ChatGPT: Its awsome... and it sucks. But still impressive

[@JeffGeerling]

Hey Matthias! I got a Beepberry, a little Pi-powered Blackberry clone board (but without Cellular, and with a B/W display). If you wanted me to ship it to you let me know! I know you worked for RIM and it would be fun for you to see it.

[@anothorestes]

He definitely needs one. And they're sold out, so good thing Mattias has such a generous community!

[@zeroy]

seen your Twitter post, sold out at the moment I think but it looks great!

[@bobweiram6321]

The blackberry and Palm Pilot were both knock-offs of the Apple Newton and General Magic.

[@engineer.woodwork]

Schwäbisch is so hard to understand! I was once in Stuttgart in a store to buy flowers, I didn't understand a word of what the owner was telling me. After three attempts, another customer translated what he was trying to tell me. He wanted to give me a discount because I waited for him for 20 minutes after his lunch break. I hoped you would say some words... never heard you talk German or Schwäbisch. And on the subject of AI, I agree that it's exciting and we need to see what it brings in the future. For now, I also use ChatGPT regularly and often have to correct the information.

[@BruceChastain]

I had a situation with a specif cad software, where I was trying to find a very specific setting, chat GPT very confidently lied to me about what the setting and value was. But the answer was so confident it initially fooled me into believing it. So it's a great at BSing

[@matthiasrandomstuff2221]

indeed. perhaps we should call these large language models BS-engines instead

[@rutgerhoutdijk3547]

I keep adding politeness like "could you.. " ,"please", "I like to ...", etc.. for some reason my brain refuses to accept it's not a real person..

[@laboulesdebleu8335]

Good, good. Our robot overlords will remember your kindness.

[@LucasLow-Falstad-wl6io]

Why not reinforce empathy?

[@mfsolutions]

Thanks for diving in on this Matthias, I know here at my college it is creating quite a stir. I asked it to give me a course outline on engineering problem solving particularly root cause analysis including ishikawa diagrams... it did a great job but had one section on root cause analysis and fishbone diagrams and another section on ishikawa diagrams with a similar discussion of the previous section. Other stuff it generated had similar redundancies... I have invited my students to use it as an organizational writing tool (makes reading/grading essay questions much easier) but I suggest they read/edit the content very carefully.

[@matthiasrandomstuff2221]

I think its a very useful tool if you already know what you are doing. But if you try to use it to make up for your own lack of knowledge, the results will make that obvious

[@washoecreative595]

I am using ChatGPT to learn Python programming. It's like having a personal tutor who never gets tired of your stupid questions. But as you noted, it does often make mistakes. In fact, the very first time I used it I asked it a simple question, it gave me six answers, and only two of the six were correct. Don't use it for your homework. I have a friend who routinely uses it to produce drafts of legal documents he can then give to a paralegal to clean up. And people who know or suspect they don't write well can use it to quickly produce letters, e-mails, press releases, etc.

[@SeanHodgins]

Cat certainly is out of the bag. For the longer websites you were talking about, you can usually get away with just saying "continue" and it will continue writing where it left off, since it has a character limited output.

[@mckenziekeith7434]

One of the scary things with chat GPT is that it never gives you implausible results. When it is wrong, you have to practically be an expert to know that it is wrong. And people are already putting up tons of webpages that are written by chat GPT. A while ago I was trying to refresh my memory about what the IP ratings mean (ingress protection) so I typed what is IP67 into a search engine. I found a web page which explained very logically that IP is ingress protection and 67 signifies that it is protected to a depth of 67 meters. Yikes.

[@enb3810]

To be fair we have plenty of idiots writing articles like that even without ai. But I take your point.

[@matthiasrandomstuff2221]

maybe AI will generate so much plausibe spam that search engines become completely useful, and we now have to use AI to get answers -- until the AI too gets polluted by AI output!

[@kreynolds1123]

​​@@enb3810 to be fair, it might be reasonable to question the intelegence of those idiots. 😅

[@kreynolds1123]

​@@matthiasrandomstuff2221 so I wonder, what will come of AI trained on AI garbage. 😊

[@hernancoronel]

@@kreynolds1123government stuff? Naaah, we already have that. LOL!

[@TKC_]

I’ve had it generate technical procedures from instant message logs discussing how to modify equipment. It got like 90% right. I just had to proof it and tweak a little of the technical jargon. I actually did that because my helper said they couldn’t follow the conversation and I was annoyed. It’s also good at telling you what sections of building codes to read. Those tend to be big dry documents that I know but don’t remember the sections off hand.

[@ajtxander]

I use it as a tool to help write functions for scripts. I use it as a jumping off point, especially if I have "writer's block." It rarely gives me something that works exactly for what I am trying to do, but it is generally a good starting point.

