Let’s try ChatGPT. Is it any good?

[@Joao-ty4nq]

Happy new year, Bisqwit! Do you mind sharing how college is going?

[@Bisqwit]

Hi! University, not college. This year the progress has not been nearly as rapid for me as it was the last year, and this is because most courses I have taken in fall 2022 have involved group works or mandatory attendance, which definitely bogs one down. In particular, there was the guided basic practical training (of subject teachers’ education), where I had to hold four math classes for highschoolers, and four computer science classes likewise - and to observe twenty-or-so classes held by others, plus a heap of other activities. During that time, I was not able to do many other courses (and a few that I started doing, I cancelled because of those mandatory attendance bullşhit requirements). The practical training was interesting experience, and it went well. But I am almost done with Physics studies, and I began Chemistry and also Philosophy and am looking forward to also try studies in Theology - all four of these being completely voluntary and not part of my degree. As for the mandatory studies, it has mostly focused on the pedagogic studies now, as most of the other stuff that I am missing was not available in the autumn periods.

[@bitterlemonboy]

@@Bisqwit How stressful is it to take so many different studies? Why don't you focus all of your work into getting your degree?

[@victorgabr]

@@Bisqwit do you aspire going into master's and PhD? Congrats for passing through physics classes, as a physics major, I know that can be supper tough. Btw, you may be the best Computer science professor everyone wanted. The quality of your CS contents is unprecedented on the internet.

[@NinjaRunningWild]

@@Bisqwit In other countries the terms university & college are synonymous.

[@Bisqwit]

@@bitterlemonboy That is a good question. One reason is that I am already focusing all my work into that, with the exception of working on the bachelor’s and master’s thesis. Some mandatory courses just happen to have prerequisites and are only available in some study periods, so they cannot be rushed. Another reason is that if I graduated and wished to _then_ continue taking other courses, I would need to apply again for some other degree. Because once you achieve what you have applied for, the study rights associated with that application end.

[@MyAmazingUsername]

"Finnish Man Bullies ChatGPT" is a more accurate title. You schooled it!

[@levonschaftin3676]

the burp caught me off guard

[@solitudesf8111]

ah fuck, spoilers

[@MarquisDeSang]

He is not Yanagi19871

[@mokopa]

@@solitudesf8111 Serves you right for reading the comments before watching the video

[@More_Row]

@@mokopa Actually not wrong.

[@klaus9503]

13:03

[@prostmahlzeit]

I really appreciate how thorough you reviewed this tool. Other youtubers task it to implement trivial student homework and do fear mongering that software developers will become unemployed soon. Meanwhile the tool is just a fancy search engine with a parrot feature.

[@ihspan6892]

Finally someone did the homework and did not repeat the nonsense people say on social media. The results are impressive, but those who say, that recent layoffs at tech companies are due to automation and chatGPT are completely in the wrong. Thank you Bisqwit as always, excellent video.

[@mokopa]

What people fail to understand about GPT is that it is a LANGUAGE model. Think of it as an English professor who feels compelled to answer every question in such a way that it makes sense - not that it should be correct. The professor is very, very good at language, but really doesn't understand programming, science or mathematics. The fact that it is able to approximate the truth and that it sometimes comes arbitrarily close to the truth is itself extraordinary. Soon the Maths professor, the CS professor, the Physicist, the Chemist will all be activated and ChatGPT will be godlike. Buckle up.

[@pavelperina7629]

Hard to say. For me it's pretty impressive. I mean it has it flaws, but if you ask it to write some code in C++ and constrain it to use certain libraries, it writes code that looks correct (yet contains a few syntactic bugs), that looks that it does something which makes sense. If you are extremely lucky, it may even work. Last week I tried to assist with writing reaction-diffusion simulation and it basically initialized array in such a way that it did not work, it choose some constants way too low making scene basically static and I've spend some time to fix it. On the other hand, OneTBB parralel loop was correct and when I asked it to rewrite some (simplified) code using AVX instristic instructions, it was obvious how to write AVX optimized loop. I have no idea how exactly it works and how my own brain works neither, but it can somehow fuse enough knowledge that one question can save watching few youtube tutorials, stack overflow and gitlab searches. What is worse - I found several papers that were likely intentionally misleading or code example with paper solving some non trivial problems where code was like 95% complete, but person writing it gave up at the most important part and choose some simple alternative hoping no one will notice.

[@pazu_513]

The AI sounds like me in my exam papers when I'm not totally sure what I'm taking about but writing everything I know in hopes of pity points.

[@nooberast5887]

Maybe it's the result of way how this particular model was trained - according to wiki, Chat-GPT was trained with use of supervised and reinforcement learning. The model was provided with conversations, where humans play as the questioner and the answerer. Then, the human trainers ranked responses that the model had created - like your "pity points". These rankings were used to create 'reward models' to further improve results. I'm definitely not an expert in ML, but it seems that model trained [itself] to maximize the reward by generating the flow of naturally-looking statements, disregarding the correctness of facts in such statements.

