> It’s true that so-called “hallucinations” are an artifact of LLM technology that can never be totally eradicated. But if you think about it, humans make a lot of mistakes as well. A calculator that gives you the correct answer 90% of the time is still an impressive technological achievement!
I know this article is meant to be dissing on LLMs, am I too caught up in the mania that this sounds to me unironically like a positive argument for LLMs?
If you're being serious, my honest answer is yes. But, then, I'm the kind of person who refuses to part with my college texts and graphing calculator because of nostalgia.
No offense is intended with this comment, but genuinely, if you see a calculator that returns incorrect answers 10% of the time as a technological achievement in a world that has had deterministic calculators for the better part of a century, I think it's worth considering that you might be a little caught up.
I don't think this point is well stated. We have things like ECC to help with math being imperfect so that feels solvable. The problem with LLMs is much more subtle. They induce cognitive errors in humans. I'm doing it right now I'm saying "them" like they have individual identities, but they do not.
If you know they're just mirrors, then you know that looking in the mirror isn't the same as exploring the world.
It turns out, though, that many people can be trapped in a hall of mirrors. They will happily buy a condom for their brain.
Right. And when we talk "to" the LLM we address it as "you" or give it commands as if there is someone there to respond.
Maybe we need a new pronoun for it. 2nd person inanimate.
YOU -> YAI?
This week I was experimenting with leaving out "you" pronouns by changing everything to passive/subjunctive and leaving out the subject.
E.g. when referring to something the AI has written or done:
The code that was written... Instead of: The code you wrote.
For something I want it to do:
May the code be modified to include the following constraints... Instead of: Modify the code to include the following constraints
I don't know if that's really helpful but it's good to be cognizant I'm not really talking to anyone.
I think it makes sense to see it positively. If one wants a calculator, we have calculators. The whole point is reaching things that deterministic calculations either can't or don't in a reasonable timeframe.
that's true. instead they (device dependent) butcher calculations by accumulating arbitrarily large numerical error and/or by using binary floating point math, which is so much better... /s
understood. my question is more, how is the advent of inorganic online content impacting the scaling of pretraining? is there some mathematical model or science on this topic? these old books being scanned, is this just about more precisely modeling human thought, or it more about getting a finer-grained picture of the world and history? any more color like that would be really interesting.
> Did an AI tell the United States to bomb that girls’ school in Iran, which killed 156 civilians?
Yes, because certainly the people who are supposed to be in charge bear no responsibility. This article does not make very good arguments. If the US was attacked by a foreign weapon based on AI, nobody would accept that argument.
> So you’re spending billions of dollars to make big machines that might guide us into socialism?
> No! Socialism is when the government collects the wealth of the entire population and spreads it out evenly so everyone is provided for.
Sigh... ah yes, the great tradition of lying about what Socialism is in order to make your unsupported argument.
If I could, I'd certainly prohibit AI training on FOSS code I've written. It is a very one-sided deal / grab: FOSS code in (for free), mostly proprietary code out (for pay to some mega-startups).
> It’s true that so-called “hallucinations” are an artifact of LLM technology that can never be totally eradicated. But if you think about it, humans make a lot of mistakes as well. A calculator that gives you the correct answer 90% of the time is still an impressive technological achievement!
I know this article is meant to be dissing on LLMs, am I too caught up in the mania that this sounds to me unironically like a positive argument for LLMs?
If you're being serious, my honest answer is yes. But, then, I'm the kind of person who refuses to part with my college texts and graphing calculator because of nostalgia.
No offense is intended with this comment, but genuinely, if you see a calculator that returns incorrect answers 10% of the time as a technological achievement in a world that has had deterministic calculators for the better part of a century, I think it's worth considering that you might be a little caught up.
> in a world that has had deterministic calculators for the better part of a century
so rewind that clock and see the above comment miraculously start making sense
I don't think this point is well stated. We have things like ECC to help with math being imperfect so that feels solvable. The problem with LLMs is much more subtle. They induce cognitive errors in humans. I'm doing it right now I'm saying "them" like they have individual identities, but they do not.
If you know they're just mirrors, then you know that looking in the mirror isn't the same as exploring the world.
It turns out, though, that many people can be trapped in a hall of mirrors. They will happily buy a condom for their brain.
Right. And when we talk "to" the LLM we address it as "you" or give it commands as if there is someone there to respond.
Maybe we need a new pronoun for it. 2nd person inanimate. YOU -> YAI?
This week I was experimenting with leaving out "you" pronouns by changing everything to passive/subjunctive and leaving out the subject.
E.g. when referring to something the AI has written or done: The code that was written... Instead of: The code you wrote.
For something I want it to do: May the code be modified to include the following constraints... Instead of: Modify the code to include the following constraints
I don't know if that's really helpful but it's good to be cognizant I'm not really talking to anyone.
I think it makes sense to see it positively. If one wants a calculator, we have calculators. The whole point is reaching things that deterministic calculations either can't or don't in a reasonable timeframe.
We have calculators that work 99.99% of the time. I am not sure of it makes sense to settle for 90% just because we could.
Where have you seen a calculator that gives the wrong answer 0.01% of the time?
That would be unheard of.
How about 1 in 9 billion?
https://en.wikipedia.org/wiki/Pentium_FDIV_bug
that's true. instead they (device dependent) butcher calculations by accumulating arbitrarily large numerical error and/or by using binary floating point math, which is so much better... /s
https://en.wikipedia.org/wiki/Round-off_error
This is a horrible article written by someone clearly upset about the state of things to process it objectively.
Please refute this argument: how can you expect 100% accuracy when your input is natural language which is fuzzy by definition?
Yes, because that would be an insanely bad calculator, 1 in 10 wrong answers?
> We’re already ripping the spines out of old rare paper books (we playfully call these “low-background steel”)
can someone closer to pretraining elaborate?
Not at all close, but this link might give you more to go on (includes more links):
https://www.snopes.com/fact-check/ai-companies-destroying-ra...
More here: https://en.wikipedia.org/wiki/Project_Panama
https://en.wikipedia.org/wiki/Low-background_steel
Replace "nuclear explosion" with "AI produced works"
In this context it means "text written before the advent of AI slop, so we know it at least came from humans and wasn't influenced by AI slop".
understood. my question is more, how is the advent of inorganic online content impacting the scaling of pretraining? is there some mathematical model or science on this topic? these old books being scanned, is this just about more precisely modeling human thought, or it more about getting a finer-grained picture of the world and history? any more color like that would be really interesting.
> Did an AI tell the United States to bomb that girls’ school in Iran, which killed 156 civilians?
Yes, because certainly the people who are supposed to be in charge bear no responsibility. This article does not make very good arguments. If the US was attacked by a foreign weapon based on AI, nobody would accept that argument.
> So you’re spending billions of dollars to make big machines that might guide us into socialism?
> No! Socialism is when the government collects the wealth of the entire population and spreads it out evenly so everyone is provided for.
Sigh... ah yes, the great tradition of lying about what Socialism is in order to make your unsupported argument.
I’m glad my many open source projects contributed to making AI so awesome.
If I could, I'd certainly prohibit AI training on FOSS code I've written. It is a very one-sided deal / grab: FOSS code in (for free), mostly proprietary code out (for pay to some mega-startups).