LLMs: Harmful to Technical Innovation? (evanhahn.com)

Absolutely. Even with a temperature parameter, at the core they are machines for following existing popular patterns.

That works provided "popular" is also "best", and sometimes it isn't.

The way I look at it is that if I am an expert in a field that the LLM is not likely to know much about, then I have a competitive advantage.

This is mostly because how current LLMs are learning stuff my memorizing stuff. Worst thing of AI through LLM is rote learning where same model is used for reasoning and memorizing. Hopefully in future it will be like small core of language understanding and reasoning which ideally can look up a reference ( through search ) and use RL to get better at the topic through trial and error

compatibility with LLMs has already been influencing my own tech choices! one such example is picking up and using Tailwind for side projects.

I've always evaluated many different component and CSS libraries, usually settling for something a bit outside of the most common solutions. for example from 2020-2023 i mostly used HP's Grommet component library, whereas most developers were using something more popular like Material UI or Bootstrap.

i had to evaluate frontend libraries in 2023 for a work project, and one of the team members suggested Tailwind, so i needed to give it a try. i found that the "ugly" utility classes were mostly the same as the bootstrap ones, which have been around a long time, and Tailwind intellisense makes decoding or remembering them a breeze. what i did not expect was how much easier using LLM's are with Tailwind. I had been using plain CSS and the LLMs do fine with that, but copying and pasting multiple files was much more annoying than having everything contained in the component as with Tailwind.

so that's another vector where LLM tools can shape technology choices, besides the winner-takes-all problem!

Good documentation and examples help both the community and LLMs. My suggestion to any aspiring language or framework developers is to spend time creating both. If you really want to turbocharge adoption you could create a tool that automatically translates code examples from popular languages into your own language (and tests that they work) and then publish them online with some helpful comments.

It sounds bad, but in reality we already choose technology largely based on how easy it is to google for.

React, angular, svelte, etc have great SEO.

Common Lisp, reason, and clojurescript not so much.

Python is popular in ML because it was already popular in the sciences. Most of the technically advanced stuff is happening in C anyway.

More and more I’m viewing LLMs as the next step in search more than anything else.

Isn't this just an extension of the culture of asking for evidence to make future decisions?

If I go to the bank and ask for a loan they will ask me for evidence of product market fit and other signals that suggest to them that I will be able to pay them back. All of this evidence is ultimately based on what has been successful and dominant and quantifiable in the past.

IMO LLMs are just the next disturbing step of optimising for the status quo, limiting risk taking, and stifling innovation.

I've been arguing for not being too trusting in the code LLMs generate. I've seen deepseek-r1 and multiple chatgpt models try to solve a genuinely trivial issue in code by deleting the unit test. And that was a well established language. Sure, it did compile. But it didn't solve the runtime bug. You cannot replace logic and experience with the statistics of just things that already exist, especially when creating something new and innovative.

Blindly generating code is deserving of the chaos that comes from it.

Learning to code, and then going a little faster on the repetitive pieces where oversight can be maintained is key.

More and more, tools like Aider seem to differentiate themselves from having taken such a position from the beginning.

This resonates with me. I used to love tinkering with "underdog" tech. I did frontends in svelte before it was cool. But for new projects, I now always default to react+tailwind, because that's what Claude knows best. And for the backend, I'm working on a SQL-only framework, because LLMs are now at almost 100% accuracy on text-to-sql.

I think it just reinforces what has always been true. If you work on a cool new framework, programming language, database, etc. your job doesn't end with coming up with a nice API, a whizzy type system or fancy benchmark numbers. You need to also write tooling like nice CLIs, IDE integration, reference docs, tutorials, high level blog posts and ideally you have a compelling use case you can dog food. With the latest models and some agentic search integration, vastly diminishes the advantage that the models get from being actually trained on specific content. It's still a little more difficult if your tech has particularly unique features, but overall it works quite well.