Hacker News Reader

Comment by recursivedoubts | Hacker News Reader

while I think there is a lot to this criticism of AI (and many others as well) I was also able to create a TUI-based JVM visualizer with a step debugger in an evening for my compilers class:

https://x.com/htmx_org/status/1986847755432796185

this is something that I could build given a few months, but would involve a lot of knowledge that I'm not particularly interested in taking up space in my increasingly old brain (especially TUI development)

I gave the clanker very specific, expert directions and it turned out a tool that I think it will make the class better for my students.

all to say: not all bad

View on HN

AI is bad at figuring out what to do, but fantastic at actually doing it.

I’ve totally transformed how I write code from writing it to myself to writing detailed instructions and having the AI do it.

It’s so much faster and less cognitively demanding. It frees me up to focus on the business logic or the next change I want to make. Or to go grab a coffee.

I would say from my experience there's a high variability in AI's ability to actually write code unless you're just writing a lot of scripts and basic UI components.

I would say this does not work in any nontrivial way from what I've seen.

Even basic scripts and UI components are fucked up all the time.

You have to learn how and where to use it. If you give it bad instructions and inadequate context, it will do a bad job.

This is the ‘you’re holding it wrong’ of LLMs.

What tool can’t you hold wrong?

Literally every tool worth using in software engineering from the IDE to the debugger to the profiler takes practice and skill to use correctly.

Don’t confuse AI with AGI. Treat it like the tool it is.

You might want to ask ChatGPT what that is referencing. Specifically, Steve Jobs telling everyone it was their fault that Apple put the antenna right where people hold their phones and it was their fault they had bad reception.

The issue is really that LLMs are impossible to deterministically control, and no one has any real advice on how to deterministically get what you want from them.

I recognized the reference. I just don’t think it applies here.

The iPhone antenna issue was a design flaw. It’s not reasonable to tell people to hold a phone in a certain way. Most phones are built without a similar flaw.

LLMs are of course nondeterministic. That doesn’t mean they can’t be useful tools. And there isn’t a clear solution similar to how there was a clear solution to the iPhone problem.

"Antennagate" is the gift that keeps on giving. Great litmus test for pundits and online peanut galleries.

You are technically correct: it was a design flaw.

But folks usually incorrectly trot it out as an example of a manufacturer who arrogantly blamed users for a major product flaw.

The reality is that that essentially nobody experienced this issue in real life. The problem disappeared as long as you used a cell phone case. Which is how 99.99% of people use their phones. To experience the issue in real life you had to use the phone "naked", hold it a certain way, and have slightly spotty reception to begin with.

So when people incorrectly trot this one out I can just ignore the rest of what they're saying...

Humans are not deterministic.

Ironically AI models are iirc.

Come up with a real argument.

YOU come up with a real argument first

"no, you're wrong" is not a valid rebuttal lolol

I pointed out the claim is both irrelevant and factually wrong.

That should be sufficient rebuttal.

Exactly. The problems start when people say it's good for everything ;)

"All the time"?

This always feels like you're just holding it wrong and blaming the tool.

It can still fuck it up. And you need to actually read the code. But still a time saver for certain trivial tasks. Like if I'm going to scrape a web page as a cron job I can pretty much just tell it here's the URL, here's the XPath for the elements I want, and it'll take it from there. Read over the few dozen lines of code, run it and we're done in a few minutes.

The AI version of that Kent Beck mantra is probably "Make the change tedious but trivial (warning: this may be hard). Then make the AI do the tedious and trivial change."

AI's advantage is that it has infinite stamina, so if your can make your hard problem a marathon of easy problems it becomes doable.

[flagged]

Couldn't help your self, could you.

Lay off him, he did the right thing

> AI is bad at figuring out what to do, but fantastic at actually doing it.

AI is so smart, one day might even figure out how to subtract... https://news.ycombinator.com/item?id=45821635

When you need to take the square root of 37282613, do you do it in your head or pull out the calculator?

Why does the AI have to be good at math when it can just use a calculator? AI tool usage is getting better all the time.

I think people generally think AI should be good at math because it runs on a very complex and very fast calculator to begin with.

Yore brain runs on physics and biology, yet here we are…

That is not the point...Its about not understanding subtraction...

I think it is like protein folding.

It will make a mess but if you drop a console.log into the browser debug console to show the AI what it should be looking for after it spent 3 hours failing to help understand and debug the problem, it will do 1 week of work in 2 hours.

I've noticed this too. The latest cursor version has a @browser command which will launch a browser with playwright, and call tools to inspect the html and inject JavaScript to debug in real-time.

When it has access to the right tools, it does a decent job, especially for fixing CSS issues.

But when it can't see the artifacts it's debugging, it starts guessing, confident that it knows the root cause.

