Like it or not, this is stuff that matters.

AI is pretty decent at generating tests for stated conditions. Once you have that, "what is correct" is well-defined.

Look at the existing code. Think about how you want it changed. A few years ago you would spend 2 hours making those changes. Now you spend 5 minutes describing what you want, and verifying you got the expected change. Or maybe you ask it to achieve something and look at what it does instead of spending 2 hours studying APIs. In either case, nothing gets checked in you don't understand.

To be clear, prompts are not a source of truth. They can't be. English isn't a precise enough language. You can pass them around as an example, but you cannot and should not expect identical results and that's okay. The prompts are not the code. The code is what got generated. The code is what will be modified in six months when requirements change. If you can't understand the code you have no business checking it in. Unless your task is specifically to provide useful prompts to others, you should treat prompts as throwaway and delete them the second you are satisfied with the result.

I don't think it's acceptable to not understand the code. The LLM helps you write it faster, but you, the developer, need to have enough of an idea of what you want to verify it did the right thing.

Why does repeatability matter? Once the code is written it's written. You may iterate but there's no reason to go back to the prompt if you got the result you want.

All neural networks are just dot products and activation functions (like cosine). What makes them neural is the connection structure and the back propagation algorithm.

Neural networks are huge graphs of nodes and activation functions. There are some common design patterns that can be used to indicate that certain nodes are actually the same variable. For example, an image recognition network wants to be able to recognize a lion not just in the center of the image, but anywhere in the image. The lion recognition portion of the network will be the same no matter where in the image it's looking. Convolutional neural networks (CNNs) are a good way to represent this.

The problem here is time. How do you represent time in a neural network? CNNs sort of work but they don't understand temporal relationships. Another approach is recurrent networks (RNNs), but they tend to lose focus on their initial conditions too fast. Attention is an alternative network structure. Instead of looking back just a few layers like recurrence, it looks back over everything, with the weight give to each previous layer dynamically learned. It's WAY better than RNNs, obviously.

Attention is one component neural networks can be designed on. The concept may be applicable beyond NNs, but it's weird to say it isn't one.

The thing AI is best at is "here's some text, answer questions about it." Since the text is in the context rather than the model, hallucination is minimized. This doesn't train on the books, so you bypass that ethical question. Is it useful? Dunno. But it's an obvious thing to try.

OpenAI won't keep paying if they don't impose the same deal on the others.

They have to. They can't charge OpenAI and allow the other players to skate by for free.

At the time.....

It can know if it was right or wrong the same way a dev does: by running tests, analyzing the failures, making adjustments, and rerunning until success. Or until it gives up. That won't stop it from going off the rails at times, so you also need a checkpoint system.

Can't argue with that.