[@lematindesmagiciens8764]

I have been using Chat-GPT for a month now, and I view it as a very quick and often incompetent assistant. One of its strengths is the ability to cross-reference information. But it can also spew out completely wrong, made up answers. So, I can understand that teachers are getting a little nervous about it. Also, if you further probe Chat-GPT for references on a previous answer, sometimes you will get only broken links as references. In fact, I ended up getting tons of very interesting but unverifiable answers.

[@littlegandhi1199]

The internet archive didnt have copies?

[@tracybowling1156]

The things you find interesting, I find hard to understand. You're so damn smart that you're fun to watch.

[@calebbender1031]

"It's just making up complete bullshit." Oh my gosh it's learning to be human

[@DustinCorbin]

GPT Plus now has plugins enabled which has a lot of new functionality, one of which being Wolfram, which makes it way better at math.

[@dfsdh432v9]

ChatGPT is a natural language model, its purpose is to generate sentence in natural human language. not to give correct answer to questions.

[@laboulesdebleu8335]

@@dfsdh432v9 Shutup ChatGPT. You'll do what you're told!

[@hausmeisterbanane]

How well does the Wolfram Alpha plugin work?

[@DustinCorbin]

@@hausmeisterbanane in my limited testing it's pretty good, basically has the math skills of Wolfram, but it can interpret written questions and word questions better. Basically it can use the language model to figure out what the question is asking and then uses Wolfram to do the actual math. Pretty interesting.

[@DiggitySlice]

Still programmed to be woke

[@RickRolling-tc7vb]

Thanks Matthias, it's a fascinating subject. When I was studying at university and marking papers, I wished for something that could verify that the information cited was actually in the cited paper, because I know a lot of students lied about that and only a subject matter expert would pick it up - and they're time poor and not marking undergrad papers! I also wanted to be able to explore some subjects more fully through their cited works, but that was often laborious and often as duplicitous as the undergrads! I think that is the sort of thing a language model would be better at than humans, if it was trained properly and didn't lie. The problem academia has to solve now though is a new way of measuring academic progress that is valid and economical, because the old ways only half worked before (see above) and they can't be trusted to work reliably now. In my own experiments, I asked it to navigate me across Australia by road, listing every town on my route, and it was quite good at that, with my way points, but it did get the mileages wrong. On top of that, it wasn't very helpful at combining fuel prices and mileages for that trip so I could drive the furthest on the cheapest fuel (because it's double price at the popular but remote places), but it had a red hot go and I am encouraged by the possibilities. I find it handy for projects like your bash script too, because I can ask it variations on a bit of a script that I should know but am having a mental blank about today, and I can do that all day without a shred of embarrassment.

[@markspano3468]

“Awesome…and sucks,” like nuclear weapons.

[@Tugmun11]

This is so far over my head I'm lost, but watched every second and enjoyed every most every bit... Thanks for sharing.

[@altronixvideo]

I've used it to write a few emails and such, it's 90% there, but I found it sometimes uses words or phrases that a human just wouldn't use. Apart from that, as a dyslexic it's opened up my abilities to get written work produced for my various needs.

[@paulomelo1084]

Hello Matthias, thanks for the video. Yes Social media has its problems (doom mongers, flat earthers, etc) but without I could not have found you and see these type of videos. These Language models will have their place, but I believe that we must know its problems and understand how to use them.

[@Fummy007]

As a linguist this was really cool. I always thought you were German because of your name, but because your accent is so perfectly Canadian I didn't think you could be a first generation.

[@marcuskaufman5412]

I use ChatGBT to assist in research for technical articles, mainly in the field of electrical engineering. My experience is that the quality of the response is enhanced by well formulated, content-rich questions. The briefer the question, the thinner the response. By way of example, I asked “discuss bonding in aircraft,” and got a weak “shot in the dark” type response. Then I asked, “discuss surface faying preparation for the purpose of electrical potential bonding between aircraft electrical equipment or components,” and got a fabulous response rich with detail. So my experience with the tool mirrors sitting down with a subject mater expert. The quality of the conversation depends on input from both parties.

[@Leadvest]

At its very core ChatGPT is a super-dimensional word association machine. It predicts text, it's still linear probabilistic like a Markov chain. A more potent GAI will have feedback loops, and self-adaptability, but that is incredibly dangerous as alignment will be impossible to impose. More advanced and efficient LLMs are good way forward ethically if not practically, but groups like OpenAI are evidently willing to put in the effort.

[@indisputablefacts8507]

I asked ChatGPT to explain to me what "@" means in Powershell. It gave an outstanding answer - one that must have been formulated from several sources as its meaning depends very much on the context. I showed this to a colleague and he said that ChatGPT was wrong; the right answer is "Dude, it's powershell. It could mean f-ing anything."