[@sophiacristina]

Do you write a lot of "sorry" on the papers too?

[@vule656]

We all did that at some point

[@ENDESGA]

*oh yes, I'm so glad you've come back to talk about this!*

[@Einygmar]

Great video! I find that a lot of people overestimate ChatGPT thinking that it performs some sort of calculations or logical analysis while in reality it simply hallucinates sentences which have the most correspondence with the input text using data from various internet articles and tutorials stored in a neurally-compressed form.

[@gnauben]

u mean like any other human being lol

[@teachies902]

@@gnauben hah! gotcha! any professional worth their salt would assert that their thinking is superior to a text hallucination.

[@Einygmar]

@@gnauben Well maybe some humans with brain damage do construct their sentences in a similar fashion but it's definitely not a general trait of humanity. ChatGPT on the other hand is just a gigantic archive that stores all of its input data using a specific compression method\approach. ML models by their nature are a lot more comparable to the widely used lossy compressed formats such as Mp3 or Jpeg.

[@EximiusDux]

Everything ChatGPT does is a literal calculation. It's a piece of software programmed by humans using calculations formatted into algorithms. It " hallucinating" an answer is "it" literally calculating how certain words relate to each other by looking it up in a database. It doesn't understand anything at all beyond having calculated a loose relationship concerning the order of words.

[@anatolydyatlov963]

​@@EximiusDux We still don't really know how understanding works; that is, we don't know exactly what happens in the brain when we have that "aha!" moment while learning something difficult. The distinction between truly understanding a concept and simply memorizing information is still not clear, or at least neuroscientists cannot yet provide a definitive answer. Because of this, it's difficult to confidently say that ChatGPT doesn't understand anything at all, since we don't even know how it works in our own brains.

[@WMKooda]

Happy New Year Mr. Bisqwit!

[@KimForsberg]

My favourite test of ChatGPT so far is when i asked it to solve a programming problem, where it gave a correct answer but the wrong solution. I ask it to run its incorrect solution, giving proper printouts and it gives the right answer still.

[@julz6077]

Do you understand what ChatGPT was trained on?

[@tschaderdstrom2145]

@@julz6077 Both smoke, and mirrors

[@treudden]

@@julz6077 its an alternate version of GPT-3 which was trained on 175B parameters

[@Katniss218]

@@treudden parameter count != training data

[@treudden]

@@Katniss218 yes ofc, but i dont know what dataset it was trained on

[@Mankind5490]

An AI apologizing to Bisqwit for its mistake, literally not even surprised lmao

[@baist1375]

Very stupid thing this ChatGPT. Always produce same text. Yes it's correct. I ask him don't repeat, don't use pattern. Just use random order of words. Build your sentences like a human. Use different style of text. Be cool guy. Don't be a dick. "I apologize...."

[@lechurrajo]

what's good about chatGPT is that you can always tell it to write a rap song using the last code reply as inspiration, or something like that. Making it ridicule itself at the end is a way to assert dominance and make the ai remember who's boss

[@gdclemo]

And that's how the war of the machines started.

[@darak2]

It's similar to the image generation AIs in some ways, as in it sometimes produces plausible results that, at a glance, seems to be very high quality, but they fall apart when you take a closer look at the details or ask it to generate complex functional structures that do not resemble very precisely the material it has been trained on.

[@mokopa]

Those AI hands. Stuff of nightmares...

[@Bisqwit]

Yeah, some AIs can create very beautiful rooms with attracting interior design, but when you look closer at some of the furniture or objects on tables, none of them make sense. Sometimes, it’s like the “name one object in this photo” meme.

[@hyoenmadan]

@Hugh Munguss Show your results, or it did never happened. And please, don't cheat with photoshop fixing. We will know you did it.

[@txorimorea3869]

Is like a dream. Even the way it generates images is similar to Wake Initiated Lucid Dream: first the image is extremely noisy, not better than a blob of monochrome points.

[@imalaia8033]

The problem with the AI is that the if you ask it something that you don't know about (like the sort thing) it will look right and it will gaslight you into believing it.

[@LukeAvedon]

Very well said!

[@nobody8717]

Almost like anything on the internet. Weird that.

[@pavelperina7629]

But it looked ok or almost ok. We asked it to convert gps coordinates (lat, lon, altitude) into cartesian. Code ended with return std::pair(x,y,z). Then we asked it find error in the code. It wrote that input is in meters (it was actually not defined) and output is kilometers (which was correct) and it should return std::tuple (which was correct). Then we asked it slap itself and give correct answer immediatelly. It wrote some excuse that it can't do anything physically. Then we asked it to add documentation, it added it in correct doxygen format and it made sense. Still, little error in the code was that earth was a sphere, not wgs84 ellipsoid. Still pretty good and I've seen many wrong answers on internet. Even paper containing major mistakes or ommiting the most important part while paying too much attention to details that are common knowledge and in many cases it was needed to perfectly understand the problem to implement something roughly according to paper, to make it work. I almost think it was kind of "copy protection".

