1. AI writing directly to assembly is useless, compilers exist for a reason. No idea who thought this is a good idea

2. productivity in LoC is a meme and can only be used by someone who don't know the history of CompSci and how "paying engineers for lines of code" story ended up. from my current experience main bottleneck isn't writing a code, it is communication, time to build containers, etc. iteration of development is reduced, but not as drastic as AI bros want you to think

3. i don't think it is useless, it increases the speed of development, but it is definitely reduces code quality without additional guardrails. i don't think it is better, it is just good for different set of tasks and feels like completely different approach to development not starting from logic and creating solution from first principles, but more like making rules by which city of people will evolve.

4. research suggests that the rate of hallucinations is decreasing rapidly over last 5 years, so let's see what we will get

5. due to jagged frontier phenomenon we can't just assume llms have the same skills as humans. they are good at some tasks, that would be a miracle 5 years ago, but are absolutely idiotic at other. currently best place for LLM agents is to make demos/MVP for presale or for throwaway code that is too simple to fail to produce

6. nah, only LLMs without RLHF are "only predicting next tokens". LLMs with RLHF are more morphed to be task solvers (this is why we see AI proving new previously unproved theorems, you can't argue that it was in the training data). this is an important technical distinction that is often overlooked by AI critics. And if we want to unravel "AI cannot create" argument - I can say that only God can create and humans stole most of the ideas for technologies. Imagine the economy in which nature patented most of it's creations and humans would need to pay for using nature's inventions

mrrr meowww