tonarinokanasan

So you're saying the problem is that it's infeasible to distribute the source code, which they already distribute to all of their developers with no problem, while there are numerous platforms that will host it for you for free if it's public FOSS?...

In many parts of the US it's typical to start driving several years earlier than that, and realistically there is no way to get anywhere other than by car. Until kids can drive, they might quite literally be unable to go anywhere or do anything without an adult to drive them. It's sprawl to an absurd degree.

Even where bikes could theoretically be used from a distance perspective, it would likely be way more dangerous and way less practical (no bike lanes and every road is full of cars, no bike parking, you're never getting to a bike shop for repairs without a car, ...)

That feels remarkably intellectually honest. I doubt if I would have replied again in that case, so I don't know why anyone was downvoting this

Ah, I see. That makes sense, but to be fair I think that was expected. I suspect they also pull the same data from every page where adsense is embedded regardless of browser, e.g., and every other company out there is aggregating the same sort of data every possible place they can get it from (shared sign ins, etc etc)

Edit: It's definitely a particularly bad look when there are several things in there that representatives for Google have apparently lied about over the years.

Forgive me for not reading all 2500 documents, but I haven't heard anything to suggest there was a bunch of sinister stuff in there -- and there's nothing implicitly evil about having docs leaked.

Or you can subscribe to gamepass for games, unlimited games

Exactly the same here. Since I swipe type, I have to imagine that would be a nightmare on Dvorak with all the vowels clustered together.

I use Dvorak, but it has nothing to do with statistics for me. When I switched to Dvorak, it felt more comfortable on my hands. My typing speed is essentially the exact same, for example, and I don't think you could find a measurable difference depending on which I use. But qualitatively -- it feels more comfortable.

While the conversion is appreciated, there's no reason to be an ass about it. OP labeled it, so it's not like it was confusing or making unnecessary assumptions about the audience. So really you're the one who just comes across as completely culturally insensitive.

I think the trouble is, what baby are we throwing out with the bathwater in this case? We can't prevent LLMs from hallucinating (but we can mitigate it somewhat with carefully constructed prompts). So, use cases where we're okay with that are fair game, but any use case (or in this case, law?) that would require the LLM never hallucinates aren't attainable, and to get back earlier, this particular problem has nothing to do with capitalism.

This is a thing that is true of all LLMs, but it seems like you're misunderstanding the core issue. It CAN give outputs like that sometimes. What we CAN'T do is force it to give outputs like that ALL the time.

It will answer "I don't know" if its predictive text model guesses that the most common response to this would be "I don't know". To do that, to simplify a little, you could imagine that it reads your question, compares that to all the text in its training data, and tries to find the conversation that looks most like the question you asked, then answers whatever the person in the training data answered. But your exact question wasn't in its training data, so if you took that mental model, and instead had it compare to 1000 similar looking things in its training model and average them, then it would hopefully do a better job of coming up with something at least close to what you actually asked. Now take it to a million, or a billion.

When we're asking questions about the real world, we would prefer for it to answer based on knowledge about the real world. But what if it "matches" data from a work of fiction? Or just someone who doesn't know what they're talking about? Or true information, but about a different subject?

It doesn't know anything. It doesn't understand anything you say. It just looks at patterns that it learned from the training data and tries to guess what words are most likely to be said in that case. In other words, "here's one case where it didn't hallucinate" and "it will never hallucinate" are not the same thing at all.

Edit: To clarify, it doesn't search its training data to answer your question, so asking "was this in the training data" is impossible. By the time you interact with it, the data is long gone. It was just used for training.

I think the bigger factor is, someone who already thinks they're great probably isn't working on noticing and improving their weaknesses. Someone who thinks they still have a lot to learn is putting a lot of effort into improving.

So, especially if they've felt that way for any significant length of time, it's no wonder which person will end up being better.