[@baist1375]

I asked him sort algo! He still produce bubble sort. I think creators wrote his model with Bubble Sort. Why?! Because!

[@willtheoct]

same for humans though

[@victorgabr]

I knew that bisqwit would break ChatGPT in 10 minutes. Man, That was a great video, as always. Keep up the excellent work. I hope you have a happy new year!

[@hasankaraman]

This video perfectly shows how fun Bisqwit is at parties xD

[@MarquisDeSang]

RIP headphone user.

[@user-sl6gn1ss8p]

so, a lot?

[@Lord_Loading]

@@user-sl6gn1ss8p depending on who you are

[@elgusanito6991]

a lot, indeed

[@zlate42]

I would have the fun of my life with Bisqwit at a party.

[@amyshaw893]

it is definitely better at some things than others. i play a lot of dungeons and dragons, and it is pretty good at generating NPCs and story beats for my campaigns that i run, and with a bit of prompting, even managed to DM a game for me, where i specified myself as the only player. In the introduxction of the world that it gave me, however, there was a line that my colleague told me he swears he has seen somewhere else. However, when i asked it to generate plausible tom scott video introductions ("I'm at _______"), it completely ignored my template and kept on generating stuff that sounded very wrong.

[@elgusanito6991]

I think because the kind of training it had its just good at generating plausible text generation, but to "reason" from its knowledge (generate enough plausability that you are truth) is just something its not really trained for, but maybe with time and more refined data it will get good at this

[@diverman1023]

You're perfectly outlining why the general populace shouldn't worry about openAI. it's merely the output of a statistical predictive model that can't surpass a human brain in many cases, but it can be used as a tool if you already have the knowledge

[@slipcurve1410]

chatgpt is an awesome liar though, so effortless and confident.

[@mllarson]

13:02 was that burp tasty? 🤣🤣🤣

[@Bisqwit]

No, not really.

[@grass8550]

ChatGPT: The calculation of the pH for a hydrochloric acid solution with a concentration of 0.2M can be done using the formula for the negative logarithm base 10 of the H+ ion concentration. pH = -log10(0.2) Calculating the negative logarithm of 0.2 base 10, we get: pH ≈ 0.70 Therefore, the pH of this hydrochloric acid solution will be approximately 0.70.

[@sanesanyo]

Great video. Highlighting exactly what I have been experiencing with ChatGPT. It makes you believe whatever it spits is the right answer and even if you direct it towards the right answer, it exponentially gets worse. I still find it quite helpful, but we have now a generation of people who thinks it is the holy grail.

[@WaszInformatyk]

13:03 🤣 That perfectly summarizes the entire encounter with ChatGPT. You can "eat" artificial intelligence's like m&m's ;-)

[@jamesking2439]

Tech like ChatGPT might be useful as a fancy search engine. It can introduce you to things you don't know, but you need to go and verify everything you get from it.

[@theseriousaccount]

It is the future of search.

[@94Quang]

hey I love your videos I'm glad you're back :D

[@DrMxy]

But what about Bob Ross oil colours?

[@Bisqwit]

I asked it to list the paint colors used by Bob Ross. It started explaining that Bob Ross used a limited palette of a colors, and it listed those colors, so I asked it, doesn’t belong to the list too? The AI apologized and agreed that yes, Bob used that paint too, and then proceeded to reiterate the false claim anyway. This happened three or four times, after which I bluntly stated: “Based on all these corrections, it seems false that Bob Ross used a limited palette of colors as the foundation for his paintings.” and the AI agreed and finally dropped the false claim. However, there was one paint color that I was not able to mention: Van Dyke Brown. The chat form outright denies the input if it contains the word “dyke”. I was able to get around by typing “Vandyke” instead. But when I asked the AI to correct this spelling for me, which it did, then I was not able to input anything anymore. Anything and everything I would type would just bring up a dialog box “This prompt may violate our content policy.” and refuse to submit it. Looks like the form denies the input, if the prompt _or_ the chat context contains a banned word. In this case, “dyke” appears to be a banned word.

[@DrMxy]

@@Bisqwit Heh that encapsulates my experience of this narcissistic, lobotomised AI too - only apologising when it is corrected and then pretending it knew all along. I also asked it about my home town, it got the geographic location right, but just made up the history and told me about people who never lived here.

[@AllenParks1]

Dalle 2 does them nicely :)

[@Nikkes02]

Good to hear from you again!

