Did pewdiepie not write a voice to text for his LLM setup?
Thing is, we can talk faster then most of us can type.
Voice + Programming is slow because of all the special symbols. But voice + vibe coding? The ability to tell your LLM to do tasks, while you focus on other parts of the code, without the need to switch tabs/windows.
What about "change the color green on this element (html page), where my mouse is pointing"... Annoying with keyboard if you need to switch windows, very possible with voice.
And LLMs are very forgiving for mistakes, unlike if you want to voice program where every symbol needs to be accurate.
People do not realize that programming as we know it, is going to change.
>People do not realize that programming as we know it, is going to change.
I saw yesterday that I had been approaching software incorrectly. It feels futuristic because it's so fast, but it's still linear. One guy making one thing at a time (with some help from the computer).
But software can now be made so rapidly, that the bottleneck is actually curation. You can now generate a hundred ideas for software and a prototype for each one in the time it takes to make some coffee.
Going through all of it is the part that doesn't scale, it's bottlenecked by the individual. That's the reward function, right? Taste, discernment.
At this point software can grow itself, it can mutate, and it can combine with other software. I think building is entirely the wrong metaphor now.
I think a better metaphor would be a genetic algorithm. You try a bunch of stuff and see what works. Then you combine the best parts.
> You can now generate a hundred ideas for software and a prototype for each one in the time it takes to make some coffee.
Yep ... thinking the old way to make software is going to end.
In the past, we made a framework, then sold that framework for clients. But we always had the issue where client X wanted Y features, client Z want X features ... And over time the framework bloats, you get issues with features that may conflict between clients. Then you start to split the framework maybe for client Z because its too much different. Now your have issues when features or bug need to be fixed...
In todays LLM world, i see it more like custom software per client, with "instruction files"...
You make a custom framework for a client, with the AI writing it based upon a instruction file, that is supplemented by custom requirements for the client. Its written for that client and only that client.
The next client, same ... the next same. If a client sees a feature that they want, you instruct the LLM to update the framework for that client using again, a addendum instruction file.
If you instruction file was written correctly, bugs are going to be on the low end, and most clients do not need constant updates to their software.
A client wants to go to a different company and can not get the source files? That company needs the database files + the original design / analysts and the new company rebuild it again into a new version.
So ironically, we are going to, to a world where custom software is very normal, and cheap.
> I think a better metaphor would be a genetic algorithm. You try a bunch of stuff and see what works. Then you combine the best parts.
Yep, put that in practice last week.
I wrote a database in barely a week and half time, and was "slow" because i made like 5 different versions playing around with clustering, different parsers, more advanced each time (regex, token, lexer-ast) and tons of other features.
When i did not like a version, o LLM, rebuild it using my new updated instructions. O, i do not like the parser as it had issues, lets make a more advance one.
We are not talking toy DB ... full insert/update/delete, joins, CTE, Window function, SubQueries, Index's, alias, ... you name it, all working correctly. If i used my instruction file today, i can make you a custom DB in a day. Two if you need something custom. If somebody told me this 6 months ago, i call you crazy lol
Normally, when you build something, you spend days, weeks into it, especially if its advanced. Your reluctant to just tear it down and restart from zero. Or pull a important component out to rebuild from zero. Because sunk cost ... Now its just a half a day work, a day at worst, and you redid what will have taken you weeks or months. It really allows for a lot more experimenting, finding what fits better as time becomes different vs you on your little keyboard typing for ages, rebuilding, making tests, again and again. When a LLM does it 10, to 100 times faster.
For somebody who is a senior programmers, your actually the most easy to adapt and get the most out of LLMs (and ironically often the most resistant to change to using LLMs). Programmers that do not adapt to the new, are going to be left behind.
Please, really, I am sure we all get it. Who is even the audience for this kind of comment at this point? Can't we have one comment section that's about how Linux is cool and good and Windows sucks? Like when we were all still real nerds instead of product hypers?
The point of my comment is that if you use AI in the CLI it can be very helpful, because they're really good with text and pretty bad with everything else.
The general rule here is that you use it for what it's good for it's actually really good.
The "typing into my terminal" is mostly for interacting with Claude Code. I wish that part worked on my phone.
Although I do use the voice typing tons for text chat, ironically.
I recently vibe coded a voice typing software (using Parakeet — your best bet is probably Handy though).
It works in my terminal. (I just changed my paste shortcut to Ctrl+V
I can now literally speak software into existence!
I made a thin wrapper around my llm() function I can pipe text into from Bash.
This allows me to make many other thin LLM wrappers, such as one that summarizes then contents of entire directories.
I have a thing called Jarvis inspired by a Twitter post, where I ask it to do anything in bash, and it just does that.
I wouldn't exactly say it's useful (I am unemployed) but I am kind of having my mind blown a little bit.
The future is already here, it's just not evenly distributed yet.