Writing code by hand is over, forever | Hacker News Reader

Writing code by hand is over, forever

eliocapella.com

22 pointsaray0757 comments

This whole LLM coding mania has highlighted something that I was previously less appreciative of: some people (self included) like writing code because it is an enjoyable, enriching experience to them. Solving puzzles in code has lit up the reward pathways of my brain since the first time I tried it, and I don’t see why that would ever change.

Prodding an LLM into solving a problem is just not the same thing. I understand that the results of prompting could be more exciting than programming to some, but the idea that manual coding is “over” feels dismissive of hobbyists at the very least. Perhaps the author is suggesting that manual programming in the industry is over, but I don’t buy that either. Thesis being, there will eventually be a cost to employing “experts” that lack understanding.

> If agents can handle the code, could we finally work directly with the logic, interfaces, and data models we want?

Aren’t high-level programming languages very refined encodings of “logic, interfaces, and data models?” Why would using natural language be more efficient at such things than special-purpose languages developed over decades?

> Aren’t high-level programming languages very refined encodings of “logic, interfaces, and data models?”

I feel like I would be hard pressed to explain what I want in English more concisely than just writing a SQL query.

What everyone is figuring out about skill-building: "But you know English, do you need to learn arcane incantations of X".

Seems like some people didnt get (yet) that code is just a better way of prompting. Some even tried to make “executable markdown” and “typesafe markdown”. Im speaking from a Golang perspective, dynamically typed languages may be different.

If you never write the code, you'll never be able to read the code effectively. Also, you will not pick up the ability to reason algorithmically. This is why I tell students that they must write the code in my classes:

https://htmx.org/essays/yes-and/

Leaving aside the power you are giving to the AI companies by entirely giving up on coding. I say this as someone who uses AI extensively, if ambivalently.

I see the same thinking in AI-art boosterism: having "good ideas" is the important and "creative" part while actually implementing them is drudgery best delegated to someone (or something) else.

If humans are in the loop at all, I don't see why "writing code" is worse than "telling the computer what to do".

With the right high-level abstractions, I can write code as fast as trying to convert that thought into English (for some tasks).

How fast can you get it into the repo, tested, deployed?

How fast can you scaffold new ideas and delete them when they don't work?

I’m a developer of 30 years, but writing a game completely with AI. It writes all the code and has something like 200 test cases now . It does well with what it understands but there are all kinds of things it fails at. Like it didn’t blink at an interactive turn-based game taking 10 minutes per turn (instead if an expected 15-30 seconds). That was caused by recomputing everything rather than caching, which I pointed out. There were 3 areas that referred to something we changed. I had to point out what was no longer coherent. AI can write code well, but it just doesn’t have the memory/context size to drive a real world project of any size. It still needs someone else to do the thinking and planning. If I didn’t code I don’t think it would have known where to look to fix many performance problems. I’m not at all worried about it taking my job.

some/most of that could easily be optimized for: set expectations for turn-time ahead of time, and it will meet them. if you had even just given it a clock to pay attention to, it would have noticed on its own, i reckon.

i'm not having problems yet in my biggest ai-only codebase, which is over 100k lines now. i haven't had any model hiccup, actually, but the smaller models take more wall time to get it done.

i have the models writing in rust, i believe it helps. ymmv.

i don't think it always architects well or anything, but it has quit building broken shit. it gets lazy, builds dumb UI, puts way too much text everywhere--but only if i don't remember to give it some better guidelines, or a visual reference.

i think with the right context, they build the right thing, and we haven't totally solved that just yet. but close enough for me; i'm so thrilled. i just want opus 5.5 to be crazy fast, like chatjimmy.

I believe both you and the grandparent will experience diminishing returns in AI performance as the projects get closer and closer to completion.

But due to the sunk cost fallacy and the belief that others have made this work you'll keep plowing money into the AI Slot machine.

I think you might be right. For $20/month (Claude Code) it’s worth it. If I’m forced to upgrade to $100/month I’ll drop it.

I got >$100 of value in about 2 minutes of effort yesterday.

I typed a design I wanted, it built it. 2 minutes of my own time, 1% of my weekly limit.

I actually think the design would have cost 4-6k for an equivalent thing by a human, and would have had them laboring for a couple weeks.

Well, my project IS complete, although I keep adding new stuff--it's going to be that kind of app, it rotates content in (minigames).

It has about 8 minigames, leaderboards, awards, achievements, async image generation, inventory, inventory-interaction-with-minigames, and various other account features, mailers, notifications, blah blah.

Literally haven't had one issue with context or micro-managing the agents on this project, iirc. I built it in a couple weeks.

Minigames are like "putt putt golf w/ daily generated levels", "slot machine," "video poker", etc. Not like "tap the fastest" or something with no complexity.

The plot labeled “Incoming bug reports over the same period.” is funny. Where does the red line come from, except arbitrarily setting 20 and 15 bug reports and plotting that? There’s no discernible connection to the actual data.