[@qeqsiquemechanical9041]

Great to see you again!

[@theseriousaccount]

I gotta say though, as a programmer, chat gpt in combination with github copilot is good enough to automate many things in my job. To leave the development and writing to an AI and simply take the role of an editor and supervisor is magical. This is the next step in computing. I can’t wait to see where we are in a few years.

[@kevinmcnamee6006]

ChatGPT just makes stuff up that sounds good. In response to some obviously incorrect answers, I asked it where it got its information. It referred to articles in some well known journals. When I looked them up, I found that the articles didn't exist. Someone should add a "fact-checker" to the next version of the code.

[@Bleenderhead]

My understanding is that ChatGPT and its ilk are all general-purpose large language models trained on huge datasets of text in a variety of fields, but they lack any explicit knowledge representation. When you ask Wolfram Alpha a question, it will actually parse the english into a mathematical expression that closely approximates what you said, and evaluate it using analytic techniques. Large language models have no such ability, and with no human-interpretable knowledge representation, no ability to add such an ability. So I don't see how they could ever become better at quantitative reasoning; no matter how much data you throw at them, they will never be able to do even basic arithmetic reliably. Humans have constructed idealized, axiomatic systems like mathematics, greatly augmenting our ability to reason about the world around us beyond gut feeling and memorization. Until AI can do something like that, it can never surpass a human.

[@PureAsbestos]

I feel like some combination of a more analytical approach like Wolfram Alpha uses with something like ChatGPT could be amazing. Like, in rough terms, it could transform your question into a wolfram alpha query, then get the results and transform that into a natural language answer

[@luna197-old]

the world has been blessed by a bisqwit upload

[@Koettnylle]

The world has also been blessed by a bisqwit burp

[@Margen67]

Umbreon needs HUGS

[@thogameskanaal]

12:55 that burp was so sudden and out of character. Why can't I stop laughing 😂😂 help

[@gusspinochet9685]

ChatGPT isn't sophisticated enough to actually be useful. Nevertheless, I'm blown away by its abilities. It does well with simple logic problems like "Bob always takes his truck to work if it is not raining. Since Bob did not take his truck to work today, what can we conclude?" to which it responds "If Bob did not take his truck to work today, we can conclude that it is raining. Bob only takes his truck to work if it is not raining, so if he did not take it, we can assume that the weather is not suitable for driving his truck. Is there anything else I can help with?" , but starts to fail if the problem has multiple parts, like "Bob always takes his truck to work if it is not raining. Bob always wears blue jeans when he drives his truck. If Bob wore shorts today, what can we conclude?" to which it simply responds "If Bob wore shorts today, we can conclude that it is not raining today." which is incorrect. Funny enough, if you ask ChatGPT some leading questions, it can sometimes correctly answer the question. For example, it responds to "Bob always takes his truck to work if it is not raining. Bob always wears blue jeans when he drives his truck. If Bob wore shorts today, did he wear blue jeans? Did Bob drive his truck to work? What can we conclude about the weather?" with "If Bob wore shorts today, then he did not wear blue jeans. If Bob always wears blue jeans when he drives his truck and he did not wear blue jeans today, then it is safe to conclude that Bob did not drive his truck to work today. Since Bob only drives his truck to work if it is not raining, it is reasonable to conclude that it must be raining today." (although I admit I had to regenerate the response once for it to answer correctly). It is still the early days of this technology, but it certainly seems as though it might one day be capable of drawing complex logical conclusions all on its own. If it can 1. not confidently generate incorrect answers (I am not surprised that ChatGPT does this so often, since GPT was simply developed for text completion) 2. chain its ability to solve simple logic problems together to solve more complicated problems like I indicated above it might become a very useful tool.

[@sporedoutofmymind]

Remember how as a kid you'd pad and pad and pad a paper to meet a word or page count requirement? Yeah, AI is a program that does that.

[@boaridaaa_lol6241]

Happy new year bisqwit

[@acdarc]

Don't let Bisqwit loose on these AI machines, because they might actually learn something and we are doomed!

[@PureAsbestos]

Something interesting that ChatGPT is relatively good at is transforming text/information. Like, if you had some instructions, and wanted them in bulleted list form. Or even translating between programming languages, seems to work alright mostly. On the other hand, it's decidedly bad at giving correct information. So yeah, it's still pretty much just a novelty right now.

[@domtom128]

13:02 LOL

[@X_marksthespot]

welcome back! it’s nice to see you still active in some way. i can’t thank you enough for founding tasvideos. you helped start a whole new art form! - LogansGamingRoom

[@eyluismi168]

Good video, my main usecase for ChatGPT is to get answers to quick questions were I know it will be faster than Google, also, If you are learning a new language, it can be helpful as you could give it an example in a language you know, ask it to translate it to the lang you want to learn and if it adds something extra, you can ask to remove it. But these examples need to be really simple. Also It just can't do math, and as you said some times it will just go in one direction and you wont be able to get anything useful. So, I think your summary is correct, you need to be an expert to find what is wrong but I'm positive about the actual status and the speed this is getting.