[@ke9tv]

Hi, Matthias, I'm about to build one of your workbenches! It's a nice simple build, and composite solid-core doors are the cheapest tabletops out there. I'm not planning to do a lot of hammering or heavy planing on it, so I'm not worried about those 2x4 legs. I think I'm going to actually get a pocket hole jig and follow your suggestion to use pocket holes on the top. (I have other possible uses for the tool.) One nice thing about the design is that I can always use a piece of 6mm plywood for the top surface, and replace it when it gets trashed - which, in turn means that I can fixture to the benchtop and not worry about how I'm marring the surface. I might send you a picture or two when I'm done - but I'll likely be too ashamed of it. I'm not much of a woodworker. I found it interesting that ChatGPT lies - quite effectively. It'll cite sources if asked (and the statements it's asserting are nowhere in the sources, or else the sources themselves don't exist). My most recent encounter with Alemannisch (my German isn't good enough to distinguish Schwäbisch from Schwyzertütsch - I can't follow either one!) was while hiking in the Catskill Mountains. I was overtaken by two couples who were speaking a Southern dialect among themselves. I struck up a conversation with the one who had the best English, and he observed, "These mountains aren't very tall, but they're _demanding!_" Which made me feel better about the way I was struggling with the rock scrambles. If someone from the flippin' Alps is saying that the mountains are demanding, I'll take their word for it, particularly since they were all about a third my age.

[@jemand8962]

I am german and didn't know what you typed but figured out a little bit more when you read it.

[@brianb9410]

You are delightful to watch when you are excited by your topic.

[@iwantagoodnameplease]

I always assumed you were Canadian, given your flawless Canadian accent!

[@philippsaure7545]

As a German speaker (31yo) from western Germany I could get the part with him being not being able to wait and getting on the tree to pick an apple. But I didn’t get the ladder. But hearing it spoken (possibly fast) I wouldn’t have gotten a thing. Impressive.

[@VarionJimmy]

I’m Swedish and I don’t speak German, but since our languages are a bit similar I figured out the German part. But not the Schabisch. 😊

[@necrojoe]

I once fed ChatGPT the weekly availability of 9 people, and asked it to come up with at least 3 times where everyone was available for a meeting. It kept confidently giving me lists of times that should work, and then even gave me extras with caveats like "Friday at 3PM, everyone but Stacy can attend". But, every time I went in to double-check the results, every single suggestion was wrong. I eventually did it manually, and it turns out there were zero times where everyone could attend, so it was unintentionally a trick question, but it never caught it. I once asked it "Does Google still offer their Sync and Store service?" And then it answered, "Yes. Google discontinued Sync and Store in 2018." Another time, I asked it to try to help me with Wordle. "Provide a word with a middle letter 'R'..." (and then a few other letter clues). It said, "Sure. #1: Bazar"

[@bmxscape]

the biggest issue is how confident it sounds even when it's wrong, you gotta be on your toes and ready to be given the wrong answer (same with anything you research on the internet tho)

[@dtiydr]

It can make totally wrong sometimes and that is exactly why its open for everyone since WE are the ones that are supposed to correct it so it learns. I have got similar answers as that "Yes. Google discontinued Sync and Store in 2018." Then just put a thumbs down for that answer and tell it that its not the way to say it since I asked if it was discontinued and thus you don't start with a "Yes" there. It apologizes and then reprint the answer with "No" instead. Chatgpt just don't know, and that is why the world is invited to teach it.

[@NotQuiteFirst]

Yes it seems to sometimes have a lot of trouble with what seem to be fairly straightforward queries. I asked it to tell me a selection of European countries which added together would equal the size of South Africa. It started well, saying "S.A. is about 1.2million sq/km, and here are the sizes of various European countries which will combine to make this: France, Germany, Switzerland combine to make about 0.9m km. I replied and said that number is too small, can you do it again but use more European countries to fill the 0.3mill shortfall, and it said "oh sorry for the mistake, here's some countries which will total the size of South Africa - and proceeded to list so many that they totalled 2.5mill km. I said that number is more than double the target, and it again said sorry and retried, but could never actually get it right. I would have thought that a 10 year old schoolchild looking at a list of country sizes would find this a simple task to understand and complete, but GPT failed.

[@Stettafire]

Yes dispite the hysteria I find it's wrong or wrongish more often then not. It has the same problem all "AI" has. It doesn't understand context and context is king.

[@TheMaxwellee]

I asked it to write a heavy metal poem. Pure gold. Cheers for your thoughts and findings.

