Hacker News

> It's tough to answer because you want to hedge for both an AI enthused employer and an AI hesitant employer with limited information about who they are and how they personally use these products.

Have you considered just answering truthfully?

Would you even want to work somewhere where you need to play a role and where they flip out when you say the wrong word you should've correctly guessed through mind reading? That sounds not like a job but a toxic relationship.

> It's tough to answer because you want to hedge for both an AI enthused employer and an AI hesitant employer with limited information about who they are and how they personally use these products.

I'm an old hat on both sides of this type of discussion from a post-grad view.

Recommendation: use it to own the conversation and to signal mutual fit. Yes, your idea of AI lover versus hesitant matters. I recommend reframing the question to pivot to your fit to the org (and org fit to you) question. Show/concisely explain how you consider whether LLMs are fit to a task and how to tell it improves outcomes.

An outcome focus and willingness to show thought process around a common use case will be a substantially strong response.

Replying as a hiring manager since this might help other post-grad job seekers:

- Any long-winded answer to a question is immediate out and has been for years.

- Not having used agents and not being able to comment on what to do and what not to do with them is immediate out since early this year.

I understand the pressure to get employed from your perspective, but differences in opinion should be voiced out and typically aren't the thing leading to rejection from the company. It's common that engineering leads seek out people with different backgrounds and views to work on the same team. If anything, answering truthfully will make you stand out from others who've responded in a generic, heavily hedged way.

still 10x better than the 'finish this leetcode tweak algorithm in 20 minutes and tell me your thought process along the way, and yes you will never need that skill in the real job but we need find out who had time to cram for the algorithm books in the last few months'

You should find out during the screener what kinds of view the executives have on LLMs. don’t wait until you’re midway through the third round.

> re-factoring a big repo of decades old fortran+C cod

Having been in academia in the past and now in software I can say with a lot of certainty that this will take a lot more upfront work than otherwise.

Academic code does not have a lot of structure. And usually lacks a lot in terms of tests. While AI is best when it can mimic patterns as well as there are tests to target.

So you will probably need to budget a few weeks to establish good patters, docs as well as testing patterns before you can seriously make it really do what you want it to do.

Consider using agent mode for some things, you are definitely missing out.

The analogy I've had for myself is that it feels like using a bulldozer to dig rather than a shovel. If you use it to dig archaeological artifacts, it can make things worse than you started. A lot of the work however, is just moving dirt around, so you are wasting time by using a shovel.

A balanced answer that’s often true these days is: you’ve found that LLMs are impressively useful in some cases but fall dramatically short in others.

Exact same experience. My background is embedded and VLSI so I hedge my bets by saying that LLM are ok for Python scripting, but not there yet for synthesizable Verilog. It is really hard to see if the "how are you using LLMs?" question is for "we are AI Native™" or a form of cheating (like in university).

Just answer honestly, and include a note that you intend to fully comply with the companies AI policies. Thats the best answer anyone can give.

I personally think "I pretty much use it as a faster and more flexible StackOverflow" is probably the most neutral position you can have on it

That's probably not going to be enough for AI maxxers, but it probably won't be too much of a turn off for anyone but the most extreme AI minners, and everyone in between will probably be fine with it.

Frankly I plan to steer well clear of any "the majority of our code is AI generated" shops for the foreseeable future. Seems like disasters waiting to happen and I'd rather let other people step on those rakes

"for entertainment value, when i'd like to see how an enthusiastic 5-year-old would react to the task."

I'd like to add that there is almost no way of "running away" from it. If I search for anything on the internet I am almost guaranteed to be handed pages and pages of AI generated content. In lieu of that I found that directly prompting for an answer tends to yield better results nowadays. Not because it's good per-se, but because having control over the prompt beats having little to no control over it though search by proxy.

It saddens me to see that high quality content is drowned in this sea of garbage to the point of being almost impossible to find.

This would be expected. The corner cases people faced with PHP throughout the decades have been well documented on the internet for eons.