A recent example: I was building up a html element out of the DOM and exporting to PNG using html2canvas. The element was being rendered correctly in the DOM, but the exported image was incorrect, and it spent 2 hours spinning it's wheels and repeating the same fixes over and over.

We're all architects now...

> AI is bad at figuring out what to do, but fantastic at actually doing it.

I've found AI is pretty good at figuring out what to do, but hit or miss at actually doing it.

I had no idea you (ceo of hmtx) were a professor. Do your students know that you live a double life writing banger tweets?

Couldn't agree more. Being a skilled operator with these tools helps you be very effective at creating new things in far less time than it would've taken you before. This is especially true if you know the architecture, but don't have the cycles to implement it yourself.

What did you use to do this? Something I'd like to do, while also avoiding the tedium, is to write a working x86-64 disassembler.

claude

Is that worth the negative externalities though? Genuinely asking. I’ve asked myself over and over and always came to the same conclusion.

hard to know

What negative externalities? Those prompts probably resulted in a tiny amount of CO2 emissions and a tiny amount of water usage. Evaporating a gram of water and emitting a milligram of CO2 seems like a good deal for making your class better for all your students.

>emitting a milligram of CO2

In the US, on average, generating 1 kWh produces 364.5 g of CO2. 1 kW may be somewhat pessimistic, but I think it's in the right ballpark for power consumption of DC inference. If processing the prompt took a minute of continuous inference (and I'm going to guess it took a fair bit more), that's 6 grams of CO2.

>What negative externalities?

Off the top of my head,

* All the ways AIs can be misused, either by people who don't understand them (by asking them for advice, etc.) or by people who want to take advantage of others (spam, scams, etc.).

* The power and resource usage of the above, both for inference as well as to deal with them.

> If processing the prompt took a minute of continuous inference (and I'm going to guess it took a fair bit more), that's 6 grams of CO2.

Neither of us really knows, but I'm going to guess it took a fair bit less. You need to consider the marginal cost of one more prompt, not the cost of running the datacenter in the first place. And with batching, hundreds of requests are executed in parallel on the same hardware. I'll concede that a milligram was an exaggeration though.

For comparison, a typical car in the US emits about 400 grams of CO2 per mile. And I know few people who beat themselves up over the externalities of unnecessarily driving one mile.

> * All the ways AIs can be misused, either by people who don't understand them (by asking them for advice, etc.) or by people who want to take advantage of others (spam, scams, etc.).

This is not an externality because it is not a consequence of using AI for good purposes. It's like saying an externality of using a hammer to build a shed is all the ways hammers can be misused.

Still, I respect you for considering the externalities of your actions. That's farther than most go. But I think they are outweighed by the benefits in this case

>This is not an externality because it is not a consequence of using AI for good purposes. It's like saying an externality of using a hammer to build a shed is all the ways hammers can be misused.

I guess we're talking about different questions. You're talking about the externalities of using a piece of technology, while I'm talking about the externalities of a piece of technology existing.

I would argue even just using the technology for good is offsetting the costs to keep the whole thong going. Essentially subsidising the bad users.

Being more specific in what you think the negative externalities are would be a good start - I see a lot of noise and upset over AI that I think is more or less overblown, nearly as much as the hype train on the other end. I'm seeing the potential for civilizational level payoffs in 5 years or less that absolutely dwarf any of the arguments and complaints I've seen so far.

> civilizational level payoffs

but first the investors must recoup their trillions, right?

Yeah, economics is weird like that, and we should probably push as hard as we can to make sure things are as fair as reasonably possible. I've got no problem with trillionaires or megacorps if the system in which they operate is fair. One thing AI is enabling is systematic uncovering of fraud and waste - we'll be able to scan omnibus bills and find corruption, track politicians, interact with the system and communicate at higher levels than before, and there's a shot at AI empowering and enfranchising people more than they've ever had a chance at before.

If AI coding alone reaches its potential, then a whole class of corporations lose their leverage in the market. In 10ish years, following moore's law, phones will have enough compute and storage to trivially run your own local search and curation of the internet. Sophisticated search engines that outclass anything currently on the market, working in tandem with a personal local privacy respecting agent. That decentralizes things in such a significant way that Google and Apple and all the gatekeepers lose their ability to manipulate and distort.

20 years from now, even if we're somehow just stuck at the current levels of AI, your personal compute will be totally sufficient to provide all the curation, moderation, proactive interaction, and search capabilities without ever having to deal with ads or bots or cloudflare.

There are lots of dangers, and ASI could wreck us, and militarized AI could be a doom scenario, or there could be overbroad stupid regulation that puts control of AI in the hands of a few corps or a government, but there are lots of bright future possibilities, and I don't see any indications that we're directly headed toward any particular doom.

Reply on news.ycombinator.com