Indeed, that's what I kind of hinted at in https://news.ycombinator.com/item?id=46442195 and coincidentally https://news.ycombinator.com/item?id=46437688 briefly after, namely that OK, one can "generate" a "solution", that's much easier than before... but until we can verify somehow that it actually does what it say it does (and we know of hallucinations and have no reason to believe this changed) then testing itself, especially of well know "problems" is more and more important.
That being said, it doesn't answer the "why" in the first place, an even more important question. At least though it does help somehow to compare with existing alternatives.
Folks think, they write code, they do their own localized evaluation and testing, then they commit and then the rest of the (down|up)stream process begins.
LLM's skip over the "actually verify that the code I just wrote does what I intended it to" step. Granted, most humans don't do this step as thoroughly and carefully as would be desirable (sometimes through laziness, sometimes because of a belief in (down|up)stream testing processes). But LLM's don't do it at all.
They absolutely can do that if you give them the tools. Seeing Claude (I use it with opencode agents) run curl and playwright to verify and then fix it's implementation was a real 'wow' moment for me.
We have different experiences. Often I’ll see Claude, et. al. find creative ways to fulfill the task without satisfying my intent, e.g., changing the implementation plan I specifically asked for, changing tolerances or even tests, and frequently disabling tests.
Yeah I feel that, if it happens your only way out is to write down a more extensive implementation plan first. For me that is the point where I start regretting to have tried implementing something using AI,.. But admittedly most of the time redacting the implementation plan and running the agent again is still faster than I could have done on my own (I try to make implementation tasks explicit in the form of a markdown file, worked pretty well so far).
I see these “you had a different experience than me” comments around AI coding agents a lot and can concur; I’ll have a different experience with Copilot from day-to-day even, sometimes it’s great and other days I give up on using it at all it’s being so bad.
Makes me honestly wonder — will AGI just give us agents that get into bad moods and not want to work for the day because they’re tired or just don’t feel like it!
Don’t downvote because you don’t like the question.
It obviously adds to the discussion: paid and non paid accounts are being conflated daily in threads like these!
They’re not the same tier account!
Free users, especially ones deemed less interesting to learn from for the future, are given table-scraps when they feel it’s necessary for load reasons.
Exactly. There's an impedance mismatch between those using the free/cheap tiers and those paying a premium, so the discussion gets squirrely because one side is talking about apples and the other oranges.
> LLM's skip over the "actually verify that the code I just wrote does what I intended it to" step.
I'm not sure where this idea comes from. Just instruct it to write and run unit tests and document as it goes. All of the ones I've used will happily do so.
You still have to verify that the unit tests are valid, but that's still far less work than skipping them or writing the code/tests yourself.
I disagree it's less work. It just carte blanche rewrites tests. I've seen it rewrite and rewrite tests to the point of undermining the original test intention. So now instead of intentionally writing code and a new unit test, I need to intentionally go and review EVERY unit test it touched. Every. Time.
It also doesn't necessarily rewrite documentation as implementation changes. I've seen documentation code rot happen within the same coding session.
One commercial equivalent to the project I work on, called ProTools (a DAW), has a test "harness" that took 6 people more than a year to write and takes more than a week to execute.
Last month, I made a minor change to our own code and verified that it worked (it did!). Earlier this week, I was notified of an entirely different workflow that had been broken by the change I had made. The only sort of automated testing that would have detected this would have been similar in scope and scale to the ProTools test harness, and neither an individual human nor an LLM is going to run that.
Moreover, that workflow was entirely graphically based, so unless Claude Opus 4.5 or whatever today's flavor of vibe coding LLM agent is has access to a testing system that allows it to inject mouse events into a running instance of our application (hint: it does not), there's no way it could run an effective test for this sort of code change.
I have no doubt that Claude et al. can verify that their carefully defined module does the very limited task it is supposed to do, for cases where "carefully defined" and "very limited" are appropriate. If that's the only sort of coding you do, I am sorry for your loss.
> access to a testing system that allows it to inject mouse events into a running instance of our application
FWIW that's precisely what https://pptr.dev is all about. To your broader point though designing a good harness itself remains very challenging and requires to actually understand what value for user, software architecture (to e.g. bypass user interaction and test the API first), etc.
No I was sharing an example of a framework that does include "a testing system that allows it to inject mouse events".
That being said mouse events and similar isn't hard to do, e.g. start with a fixed resolution (using xrandr) then xdotool or similar. Ideally if the application has accessibility feature it won't be as finicky.
My point though was just to show that testing with GUI is not infeasible.
Apparently there is even a "UI Testing for devs & agents" https://www.chromatic.com which I found via Visual TDD https://www.chromatic.com/blog/visual-test-driven-developmen... I can't recommend this but it does show even though the person I was replying with can't use Puppeteer in their context the tooling does exist and the principles would still apply.
> My point though was just to show that testing with GUI is not infeasible.
Indeed, which is why I mentioned the ProTools test harness and the fact that it took 6 people a year to write and takes a week to run (or took a week, at some point in the past; it might be more or less now).