[@kenan2386]

AI isn't taking over anytime soon with this guy

[@galzajc1257]

I program in Wolfram language (Mathematica) and Python. It turns out, that it can't really write Wolfram code, but it writes quite a great Python code (like 1 error/20 or 30 lines, also uses okay algorythms). It is probably the case, that it had more traning data for the most popular programming languages and it can code in them better than in others. So for Python: yes it still makes errors that you have to manually fix, but that is way faster than coding everything, because the code it produces is quite readable and easy to check. But math and science problems: yes we are probably still many years from that. Maybe it could be trained writing math proofs in proof assistants, so that the results are varified. Because honestly many many formal math proofs i've learned are just using some deffinitions and properties tovarify other properties with a basic logic and calculations. So it might be possible to do that. But with all the things, that require some spacial reasoning and geometric intuition, it's hard to imagine, that a language model would ever do that. There are probably better ai alternatives than a language model.

[@_polpon]

Glad you are taking on this subject, it is clear that AI isn't advanced enough, but I also think it's clear that it is advanced enough that we are beginning to actually test it. The revolution can begin small, and I think that's what we are seeing, the beginning of the revolution. It already exceeds most humans in the capability of making art, soon music will follow. More precise topics, like maths and chemistry will take a long time, math and chemistry is just better suited for conventional logical algorithms. But it will probably come as well in the future.

[@eduardovalenzuela2391]

@bisqwit i dont know if you were giving up on the chat bot at this point 12:57 , but it gives a clear glimpse of the capabilities of the tool, not to its extent but rather to its entertainment purpose

[@AlyphRat]

I (somehow) managed to make a .NES homebrew where a bunch of 16x16 tiles are printed on the screen with the help of AI. It even taught me which compiler to use, how I should manage the folders, and how the 6502 ASM programming language works. Despite making some mistakes, the mere fact that I made something work with its help already makes it amazing. (Not only that, if you ask it to comment the code, it actually does explain what every line/function does) Edit: Just watched the video to its fullest extent, just like wood to the fire, it will only be efficient if you keep adding more and more context to what you want it to say. So there's that too.

[@CandidProle]

I never signed up because they wanted my phone number. Looks like I'm not missing out on much. Thanks for the video!

[@xtdycxtfuv9353]

Do you want to use my account?

[@BRUXXUS]

Yep! As someone with extremely limited coding knowledge, any code I asked it to write, no matter how simple, was always broken. And since I don’t have much experience, I had no way to know how it was broken. It was fun using it to make recipes on a limited number of listed ingredients, though!

[@dandymcgee]

@Hugh Munguss "someone with coding knowledge" could mean the same thing as "someone with extremely limited coding knowledge". Your statement is ambiguous, please rephrase. :D

[@MESYETI]

i got it to write pong using C++ and SDL2 and it worked

[@hapashreguiem7460]

Bisqwit destroys a robot with fact and logic

[@dobrescusilviu4189]

"Simple graphics demo" - yeah right, simple. Awesome video btw.

[@triularity]

It's a good thing to know when an AI becomes sentient and tries to destroy humanity, it will probably get stuck trying to repeatedly use a "matter disintegrator". So as long as no human is foolish enough to tell it matter disintegrators don't exist, we should be fine. =)

[@pasikavecpruhovany7777]

I mostly agree with your conclusion, but when you know what the answer should look like but are not highly proficient in given language, AI can save quite some time. In C++ and VHDL it did terribly but for scripting and web stuff, I'm quite hopeful. It can give you 100 lines of all-right-documented code in a minute. IMO, chatGPT works much better for that purpose than copilot, even with the ux inconvenience. Examples: * Fitting measured data with sin function and plotting in polar coordinates (altrough did try to argue with me that sin is a polynomial) (Python) * Some simpler web-scraping tasks (eg. tables on wikipedia page, ) (Python) * Bulk sorting and renaming of files based on weird rules (in my case, contents of an image) - (Bash and Python) * Turning bullet-pinpoints paragraphs, padding out any kind of text. * Brainstorming ideas (For illustrations to put in a presentation about given abstract topic to make it less boring, brainstorming AI image generation prompts, etc.)

[@heinkel3460]

Will you test Google Bard?

[@Bisqwit]

I may, at some point. Recently I have been trying out various LLMs that can be run in a workstation locally, so that may also become a video topic.

[@Saw-qv3bl]

I believe Chat GPT is better when your questions can have varied answers, not one single single solution. That is why it can be good at designing games etc.

