Writing code without an LLM is akin to cooking a meal yourself, as opposed to something in between hiring a private chef and pulling a slot machine lever. When I cook a meal for myself or my loved ones, I become deeply familiar with the ingredients and the recipe. I am a good cook. Every decision, small and large, I make myself. Using feedback, I can tweak the components and the process.
However, if I order food and call it my own creation, I’ll be speechless when my nut-allergic friend asks me if I used peanut oil or olive oil. And does it taste good? The end result is subjective, not binary, like your three similes. What if the food sucks? Must I now micromanage the chef? I don’t want to make them cook it again and again, hoping each time it’ll taste little more like how I want it to. I’d rather just cook it myself; I am a good cook, afterall.
Well to me using an LLM is akin to using microwave for food, and programming is getting your hands dirty in the art of cooking. I used to do the latter but the pace is too much, so I'm forced to go for the former. I still miss the days where I can cook myself
This is a good analogy. Also I hate cooking; it feels like a chore and waste of time. Given the means, I prefer to have someone else do it for me so I can focus on more interesting things.
If the analogy is good, does it mean AI boosters hate programming, and feel like it's a chore and waste of time? Yet they keep calling themselves programmers.
I dislike cooking; I don't call myself a chef, nor go on cooking forums saying they're all idiots for churning butter as a hobby, and telling them they are luddite dinosaurs just because I bought myself an Instant Pot.
And clamoring that instant pot is capable of doing the job of 10 chefs, while conveniently forgetting that chefs using the instant pot is way more useful. And then congratulating each other on new way of adjusting the temperature and timer of the instant pot.
I don't see it the same way. I see it as akin to building a machine which can do the cooking for you. There is a lot of experimentation in terms of how you can construct that machine and how you can set it up, and for some people the payoff is in exactly this process.
That's me. I've just moved the puzzle to figuring out the CC optimal "language" and mode of working with an LLM. At the same time, I haven't lost my love/expertise... Yet.
I explained it yesterday that it's no different than Linq powered by EF Core. You still need to understand the database and how SQL will work, or should. If you don't, you're queries will be slow and inefficient, hazardlessly oftentimes.
LLMs are an impressive abstraction, but they are not magic...yet. or maybe it is, but you still need to figure out the incantations. Not that I like to uptick and witchcraft, so I hate that analogy.
My doc didn't do the full set of test/exam for an hEDS diagnosis, so technically it's just generalized hypermobility spectrum disorder based on past medical history and a near "perfect" beighton score. It could be hEDS though, as a lay person reading the diagnostic criteria I either fit it, or are borderline.
Injury was to my SI joint, I've historically always irritated it lifting, but I set a new PR for deadlifts and it was debilitating for 3 weeks before my other half made me stop being stubborn and see a doctor about it.
I've also had my left shoulder joint surgically repaired after multiple dislocations.
I find very little joy in trying to wrangle the blackboxes that are LLMs. The undeterministic nature of them frustrates me, and feels nothing like the software engineering I know and love. However, I know I’m in the minority here, as almost everyone else in the industry I’ve talked to seems to love using them.
I agree with you. It’s actually similar to the frustration of using ill-documented, ill-behaved libraries or tools that are either closed-source or otherwise inscrutable. They may even be deterministic, but what’s frustrating is when you can’t reason about their behavior, can’t predict what will happen when you change x to y.
A lot of people here agree with you, and I definitely get what you are saying! It's a personal thing in the end, as people enjoy different things even in the same areas of expertise, and there are definitely a LOT of things to criticise AI about (and the companies around them).
I absolutely love deterministic, reliable, robust software so I find the challenge of non-deterministic agents completely fascinating and exciting. Programming for the user has always been a challenge (especially in the browser where the environment itself isn't consistent), and now we have the same kinds of challenges across the stack.
I'm enjoying that aspect of it so much. How can we make it safer, more reliable, predictable? I don't think we'll have totally satisfactory answers for a long time, and it'll be a lot of fun to explore the space until we get there.
I think what really kills me is allowing non-deterministic agents to write non-deterministic code and not really being critical of the process or the results. That's poison for our industry, newcomers to the field, and frankly the profession itself. I'd like to see us take this more seriously.
Do you find joy in using LLMs to write software? I tried using Claude/Cursor/CodeX/etc. for personal projects and experimentation, and I found no joy in it. I learned nothing, and when my MVPs were complete, I only had a shallow understanding of how the code that powered them worked.
I do, but I also use LLMs in a manner that seems drastically different, from most folks here.
I use the standard $20/ChatGPT Pro sub, and run Thinking 5.5 as a chat interface.
I use it like a "trusted personal advisor," as opposed to a "black box employee."
I'm intimately involved in almost every step of the development process. Most of what I ask from the LLM, is function-length snippets.
