I confirm the same experience. I tried to port the Dynamic Range Optimization code in OpenCamera from Java to mathematical formulas (i.e., the spec), with the intention to translate that into Python with NumPy, so that I could run it on my own images not coming from my phone camera sensor. Tool used: just a chat with ChatGPT with the relevant files uploaded, research activated, and questions asked where I did not understand or doubted something in the response.

Result: ChatGPT faithfully and correctly reverse-engineered the initial highlight pre-compression step and then said that the rest (the real thing!) is too complex and not important anyway. I did not pursue it further.

Then, when it’s time for the monthly payment, let’s write “Paid” in the chat and request the token.

AI tools are great, but they require a lot of oversight. Developing a product is a process. People are generating so much code that they can’t keep up with code reviews. What they really need to do is spread the work out over several days. If your project is solid, you’ll get positive results anyway. (I’m not talking about the MVP stage here.)

There's multiple ways to skin a cat, human.

Not had deleted tests but had a lot of "this isn't working I'll insert a skip".

I gues a fix for that is a hard check for skips

Did the author say which model and harness handled the first attempts? Codex at the end ok but what did he try the rest with?

8 billion tokens? What did that cost?

Very human

believing that oneshot prompt is enough specification is pretty delusional.

I keep seeing people talk about the power of these SOTA models, yet keep reading the types of prompts that make no sense to anyone who understands the ludicrous number of decisions that would need to be made.

Tends to be a problem. I've tried to mitigate these problems by using either external harnesses (aka GitHub actions that are "fixed" based on known-good) or by using n-number of witness agents (e.g. Kimi/Qwen/whatever <=> Claude/OpenAI/Google). Generally sucks more time and energy (and now token/$).

that being said, I still have a "fix the code, not the test" line somewhere in here...