"The model refuses to follow my specific word detail prompts" and "The model refuses to perform hacking attempts" are on the same side of the model refusing to do something baked into it though.
They can, but what would make them think their initial misunderstanding is even wrong? While some might think "surely the system isn't like that..." many would default to "of course they want to steal my money" etc.
Not to mention the amount of misinformation that can get passed around
I think because the gradient is simply too confusing for laypeople to understand.
Even for a simple system like US social security that has a gradient. For every $2 you make over the limit, you lose $1 in benefits. I've heard countless times misconceptions of people thinking they'd be losing money (as in literally having less money net) by working.
I think the methodology is interesting, but you can probably design a similar, and probably better loop for "diving" into a topic and going through examples than is presented.
I have to imagine they just asked Claude to dive into a topic and generate example problems along the way.
As for the content... a lot feels like knowledge overload and concepts are introduced without explanation or "why". It basically says "here is a training loop" and never answers my immediate question of "...for what?" It also introduces random concepts like setting the seed that don't look like they're even in the final training loop it provides.
As a comparison, Pytorch itself has docs that go over the entire training loop as well. And it explains why you're loading the data its loading. With examples that run instead of being scattered on.
That license IS any copy left license, but nobody will trust a random fork is my understanding.
I can agree though that the network effect of "everyone else uses X" is pretty unfortunate and has led to a lot of other issues too, so yes very disappointing
Representative democracy was supposed to be the solution for this. i.e. "I dont know what the best way to do age verification should be, but it's probably important"... and then voting for the candidate that has shown ability, intelligence, and a good nature to look at the issue in your best interest and handle all the edge cases without screwing you over.
What we've lost is the integrity of this system entirely. No candidate can truly be trusted to implement an age verification system without first getting a check from big tech
You are correct, but especially in current day that analogy is quite bad.
I expect Median LoC might be very high with the average developer using AI these days... but the dev who is making atomic changes that are fixing the AI output is probably tiny LoC but way more important
That does not sound at all reassuring, that the only safeguard is the system blocking access and that the API has no safeguard.
Its also easily possible to have sensitive files misplaced, especially for a general non-technical user that would be the one falling for a browser hijacking attack
I would also imagine the majority of people used to RLHF on AI could have English not as a first language. I have no expertise in this area, but I feel like the tuning at that phase would also affect the "voice", "tone", and "style" that gets put out.
Because it's not blatantly obvious what is a B versus a C until after the fact.
So you
1. Reduce morale of Bs with potential. Layoffs or just reducing opportunity affects the whole team
2. Enforce stereotype threat on Bs you've miscategorized, reducing their performance and turning them into Cs.
It probably affects the As too, but I know I'm not there yet, so I can't speak on it.
This just sounds like your Linkedin went stale for some reason or other. Do you have the Open to Work status? Or even just be active on there, apply to jobs, like posts or stuff like that?
I get plenty of messages from "recruiters" through LinkedIn. They're not particularly great, but I also only have a few years of experience
That's like those stories of LLMs saying "I fixed the vulnerability in your app" by deleting the project entirely