Aside from the rest of the (interesting!) nuanced discussion going on the comments here -- I really like his idea towards the bottom of combining the colors for numbers and strings into one.
And, maybe I'm missing something, but to me it seems obvious that flat top part of the S curve is going to be somewhere below human ability... because, as you say, of the training data. How on earth could we train an LLM to be smarter than us, when 100% of the material we use to teach it how to think, is human-style thinking?
Maybe if we do a good job, only a little bit below human ability -- and what an accomplishment that would still be!
But still -- that's a far cry from the ideas espoused in articles like this, where AI is just one or two years away from overtaking us.
The standard way to do this is Reinforcement Learning: we do not teach the model how to do the task, we let it discover the _how_ for itself and only grade it based on how well it did, then reinforce the attempts where it did well. This way the model can learn wildly superhuman performance, e.g. it's what we used to train AlphaGo and AlphaZero.
As you suggest, I've had a moderately successful time trying to get AI to write its own Sublime Text plugins so our favorite editor doesn't get left behind, so might be cool to try with this too?
If you haven't seen it, this "Linux Touchpad like Macbook" project is related, the last/best effort I've seen in this direction. Here's a random update from a few years ago:
I know there are a lot of conversations going on about the dangerous elements of LLMs, but it’s nice to read a story like this alongside them — it’s a reminder of the remarkable potential of this new technology.
Could you have found out somehow else about the severity of the problem? Maybe, sure! But the fact that you had this LLM to ask, and it so easily understood whatever info you gave it, asked clarifying follow-up questions, and gave you info/directions in the way you needed to hear them — it’s such a new way of interacting with computers, almost like an API but made for humans to use, and cases like this is where that value shines most clearly IMO.
Thank you for sharing this feedback! Yes, the big unlock for me was using the LLM to explain things to me on my own terms and that's what ultimately gave it more persuasive power over even a simple (non-personalized) google search
I'd like to read this, but it seems to be behind a paywall?
(There is a reasonable amount of the article before the paywall, and since the other comments here don't mention it, I am wondering whether they just thought the article was very short and then ended abruptly, lol)
They're both constants, so yeah, they should be!
reply