On language models and intuition
For the past week I have been working closely with Claude Code, migrating a crucial part of my service to make it faster and less resource hungry.
It was not "vibe-coding", though. It was a deliberate step-by-step process of carefully going over the existing functionality and porting it to a new platform.
This process allowed me to gain some new insights into how we work with LLMs. Even though it could help immensely and do the things I commanded it to do, it really lacked creativity.
And more crucially — intuition.
All of the ideas about how to do major strategic steps I had to tel jt myself. I had to be creative and give it suggestions to try them out.
It was really like partnership with a robot: it can do many things, some even better than me. But overall it couldn’t completely replace me.
And I think this is the most mysterious part of all of these new technologies which hasn’t been really solved and might not ever get solved.
It is good to be aware of this as to not delude ourselves that these systems are capable of more than what they really are. And at the same time, it is foolish to dismiss them as useless word-churning machines.
The truth is, as usual, somewhere in between.