AI coding diffuses the knowledge of common 80% of programming. So it should make programming more accessible actually. As it can remix well it basically can extract for you any implementation it has seen on github which I believe is great for humanity progress.
Examples of bit harder things I can do thx to AI:
* write code in less known languages
* do research about things to implement i had no clue about
* vibe code games - this is good example of what I say below. At the moment you need to iterate: write feature, optimize code, write feature, optimize code - because it takes crappy implementation that does job but then is totally capable of fixing it
If someone would ask me for example if for python I prefer UV or AI I would tell UV for now but python had ultra bad tooling.
I believe that LLM have more potential than they currently show. As someone explained they don’t use knowledge links properly due to prediction objective. Eg they don’t attack problems from “tangents” just go in straight line, RL somewhat fixes that but it’s not there yet - it still just trick with perhaps, let me consider etc. This might get improved upon IMO and some new aproaches add tree search on top.
To add on top of it. As anyone in ML know the hardest part is preparing data. If the AI hype will lead to good permissive datasets, we will get a great progress once someone fixes algorithms.
Considering the need for LLM boilerplate coding as a language and tooling failure is an interesting observation. It begs the question why there is still such a thing as boilerplate. I mean I understand how programming languages work and why boilerplate is essential, but at the same time LLMs demonstrate that brevity is possible without loss of expressivity or legibility, albeit in a very roundabout and costly way. Perhaps there is a programming paradigm that strikes a balance between these opposites, especially if you're willing to spend more computational resources on it than traditional compilation (but with a complexity ceiling orders of magnitude lower than LLMs).
This just reflects the type of coding that George is doing. But the VAST majority of code written in the world is CRUD, Forms, Scripts, etc that AI aka "English" is a perfectly reasonable fit for. I mean, I use AI to write code hours a day and I don't think I'd let it drive my car for me.
I think George is right about vibe coding. It's great for rapid prototyping, but in my daily programming on a production codebase, PRs must be reviewable, so I have to understand what I'm changing even more so.
AI coding diffuses the knowledge of common 80% of programming. So it should make programming more accessible actually. As it can remix well it basically can extract for you any implementation it has seen on github which I believe is great for humanity progress.
Examples of bit harder things I can do thx to AI: * write code in less known languages * do research about things to implement i had no clue about * vibe code games - this is good example of what I say below. At the moment you need to iterate: write feature, optimize code, write feature, optimize code - because it takes crappy implementation that does job but then is totally capable of fixing it
If someone would ask me for example if for python I prefer UV or AI I would tell UV for now but python had ultra bad tooling.
I believe that LLM have more potential than they currently show. As someone explained they don’t use knowledge links properly due to prediction objective. Eg they don’t attack problems from “tangents” just go in straight line, RL somewhat fixes that but it’s not there yet - it still just trick with perhaps, let me consider etc. This might get improved upon IMO and some new aproaches add tree search on top.
To add on top of it. As anyone in ML know the hardest part is preparing data. If the AI hype will lead to good permissive datasets, we will get a great progress once someone fixes algorithms.
Considering the need for LLM boilerplate coding as a language and tooling failure is an interesting observation. It begs the question why there is still such a thing as boilerplate. I mean I understand how programming languages work and why boilerplate is essential, but at the same time LLMs demonstrate that brevity is possible without loss of expressivity or legibility, albeit in a very roundabout and costly way. Perhaps there is a programming paradigm that strikes a balance between these opposites, especially if you're willing to spend more computational resources on it than traditional compilation (but with a complexity ceiling orders of magnitude lower than LLMs).
This just reflects the type of coding that George is doing. But the VAST majority of code written in the world is CRUD, Forms, Scripts, etc that AI aka "English" is a perfectly reasonable fit for. I mean, I use AI to write code hours a day and I don't think I'd let it drive my car for me.
I think George is right about vibe coding. It's great for rapid prototyping, but in my daily programming on a production codebase, PRs must be reviewable, so I have to understand what I'm changing even more so.
maybe a new language is due - like a revamp of applescript