With the NES there are all sorts of weird edge cases, one of which are NMI flags and resets; the PPU in general is kinda tricky to get right. Claude has had *massive** issues with this, and I've had to take control and completely throw out code it's generated. I'm restarting it with a clean slate though, as there are still issues with some of the underlying abstractions. PPU is still the bane of my existence, DMA, I don't like the instruction pipeline, haven't even gotten to the APU. It's getting an 80/130 on accuracy coin.
Though, when it came to creating a WASM target, Claude was largely able to do it with minimal input on my end. Actually, getting the WASM emulator running in the browser was the least painful part of this project.
You will run into three problems: 1) "The Wall" when any project becomes large enough, you need the context window to be *very* specific and scoped, with explicit details of what is expected, the success criteria and deliverables. 2) Ambiguity means Claude is going to choose the path of least resistance, and will pedantically avoid/add things which are not specced. Stubs for functions, "beyond scope", "deferred" are some favorite excuses to not refactoring or implementing obvious issues (anything that will go beyond the context window, Claude knows, but won't tell you will be punted work). 3) Chat bots *loooove* to talk, it will vomit code for days. Removing code/documentation is anathema to Claude. "Backward compatibility", deprecated, and legacy being its favorite.
This sounds exhausting, once the thrill of seeing code rapidly generated wears off, I wonder if it's even worth it. If someone was going to use code they didn't write, why not just pull down some open source implementation from somewhere and build on top of it? It's basically gets you the same thing but without the LLM hassles, and you can start building on a more sane foundation.
Give it copy paste / translate tasks and it’s a no brainer (quite literally)
But same can be said of humans.
The question here is, did it implement it because it read the available online documentation about the NES architecture OR did it just see one too many of such implementations.
Indeed, the 'cleanroom' standard always was one team does the RE and writes a spec, another team that has never seen the original (and has written statements with penalty clauses to prove it) then does the re-implementation. If you were to read the implementation, write the spec and then write the re-implementation that would be definitely violating the standard for claiming an original work.
It’s a shame that the source code isn’t commented and documented more. At the very least, I would see it being helpful to add some documentation for every CPU op code being emulated.
Forbidding LLM to write comments and docstrings (preferrably enforced by build and commit hook) is one of the best "hacks" for using that thing. LLM cannot help itself but emit poisonous comments.
I used to worry that using LLMs to code would let them use my code and train on my hard work. Then I realised how bad my code is, so I'm probably singlehandedly holding off an agi catastrophe.
Meh. No human has written the horrors llm produces. At least I am yet to see codebase like that. Let me attempt a theatrical reenactment:
// Use buffer that is large enough to hold any possible value. Avoid using JSON configuration, this optimizes codebase and prevents possible security exploits!
size_t len = 32;
// this function does not call "sort" utility using shell anymore, but instead uses optimized library function "sort" for extreme perfomance improvement!!!
void get_permutations() {
... and so on. It basically uses comments as a wall to scribble grandiose graffiti about it's valiant conquests in following explicit instruction after fifth repeat and not commiting egregious violence agains common sense.
And since it's vibe coded, no one knows what the opcodes are. LLM won't remember. Human has no comments. Human can't trust post-hoc LLM-generated comments because they're poisonous.
If function of vibecode is not self-evident, dispose of it.
Or, to put it differently, having vibe comment does not free you of responsibility to inspect actual vibe code.
If code contradicts comments, LLM is as likely to go by comments. It is bad enough to have heaps of dead, unused code. Comments make everything much worse.
Even if you try to get them to not, they will still overcomment the code. Or at least overcomment it from the perspective of a human. From the perspective of the LLM, I suspect the comments are necessary for it to be able to get the code output correct.
It's also a discoverability tool. If the code has good docstrings and decent naming for functions/variables it's a lot easier for the LLM to find the correct places to edit.
I tried this a while back using gemini 2.5 pro, round about the time gemini cli was released. I never got the emulator to work in the end, so I dropped the idea.
So this is impressive for me in terms of how fast things have progressed.
Heck, when Satya Nadella wanted to demonstrate Copilot coding, he had it emit an Altair emulator. I guess there's little room for creativity in 8-bit emulator design so LLMs can handle them well. https://thenewstack.io/from-basic-to-vibes-microsofts-50-yea...
This is a good point. I wonder how much NES emulator code is in Claude's training set? Not to knock what the author has done here, but I wonder if this is more of a softball challenge than it looks.
WASM and the performance seems catastrophically bad (45ms to render a frame on an M4 laptop)? It would be much more impressive if Claude could optimize it into something that someone would actually want to play? Compare this to a random hit from Google, https://jsnes.org/ which has sound, much smaller payload, and runs really fast (<1ms to render a frame).
The cost of slop is >40X drop in performance? Pick any metric that you care about for your domain perhaps that's what you're going to lose and is the effort to recover that practical with current vibe-coding strategies?
It demonstrated the capabilities of an AI to a potentially on-the-fence audience while giving the author experience using the new tools/environment. That's solid value. I also just find it really cool to see that an AI did this.
Yeah, it shows the AI is not capable of writing maintainable projects. I'm off the fence. And its cool you find it cool, but reducing the problem space to that of a toy project makes it so much less impressive as to be trivially ignorable.
The new LLM (pattern recognizer/matcher) is not a good tool