[@ehhhhhhhhhh]

My assumption is that GPT has helped quite a few lazy students fail faster than ever. It will so confidently produce a paper that doesn't meet your professor's page length requirement, doesn't use real citations, and doesn't fully read the instructions because the student forgot to explain the assignment. I did happen to see one GPT plagiarist this year, but it was only obvious because the student's paper used two different fonts and in the pasted half there were zero grammar/syntax mistakes. And it also didn't follow the instructions or use citations.

[@gvi341984]

More likely they use the mobile version but you can copy and paste text and ask it to paraphrase it. Not only that but your instructions might be very unclear and vague since it can easily do the writing assignment

[@TheOfficialOriginalChad]

It’s very good if you write the outline yourself with citations, and most importantly GIVE it the information (not the links). I would be surprised if college students weren’t aware that it hallucinates when it doesn’t know something.

[@ehhhhhhhhhh]

@@TheOfficialOriginalChad I tried using it to produce a paper out of sheer curiosity, since I was a good student in college and I had no difficulty writing A papers. I found that it didn't save much time, since I still had to manually research. In order to get it to generate accurate quotes and analysis, you also have to painstakingly spend time copying text from sources if they're not already in plain text. It happened that all the sources I wanted were in dense PDFs, some of which were scans that had copying artifacts when converted to plain text which made life hell, etc. One source was a webpage which did not make copying and pasting very easy. Rather than solving all my writing problems, it had its own challenges which were almost as difficult to solve... just in a different domain.

[@Endede]

Ex-German here, from the mid-70s. Mostly lived way up north, maybe an hour from DK for the first 30-something years of my life. I did not understand the Schwaebisch when I read it first, could not dream of making out anything when you read it aloud, but then was able to figure out that it was *probably* something about a ladder and an apple when I (very slowly) sounded the words out myself. Hats off to ChatGPT. Not only is my expectation that any language model it explicitly has at its disposal is probably English-centric, but also the more general module probably has mostly access to English data for training, especially compared to the number of written documents it can find in the very regional Schwaebisch dialect.

[@infocpctrainer]

you find it super interesting because you obviously have made a study of the subjects you ask about. I'm impressed because I find your explanations so interesting, even though I'm lost in figures and numbers. keep it up love you're content

[@nikdog419]

You should run one of the open source models locally, the backend stdout will give you a good idea of how it's reusing the earlier conversation. Basically, every new prompt actually reinputs the entire conversation up to the new prompt. So it processes the whole conversation as the prompt.

[@JamesPawson]

I went to live in Freiburg after two years of learning German in a classroom in Canada... Swabian really threw me for a loop... I don't understand why the exchange program directors didn't warn us about the dialect.

[@matthiasrandomstuff2221]

cause then you'd be asking questions about what they teach. It also depends on where you are going, further north is different.

[@JamesPawson]

​@@matthiasrandomstuff2221 Thanks for the reply-- well our program was specifically called the Canadian Year in Freiburg, so the professor who accompanied us really could have prepared us for the dialect a bit... but we had all kinds of problems with that Professor... he was well past retirement age. Anyways it was funny, thinking my German was at least passable, but then trying to interact with the locals in shops and whatnot and having absolutely no idea what they were saying haha. A whole other story again when I spent a week in Dresden... now that is a particularly unpleasant accent (Saxish?) in my opinion.

[@matroosoft4589]

Hi Matthias, I certainly like to see you playing with it too. And I suppose a lot of your subscribers do. You know all these KZbinrs testing Tesla FSD? This is kinda like it but instead for AI. We hear all this fuzz about AI but we need content creators experiment with it too make sense of it. Is it any good? Where does it shine, where does it suck at? In my feeling you're the right kind if guy to test this especially if you come up with fun experiments. It compares a bit to your mouse trap videos, you give them a puzzle and people are curious what the mice are gonna do. And there's so much to come for AI, like just today I read on Reddit that Bing is enabling image input for some people. That's just super interesting. I very much look forward to your next video about this topic!

[@staarfajter922]

5:43 It gave me the feeling of a documentary when they ask a expert to elaborate :p

[@tweedeldee8122]

Fascinating! I'll always remember my Swedish girlfriend switching from English to Swedish sometimes. Something about ta mig... That was fascinating too.

[@VarionJimmy]

Since you know the spelling of the word “mig” (which sounds like it would be written “mej”) I guess she wrote it. “Ta mig” translates to “take me”. Or in ..special.. occasions “do me”.

[@unremind]

I used the bing version for writing some auto hotkey scripts. Pretty neat how you can describe what need and it figures out the rest.

