One major point of disagreement: LLMs are not necessarily better at writing React because they have seen a lot of React. In fact, some benchmarks which ask the question: “which languages are LLMs better at” often show that LLMs excel at certain rare esoteric languages like Elixir, and can often solve more benchmarks problems when they write Elixir compared to more popular languages like Python. Even though the LLM has been trained on way more Python. So what makes LLMs better at a certain programming language isn’t how many examples it has seen. Frontier models can write very good code in a new programming language it has seen for the very first time, just from a good spec.
The real question here is: what features of a programming language make it a good language for Large Language Models? And how would you design a framework that’s not meant to be written or read by humans, only AI? That is a very interesting question to me. I theorize that LLMs are good at writing Elixir because it doesn’t have OOP concepts that require abstract models “in your head” (the LLM doesn’t have a head) and physical world metaphors that only humans have experience with. Instead Elixir’s primary tool is “pattern matching”, which works on literal text patterns instead of abstractions, and that’s what makes LLMs good at it. The question “what features of a language make it good for AI” is very interesting to me and I foresee a lot more research in this area.
Fair point, but benchmarks only give you a partial view of what producing reliable, prod-grade code is about. I tend to believe that, in the real world, the ecosystem seriously influences agent performance, like library versions, error messages, tooling, the accumulated debugging knowledge around a framework. On that front, React seems unbeatable.
It’s about the quality over quantity. I’d bet there’s significantly more bad React published online than Elixir. LLMs don’t know what’s bad and what’s not, only what’s more likely. How much bad Elixir is published online?
I’d also bet the nature of the language itself would only affect how few samples it can train on before appearing competent
Stop trying to recreate the desktop experience in the browser. If you want a desktop experience in the browser then simply deliver a desktop application through webassembly. Problem solved.
End users really don't care. Things are the way they are because designers are just sniffing their own farts.
One major point of disagreement: LLMs are not necessarily better at writing React because they have seen a lot of React. In fact, some benchmarks which ask the question: “which languages are LLMs better at” often show that LLMs excel at certain rare esoteric languages like Elixir, and can often solve more benchmarks problems when they write Elixir compared to more popular languages like Python. Even though the LLM has been trained on way more Python. So what makes LLMs better at a certain programming language isn’t how many examples it has seen. Frontier models can write very good code in a new programming language it has seen for the very first time, just from a good spec.
The real question here is: what features of a programming language make it a good language for Large Language Models? And how would you design a framework that’s not meant to be written or read by humans, only AI? That is a very interesting question to me. I theorize that LLMs are good at writing Elixir because it doesn’t have OOP concepts that require abstract models “in your head” (the LLM doesn’t have a head) and physical world metaphors that only humans have experience with. Instead Elixir’s primary tool is “pattern matching”, which works on literal text patterns instead of abstractions, and that’s what makes LLMs good at it. The question “what features of a language make it good for AI” is very interesting to me and I foresee a lot more research in this area.
Fair point, but benchmarks only give you a partial view of what producing reliable, prod-grade code is about. I tend to believe that, in the real world, the ecosystem seriously influences agent performance, like library versions, error messages, tooling, the accumulated debugging knowledge around a framework. On that front, React seems unbeatable.
It’s about the quality over quantity. I’d bet there’s significantly more bad React published online than Elixir. LLMs don’t know what’s bad and what’s not, only what’s more likely. How much bad Elixir is published online?
I’d also bet the nature of the language itself would only affect how few samples it can train on before appearing competent
98% of current front end web development deserves to be hit by an asteroid, not sure that LLMs can deliver though.
Stop trying to recreate the desktop experience in the browser. If you want a desktop experience in the browser then simply deliver a desktop application through webassembly. Problem solved.
End users really don't care. Things are the way they are because designers are just sniffing their own farts.
> The asteroid currently hitting front end web development
...is the same asteroid currently hitting all forms of knowledge work and even more theoretical pursuits (math research, etc).
Different sets of people are just noticing/reacting to the impact at different speeds, but we are all standing in it.
LLMs are a disease. Why would I ever open information to the public now -- short of a poison well?