@kklapsas.bsky.social
On the same vein I dislike "but homosexuallity/changing sex etc also exists in nature so it's natural" which is completely irrelevant. Like if fish didn't change their gender or whatever would it then be ok to discriminate against trans people?
August 15, 2025 at 11:11 AM
On the same vein I dislike "but homosexuallity/changing sex etc also exists in nature so it's natural" which is completely irrelevant. Like if fish didn't change their gender or whatever would it then be ok to discriminate against trans people?
I bet a computer program (though not a simple one) could differentiate this from true randomness. Eg you would very rarely get long streaks of zeros because a lot of very common words have one or three letters. But yeah when choosing where to order from eg, I just take the parity of a random word.
August 11, 2025 at 1:16 AM
I bet a computer program (though not a simple one) could differentiate this from true randomness. Eg you would very rarely get long streaks of zeros because a lot of very common words have one or three letters. But yeah when choosing where to order from eg, I just take the parity of a random word.
What I do is generate ordered pairs of random longish words that I can't immediately see their letters parity, until they have a different parity to each other. Then I pick 0 if the first one is even and 1 if it's odd. Definitely not quick though.
August 11, 2025 at 1:02 AM
What I do is generate ordered pairs of random longish words that I can't immediately see their letters parity, until they have a different parity to each other. Then I pick 0 if the first one is even and 1 if it's odd. Definitely not quick though.
Eh I don't know, I don't disagree with what you say but I still think the attention mechanism is genius and going from there to a transformer is also a very non obvious and inventive step and reducing it to "matrix multiplication" is pretty dishonest and/or stupid...
August 10, 2025 at 12:16 AM
Eh I don't know, I don't disagree with what you say but I still think the attention mechanism is genius and going from there to a transformer is also a very non obvious and inventive step and reducing it to "matrix multiplication" is pretty dishonest and/or stupid...
I didn't say it's wrong, I said it's bad. What it says (ok minus the tokenisation, whatever) applies to very simple neural networks as well so it's not the main thing about LLMs. It's like dismissing an operating system to "it's just a bunch of nand gates".
August 10, 2025 at 12:01 AM
I didn't say it's wrong, I said it's bad. What it says (ok minus the tokenisation, whatever) applies to very simple neural networks as well so it's not the main thing about LLMs. It's like dismissing an operating system to "it's just a bunch of nand gates".
If you've been teaching classes I really hope you understand that LLMs do way more than matrix multiplication...reducing a transformer to that is like reducing an operating system to "they do a bunch of nand gates", accurate but completely misleading
August 9, 2025 at 11:21 PM
If you've been teaching classes I really hope you understand that LLMs do way more than matrix multiplication...reducing a transformer to that is like reducing an operating system to "they do a bunch of nand gates", accurate but completely misleading
I mean, not really, the first post is bad too, what it describes is basically a perceptron which can't even tell handwritten digits apart. The transformer architecture is both very crucial, much more complicated and much harder to invent.
August 9, 2025 at 11:15 PM
I mean, not really, the first post is bad too, what it describes is basically a perceptron which can't even tell handwritten digits apart. The transformer architecture is both very crucial, much more complicated and much harder to invent.
Except LLMs literally don't see the characters, it's impossible for them to count them (because of tokenisation). Like, there are many shortcomings in LLMs but not counting things that are hidden from them is not one of them.
August 9, 2025 at 11:04 PM
Except LLMs literally don't see the characters, it's impossible for them to count them (because of tokenisation). Like, there are many shortcomings in LLMs but not counting things that are hidden from them is not one of them.
Look up "Sleeping beauty paradox". It is a problem that works in a classical fully determinist universe where this kind of uncertainty still arises for (basically) the same reasons.
June 22, 2025 at 3:06 PM
Look up "Sleeping beauty paradox". It is a problem that works in a classical fully determinist universe where this kind of uncertainty still arises for (basically) the same reasons.
Imo, ratings should be reflective of emotional reactions not population statistics. Like, consider all possible attractiveness levels, rate the worst 0 and the best 10 and linearly-ish for the rest, irrespective of how many people fall in each bucket. In my scale it turns out most people are 6-8.
June 11, 2025 at 10:42 PM
Imo, ratings should be reflective of emotional reactions not population statistics. Like, consider all possible attractiveness levels, rate the worst 0 and the best 10 and linearly-ish for the rest, irrespective of how many people fall in each bucket. In my scale it turns out most people are 6-8.
Maybe it's the MAGA body in terms of "the body they idolize" not "the body the have"?
April 28, 2025 at 11:35 AM
Maybe it's the MAGA body in terms of "the body they idolize" not "the body the have"?
So dyslexic people's brain is bad, got it!
April 25, 2025 at 2:07 AM
So dyslexic people's brain is bad, got it!
2 is the preferred but sometimes the only available sheet is in concert pitch so you have make do. BTW if the concert chord is C, a Bb trumpet will play a D, not Bb.
March 12, 2025 at 9:31 PM
2 is the preferred but sometimes the only available sheet is in concert pitch so you have make do. BTW if the concert chord is C, a Bb trumpet will play a D, not Bb.
I think he's making a joke about how Democrats had the ability to enact change when they had the office but chose not to do it.
February 27, 2025 at 11:06 PM
I think he's making a joke about how Democrats had the ability to enact change when they had the office but chose not to do it.
Tsp actually is harder han 3sat, (as far as we know) it's just that the difference is (only) polynomial. This difference is not trivial but they are still in the same complexity class because it's closed under polynomial differences.
January 10, 2025 at 11:54 PM
Tsp actually is harder han 3sat, (as far as we know) it's just that the difference is (only) polynomial. This difference is not trivial but they are still in the same complexity class because it's closed under polynomial differences.
Do you find it intuitive why SAT is NPhard? (like 3sat with any number of literals per clause instead of 3). If yes then you can look up the proof that you can reduce any SAT to a 3SAT (fairly simple). If no then you have to squint a bit until you see that a SAT instance simulates a Turing machine.
January 10, 2025 at 11:44 PM
Do you find it intuitive why SAT is NPhard? (like 3sat with any number of literals per clause instead of 3). If yes then you can look up the proof that you can reduce any SAT to a 3SAT (fairly simple). If no then you have to squint a bit until you see that a SAT instance simulates a Turing machine.
Ha, I had almost this exact conversation last week actually. Nice.
January 9, 2025 at 11:03 PM
Ha, I had almost this exact conversation last week actually. Nice.
I think this has been tried and it was just that it was more trouble than it's worth, I remember Karpathy had a talk where he explained why.
December 24, 2024 at 2:42 PM
I think this has been tried and it was just that it was more trouble than it's worth, I remember Karpathy had a talk where he explained why.
Yeah i think it's just much more expensive and you need much more data which are harder to get than just text. Should happen eventually though.
December 24, 2024 at 2:38 PM
Yeah i think it's just much more expensive and you need much more data which are harder to get than just text. Should happen eventually though.
Has been done rather hilariously here!: youtu.be/Hg5dAwpftRk?...
Guess Who, But With VIBES Only | House Rules
YouTube video by No Rolls Barred
youtu.be
November 23, 2024 at 10:00 PM
Has been done rather hilariously here!: youtu.be/Hg5dAwpftRk?...