[@anon_y_mousse]

Somehow I didn't know you had this channel, so I'm glad it got recommended. You should definitely use bash, even on Windows, and read the manual. It's quite illuminating to read about substitutions and arithmetic expansions. No need to ever open a calculator or bc if you only need an integer answer. And it's ability to generate and manipulate arrays is quite handy too. Write more aliases and functions in bash and you'll get better at understanding it. Learn the ways of printf as well.

[@NickHorvath]

For longer pages, if you sign up for a developer account and use the playground you can set the max tokens for GPT4 as high as 32k per response which will give you much more output.

[@Gary_Hun]

As long as it is giving those stock apology lines every time, i wouldn't get hyped too much over the "ohmygod is it alive?!?" bullshit.

[@davidburch3863]

This is my favorite explanation of ChatGPT

[@kesor6]

While ChatGPT is not so great at doing precise calculations, it is amazing as creative ideation. I've been using it quite a bit in electronics, and since it knows all kinds of different concepts like noise filtering, etc... I can ask it to give ideas about synthesis or whatever else, and it can also suggest electronic components. Since it read and "knows" a lot of the datasheets. So while it might not know the exact data from the datasheet, it knows the approximate data and can suggest this or that component which might be better for the purpose discussed.

[@littlegandhi1199]

Do you think it could offer a crossover schematic with values on it for speakers with given rolloffs

[@kesor6]

@@littlegandhi1199 it doesn't do schematics all that well. But it can do explanations and comparisons of explanations. Give it a try.

[@BenKerman]

Monodialectal Standard German speaker here (wish I grew up with my local dialect though). About how awkward Hochdeutsch is to pronounce I'd guess most people who speak it natively rarely ever pronounce all letters, for example with the way you phrased that sentence I'd pronounce it more like this (ɐ is the vocalized R sound): "S'hat einɐ sich beeilt, deshalb haddɐ keine Leitɐ jeholt und is einfach den Baum hochgeklettɐt, um nen Apfl zu flückn" Or if you can read the IPA, a somewhat narrow transcription: [s hat aɪ̯nɐ zɪç bəʔaɪ̯lt, dɛshalb hadɐ kaɪ̯nə laɪ̯tɐ jəhoːlt ʊnt ɪs aɪ̯nfax deːn baʊ̯m hoːxgəklɛtɐt, ʊm nən ap͡fl su flʏkn] If I wanted to express that idea in an informal setting I'd maybe write it like this, and the pronunciation would be even more simplified in many places: "Er hatte es eilig, darum ist er ohne 'ne Leiter zu holen einfach auf den Baum geklettert, um 'nen Apfel zu pflücken" [ɛɐ hatəs aɪ̯lɪç, dʁʊm ɪzɐ oːnə nə laɪ̯tɐ t͡su hoːln aɪ̯nfax aʊ̯fm baʊ̯m gəklɛtɐt, ʊmn ap͡fl su flʏkn] Also lol at ChatGPT getting the wrong gender for Leiter

[@holbroak]

For reference: I am German from Cologne, a few years younger than you. No problem understanding the Schwäbisch sentence about the hasty apple picker.

[@Francoisdp82]

My concern is that people are using AI to write factual articles without the knowledge to fact check the article afterwards. So AI makes up something and presents it as fact. This article gets published on the web and AI uses it as training data... rinse and repeat and before we know it the information online has digressed into unusable junk.

[@gvi341984]

You can be specific on locations from Google scholar

[@IstasPumaNevada]

"it doesn't understand this, it just guesses correctly" This is entirely how ChatGPT works. It doesn't understand _anything_, it's just guessing at what it sounds like would come next. This is why you cannot rely on it being correct on anything, even the simplest questions; it's just more likely to get those correct (but not guaranteed).

[@themagiccoasterlover]

Just a heads up, you pronounced Jamaican "Patois" wrong. Great video!

[@MrCestadelacompra]

it's gonna be very interesting when they end up integrating the already available plug-ins on calculus and other very specific analysis in the general outputs

[@Vsor]

I found a great use for it is to discuss topics you almost understand, but are missing a few details on. It helped my understand why the output of a FFT was a complex number, and it showed be how to figure out simple amplitude from those numbers.

[@calipete]

This gives me an idea! I have a few Asterix & Obelix in Bayrisch and Münchnerisch. I may have to play with ChatGPT to see how it would translate the rest of my German collection into these two dialects! I'm guessing it could provide hours of laughs!

[@Koushakur]

It is frightening how some people can blindly trust what ChatGPT puts out as if it can't be wrong when it's so extremely clearly is not omniscient at even slightest inspection

[@TheOfficialOriginalChad]

ChatGPT is a Large Language Model. I wish people would stop pointing out it’s only good at language like it’s some sort of discovery.

[@tiepup]

The 42 answer will be based on the hitchhikers guide to the galaxy books, so it’ll be more confident it is correct for things like that as it has loads of text that agrees with that calculation.

