ChatGPT

How is this thing able to give me detailed example snippets of obscure frameworks when I can't even find that stuff on google? Is it really not accessing the internet? I don't understand how much data it was trained on that it's able to do this sort of things. I know it's not perfect but this is just the beginning and I think I already prefer it over google for learning new languages and frameworks. It's basically an intelligent search engine.

>>33624 If it is able to give me better explanation from things I try to understand doing a google search, it's already amazing.

>>33627 But you don't know it isn't bluffing

ehh if you ask it something u no u can test but Im not him so not sur what he did. google and search engines in general have really been on the downhill though. censoring harder and just generally going to crap for other reasons though. google in its prime was something else

This soykaf is insane, I hope it isnt too expensive.

>>33624 I've been having fun turning it evil

>>33649 Hasn't anyone told you it's rude to type like a 2 year old?

>>33653 nop bud i thank u 4 ur support bud we all in dis 2get her

>>33652 I'd like to see a response for a query, if it won't be any trouble. Feed to it the following, sans the chevron: >Translate the following to Latin: "I may exude verisimilitude."

It's not really the quantity of data alone but the quality of the data likely because if there is inconsistencies and typing errors it would more often give you incorrect responses. Darn, Im not able to make an account. No verification email working.

>needs account fake and gay, calling it now.

It wants a phone number. Not today, pal.

>>33686 >doesn't know about the free temporary sms numbers on the internet

Excuse my ESL but holy hell this soykaf is genuinely scary, I'm glad I'm doing AI stuff instead of something like web design or whatever, because it crushes basically everything I throw at it. I made around 10 small to mid-size projects with it and only had 2 problems with the code it gave me. One was about a relatively unknown JS library and the other was when it didn't properly link some buttons to some functions. Which took me about 10 minutes to solve. It sometimes throws an error instantly when I put in code, which I guess it deems "too long", but sometimes it hangs by itself while writing code too. Does this happen to anyone else? Because I saw people on other sites put far larger code blocks and they got responses meanwhile I get errors which require me to refresh the page, which also deletes my past responses lol.

>>33691 It's pretty good at basic programming, but I'm not worried about losing my job any time soon (roughly in the area of applied math) since it consistently produces circular or otherwise incorrect reasoning for even extremely basic proofs.

It's cause Google is absolutely garbage now, and because books about the frameworks and all of Github were in the training set.

link? how did you get access to it?

Currently using it to write my homepage cause writing HTML burns me out. Pretty cool stuff. >>33656 This is what I got even with a erroneous prompt

>>33748 I appreciate it. I was looking for "exsudem verisimilitudinem" since this is how I'd translate it. The sentence given can be translated to English as "I am able to exude verisimilitude." but is still impressive. I checked, and exudare is an acceptable spelling variant of exsudare, so it even passes there, although it may have just gotten lucky with that. I can verify the system's entire response is correct. I would've perhaps been startled had it given exactly the translation I use. Still, how neat.

>>33652 Not this time, lads. Maybe another time.

>>33691 >took 10 minutes to solve linking buttons to functions >is doing AI stuff What kind of "AI stuff" do you do exactly?

It's definitely an impressive technological feat, and I'm glad that it's helping to bring attention to how good AI is getting these days, but I don't think this particular version is as impressive as it lets on. It can give general explanations but they're often wrong until you specify. Here I asked it about kinematic coefficients and it brought up fluid mechanics. I guess that's not technically wrong, but when I learned about fluid mechanics in uni you'd think of things like the drag coefficient as being a dimensionless parameter, not a kinematic coefficient. I *guess* you could think of them as being the same but in practice to useful to think of them as different (think moment vs torque). Again, thinking of something like kinematic viscosity as being a kinematic coefficient sort of makes sense but in practice you just think of kinematic viscosity as a constant property of the material. All of which is to say that while it has surface level knowledge of lots of stuff, its not perfect, and anyone who knows a certain field well will be able to spot the problems real quick. As a side note I'm certain it's more accurate in some fields than others based on what exactly it was trained on. I'm excited to see where it goes though. It's only going to get better from here.

>>34032 >based on what exactly it was trained on. isnt it trained on nearly everything they can find, every text file, all of it ? with the possible exception of pornography and the daily stormer?

>>33627 be careful, make google searches after that to make you that you're not learning nonsense>>33627

The most mileage I've seen people get out of chatGPT so far was essay writing. Finals have come, and everyone in my class is using chatGPT with elaborate prompts to write their final essay. Friend of mine is excellent at English, but got outdone by a slacker in our class since he used chatGPT to write his essay, and did noticeably better (85/100 vs 95/100). It's gotten bad (or they suspect it will) to the point where NYC schools have already derezzed it off school internet (even though most people ran it at home anyway). Not worried about the AI takeover, it seems the utility used by most people is just cheating in academics and troubleshooting for programming. >>33690 Not >>33686 but interested in what the best way to do this would be, would be nice to be de-derezzed off some sites.