It's made a huge difference in the velocity and scope of my work.
I have learned that I need to be very careful, though. The LLM sometimes really borks things, and I have to rip out the garbage, and rewrite the code, myself. I can't even imagine the quality of "vibe-coded" software.
Interesting, why do you think it's an attack? I was skeptical, and I felt there is a common theme. if I'm wrong, I sincerely apologize, but I can't brush off the feeling.
But hey, maybe I'm really over-thinking, so I'll go off.
I'm not advertising jack. Everyone else mentions the LLM they use, along with all kinds of details, like token costs, etc. I mention it to frame the context. I don't think that using ChatGPT impresses anyone, around here.
It is not a common theme. However, when I do post something like this, I find a lot of kindred spirits, so I guess we are the "quiet ones."
Most people -the vast majority- use agents and IDE integrations; sometimes, in amazing ways. It's a very different way of using LLMs from the way that I do. Maybe the way I use it is considered "quaint," and people don't want to admit it, because they are afraid folks will make fun of them. I don't really give a rat's buttocks. I'm retired, and long past the need to feed my insecurity by accepting the judgment of others.
I am big on checking out people's profiles, when I am interacting with them. Sometimes, it makes a big difference in the way that I approach them. That's why my own profile is packed full of information. I'm not showing off -many folks here, are a lot more impressive than I am- I just want people to know who I am.
But that doesn't prevent the usual Internet Ready, Fire, Aim approach.
And one habit that I deliberately foster, is not engaging folks that want to attack me, beyond one or two mild responses. Once I say "Have a Great Day!", we're done. You can add whatever last word gives you good feelz. I won't respond.
I also don't attack. I respect this community, and engaging in troll-battles, just makes it ugly. And I could be really good at trolling; I just feel as if you can't shovel shit, without getting it on you.
We're good, here. I checked out your profile (and GH profile).
We probably have a lot more in common, than differences. I'm always glad to find people to interact with.
I know that my HN persona is a bit "stuffy," because I'm going out of my way, not to be abrasive, and to contribute to the community, but there are folks that absolutely hate me -in a Commissioner Dreyfus kind of way-, and I'm not really sure why. Maybe it's the Apple thing.
Eh, whatevs. I'm "on the spectrum," and got used to people disliking me for no reason that they can even articulate. It used to really bother me, but these days, it's just background noise. I'm actually a fairly decent chap, and probably worth getting to know.
Glad to hear! And I understand, yeah there are weird people who just want to make things difficult for everyone else.
And, I really believe you are, and glad for having this conversation, This is something, I believe we all fell in love with the internet heh. As I'm going through a bad patch of life, this made me smile. Thank you.
ChrisMarshallNY has consistently posted like this about their work/hobby post-retirement for some time, and often with a bent on 'I do things my own way'. Didn't seem like advertising to me. Including a version number is hardly unnatural extreme effort in a comment, surely?
I like building stuff. I don't care about the code. I convinced myself over 20 years that I liked coding to get myself through the drudgery of corporate work but the reality is the building was the important thing to me. I'm able to build things quickly with AI.
A carpenter can build a house with commercially produced dimensional lumber without having to harvest the trees, cut them into boards of standard dimensions, kiln-dry it, and transport it to the building site. We still say he "built" the house.
At some point, though, it's fair to ask "what did I add?"
If you are directing other people (or maybe robots, someday) to do the actual building, you can't really claim to have "built a house" even though it all happened under your direction.
I guess bottom line for me to feel I had a real part in "building" a thing would be whether I added anything I am specifically qualified or skilled at doing, or did I do what anyone could have done?
I plant a seed, a human being carries it to fruition for 9 months through her determination, love and literal life blood, and I can claim I brought the seed to life.
He plants an idea and patiently coaxes a machine to spit out the code that matches it, you're giving the machine the credit.
If he was constraining the AI's output, his input was meaningful.
Did you ever bring anything to life? Whatever it is, you are also relying on the work others have done before you and standing on the shoulders of giants.
I guess some people only enjoy the destination and don't care how they got there. These people seem to enjoy AI more than the people who want to enjoy the journey along the way.
What? For the record, I am not saying that you shouldn't ever use AI. Plenty of good programmers use it as a tool to boost their productivity or to validate things.
But if you are using AI to write all your code, I think you are missing a lot. It is like when you use LLMs to write your whole paper for you, rather than to check your grammar or offer critique of something you actually wrote.
I'm curious of the places you've found joy while writing software traditionally. For me, it has been in reasoning about the system, debugging issues, and discovering what works. The iterative process of eventually coming to a more complete understanding, as you stand on and build off of your prior understanding.
All of those elements are present for me while using AI to augment my output. I have started using voice to interact with my coding harness though and I think that has maybe influenced my opinion. I also don't let things go fully autonomously and look at the diffs along the way.
