Post Ak3nXSP1u44pcm49nU by mikerotch@poa.st
(DIR) More posts by mikerotch@poa.st
(DIR) Post #Ak3epjIGwUXhdUXkw4 by Ghislaine@poa.st
2024-07-18T14:39:33.155910Z
16 likes, 7 repeats
(DIR) Post #Ak3ermw2NIoPVb7o36 by nugger@poa.st
2024-07-18T14:41:28.640359Z
2 likes, 0 repeats
@Ghislaine REMINDS ME OF THAT HORRIBLE TRAGEDY
(DIR) Post #Ak3es6IMOWxRYKiOyu by tyler@nicecrew.digital
2024-07-18T14:40:36.136307Z
1 likes, 0 repeats
I don't think that's true hold on let me check
(DIR) Post #Ak3esAvr9cvZvzE06C by DurpState@poa.st
2024-07-18T14:40:39.515699Z
1 likes, 0 repeats
@Ghislaine aww, it's retarded
(DIR) Post #Ak3exbFIkS7wzauH3Y by tyler@nicecrew.digital
2024-07-18T14:41:55.644069Z
1 likes, 0 repeats
Oh what the hell
(DIR) Post #Ak3f2YgtftcCW5B9Bw by white_male@poa.st
2024-07-18T14:43:25.760949Z
0 likes, 0 repeats
@Ghislaine Best thing, these systems can't learn. The stupidity baked in at the start stays with it forever.
(DIR) Post #Ak3fLBRwJEvnwA40jQ by supersid333@gearlandia.haus
2024-07-18T14:46:46.994609Z
1 likes, 0 repeats
@Ghislaine AI is great, it's the most comfortably wrong thing on the planet.
(DIR) Post #Ak3i51tKNXa5WPrZLs by petra@poa.st
2024-07-18T15:00:37.689259Z
1 likes, 0 repeats
@Ghislaine If you ask ChatGPT to explain this calculation at the bit level using the IEEE 754 Floating-Point Standard it's clear why this happens. The computer is a tool. If you don't understand your tool then that will bite you in the ass some day.Years ago I was working on an algorithm at a telecom's company. It worked fine in Matlab (standard engineering s/w for signal processing + tonnes of other shit now). Matlab used floating point. The results matched the theory. Beautiful.Then we implemented it in the fixed point signal processing TI processors we were using at the time. Completely broke the algorithm. The programmers loved this because it showed the stupid engineers couldn't do their jobs. Literally happy at the set back in spite of it screwing the company.It was a basic fix. You just had to understand and compensate for the differences between floating and fixed point. Engineering work. It's what we get paid for.But the fix was never implemented because the programmers decided it was fundamentally broken and engineers would never again touch "their stuff". Had lots of chinks starting to infiltrate by then, and the pajeet wave was in its infancy but that was there, too.Yes, the company was retarded. Or was by that point. It had a good history until shortly before that point. And it failed spectacularly a few years later because it was just so far behind on newly evolving telecom techniques straight out of research labs. Because programmers couldn't read technical literature (maff is hard!).Good riddence.Understand your tools, boys.
(DIR) Post #Ak3iKB3kZer6IexeSm by JoshuaSlocum@poa.st
2024-07-18T15:20:13.598812Z
0 likes, 0 repeats
@petra @Ghislaine had a look around when i saw this (it's clearly a thing) and one of the explanations is that the LLM is reading it as you would version numbers due to how it tokenizes the inputi.e., v9.11 is newer (larger) than v9.9
(DIR) Post #Ak3kb92BYal6qlRoRM by petra@poa.st
2024-07-18T15:23:19.676835Z
1 likes, 0 repeats
@JoshuaSlocum @Ghislaine Huh. I need to look at its detailed explanation regarding floating point.I didn't and that's my mistake. I've seen these things hallucinate completely plausible answers that are completely incorrect. Which is a whole other fucking issue. Thanks.
