flying_sheep
Joined 9 years ago
Joined 9 years ago
I'm not arguing that all software should be vibe-coded.
My point is that software that was economically infeasible to develop before LLMs is now becoming feasible.
Such software typically has limited economic value. People may find it useful, but they weren't willing to pay enough to justify its development cost.
The software you're referring to is generally high-value and safety-critical. People are willing to spend enormous amounts of money to ensure its correctness and reliability.
These are different markets, with different economics.
There are tons of projects that were economically infeasible to write, and now become feasible. Something that needed lots of human time but almost no one wanted to sacrifice anything, including money, to get them. But now we have LLM subscription in which per token cost is almost free.
They are fun to create and use.
What I mean is that we can bring some web tech to terminal natively. We don't even need a separated shell. Security and bi-directional communication is built by default because of UNIX socket. But we still need to think how to handle stuff like cookie, local storage, external CSS / JS, ...
This situation may be naturally evolved, or may be deliberately shaped by people or institutions. In any way, negative-sum can drive people to do irrational decision because the game is not beneficial in all cases.
So we end up with this pattern: we capture the useful order for ourselves, then push the mess, waste, heat, and instability onto the environment.
This is the fundamental nature of human. We can't change it without higher-order regulation. Even in AI training, we assume there is one true distribution and optimizes for it.
We created these systems. And unfortunately, most of us are the unwanted in these systems at the same time.
This is true until certain point, unless the requirement / contract itself has loophole which the attacker can exploit it without limit. But I don't think this is the case.
Let's say, if someone found an loophole in sort() which can cause denial-of-service. The cause would be the implementation itself, not the contract of sorting. People + AI will figure it out and fix it eventually.
I still don't know how to create a native app so inefficient that, it needs to take more than 500 milliseconds to open a directory
And the AI response in the Google Search still has factual errors occasionally...
On the contrary, the ChatGPT has been doing great. The accuracy, recall, precision, or whatever metrics of ChatGPT exceeds Google Search by a large margin. Or close to or exceed what a human expert can do