Swift, not so much. It's relatively new. Looking at AI's abilities like an engineer's career span scaled about 10-20x of time makes it make a bit more sense.

It's going to be worse at newer/niche things, intuitively - which is only going to get worse as it "learns" from garbage outputted by other LLMs moving forward.

That's just one way to use LLMs though. Recently on a flight I could not figure out how to connect my wife's earphones (i.e. put them in pairing mode) to my macbook since I was used to the old Airpods Pro case. So I asked Gemma4 26B A4B (offline, LM Studio) and was told to use the 'two tap on front of case' gesture, which worked. This situation would have been significantly more frustrating without (local) LLMs. I'm essentially carrying around a basic "how to" on everything, inaccurate though it may be, it's better than nothing.

My experience was different. I found it extremely good at fronting technology like react while I had to hand hold it for the backend tasks. Even with fable it was the same.

Well Apple just released a bunch of Agent Skills. I tried it on my macOS apps and I noticed some improvements codewise and updated some deprecations I didn’t know existed in Swift.

Would you describe yourself as more skilled at frontend engineering or at backend engineering?

In my experience the language has become irrelevant for me, I created a system like mix of revenuecat and firebase and I’m not even sure what language which part is. It has client side libraries that are swift and kotlin, the Identity management is Swift but the iAP/Subscription tracking is go IIRC. It’s all integrated somehow and works very well.

Which LLM though? Models can still be significantly different in their capabilities.

I’m going to guess you are better at frontend than backend.

The classic AI Gell-Mann effect.

Might be because there are less Swift projects to train with.

But I've seen Claude write crazy code in Python and JavaScript, too

The real best case scenario is using LLMs to help build deterministic systems. Instead of asking an LLM to do some task that you know will be repeated, instead ask the LLM to build a program (Python script or whatever) to do the task.

I am seeing similar things in just regular tooling and development. Things that can be solved deterministically or what would have been a simple CLI 5 years ago are now an LLM integration.

Instead of using the LLM to create deterministic tools, we are using LLMs to replace them. It's completely backwards and I don't know why people (especially high ranking people in my company at least) seem to think that this is the way forward. No, I don't want a whole CI pipeline that is just LLM prompts. Yes it's very easy, but it's expensive, slow and prone to failure in ways you can't even predict.

Same things like using LLMs for the code review process. What would have been a simple linting rule is now a pass with an LLM rather than using the LLM to create the linting rule, which it is absolutely excellent at creating.

My management is pushing for us to come up with ideas on where we can use LLMs in our product. The whole team has been very resistant for this exact reason. Anything we can think of will only make things worse, and we’ve already been told anything above a 1-2% failure rate is unacceptable. If anything we need more structure and standards to hit that, not less.

Basically, folks nowadays think that this article[1] was aspirational rather than a cautionary tale.

[1] https://thedailywtf.com/articles/Classic-WTF-No-Quack

> replacing deterministic systems in their support flows

The issue is, they don't want to provide "better" support but "cheaper" support. Imagine a trained agent that understands the big picture. Now imagine a company investing in humans to use AI to retrieve knowledge that the human can easily identify as being relevant or not, and using that knowledge to better aid the customer.

Right now AI is being sold as a "we don't need support personells" instead of "how can we provide better service." For a lot of products, better service will probably not matter as "cheaper" products will win most of the time.

Most people don't want to pay for better. They want to pay the same for something better, which is what companies are not investing their time in figuring out how to use AI properly for I think.

That's the completely opposite of what people should do. The laborious task of programing logical work flows is the only reason AI is useful for me.

Yeah but did number go up? Can CEO check a box to show investors?

Now that’s real value.

As a contractor who built a lot of predictive systems and workflows in last three years I can tell you that quite often there is a specific request to put AI into it even when it is not needed and would objectively make the system worse, slower and more expensive.

The AI psychosis is a real thing.