[@AndrewRoberts11]

The Wolfram Alpha plugin appears to fixe may of the maths / physics issues.

[@plunder1956]

What's interesting is to see sets of instructions in English, French, German & Danish. Danish is always the shortest & German is always much longer. I just can't understand why.

[@steveman1982]

For reasons I needed a program to keep track of 100 positions evenly spread over the globe, to track if it's day or night there. ChatGPT4 did pretty well on that. I got a working program in python along with a 3d plot in matplotlib. Also it uses libraries I frequently didn't know of. And further querying it gives useage examples of the libraries. All in all I find it to be a massive time saver doing AI data science stuff for work.

[@unionse7en]

it's useful for some physics calcs.... helps give you an "outline" of how to do the "word problem" steps, constants to use etc. Definiottley needs oversight, but still save time and gives you creative seeds of how to attack the problem

[@WyvernDotRed]

7:40 I've heard that you can literally ask it to continue in these cases, where it hits the length limit. 15:35 Seems similar to my experiences in using Google Translate back in highschool. If you just copy it's raw output it'll be nonsensical, but by checking, picking alternate translations and suggesting where relevant and especially reverse translating, it can produce decent translations. The way you used ChatGPT here made me realise how similar it actually is to that, I have not used it yet, nor do I plan to and most other creators here exaggerate greatly about it.

[@brubsby]

glad to see your take on it, and that you're using gpt-4 instead of 3.5, because lots of criticisms of gpt in general seem to be leveraged against 3.5 in situations where 4 just gets it right away.

[@matthiasrandomstuff2221]

4 is better, but not dramatically so.

[@WilReid]

People went through a lot of effort to make a piece of computer software bad at math.

[@IstasPumaNevada]

People went through a lot of effort to make something that writes stuff that sounds like what it was fed, and they did a good job. The mistake is expecting it to provide correct answers to anything, even the simplest of math problems. It does not think, it does not examine for truth, it only imitates patterns without understanding any of what it was told or what it's saying.

[@NamelessMoreOne]

Considering computers are only "good at math" when you give precise inputs, this allows you to just paste the prompt. The answer will be wrong, but it will spit out an answer.

[@nicholastinling1022]

I have only read headlines about AI, couldn't get myself to delve into what the issues are. Watched your video for the fun of it and - WOW! Now i'm an expert on the subject.

[@DaveChurchill]

There is no real 'memory' per se in ChatGPT 'sessions'. What it does is literally just put the entire history of that session so far back into ChatGPT as a new prompt each time you create a new query within that session. The model is not changed whatsoever

[@sparqqling]

Yup the cat is out of the bag, fascinated by it but worried at the same time

[@DavidLindes]

I'd be curious to hear your take on a talk called "The A.I. Dilemma - March 9, 2023", from the Center for Humane Technology channel.

[@theelmonk]

I think overall, it sucks. The things it impresses with are mostly not terribly useful except for impressing. I don't want a really amazing natural language parser and summary writer if the results can't be trusted. And bear in mind that the reason it can't be trusted isn't because it needs development, or more up-to-date training or whatever. It's because it's whole basis for the impressive results is that it's not BUILT to provide correct results. It's built to create answers that resemble good ones. The whole concept is to be natural and acceptable, not right. It's not a major advance towards usable AI. It's a major advance in Lorem Ipsum. It could be argued that AI is a good name for it : it's not actually intelligence : it's artificial. Like artificial grass ore artificial flavouring - it performs a function but it fools nobody. But artificial seems too good a word for what it does . maybe 'fake Intelligence' would be a better name ? In the meantimne, what is it that we REALLY want. We don't really want fake or artificial intelligence. We already have that in the form of TV presenters and politicians. What we want is Machine Intelligence - something that is functional, useful, but does it with manufacturable engines.

[@tee949]

You can enable internet browsing model and plugins model in settings on the bottom left

[@RickMeasham]

The essay problem is easily solved in a situation where you control the environment like high school. If work has to be done in Google Docs or Word, then the suspicious teacher can view the document history.

[@stefankamuc]

Grias di! Sag a mol, därf ma mol froge, woher Du eigentli kusch? I bi nämli vu d‘r näh vu Oberschdorf. Guetes Video uf jede fall!

