gwd

27 ms·

331.
by gwd 1y ago
Not sure how we got from Scott A being a rationalist to trains, but since we're here, I want to say: I've taken a Chinese train from Zhengzhou, in central China, to Shenzhen, and it was fantastic . Cheap, smooth, fast, lots of l
332.
by gwd 1y ago
Chinese Pinyin use q for one of their two ch-like sounds -- we could do the same. Q is even more redundant than C, because not only is it identical to K, it's always followed by a U.
333.
by gwd 1y ago
In the case of K, U, and W, those were borrowed from Greek, weren't they? Not sure where J came from. One question is why we borrowed K when we already had C. In modern English, C is more or less a completely superfluous letter, addin
334.
by gwd 1y ago
It sounds like you're agreeing with the article? From TFA: > Over the past year, we've worked with dozens of teams building large language model (LLM) agents across industries. Consistently, the most successful implementations
335.
by gwd 1y ago
Just to emphasize your point, below is a workflow I wrote for an LLM recently, to do language tagging (e.g., of vocab, grammar structures, etc). It's very different than what you'd think of as an "agent", where the LLM
336.
by gwd 1y ago
I think a better way to say it is that the brain doesn't commit to long term memory things that it doesn't need. I remember hearing about some research they'd done on "binge watching" -- basically, if you have two
337.
by gwd 1y ago
> I think the point the author misses... It looked to me like the author did know that. The title only says "Fine-tuning", but immediately in the article he talks about Fine-tuning for knowledge injection , in order to &quo
338.
by gwd 1y ago
> Through extensive experimentation across diverse puzzles, we show that frontier LRMs face a complete accuracy collapse beyond certain complexities. Moreover, they exhibit a counterintuitive scaling limit: their reasoning effort increa
339.
by gwd 1y ago
I mean, yeah, most of the time what you do is add more text and bubble up. But: 1. The very fact that adding more text isn't really any more verbose than not encourages you to add more text, making errors more informative. 2. A non-ne
340.
by gwd 1y ago
I agree with the premise; but the lists sound suspiciously AI-like. "Understanding what's worth making in the first place": Pretty good "Evaluating quality": Could be a lot better.
341.
by gwd 1y ago
First of all, how the links are made permanent has nothing to do with the principle that they should be made permanent . There seemed to be two principles at play here: 1. Links should always work 2. We don't want to store any user
342.
by gwd 1y ago
I was going to say, when you send bulk mail with Brevo, they explicitly prompt you for what the "preview text" should look like, and show how it might look on an iPhone Mail or desktop Gmail.
343.
by gwd 1y ago
That's a failure of the joke not to set it up -- one of the "top corners" of the square is missing. Chickens normally don't make an effort to get to the "other side" (as far as we're aware anyway). To mak
344.
by gwd 1y ago
Basically yes, once the "problem" gets too big, the LLM stops being useful. As you say, it's great for automating away boring things; as a more complicated search & replace, for instance. Or, "Implement methods so t
345.
by gwd 1y ago
> Terms like chancery, Michaelmas, Lincoln's Inn Court, Lord Chancellor, etc, are a foreign language to an American but pretty obvious to an educated Brit in London, especially if they have familiarity with the court system. This wa
346.
by gwd 1y ago
Sure, and 10db means "10x more power". But: 1. 10x more power than what? It changes, and you Just Have to Know. 2. It's 10x more power; so if you're measuring power, like pascals, then 10db means 10x more pascals. But
347.
by gwd 1y ago
> Aspect ratio is a ratio. Right, but in this case they only give you one of the two numbers. Imagine being told that your TV had an aspect ration of ":16", and you just have to magically know what the other number means in th
348.
Why exercise drops off sharply at age 49(axios.com)
5 points by gwd 1y ago|0 comments
349.
by gwd 1y ago
Because sci-fi is about thinking about how technology affects us as human beings, and the conference was about how AI will affect us as human beings. He was invited presumably partly just to draw people to the conference, partly because he&
350.
by gwd 1y ago
That quote sounded terrifying. It reminds me of The Incredibles, where (spoiler) the villain recruits superheroes to try to defeat his "out of control robot", in order to make it invincible. I think we want AI to have an "
351.
by gwd 1y ago
One of the frustrating things I experienced trying to do self-study using Anki and self-created "extensive reading" was that I found I was spending a huge amount of my time doing things to enable me to study , rather than actual
352.
by gwd 1y ago
At some point, in response to someone describing another comment as made by "a random anonymous person on the internet", I said "I don't have my name in my bio but I don't think it would be particularly hard to find
353.
by gwd 1y ago
This is one of the reasons I ended up writing my own language study system. I haven't used the FSRS thing mentioned above, but I've heard it's much better, and having looked at the algorithm, it seems like it should be bette
354.
by gwd 1y ago
> It’s the space in between reading/understanding something and the SRS. ...If someone knows of such a tool, I would love to hear about it. I have a slightly different system I'm developing: Rather than reviewing with flashcard
355.
by gwd 1y ago
> When 2:50 rolled around and your meeting was supposed to end, do you think people actually ended the meeting? Noooooo. Absolutely not! At U of M, they solved this problem by having classes officially start 10 minutes after the time the
356.
by gwd 1y ago
> I wonder if they go back and add details in or plan from them way back at the start. I heard a rumor that what Agatha Christie (and perhaps other mystery writers) would do is to write the entire story with no perpetrator in mind; then
357.
by gwd 1y ago
> Why have two people linked the same youtube video as if that's some kind of response to a graph? Because the "but" in the sentence implies that leveling out means there's no problem. Filling out the missing pieces:
358.
by gwd 1y ago
> which I'll never use. Your younger co-workers will have been educated in public schools you paid for. If you're an employer, the people you hire will have been educated with public schools you paid for. When you retire, the
359.
by gwd 1y ago
Birth control was available to the Baby Boomers's parents, and the result was... a baby boom. My two grandmothers had 7 and 5 children respectively in spite of the availability of birth control.
360.
by gwd 1y ago
> population there has almost perfectly levelled out at just over 51m This video would beg to differ: https://www.youtube.com/watch?v=Ufmu1WD2TSk&pp=ygUWa3Vyemdlc... That said, I am inclined to disagree with his take

More ›