Even if the numbers made sense it would be pretty damning for OP's company. He's outright admitting that Filestage's engineering team is so dysfunctional that the code they were producing is even worse than slop?

I have a hard time believing the author’s team is doing all their coding with AI, with only the $20/month tier of subscription. When I experimented with writing zero code by hand, my token spend shot up to more like $300/mo.

I’m also not ready yet to conclude that that option, as opposed to a more modest version of AI assisted coding, will remain open to me indefinitely. We now have financials confirming that Anthropoic, too, is running a -200% profit margin, which leaves serious cause for concern about whether either GPT or Claude models will still be around in another year or two. And the judge is still out on whether local models can perform well enough to actually do all my coding for me.

I’m confused how LLM code is so different from open source code, from the perspective of the developer.

Like if someone gave you a magic wand that let you conjure a mature open source library that did most of what you needed for your project, that would be amazing right? Would you refuse to use it because you miss writing your own database or API framework? Or would you just get to build way more ambitious things?

Software is already stitching a bunch of things together in most cases, why not stitch together things made by a machine instead of by an open source developer? What percent of developers primarily work on bare metal algorithms? .0001%? Everything else is stitching together existing pieces at some level of abstraction.

The real difference seems to be the non-determinism, not that the code is produced automatically. And I think a lot of the work right now that people don’t like seems to be people dealing with the non-determinism, not that so much code can now get generated.

I have read accounts of people have similar freak outs about higher level languages, object oriented programming and web services, and all of those just made more demand for software because the potential of software has not nearly been reached. I wonder if the extreme growth of the field has resulted in most people feeling like those huge changes that reduced the amount of code that had to be written per unit of functionality have been around since the dawn of time, and now that another level of code generation has been reached it’s the end of the world (again).

Figuring out how to fit together off-the-shelf parts to solve a problem is (was) fun

You're still allowed to do that, even if you don't want to write the glue code.

Well cause I don't really care about the code so much as I care that there's someone or some group maintaining it. You've never looked at the last commit time when you want to include some library in your project? I need proof of some buy-in from someone else. That someone is putting their time on the line. Otherwise I know using that package is gonna cost more time than its worth.

Right, but this is a problem with using it, not the problem with it destroying software as a profession.

If I wanted a pyramid, I'd want to build it out of stone blocks and be sure that each block is actually stone. I wouldn't want to get halfway and find out that they were actually fragile.

Trusted engineers (of open source projects) can deliver me stone, both the authors and the community (often >100,000 devs) have checked that a specific block is stone and can bear the load.

AI is proven to output all sorts of random stuff. Sometimes a pebble, sometimes foam. It almost always looks like stone to most analysis. However as soon as it's load bearing it's likely to crumble.

Agentic agents have a habit of rebuilding the pyramid and randomly replacing stone blocks with "something else" but we don't know what just yet.

I question the epistemology of people who did not realize this by around late-2024. (Sonnet 3.5 was out and could do some things okay.)

It's defensible to say "October 2026 is not yet the time when we should be 100% deferring code writing to AI (let alone not even looking at the code any longer)". But the writing is very obviously on the wall. It's likely not going to take longer than 10 years.

I adopted Copilot, Claude and Codex quickly, but I remained skeptical how far they would get us ... until the beginning of the year.

Then we saw the capabilities really take off, and it was obvious how things would go.

I am not claiming to be a great forecaster, or anything, but to me "it sucks now, it's gonna get better, and then it's gonna get even better" was already apparent to me as highly likely starting around GPT-3. (The rationalist bloggers apparently already knew it as of GPT-2. Scott Alexander and Gwern, for example.)

I did not know when it would become so practical that I'd not feel much desire to hand-write code, but I was pretty confident it was coming.

Maybe handmade code will become like handmade watches, a luxury item or a hobby activity. But I fear that the loss of understanding will lease to worse software and more frustrated users. We all have to see how easy it is to maintain auto-generated code.

So many of these pro-AI articles lean on references to "people used to be scared of calculators!" and the like, but it's a nonsense comparison. Calculators weren't predicated on monetizing prior work at the cost of devaluing future work. Functioning calculators are at least as accurate as a human, generally more so, and they do not produce incorrect outputs when given correct inputs no matter what. And calculators don't lie to you about what they've done, even if the result will make you sad. The comparison doesn't hold up.

(Also, I think some of the concerns about calculators at those times actually were valid, even though I wouldn't argue that society should not have adopted their use.)

Just put the code in the bag.

Complete hogwash and nonsense article written by a hack.

I don’t care, I’m still gonna do it

i ain’t reading all that

i’m happy for u tho

or sorry that happened

Deleting code by hand will thrive

i'll be deleting code for a couple of model generations. the new, better model will delete bad old code written by older models.

and then, at some point, that will cease to be able to happen. they'll be good enough.

i predict that happens before March 2027.

until the model is unavailable

Sentiment is mostly correct, but the writing is terrible.

Reply on news.ycombinator.com