With inexperienced or non-technical people, talking to them about AI can be very confusing, as a LOT of their "AI" usecases are basically they didn't realize or know how to write a program for this straightforward logic.

models will get smarter, this wont be an issue

Maybe this is because I live in Wyoming, but "AI is not ubiquitous, there are some people, like Vegans, who eschew it" is not the most compelling argument.

I assume for a lot of people, an llm is going to produce higher quality results for most knowledge tasks than they could do on their own. I think that's okay

> Anyone who does a Google search gets a satisfactory looking answer as the very first entry.

Google has search results still? I don't use Google much anymore (thanks Kagi), but this is what ends up showing for me, I don't even see any search results anymore: https://i.imgur.com/eHIA2Df.png It seems like it's 50/50 on page reload if the LLM-reply UI expands automatically or not, which covers my entire screen. I guess Google is doing some A/B testing perhaps.

What Im question is how is Google increasing Price-per-Click each year if people are clicking less and less on the links below the AI search result

Even if we accept that all people are satisfied with the AI search overviews, that would still only be everyone for one thing.

When was the last time you used Google? The first entry (and a few after that) is always spam.

Anyone who does a search and accepts the first answer just doesn't care much or is incompetent. Anyone with any critical thinking whatsoever does way more than that if they want a correct answer.

> But ironically, I used AI to build it

This is a pattern I encourage - the AI might not be reliable, but with coaching, it can produce reliable tools. `colordiff` was causing issues with `less` when I was looking at diffs (character encoding issues I think), and when I asked Kimi K2.6 what to do, it built me a rust command-line diff tool in one shot that I've been using ever since (it even downloaded rust, wrote the tool, and compiled it).

Have you seen jscpd? What does your tool do differently?

That's funny, Google Gemini and AI mode in search has replaced my ChatGPT prompting, because I know Gemini will correctly cite sources (as of course it's by Google) rather than hallucinating.

Also, Gemini is free or at least has much higher usage limits than ChatGPT or Claude, and it's well integrated into Android and soon Apple with their new Siri, so things like circle to search just work well.

In my non-tech circle, most people don't even realize how the internet is running literally everything. Even if we start to use mass scale AI for something, they wouldn't realize or care much about it. They at best turn on the TV to watch netflix or look at the phone to send messages on whatsapp. If all of that went away tomorrow, they'd be inconvenienced at best and then go on with their day to day life. This feels like we are literally all in our IT echo chamber where we throw stuff on walls and go crazy, while the world is sunshine and rainbows, always been.

You'll find it hard to pin down what you mean by "everything" otherwise you wouldn't have said that. Nobody uses the internet for everything.

Local models are highly likely to dominate in the long run as "good enough" inevitably becomes trivially cheap. This is a very different pattern of incentives and adoption compared to the internet.

I think it's more similar to the advent of personal computers. They had a brief surge and then turned into something else (smartphones, cloud, etc.) for all but a few niche cases. AI is not changing the consumer landscape. It's getting absorbed into existing platforms where there's a clear use case and benefit. It's just another expected software feature. This is far from the first time people have rejected a "personal assistant" concept and they'll just keep rejecting it.

Embarrassing NPC behavior to throw in the towel on working because you lost your crutch.

I have no doubt that as AI gets more expensive, my employer would lay off more developers to pay for more AI tokens, until there are very few developers left. And the hilariously sad part is, the current developers keep training the AI to do their job. Eventually I expect they will lay off almost all the developers. It really feels like we're going to be stabbing each other in the back just to be the last one to get let go.

I don't think AI has any real value for software development, personally. The quality just isn't there, unless you invest so much effort that you may as well have written it yourself. But the market can stay irrational longer than you can stay solvent, and even though I think the industry will get over the idiocy of having LLMs write software, there's no telling how long that will take. So it's a scary time to work in tech even if I think the trend will ultimately reverse.

Yes I believe software benefits uniquely, just like building tooling and automating software have long been easier in software than other domains. Humans defined all the rules of the world you live in, humans wrote strict rules in methodically parsable formats.