[@MrBrax]

"earth is round evidence"

[@Bisqwit]

Yes, I wanted to see what it says if I ask “what evidence is there to think that the Earth is not flat?”. It listed seven acceptable reasons, after which I asked it to conjure contraevidence, it essentially refused, while explaining reasons that some people might think it’s flat (because of the sheer size). It was not a particularly interesting discussion. I tried it out for a curiosity, because as an AI, it has absorbed a lot of information, both true and false, and I wanted to see whether it can play both sides.

[@Mycon]

Great video. I asked the model for completely new pangrams instead of the usual 'quick brown fox..', it failed miserably. But I do find it useful when programming with higher level languages using libraries like C# Winforms, to figure out approximately how to do things. I would never use it for C/C++ since there I usually write my code from scratch.

[@xtdycxtfuv9353]

I just learned what a pangram is.

[@otto_ueue]

The senior is back! 😃

[@antonwit425]

This is not about current video, but I would like to ask you for a video reference, you were i think in the past trying to shrink some files and you were bruteforcing it by changing the order of the files to see which combination produces smallest result. Would you be able to reference me to that video?

[@spider853]

ahaha, what was at 13:00 ? It's 2AM and I'm laughing my ass off

[@Tigrou7777]

It will be interesting to have another look at with ChatGPT 4.

[@MissNorington]

My best 14 minutes of 2023, by far! 😄

[@COMATRON.]

13:02 - nice burp mate

[@TimVerweij]

That was very entertaining, thanks! It was very similar to my own experiences with ChatGPT.

[@vicmac3513]

1. Millaiset 4 atk-oppituntia pidit lukiolaisille? Oliko tuskallista? 2. Onko uusi työjuhtasi aiheuttanut ongelmia tai saanut päivityksiä? Erityisesti Corsairin virtalähteen kuulumiset kiiinnostaa.

[@Bisqwit]

--Please don’t use machine translation. Without human oversight it creates eerie translations that are slightly off.-- Ilmeisesti et käyttänyt konekäännöstä. 1. Pidin tunnit yläasteella, en lukiolaisille. Käytiin Tie koodariksi -kurssia. Ihan ok, kaikki eivät vaan oikein jaksaneet keskittyä (valinnainen tietotekniikka, valittu siksi että pääsee helpolla ei siksi että asia juurikaan kiinnostaisi). 2. Uusi ja uusi… Corsairin virtalähteiden kanssa ei ole ollut mitään ongelmia. Seasonicin SSR-1000FX piti palauttaa kun se ei yltänyt lähellekään mainostettuun suorituskykyyn.

[@vicmac3513]

@@Bisqwit En käyttänyt konekäännöstä, luulin vain eträ 'high school' tarkoittaa lukiota eikä yläastetta. Hyvä, että vikaherkkä malli on kohdallasi toiminut. Tsemppiä kevään opiskeluhaasteisiin!

[@Bisqwit]

Kiitos! Ihmetyttää kyllä väite vikaherkästä, kun mä en muista, että olisi ollut yhtään ongelmaa Corsairin virtalähteiden kanssa 20+ vuoden aikana. Paitsi joskus muinoin saattoi tuulettimet ruveta ynisemään.

[@andrewdunbar828]

Much, much more in-depth and insightful exploration of ChatGPT than I've seen anywhere else. Thanks!

[@cheetobreath123]

I have been using chatgpt to kinda supplement studying. I recently applied to a job that requires the usage of C, but my C is not that good. So I would practice and study and if I got stuck would ask chat gpt to kinda fill in the gaps kinda like stack overflow. It’s an interesting tool but as shown in the video is very adamant about it’s responses even if it’s wrong. Also I wanna thank you in general cause you gave me courage to start to learn C and CPP (btw I got the job just waiting on the actual package details)

[@Terszel]

I think if they ever release a way to quickly finetune these more powerful versions of GPT, it will succeed as a very useful helper. You mentioned studying and that's a really great use case, if we can at least get to a point where ChatGPT can reason about the validity of its answers, it could be used as an interactive mentor, whether its learning a new language, teaching yourself a new subject, or helping you on something you're stuck on

[@plateoshrimp9685]

I wonder if GW-Basic code might be somewhat uncommon in it's training data vs. languages that are still in professional use.

[@Bisqwit]

Most definitely.

[@codegenprime3362]

Happy new year

[@Bisqwit]

Same to you!

[@neoqueto]

ChatGPT can't do algorithms very well, but it can do CRUD.

[@Bisqwit]

Do you mean I should test how it fares with SQL queries?

[@neoqueto]

@@Bisqwit it ain't the worst, at least with basic queries. JavaScript is its forte though. And of course it can do regex way better than most humans lol.

[@Mycon]

So why do you think it was able to answer that one question previously, but failing every time when you tried again? Did they update the model? Or do you think it has become so popular that your instance of it isn't given enough resources to come to a solution? (I don't know if this is plausible) It definitely displays too much confidence in its answers.

