Contrarily, I think the reality is that most of us couldn't care less about this AI soap opera.
It sounds absurd, but some are watching such a procession take place live as we speak.
Still useless for my day to day coding work.
Most useful for whipping up a quick bash or Python script that does some simple looping and file io.
I work in tech and it’s my hobby, so that’s what a lot of my googling goes towards.
LLMs hallucinate almost every time I ask them anything too specific, which at this point in my career is all I’m really looking for. The time it takes for me to realize an llm is wrong is usually not too bad, but it’s still time I could’ve saved by googling (or whatever trad search) for the docs or manual.
I really wish they were useful, but at least for my tasks they’re just a waste of time.
I really like them for quickly generating descriptions for my dnd settings, but even then they sound samey if I use them too much. Obviously they’d sound samey if I made up 20 at once too, but at that point I’m not really being helped or enhanced by using an LLM, it’s just faster at writing than I am.
The 3rd one still being quite involved, but leaps and bounds easier than 5 years ago.
I spent 2023 developing LLM powered chatbots with people who, purportedly, were very good at prompting, but never saw any better output than what I got for the tasks I’m interested in.
I think the “you need to get good at prompting” idea is very shallow. There’s really not much to learn about prompting. It’s all hacks and anecdotes which could change drastically from model to model.
None of which, from what I’ve seen, makes up for the limitations of LLM no matter how many times I try adding “your job depends on Formatting this correctly “ or reordering my prompt so that more relevant information is later, etc
Prompt engineering has improved RAG pipelines I’ve worked on though, just not anything in the realm of comprehension or planning of any amount of real complexity.
I do claim that I have a tendency to be quite right about the "technological side" of such topics when I'm interested in them. On the other hand, events turn out to be different because of "psychological effects" (let me put it this way: I have a quite different "technology taste" than the market average).
In the concrete case of LLMs: the psychological effect why the market behaved so much differently is that I believed that people wouldn't fall for the marketing and hype of LLMs and would consider the excessive marketing to be simply dupery. The surprise to me was that this wasn't what happened.
Concerning NVidia: I believed that - considering the insane amount of money involved - people/companies would write new languages and compilers to run AI code on GPUs (or other ICs) of various different suppliers (in particular AMD and Intel) because it is a dangerous business practice to make yourself dependent on a single (GPU) supplier. Even serious reverse-engineering endeavours for doing this should have paid off considering the money involved. I was again wrong about this. So here the surprise was that lots of AI companies made themselves so dependent on NVidia.
Seeing lots of "unconventional" things is very helpful for doing math (often the observations that you see are the start of completely new theorems). Being good at stock trading and investing in my opinion on the other hand requires a lot of "street smartness".
Rather: cynicism and a form of intelligence that is better suited to abstract math than investing. :-)
Maybe I'm doing it wrong?
I've been writing code for ~30 years, and I've built up patterns and snippets, etc... that are much faster for me to use than the LLMs.
A while ago, I thought I had a eureka moment with it when I had it generate some nodejs code for streaming a video file - it did all kinds of cool stuff, like implement offset headers and things I didn't know about.
I thought to myself, "self - you gotta check yourself, this thing is really useful".
But then I had to spend hours debugging & fixing the code that was broken in subtle ways. I ended up on google anyway learning all about it and rewrote everything it had generated.
For that case, while I did learn some interesting things from the code it generated, it didn't save me any time - it cost me time. I'd have learned the same things from reading an article or the docs on effective ways to stream video from the server, and I'd have written it more correctly the first go around.