glhaynes

I feel like maybe the answer is browser-integrated LLM plugins tasked with "I don't ever want to see anything having to do with this topic of argument ever again."

Do we have reason to think that they scrape them more often than they scrape other stuff? (Honest question, I have no idea.)

Doesn't Stackoverflow exist because people can in fact not read the documentation or source code, at least not to the point where it helps them understand their problem.

Yes; not sure how this is relevant to LLMs. One of the most impressive things I've seen change in the last few months is that they don't think twice about cloning a repo of linked-to-my-app code to analyze it to determine a solution. Or even disassemble closed-source object code to find an answer!

With agentic speed/capability, the calculus of "eh let's experiment and explore from the outside a little first" vs. "let's just read the dependencies (whose codebases I'm not familiar with), potentially including the binary itself, and figure out exactly what's going wrong" changes. So I think what's valuable in terms of ancillary info for technical topics is changing.

Anyway, I'm learning so much more so much better than I ever have before. Turns out the ultimate slop tools are also the ultimate learning tools if you use 'em right.

Weak models tasked with review can catch a decent amount of the mistakes that weak models make and help them be much better, especially if you have them verify against authoritative sources. Strong models make far fewer mistakes to begin with. And, for the mistakes they do make, a swarm of reviewers (same model or somewhat weaker, reviewed by the stronger model) can really help reduce the error rate further.

Could gone to some great parties.

Did that instead.

— Elon Musk

You're using a rhetorical device to imply that they're "falling for" a narrative. "Falling for" only follows if that's why they are making the prediction; this construction doesn't allow that they may have come to this belief on their own.

But, yes, it's all a bunch of vector draw/fill commands. The first few AGI games showed that process happening; and, as a little kid having one of their first encounters with a computer, I found it fascinating to watch. Later games drew it in an offscreen buffer and showed it all at once, which looked more "professional" but took away some magic.

Here's a video showing the drawing process: https://www.youtube.com/watch?v=v5DSFvWrNWA. Magic!

EDIT: I had an agent make a guess and it came up with reasonable looking numbers that said an RLE bitmap would be a small handful of times larger. But it also interestingly pointed out that the vector commands encoded priority (i.e., what you can walk behind/through) whereas that'd have to be a second data set for a bitmap approach.

Of course that just leads to: what’s the best way to achieve that goal? Through elegant code or adding lots of tests? Which is a debate from long before LLMs existed.

That said, I have to recognize that this may be partially because of my personality. I don't "do great" at mixers like this. I'd rather go home and be with my family than drink beer—regardless of label—in a corporate setting. People describe me as charismatic and engaging one-on-one, but I'm awkward and unhappy at a big crowd event.

But there are other people whom I think get a lot of value and connections out of them! So it's kinda hard for me to say.

Downgrades in quality, though, stick out like a sore thumb. "I didn't really like going to these things before, but at least they had good beer." It can also be a real "it's the thought that counts" sort of thing. When you show me that you're willing to spend less on me, it sends a signal, sometimes stronger than if you'd never spent anything on me in the first place.

It's just much more visually interesting than a page full of perfect burgers. Each one looks like a unique thing from the real world; they don't "look AI", as the kids say these days.