[@Bisqwit]

They do update the model from time to time. It can also be affected by minute changes in the input: in this case, slightly different variable names.

[@0xdeadbeef131]

13:02 YO what xd

[@alessi4249]

13:02 The burp was very jarring. Great video!

[@davidporterrealestate]

13:03 on a loop - yes i did

[@DoubleGlog]

Great Burp @13:00. Wonder what ChatGPT has to say about leaving it in editing?

[@NickNorton]

I think with this level of interrogation the ai will eventually require better Pay and Working Conditions. lol

[@Terszel]

GPT will always be near useless when it comes to calculated answers, whether its dates, math, etc. It needs some metaprocess on top to reason about the information it generates to error correct itself. Without that it'll always just be a nifty conversational toy

[@MMetalRain]

While ChatGPT conjures elaborate lies, I think there is also use case where you don't know much about the question domain, ChatGPT may guide you to some interesting direction. You still have to evaluate yourself if that direction is correct, but at least you might get unstuck.

[@Bisqwit]

Indeed. One day, I asked it to first reverse-transliterate and then translate a few Hebrew song names that I have (I teach Israeli folk dance and remember about 300 dances by heart, each of which is made to a particular song), and it actually gave really good and insightful translations that I was happy about.

[@stavinke]

My favorite part is 13:02

[@m4rt_]

I started re watching your videos a few hours ago... and suddenly I see this show up... what a coincidence.

[@josemariavildoza9920]

AI response to my existential questions: 13:02

[@ivomdcoliveira]

Hello Bisqwit, why did you stop making videos? What is your opinion of chatGPT as a programming helping hand? Surprised you did not test it with code.

[@tarilonte]

I've got this code on the 1st prompt: 10 DIM values(40) 20 FOR i = 1 TO 39 30 FOR j = i+1 TO 40 40 IF values(i) > values(j) THEN SWAP values(i), values(j) 50 NEXT j 60 NEXT i

[@Bisqwit]

Nice. That would be the correct way to implement the Exchangesort algorithm, which has O(n²) complexity for both best and worst case, unlike the more common (but still inefficient) Bubblesort algorithm which has O(n) best case.

[@boaridaaa_lol6241]

Mostly I use chat gpt for ask code that based documentation for example how to create window without border and close button. In win 32

[@Spikeypup]

Fascinating. Seems like a lot of effort to get a lot of wrong answers, but intriguing, nonetheless.

[@Bisqwit]

It does have some capability to acknowledge its own mistakes and correct them when asked, but equally often, it apologizes for the mistake and then proceeds to commit the same mistake again. It _does_ sometimes give right answers. It’s just that the questions I ask it are usually ones where it gets things wrong.

[@OggieSutrisna]

what linux are you using?

[@Bisqwit]

Debian

[@Lord_Engine]

Wintermute it is not. But it's pretty good at gluing things together wrong, I'll give it that.

[@JoLiKMC]

Oh, now you're just tossing that big brain of yours around like a blunt weapon. (; Glad to see a new video from you, Bisq! I hope you had a good year full of positivity, and I hope this year is even better for you! *Edit:* 13:02 - … that better have been a sound effect. ):

[@Bisqwit]

Thank you Joseph! Nope, wasn’t sfx.

[@JoLiKMC]

"No, wasn't sfx." I think Bubs from _Homestar Runner_ said it best, then… "Ugh, you _nasty!"_ (;

[@conkerconk3]

ChatGPT helped me learn about 3d rotation (Primarily how to rotate a point using raw math, quaternions or a matrix). It can definitely be a great learning tool If you're curious what it gave me, here is one of the answers I got on how to rotate a point along an axis in a specific angle (i modified it for my Vector class, originally it was using arrays) in java: double x, y, z; double cos = Mathcos(angle); double sin = Math.sin(angle); x = (cos + (-cos) * axis.x * axis.x) * this.x; x += ((-cos) * axis.x * axis.y - axis.z * sin) * this.y; x += ((-cos) * axis.x * axis.z + axis.y * sin) * this.z; y = ((-cos) * axis.x * axis.y + axis.z * sin) * this.x; y += (cos + (-cos) * axis.y * axis.y) * this.y; y += ((-cos) * axis.y * axis.z - axis.x * sin) * this.z; z = ((-cos) * axis.x * axis.z - axis.y * sin) * this.x; z += ((-cos) * axis.y * axis.z + axis.x * sin) * this.y; z += (cos + (-cos) * axis.z * axis.z) * this.z; return new Vec3d(x, y, z); I only tested this on the forward axis (0,0,1), on the vector (4,1,0). (4,1,0).rotate(Vec3d.FORWARD, toRadians(30)) results in (2.96, 2.87, 0) which makes sense (the rotation should be anti-clockwise as per the left hand rule of forward Z facing into the screen, opposed to the OpenGL default of out of the screen)

[@GameBacardi]

I would say, this AI give more headache to stackoverflow than user who ask help for problem.