(DIR) Post #Ak3krODgrhIsa3ZigC by JoshuaSlocum@poa.st
2024-07-18T15:48:38.422807Z
0 likes, 0 repeats
@petra @Ghislaine oh i'm sure it's an AI retardation, the math part of it is clearly jacked upi found it interesting that (possibly) the AI was looking at it as version numbers
(DIR) Post #Ak3mA2gdFVMnFjZQxs by lain@lain.com
2024-07-18T16:02:28.897051Z
0 likes, 0 repeats
@Ghislaine this is probably a tokenization issue, pretty interesting
(DIR) Post #Ak3nEDq3pNq6x8XOng by petra@poa.st
2024-07-18T15:53:16.294287Z
1 likes, 0 repeats
@JoshuaSlocum @Ghislaine But python really does give the wrong answer. Just run python in a terminal and type in the equation. You get the incorrect answer. It really is due to some finite precision issues in representing real numbers as floating point binary. This is normal behaviour but surprising if you don't understand how it works under the hood.We had a whole term of this numerical analysis stuff back when. Conte and DeBoor. Using Fortran no less.
(DIR) Post #Ak3nUpHGOo0rhUB9Yu by JoshuaSlocum@poa.st
2024-07-18T16:18:10.547230Z
0 likes, 0 repeats
@petra @Ghislaine yeah, stuff like the 000000009 is why those in the know do money calculations by converting to centsit's like the FDIV bug and the Superman III plot device had a retarded baby
(DIR) Post #Ak3nXSP1u44pcm49nU by mikerotch@poa.st
2024-07-18T16:18:07.934757Z
1 likes, 0 repeats
@petra @JoshuaSlocum @Ghislaine I am very glad that I almost never have to deal with floating point operations in software, and on the rare occasions when it I do it's only for display purposes and never for critical business logic.Dealing with integers and their tricky issues is plenty of complication in my life.
(DIR) Post #Ak3ns9CeTxVo9v8t84 by petra@poa.st
2024-07-18T16:21:56.133866Z
1 likes, 0 repeats
@mikerotch @JoshuaSlocum @Ghislaine Early in my career I was working in a group with two guys that implemented digital filters. I'd been taught how to design them in school but it always assumed infinite precision real numbers. Which is fine. It gets the fundamental theory down. But I was blown away when I found the filter coefficients as numbers looked nothing like the idealized values. There was a whole layer of tricks you needed to get it all to work in finite precision. But it did work.
(DIR) Post #Ak3ofYeYeEobHSkvNQ by mikerotch@poa.st
2024-07-18T16:26:11.624143Z
1 likes, 0 repeats
@petra @JoshuaSlocum @Ghislaine If you're implementing in C or C++ then signed integers overflow... probably. Usually. Depending on the OS, compiler, and optimization level.Unsigned integers always overflow with well defined behavior, meaning your math is predictably modular (so you better know to expect that).Many languages outside of C and C++ don't have unsigned integers though.
(DIR) Post #Ak3pQRGHwlgry0l5do by petra@poa.st
2024-07-18T16:32:12.982479Z
1 likes, 0 repeats
@mikerotch @JoshuaSlocum @Ghislaine Yes. This can all be handled well and predictably by engineers who understand what they are doing. But it's a surprise to people who don't.Like being able to rearrange the terms of a convergent infinite sum in certain cases and change the answer it converges to by changing the order of the terms you are summing. It's magic, or a contradiction or a paradox ... No it's not. You're just confusing the behaviour of finite sums that you are used to with infinite sums which you are not. It's what will perk people up in an introductory analysis course and understand that infinities really are subtle and interesting.
(DIR) Post #Ak3pQShcad9GR4YQsq by mikerotch@poa.st
2024-07-18T16:35:39.889715Z
1 likes, 0 repeats
@petra @JoshuaSlocum @Ghislaine > This can all be handled well and predictably by engineers who understand what they are doingIt can be handled well, but it still trips up a lot of people.Explicitly testing for signed integer overflow in a way that prevents the compiler from ever optimizing away your test is definitely in the top 10 of most difficult C/C++ programming challenges.