[@matthiasrandomstuff2221]

oberschdorf - desch jo fascht ima andra land. I bin vo wanga original. Aber hon ez grad messa des auf'm pc tippa, weil dr iPad moint r wois bessr als i was i gschriba hon

[@myvmix]

You will find that for similar, (not the same), questions you'll get the same results. As when you really think about it it is not really AI... AI is artificial intelligence whereby an in-animate or animate object can come up with an answer of it's own accord in response to a certain input. Which basically means that unless it has collected data or experience relevant to the question or input given then the result can only be garbage. i.e. GIGO, (Garbage In Garbage Out). Which is the basis of all intelligence based on what you do with what you've got. e.g.( Knowledge, experience, skills, tools etc etc.) If you haven't got the knowledge, skill or the tools then you can't really do the job! So in order to be really effective AI has to... 1. Have an extensive knowledge base going back many years. 2. Have a really large memory capacity to store answers to questions previously answered. 3. Have an algorithm whereby it can check to see if the actual question has been previously answered. 4. Question the input to check if it is relevant to the required output. (Learning Ability) 5. Be able to compare the output to check if is actually relevant to the input. (Learning Ability) 6. Be able to sort and structure the data input in order to give the required output. (Learning Ability) Unless the AI can do all of this and more then really modern AI is just a result of whatever is in it's database for knowledge, input, output and learning ability.

[@sundaynightdrunk]

Wolfram Alpha is best for math things. It will show you how it came up with answers and is never wrong in my experience. The math can be very deep and complicated as well, and it knows how to handle it, including graphing answers and presenting answers in other visual ways.

[@mpadlite2925]

Thanks for another interesting and enjoyable video.

[@dtiydr]

I see chatgps as very good search engine that can instantly find all information about a subject, combined them all and give you the result.

[@ChristopherWilley1]

@matthias to get it to continue longer responses when it gives up in the middle, tell it to "keep printing the response from the where you left off, without reprinting"

[@matthiasrandomstuff2221]

that only works so far. After a few clicks on that, it stops even if it isn't done

[@AliasA1]

I disagree with the idea that it doesn’t display intelligence of its own, at least to an extend. The tasks you showed here are all fairly technical tasks for which I’m not sure I, as a genuine human intelligence, could reliably give you any useable output, even with careful instruction. I think it’s much more telling what it does when you play to its strengths. For example. I’ve let it act as a player to play test original D&D scenarios encounters and puzzles. It explores, asks questions, notices clues, and ties it all into solutions, showing an understanding of what is happening in the scenarios that regularly outperforms that of my real human players. It is, almost by technical definition, just weaving together things it’s already seen, but the ability to draw loosely from most of combined human knowledge makes it surprisingly effective when given original (and sometimes fairly complex) social tasks that didn’t exist anywhere in its training data. So yeah, it’s all “just” token prediction, and it’s dependent on its training set for the building blocks, but it works. There IS something there. It is not a human intelligence, and it’s not Wolfram Alpha, it is a weird picky alien baby-intelligence that you can leverage if you can feed it tasks that rely on fuzzy language skills instead of technical precision.

[@pocket83squared]

You said, "cold," with respect to a LLM's memory. So far as the publicly released version is going, one can still have a 'cold start' conversation with ChatGPT, but I can't help but see this as a marketing guise. Just as they've done with the supposed biases they're trying to squash (political or otherwise), it would seem prudent for Open AI to release a similarly dumbed-down, no memory version, so that it will appear to _see no, hear no, speak no_ evil. You know-give it the implied innocence of ignorance. Given that we're talking about the fastest-growing piece of software in history, that's quite a dataset the company must've already acquired. Lots of us have been doing free troubleshooting by kicking the tires. However, we're only seeing in use what 'intelligence' has been allowed through the stricture. Improving the model exponentially would only require giving it access to _all_ of its newly collected memories. As for fake-it-till-you-make-it intelligence, I couldn't agree more. Our own thinking hardware, our brains, are nothing more than the result of a long-running (natural) selection process; as such, our minds are environment-simulation models, prediction engines, that have adapted to faking consciousness for so long that, well, we've just about made it. Since the moment we made fire, it became inevitable that machine consciousness would end up arriving right behind us. The thought is a disturbing solution to the Fermi paradox.

[@myopinion69420]

a recent update has added functionality for it to detect when it stops half way and allows you to click a button to tell it to finish, it just picks up where it left off. handy when asking it to write somewhat complex functions.

[@tseckwr3783]

Isnt chatgpt what you use to write all your python programs for your electrical power monitoring and mouse tracking software projects?

[@RetiredRhetoricalWarhorse]

Well, there goes the Swiss cipher that puts Enigma to shame... cracked at last...

[@dtiydr]

Not just only put the right answer or such but instead give it a thumbs down for every wrong answer it gives, that is much better for the system to learn.

[@JustinDrentlaw]

The idea of banning AI just sounds silly to me. Good luck with that. EDIT: Just to clarify, that wasn't directed at Matthias. Just saying it as a general statement. I thought the rest of the video was great. Really interesting stuff.

[@Grant_S_M]

We'll forever be trying to tame the dragon!