[@AT-zr9tv]

I really enjoy your dry and geeky humor. You must be a fun friend to have.

[@ArnaudMEURET]

This is yet another terrifying proof that chatGPT can be the best lawyer in the world today ! 🤦🏻‍♀️

[@srccde]

5:27 Why do you calculate the base 10 logarithm? It doesn't state that in ChatGPT's answer. The natural logarithm of 0.2 is -1.60944, which is still off but close enough. The rest of the AI's answer is still... irritating.

[@Bisqwit]

Because in chemistry, most logarithms are done in base10 (not natural log), especially those related to pH calculations. The notation varies. Sometimes it is denoted by log, sometimes lg and sometimes log₁₀. Chemists often just write it as log. In contrast, the natural logarithm is sometimes denoted by log, sometimes ln. Chemists usually write it as ln, to make it distinct from the decimal logarithm.

[@srccde]

@@Bisqwit Ah okay, thanks for the clarification!

[@calm_compost]

And the AI apologised for his mistake...😂😂

[@muckymcfly]

@13:02 - Did the A.I. make you do that or was it a mistake 😂😂

[@smithwillnot]

FINALLY, someone with same experiences with chatGPT as me.

[@TurboBaldur]

I too have tried to ask it many questions, and the result is always 100% confident, but sadly 100% nonsense.

[@smithwillnot]

@@TurboBaldur I guess in that way it's kinda like talking to overconfident person IRL.

[@ScaryHutmanPictures]

I kinda miss your old programming videos with chiptune music in the background. Any plans to bring those back?

[@Bisqwit]

I kind of debated whether I want to add music on the background of this, but I was running short on time and some people have written comments in which they wish there was no music, so I published this like this.

[@ScaryHutmanPictures]

@@Bisqwit I see, and how about the programming/speedcoding aspect?

[@Ri0ee]

Asking a simple question, something like "is 40 greater than 60?", already proves chatgpt is no more than language model

[@34125867]

What's that supposed to prove or say? It just says "No, 40 is not greater than 60." for me

[@Bisqwit]

There do exist equally simple questions that the AI randomly fails, though. But it has inference skills that sometimes amaze me.

[@flippingbananas-1514]

13:02 Lol the burp out of no where

[@garryiglesias4074]

I have a huge problem with those AI... I mean, it's not LOCAL... So by trying it, by interacting with it, it LEARNS from you. So you give away human skills it didn't have so now it is a better SERVICE for its owner... You give away, for FREE knowledge, often high quality knowledge to a proprietary machine which will be used through a LICENSE (and a price, eventually)... I'm ok to have a LOCAL and PRIVATE assistant, I'm not ok to have a proprietary assistant which will resell my skills to "anybody who pays" another company. And I'm surprised by some of the proposition Visual Studio is now giving me on proprietary code owned by my company, MS is grabbing private code base with Visual Studio and even if it could be set off, first I have NO guarantee it really goes offline THEN as it's not the default, it basically pumps professional workers code. The Open Source problem is already one, as they broken licenses, BUT it's worst because they stealth protected code, and "high level" (highly skilled) programmers knowledge for free. People should not even "play" with that.

[@Bisqwit]

Currently the issue is that this kind of AI requires more resources to run than any private person can feasibly commit. You don’t need a beefy computer with a bleeding edge overclocked RTX 4090 card, you need like 50 of them. You could rent time from a GPU cloud service such as Lambda, but it won’t be local.

[@garryiglesias4074]

@@Bisqwit Yes, that's another problem but in IT people tend to put their head in the ground but we are at an age of resource (physical resources, fuel) decreasing availability. And a skilled programmer (/engineer) will cost a LOT less resources (some salad, water, and fruits), than the energy hole that those AI are. The illusion comes that 1 liter of oil cost "nothing" compared to the power it hosts. So now we live with thousands of mechanical slaves working for each of us. Which means those thousands slaves cost nothing compared to a qualified/skilled worker. It won't last forever. If we had to pay the "real price" of fossil resources, related to the whole Humanity History, we should pay more. Also, humans still have (true) creativity for them. Those AI are just models upon existing, documented knowledge... They won't invent Rugby while they're asked to play Football (real football -> soccer). Sorry for my clumsy English, I'm French.

[@Bisqwit]

I should also add that you _can_ run AI locally. I have done so several times since I last replied. However, the magnitude of AIs that you can run locally is defined by the RAM in your GPUs. With a RTX 3090 (that I have), you can run 7g parameter models relatively conveniently, but not much larger.

[@garryiglesias4074]

@@Bisqwit Yes, many solutions are emerging for local usage.