Finished by AI
It's not exactly analogous to a hammer, even though I agree that it is a tool. If a carpenter got the framing all wrong, the hammer wouldn't fix it. If a prompter got the instructions all wrong, or didn't supervise enough, or didn't verify enough, the AI still makes something that appears to pass initial muster.
(Technically not "makes", but I always feel like it's better for both sides to understand each other more completely in order to have a more fruitful debate)
Maybe this is one of the better times to encourage "literate programming" all over again?
If the majority of the source is detailed prose comments instead of "code", you would be able to tell. I think you're correct that this is an information problem (like the law of small numbers).
I strongly believe that if a new project is interesting enough for someone to be reading the source of a literate codebase, the prose that goes with it will be distinct enough that the first publisher would necessarily have to be human.
The project is quite transparent about it, saying it was "proudly vibecoded". The entire "not sure what was hallucinated" narrative seems stuck in 2025.
Some friends of mine play old games that are on community life support with official servers having been shutdown years ago. I sometimes support them when they have some tech issues and whenever I stumble upon documentation of these projects being spread over forums and vintage 2010 era gaming portal websites I'm glad that we have AI now.
Oh, I didn't spot that in the submission, is it behind links? I read a few paragraphs of the description and left because it didn't seem organic. My bad.
> whenever I stumble upon documentation of these projects being spread over forums and vintage 2010 era gaming portal websites I'm glad that we have AI now.
I love AI! The submission title just gave me hopes and the non-organic feel was a bit of a bummer. But nothing against the project itself.
Nothing has changed in that respect since 2025, despite the many claims to the contrary. LLMs still hallucinate quite regularly.
The issue with using AI as a tool to build software is that the inspection step is not required. I'm not sure an AI-generated description is a sign of bad inspection processes. I think maybe in the case of a game mod, the description is not as important as it would be for some software that someone is selling.
I guess my point is that quality can't be assessed merely by whether something was built with AI. A trustworthy team will have carried out inspections along the way. The question is how can we know that?
Wrong. Lack of inspection is a process issue symptom, not an AI issue symptom.
Never has been, except in rare instances.
There is no agency in the model.
The humans committing the code are still responsible - through and through.
Use the saw correctly.
To create something and to be responsible for it are two correlated notions, but they are not one and the same.
> There is no agency in the model.
Of course there is. Agency is what allows it to be useful. If it had no agency, you would have to ask it to do every step of what it does. Why does asking Claude to "make me a website" work? Because not only does it spit out the HTML, CSS and JS, it also starts trying to test it by starting a local server and so on. Agency is what has made LLMs so useful - the ability to actually execute the bash commands they spit out and react to them. Without tool-use and chain-of-thinking - effectively the two features which allow them to express agency through their reactions - they are truly little more than fancy autocomplete.
> These models CAN direct themselves - in the same way a circular saw can.
This is a hopelessly shallow comparison. You're really comparing an entity that reacts to its own output with what is effectively a blade stuck onto a motor. A circular saw has exactly one job and that is to cut whatever comes in reach of its blade. An LLM given the same inputs and access as a human developer can do their entire job end-to-end. Perhaps the long-term quality won't be as good - but the LLM can definitely do it.
I said commit.
An LLM has exactly one job: autocomplete with the text that it thinks comes next, based on what was input.
Anthropomorphize it all you want - that’s what the models do.
The rest is humans making offerings, humans taking offerings.
Agency is what has made AI service providers useful.
Just throwing a single model at a task with its own autonomy is exactly what’s been repeatedly highlighted as the largest weakness in LLM output and security.
…because, as a technology, it’s a probabilistic parrot with no agency. Just some artificial agency strapped on - sometimes.
Said another, perhaps simpler way: if you instruct a chatbot “make me a marketing website for a store” (and it’s not great), you’re an asshole. Not the chatbot. Not the industry or FOSS community providing them.
Sich doch einmal wegbegeben!
Und nun sollen seine Geister
Auch nach meinem Willen leben.
Seine Wort’ und Werke
Merkt’ ich und den Brauch,
Und mit Geistesstärke
Tu’ ich Wunder auch.
------------------
I like Goethe's German. In general, German does not feel very song-songy to me, but Goethe wrote some really good and catchy poetry.