Industrial invention used to be judged by actual usefulness. If we take space exploration as an example, at every stage, we have actual usefulness for what the technology so far was able to accomplish. Same with computing. But now we're two years after the first release to the general public of LLMs and no one can provide an actual, reliable use case for it. Instead, it's mostly about destroying values, hindering workflows, and crippling the internet.
A few years will start to show if what AI is producing is a lability, or if it is just fine, methinks. If you haven't felt any pains from its legacy cruft by that point, it isn't apt to magically appear a long way down the line.
> If we take space exploration as an example, at every stage, we have actual usefulness for what the technology so far was able to accomplish.
When Archytas propelled a wooden bird with steam, the first recognized step towards building the rockets that got us to space, what usefulness was found in that?
It offered a basic foundation on which we were able to consider other solutions and the concept in general. But couldn't you say that is, at very least, where we are at with LLMs right now too?
Let's ignore the hype and marketing and truly consider LLM as the first stage of "something". The truth is that I've heard a lot of speculation, but no inching towards any of the speculated objective on a practical scale.
For anything else, any such first stage was usually constrained on a very small scale while the glaring issues were ironed out. Now a lot of us are being forced to tolerate what is essentially a prototype or toy in tools and systems that were fine before.
Perhaps LLMs are in the Model T stage? The experience is terrible. Its slow. The roads are muddy and full of ruts. You are bound to break your arm when you spin the crank. But, still, it offers something ever so slightly more than the horse that you put up with it anyway.
More like a Model T that goes fast in the wrong direction, don't respond to the steering wheel, decide to reverse on random occasions,... The road may be bad and the horse is not that great, but no one would hop on that particular model unless for entertainment purposes.
There is nothing fast about LLMs. That is their biggest pain point of all when used in an engineering context. Even if they offered perfect production, you can still do it faster yourself by a significant margin.
> but no one would hop on that particular model unless for entertainment purposes.
Sounds about right. It might be the "in" thing in tech circles, but the people I encounter on an everyday normal basis have never used LLMs. Many of them don't even know what an LLM is. And that's with LLMs being freely available! Imagine if you had to pay to use them. It would be a handful of people at best.
> Natural language is a terrible interface to interact with a computer
And period. It is terrible when interacting with people to. Its only saving grace is that it is widely supported.
> You fundamentally cannot get it significantly better.
It will when we flip the script. The trouble right now is that we're trying to do it backwards. Eventually we'll settle on a programming language as the de facto language of LLMs, that can be translated to natural language for interfacing with the pesky humans still running in legacy mode. That is when transformation will occur.
Industrial invention used to be judged by actual usefulness. If we take space exploration as an example, at every stage, we have actual usefulness for what the technology so far was able to accomplish. Same with computing. But now we're two years after the first release to the general public of LLMs and no one can provide an actual, reliable use case for it. Instead, it's mostly about destroying values, hindering workflows, and crippling the internet.