AI Coding(geohot.github.io)
geohot.github.io
AI Coding
https://geohot.github.io//blog/jekyll/update/2025/09/12/ai-coding.html
8 comments
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.
I agree pretty wholeheartedly but as someone who was, "trained as an artist first and a computer guy second" the ability for the LLM to provide functional boilerplate and, "hunt down" those weird problems that can only usually be resolved by reading sometimes hundreds of pages of documentation. I love the comparison-- "just as compilers replace programmers and spreadsheets replaced accountants..."
The one use case where I'm really excited for though is when we (perhaps we have) to where my nieces and nephews can tell the machine-- "I have an idea for a video game could you help me make it?" That to me is the natural extension of the movement to things like Roblox and Minecraft. Ultimately though the, "goal" is to get them to, "real understanding."
The one use case where I'm really excited for though is when we (perhaps we have) to where my nieces and nephews can tell the machine-- "I have an idea for a video game could you help me make it?" That to me is the natural extension of the movement to things like Roblox and Minecraft. Ultimately though the, "goal" is to get them to, "real understanding."
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.
people want easier and if you hold a shiny object in front, most will lunge for it. Developers have been benefiting greatly for some time now. High paying jobs, no ceiling in business, praise from colleagues, respect from family, etc. Non-devs close to it now have this opportunity.
No one knows what the impact of AI code gen will be on developers. But if AI code gen under delivers and the demand on truly talented devs rise, then the ones 'doing the hard work' now and maintaining their craft by human coding will win big.
No one knows what the impact of AI code gen will be on developers. But if AI code gen under delivers and the demand on truly talented devs rise, then the ones 'doing the hard work' now and maintaining their craft by human coding will win big.
maybe a new language is due - like a revamp of applescript
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.
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.