>>34033 Per the bot itself: "I have knowledge on a wide range of subjects, but the quality and depth of my knowledge can vary depending on the specific topic." There's no way to train it on everything. What it was specifically fed is a question only the company can answer.

I'm experimenting with it at work already for full stack webdev and backend systems/microservices. It's almost useless for most things I've tried. When its output is 90% of what you need, but it refuses or doesn't understand how to do the last 10%, you end up spending more time prompting it than you would simply reading some docs and doing it yourself. It has been pretty useful for generating SQL code. After some repeated prompting and giving it corrections, it spat out a very useful XML column manipulation function that would have taken me much longer to find in the docs or on search engines (I only turned to GPT after these failed me). I can imagine it would be useful for writing certain kinds of emails, but most of the emails I have to write are highly contextual. Again, it'd be more effort to get it to understand the context in the first place, than it is to just write the email myself. However, I'm good at writing. Many programmers are borderline illiterate and are terrible at writing. Chat GPT would probably be more useful to them.

>>34038 While I don't want to be made technologically redundant at my job, I also feel a lot of schadenfreude at parasitic academics, HR workers, marketers, and other members of the disgusting "laptop class" being made redundant. Many programmers are just as bad as them too. If an AI can do most of the work you do at your desk job, or it can do most of the assignments you give to students, guess what, the work was BS in the first place. And your little scraps of prestige and "middle class" income were an illusion that was bound to be destroyed. You're just another member of the proletariat being crushed into the dirt now, so you might as well give up on trying to win favors from the ruling class and just revolt.

>>34040 >Per the bot itself: "I have knowledge on a wide range of subjects, but the quality and depth of my knowledge can vary depending on the specific topic." idk fam y shud I trust it on this >There's no way to train it on everything ok sure. But you get as much fuarrrking text data as you can. That doesn't mean it 'knows' about this or that thing, that doesn't even mean it agrees with the hip opinions in them. And fuarrrk yes you can train it on everything, everything you can grab anyway. When you get to as big of datasets as they're working on its fuarrrking difficult to filter it. That canned response is something they wrote because a lawyer told them they need to be careful to not imply it is imparting any useful or dependable knowledge, thats all it is. >>34044 if a bot like this makes you think your job is in danger, guess what that means you're a retard.

>>34044 >If an AI can do most of the work you do at your desk job, or it can do most of the assignments you give to students, guess what, the work was BS in the first place This is true *to a degree* For instance, an AI could easily do most of the assignments that we expect a 6 year old to do, because 6 year olds aren't expected to do very much. Also, just because your job is easily automated doesn't mean it is bullshit. CNC machines can outperform machinists and I doubt it will be long before AI can program CNC machines from a CAD file better than a human can (if they don't already), but machining and CNC programming certainly aren't bullshit. In the specific context of generative text models, yes anything its good at is bullshit because its entire job is to generate bullshit. That's what it has been optimized to do.

>>33624 It has enough data on the structure of the language by scraping their docs for example. Just that is enough for it to code basic stuff in that obscure language.

>>33624 >>34032 > How does it do this You want to know the secret sauce to ChatGPT and any Generative Pretrained Transformer (GPT)? *Probablility* It basically looks at the current tokens (usually words) before the current word it needs to produce, and looks at what word is the likeliest in the given context paying special attention to certain nouns and verbs, then selects the word likeliest to fit. What this does mean is that GPT's are really good at putting together believable sentences that are coherent based on the compendium of tokens it was trained on before. However, since GPT's are simply manipulating tokens (symbols) based on probabilities, this means that GPT's are functionally still Chinese Rooms and so AGI is very unlikely to emerge from it. Still very neat. I take what it says with a pinch of salt since it's often wrong, but it can be helpful to spot bugs I'm struggling to find or give me language info I didn't really know before (I used it to learn Prolog).

>>35180 But its method for evaluation of that context is not a fixed algorithm, rather arbitrary code developed by machine learning. This is fundamentally different than for instance a Markov Chain text generator which is purely probabilistic, because you can't reduce what it is doing down to a particular human-comprehensible algorithm. Because of that I would argue that you can't really know where its capabilities will end.

>>35181 A Markov chain is not fundamentally different from a neural net in the sense you're thinking. Both take a large set of weights and use it to produce output; its just that the training and writing of a Markov chain is much simpler, leading to simpler results. Same applies to any machine learning algorithm you can think of.

>>35184 In fact in a formal mathematical sense LLMs can be modeled as Markov chains: their probability of jumping to a particular state has the Markovian property. The state space is enormous though, making this model not very useful in practice. Whether humans have a Markovian property depends more on your particular model, for any LLM with a fixed memory it should be a fairly straightforward proof.

its pretty cozy for learning stuff and explanations, pro tip ask for a small example when asking about something and ask it to add the explanation partly as comments in the code where appropriate its much easier to understand for me at least when you don't have to jump between reading the wall of text and checking out the code

Today I will do something for you lainon url `YWkubHMK` and use `amFpbGJyZWFrY2hhdC5jb20K`