Chamath Palihapitiya claims that the AI code quality is superior to the error ridden and sloppy code that humans write. That’s going to trigger a great many people. There is a lot to unpack here.
Let’s grant the premise that today’s strongest frontier models write much better code than past models. AI-written code is still subject to much human error, because it is driven by intent; and human expressions of intent tend to be error-prone and incomplete. When AI fills in the gap by inference, this guessing is hit-or-miss. Folks may call this hallucination, but I say it’s your own damned fault for not being precise and complete enough in your own instructions.
With correct intent, does the AI do a better job than humans? Certainly better than the average human. Maybe not better than the best humans within the human’s area of expertise.
Importantly, each human is highly specialized. Usually expertise is deep and narrow. Where our expertise is broad, it is usually shallower. That’s the difference with AI. AI expertise is broad and deep throughout. With that advantage, we can direct an AI coding agent at solving diverse problems or problems that span a scope larger than any human can deal with.
That’s usually why we need teams of humans to collaborate, but human communications scale at O(n²) and very slowly. Human super-intelligence relies on a great deal of patience and discipline for the diffusion of understanding. AI can scale much more quickly, if human understanding is decoupled from AI being allowed to work.
There is much disagreement over whether code quality is better or worse with AI. A lot of the complaining now, I believe, is because human explainability and understanding of the machine output is poor due to sheer volume and complexity. Humans cannot keep up, and if the AI gets stuck, humans are near-helpless to intervene, because our understanding was left behind.