviously it's a bit more challenging to get help from an LLM.

One of more recent attempts at using LLM code assist was to try to fix a bug in a Swift SSH Agent's connection handling that was causing hangs. I know zero Swift, much less the networking frameworks. So I pumped the output of `tree` on the git repo into the LLM, asked for which file likely handled connections, and it found it right away. That's probably 15 minutes saved. Before putting in the file I asked for likely reasons for deadhangs, got that list, then put in the Swift file that handled connections, and it pointed to what the likely problem was. That's probably 1 hour+ of reading documentation to try to figure out what the code was doing wrong with the networking framework, assuming the LLM was not hallucinating. And that "not hallucinating" probability is high enough in my experience that I spend >50% of my time trying to verify I'm not getting bullshitted.

The LLM proposed a fix (~10-20 minute savings), but even as somebody who doesn't use Swift it seemed like >99% chance that it had just introduced a bunch of race conditions in the data structures it used to track connection status. So I asked about it, and it said "Oh yeah of course how could I forget" and then significantly complicated the solution with something that I thought looked like it probably worked. But was the LLM just being obsequious or was it correct the first time? So hard to tell...

So in about 20 minutes I probably accomplished in a language I didn't know, in a code base I didn't know, about 2 hours+ of learning.

But if I knew the language, it would have saved me very little time, and may have cost me some time.

Searching is an area where LLM technology excels. It makes sense, given their structure.

Of course, you have to find a company willing to spend more money on worse (for them - less ads) search, and it's won't be Google.

The results aren't always accurate, but neither is Google...

Most of my LLM experience is pulling off the mask with a label of "AI" only to find another mask underneath labeled "information retrieval".

I agree it's impressive and stuff, but I wouldn't consider a JS POC as a serious project. I have never done that in my whole life and would rather see results from a 10 years old application with a million lines of code of C++. That's would be realistic. What you did is refactoring a pet project and I don't know why we're wasting $billions for that.

The killer app is entertainment. Since LLMs emerged people have consistently loved getting them to say whatever they want or roleplay with them. Once integrated into games, it will be very fun to have natural conversations with the inhabitants of game worlds. Imagine a goomba talking to Super Mario before he stomps its head or delivering your pithy one liner in response to some final boss’s villainous monologue.

and strangely in two years no one has demonstrated anything like this that people found of value. In fact for all the "entertainment" sectors it has been injected into, we have gotten poisoning of self-publishing and soulless generic pornography. Remember the twitch channels that had AI generated content? Where did those go? Surely by your rationale the market would have taken over by now. Surely there would be something.

It's almost like entertainment requires some humanity and thought and true creativity behind it.

The primary reason is because LLMs are expensive computationally and financially.

In an open world game, it’s trivial to assign memories and facts an AI learns about its world from interactions or in response to game events. All an LLM has to do is be fine tuned to take data from that internal knowledge base and express it as natural language text, in order to have intelligent and useful conversations with a player. It’s not difficult.

This only works if your game doesn't use the GPU. Then there's the whole problem with nondeterminism. But I'm sure when those small problems are solved people will use this technology in games /s

by giving the models durable memory, they become "agentic". in any case, they don't make as much of a mess when their output is getting written into git.

Reply on news.ycombinator.com