Depends on how you use them. I'm a detail-obsessed perfectionist. I believe these qualities are what have enabled me to produce better software than most people. I use LLMs the way I can without violating these principles.
LLMs have gotten so good at coding I often find it exhilarating. I'm mostly doing experiments in accelerating machine learning in hardware. I can do way more experiments now. As an example: Just this afternoon working with claude to experiment with using a GA to learn a non-uniform 1D CA rules to calculate the popcount function where the GA is running in an FPGA to speed it up. It's non-uniform in that each cell gets it's own 8-bit rule.
I asked the LLM to explain, rephrase, or rewrite things until I was happy. Some examples :
I asked for examples of how the algorithm worked. I asked for examples of how to call the code. I asked for a happy-path unit test and a simple error-handling unit test. I asked it to rewrite something as a pure function. I pointed out an obvious race condition and told it to guard against that issue. I asked it to rewrite a function in the style of this other function. I told it to separate one function into two separate functions that handle the first step and the second step separately.
Etc etc.
If you don't understand it, ask for more or better comments, or better variable names, or cut down the scope into a smaller section, or more examples.
Edit: also I almost entirely leave the LLM in read only mode... I tell it to make the smallest change possible, and tell it I will only copy paste it in its proposed change when I understand the change and where it needs to be made. That way it's my hands on the keyboard, interacting with the code by making recommended changes... 80% of the code is touched by me (via copy-paste) most-of-the-way before I will 'git commit'.
Sure, there was one recursive folder descent function that found the most recent file modification time that I didn't fully understand, but it's self-contained in a function, I don't care to learn every corner of file modification times, and it appears to work, so I left it as is for my static site generator.
And there’s good reason for that. Anthropic, OpenAI, Salesforce, and so on have aggressively marketed LLMs as better than humans at working. It’s no surprise when we find out something is build using an LLM, we expect it to match the marketing.
But what constitutes "better than humans at working"?
Zero defects? Because you can always find at least one defect. But people don't naturally think statistically, so they grasp the thing that confirms their bias and then hang on tenaciously.
You'll notice the incredible amount of vitriol resulting from a purely cosmetic bug (which, it turns out, results from a missing TERM env in the base image - Claude is very conservative when it can't determine utf-8 support with 100% certainty).
Do you have a source for your “millions” claim? Every source I can find says there are somewhere between 60 and 70 thousand immigrants currently in detention centers, based on internal ICE data. That’s still a fuck-ton of people to cage up, there’s no need to hyperbolize.
There’s an optimal number of employees required at any productivity point.
Why don’t Google hire 3 times the number of developers? They have the money right? What’s your logic for not hiring more?
Hiring and firing people aren't symmetric actions.
They're asymmetric because hiring more people costs more than just the salary. For example, some folks' entire jobs are to recruit and hire people. Once they are hired, you have to onboard them, etc. So the more you hire, the more you have to pay the folks with supporting roles (either directly or by way of them not having infinite time/capacity).
Firing people isn't free, either. It comes at the cost of bad PR and severance, but the latter is voluntary and calculated by the company, and the former is quickly forgotten by anybody that matters to a publicly traded company (investors).
That means not hiring those two people in the first place is usually cheaper than firing them later.
To the original point: Cloudflare isn't hiring fewer people; they are firing people. If they are trying to grow (like every single investor is counting on them to do), then why would they fire people (the cheaper action) now when they would likely need to hire people (the more-expensive action) later in order to meet that increased growth?
The charitable answer would be that the people they are firing were deemed unable to adapt to using AI for all of this supposed increased productivity. But Cloudflare aren't saying that. In fact, they're saying the opposite by stating it's not about individual performance.
your's is a caveat against my larger more correct point: there's an optimal number of employees needed at any given productivity point.
its true that hiring and firing are asymmetrical, and CF has shown that they are willing to bear the brunt of the asymmetry and fire people despite the downsides.
that asymmetry lies doesn't disprove the original point: cloudflare simply doesn't require the _same_ number of people to work for them with AI.
if you disagree with this then you believe that companies should only have monotonically increasing number of employees which is quite ridiculous a claim
Unfortunately I think we’ll see more and more of this as companies continue to encourage their employees to use LLMs everywhere and for everything. Eventually they will have to come to terms with the cost of such mandates, and it’s either ask your employees to “use AI less,” or it’s let some percentage go and continue to let the rest burn tokens.
I am perplexed at how it can cost so much. I have been using AI every day, all day for a few months now and I have not even gotten to spending $300 a month. I use Cursor for teams, so we get ~$80 of usage for our ~$40 per member, then we pay Cursor's upcharged API rates from there, and I STILL don't spend more than $300 a month, if that. What the hell is everyone doing with their fucking tokens?