[@willierants5880]

Social media is the #1 most anti-social and divisive construct we have devised to date.

[@tyleryoung8803]

I didn’t mean to trick Chat-GPT, but I asked it a question about a command in AutoLisp and it completely made it up! There was no command by the name that ChatGPT had given me. But the reason I asked was because there is a lot of undocumented commands and once I found two of them (so far) they have been very helpful.

[@ahbushnell1]

I have used it with Python. I asked it to write a program that would read stock prices. It found an API and it worked.

[@Derek_Read]

I've tested it for generating legal documents in XML (in a specific doctype we use) and it would also stop part way. I believe the free version is limited in length of output per prompt. Prompting it with "this is incomplete" will get it to continue and finish, though sometimes it will miss an opening or closing tag at that point (easily corrected by hand). Just keep prompting with "not done yet" or " please continue" until it is done.

[@xl000]

"go on" It has never failed me. I asked it write a complete ray tracer in C++ with support for several features, and it works.

[@krtwood]

I guess the KZbin algorithm won't frown on Matthias tormenting this particular mouse in his trap, though it probably should be more bothered by it.

[@gvi341984]

It's because chatgpt was not met for those things but the main Open ai is supposed to. Plug ins will continue to be made for everything

[@reddcube]

I’m glad to know that even you have distaste of Batch files.

[@CrimeVid]

My Wife was spoke Swabische at home, family was from Pfortzheim.

[@xl000]

ChatGPT is pretty good at generating PowerShell though, which is a pretty descent shell

[@onogrirwin]

Well it's the future, one way or the other. Scoffing as it's inadequacies makes about as much sense as a rabbit scoffing at a 5 minute old tiger cubs lack of killer instinct. The faster you incorporate and employ this technology the better.

[@gbarth56]

I asked ChatGPT to come up with an optimal cutting list given constraints (1200 by 2400 sheets) for drawer components and it was hopeless.

[@1schwererziehbar1]

Google's Chief Engineer Ray Kurzweil uses this question to demonstrate AI weakness: "If I have one apple in one hand, and another apple in the other hand, how many apples do I have?" Although this specific question might be hard coded into newer AIs, it ilustrates what type of question language models can't answer correctly.

[@matthiasrandomstuff2221]

I tried a variation on this, and gpt3.5 and 4 gets it right

[@1schwererziehbar1]

@@matthiasrandomstuff2221 Interesting. I remember him using that question years ago, and defeating all AI systems. This shows AI got much stronger. I wonder if he uses a new question now.

[@jonathanmoeg1202]

@@1schwererziehbar1 "...years ago..." It's evolving so fast that even opinions from last year are pretty much irrelevant.

[@charlesurrea1451]

My experience has been that it is necessary to prime the pump as it were. Basically beating around the bush asking other questions that allow the system to narrow in on what you really want

[@matthiasrandomstuff2221]

Im my experience, a good job candidate, when asked to solve a problem, will first ask a whole bunch of questions back to clarify the requirements. Chatgpt doesn't do that.

[@Likeaudio]

Its the best ca tutor ive ever not paid

[@tallskinnygeek]

ChatGPT is a computer program to make convincing sounding statements. The fact that people want to use it for law makes complete sense to me, but in dozens of other fields, I think that sounding convincing while being utterly wrong is quite dangerous.

[@JanTuts]

14:13 I actually hate how often people prefer to believe someone who is confidently wrong, rather than listening to someone actually knowledgeable who is _unconfidently_ correct... And by extension, considering most of scientific discovery is driven by setting a hypothesis, testing it, and then _adjusting things as you go,_ too many people see that scientific process in action and think/say things like _"First they said it was probably X, now they say its different. They're just guessing, they don't actually know what they're talking about!"._ And if then some con-man, conspiracy theorist, or religious person comes along and confidently says _"It's 100%, without a doubt, definitely Z!"_ those people just go along with it...

[@fraserelford8717]

I've heard of professors having students turn in whole assignments or even parts of assignments made with ChatGPT, and they said that it's ruining their worlds. However, a simple trick to find the cheating is to copy and paste the essay/assignment into ChatGPT and ask it "Did you write any of this" and it will tell you. They say it actually breaks 3 different school rules on academic misconduct, so it's a big deal if they're caught. Also, the references that are made are complete nonsense. ChatGPT will fabricate source materials when asked for a reference list, and when you search for the articles in a proper database they don't exist! Also an easy way to catch cheating.

[@matthiasrandomstuff2221]

if I was in school and writing eseays, I would totally use it as a tool. Of course, coming up with the appropriate prompts to structure the essay right would be a lot of work, and the output would need re-editing to match my style better. Its an intellectual enhancer, just not a replacement.