The moment you have to interact with the physical world or humans (psychological, imaginative, aesthetic, etc), there are often undiscovered or changing rules—or no rules at all. Or systems are subject to perturbations beyond a defined scope.

The other thing I believe is software developers are experts at doing the things that allow them to make doing those very things easier and more automated. And they do this in public, perfectly documented online.

Both because of the things I described above and because software developers have created the largest machine-accessible training set for plying their trade of any trade, ML—that is ultimately interpolating massive datasets to do things—is unsurprisingly uniquely successful for software tasks.

I've been thinking about this, and I think software is uniquely knowledge work that has the most defined structure and least personally interaction. Hell, some of the software I write is for machine to talk to other machines. It's not surprising such a closed system is so amenable to AI, and other knowledge workers are not getting the same benefits.

Software has huge and detailed code repositories ripe for training use. There's just enough inference in current models to remix that code in useful ways for the most popular languages.

The less popular a language, the more models struggle.

Writing, UI, and presentations have similar knowledge bases.

Outside of those, quality becomes much more hit and miss. If you ask for a recipe you may get something good, or you may get something completely inedible and random.

"Domain specific knowledge" really means "strong foundations and relevant abstractions" and LLMs just don't do that reliably.

That's a decent article. My only issue is it seems heavily biased at the end, or at least he seems to misunderstand what the 'A.I. types in Silicon Valley' are doing.

> Computers should adapt to people. Asking people to make themselves more legible to software — to turn themselves into a database — is a doomed idea.

I've been in software a long time, and I do sort of see this trend, but I think it's because these are tools that build other tools. The interface has always been a 'best I can do for now' thing, with the focus on doing things that are useful. Computers were just calculators in the beginning, which led to more complex calculators, instruction sets, programming languages, operating systems, GUIs, interconnectivity, etc.

What people are doing today is experimenting, like they always have. They're putting their experiments out there so that others can use them and build on them. Some will use those tools to build other tools, and some won't. But over time, the experiments that work will get distilled and turn into real products that people who 'do not yearn for automation' will still want to use, so it seems like the value is there.

I guess the real question is whether they will create value that offsets the near-term costs, because I don't think the billions in investments are sustainable, and I'm not convinced the centralized data center paradigm is the right way.

Software engineers aren't even all using AI, contrary to frequent claims here that they are. There are very many who have tried it, found it didn't add value to their work, and aren't using it unless FOMO-driven managers force them to.

Great examples. I think people not using AI for issues like these lack imagination or more charitably, simply don't know that it works so well for these. Especially non-technical people can find great value out of AI, not just SWEs.

I think what people mean by everyone varies a lot, which is why I wanted to draw attention to more specific numbers. For example, in the Datos data cited[1], on desktop 86% were using traditional search engines >10 visits/month vs. only 21% for AI chat tools. That is indeed a very significant percentage, but more than 4x less than search and (at least I) wouldn't say that ~1/5 is "everyone."

[1] https://sparktoro.com/blog/new-research-20-of-americans-use-...

yeah exactly my thoughts. everyone didn't mean literally everyone.

and for the ones that are using it (especially the paid subs). the lure is undeniable.

Kinda like how quickly cellphones popped up and changed everything.

It's funny lately I've been seeing the cursor advertisements all with some premise of regular young person wants to develop an app and the ads really do focus on the simplest of premises: the only ones I've seen in these skits are essentially variants on the "todo app" web app tutorial

the tech is pretty good at helping identify simple bugs when they happen and to write short sections of code given very explicit instructions but yeah I have yet to see good examples of short one sentence ideas turned into a working product that looks better than anything that could be a UDemy tutorial app.

How much of that use is driven by corporate mandates to use AI anywhere and everywhere (even when it's a terrible fit)?

It's a retrospective analysis of an assertion made by NYTimes. The original headline wasn't clickbait, just presumptive, and even so, it's a pretty significant publication that spends a lot of time on the HN front page (alongside you, I'll add). I think it's perfectly fair, and nowhere close to a strawman, to deconstruct that claim a year later.