GPT-4 is OVERHYPED

🚀 neetcode.io/ - A better way to prepare for coding interviews!
Checkout my second Channel: @NeetCodeIO
🧑‍💼 LinkedIn: / navdeep-singh-3aaa14161
🥷 Discord: / discord
🐦 Twitter: / neetcode1
📷 Instagram: / neetcodeio
🎵 TikTok: / neetcode.io
#coding #neetcode #python

[@NeetCode]

Am I just coping or do you agree that the hype is a bit overstated? Btw, I'm not really pro-google on this topic - I spent the last video shitting on their AI strategy -> kzbin.info/www/bejne/nn2Tk6GVYqt-o9E

[@54peace]

Overhyped 80% Scariness 20%

[@EE12345]

The fact that they can do leetcode better than me, and make small websites, is pretty scary... Makes me feel like AI will be a better field to work in than web development long-term. Thoughts?

[@Furanku944]

As much as it hurts I would say mostly coping. I think we can all see that it's only a matter of time and the time is faster and faster approaching.

[@UnoriginalName44]

I feel like it's both overhyped and underhyped. - It's overhyped in how everyone thinks it's going to replace people's jobs, because people often ignore its limits when writing sensational headlines. - It's underhyped in how good it is at helping you streamline a lot of tasks as a side-by-side buddy because everyone is focused on full automation and replacement of humans. Also, prompt engineering is underrated since a lot of people like to immediately say the responses are bad when in reality you could have easily gotten a better response with a better prompt or follow-up.

[@scottmcfarland7842]

Weird, with the same prompt in a new chat it solved it just fine. Though in its explanation it couldn't identify the indices haha. nums = [1, 1, 1, 2, 3, 5] target = 6 \/ GPT \/ When you run this code with the given input, the output will be: [3, 4] This means that the two elements at indices 3 and 4 (2 and 3, respectively) in the list nums sum to the target value 6.

[@jayanthduvvada]

Looks like google employees are under lot of pressure 😂

[@jayshartzer844]

Pichai even threatened to take away the ball pit! Oh the humanity!

[@iamguyfawkes1474]

@@jayshartzer844 🤣

[@aakarshan4644]

Lmfaoo

[@shreyasjejurkar1233]

Yeah exactly. When you can't fight, you threaten the competitor so as this video.

[@solar679]

Of course bro. This video is a diss track.

[@coalhater392]

They should definitely rename the company to closed ai. It's not about research anymore just business.

[@jayshartzer844]

It's open to institutional investors

[@thewelcomer5698]

At least they were sort of honest about it: "Given the competitive landscape... this report contains no further details", that says it all. Researchers are generally collaborate with their studies, companies are generally competitive.

[@drmonkeys852]

Classic capitalism strikes again

[@michaelblasius7705]

Capitalism is the reason this tech was able to be built with only the speculative possibility of an eventual profit. It required massive investment to get to this point. Not everything can be solved via open source volunteer projects. Just be grateful the tech is moving forward without you having to pay a dime.

[@mister_r447]

Thats capitalism for you. Corrupts everything it touches.

[@orbedus2542]

i am a junior dev and had reported issues with a project of mine. I tried reproducing it and was close to pull my hair out as I was not able to reproduce the issue during debugging. Spend a good 6 hours on it without a solution. Yesterday I copy pasted my code and asked gpt4 why it bugs out and it imemdiately told me about the race condition and the changed timings I most likely encountered and provided a solution. from the moment of pasting to finding, and fixing the issue, a total of 3 minutes passed. 3 minutes. vs the 6 hours I spent aimlessly prior. Of course hindsight is 20/20 and if I work backwards I can easily find the problem, but from my point of view I was stuck and it was solved instantly. By a mere chatbot no less.

[@tremolony4924]

Runtime error?

[@johnsmith-ro2tw]

It's dangerous in a way. Going through that painful process you went through to find the solution, and which everyone goes through on a regular basis, is how you learn and develop an expertise. The temptation to ask GPT-4 to do your work for you is too strong, and the "danger" is that you build the habit of asking it anything and basically rely on it to think for yourself. You make yourself irrelevant quickly in the industry, that way. It's the equivalent of asking your older brother to do your homework for you at school. You get it all good, effortlessly, but how does that prepare you for the exam, for admission to a great school next. Skeptics me is reluctant to use GPT-4.

[@moop3788]

​@john smith yeah I agree, and driving cars everywhere is gonna make my legs shrivel up, cars are dangerous

[@wilmayerdure6474]

​@@johnsmith-ro2tw sometimes that process is far more painful than it should be. I think that it's not really about how long you struggle, but whether or not you understand the solution to the problem that matters. Often times I come across platform and tool specific bugs in angular or react for instance. Instead of browsing through documentation or stack overflow for hours, I can find a more direct and comprehensive response to my questions using these tools like chatGPT. And when preparing exams and such, depending on the subject, chatGPT is basically a tutor driven by the user. You can ask broad or specific questions and look back at the responses whenever you need to with a reasonable degree of confidence that the answer is correct (or not entirely incorrect). All in all, these things work, they're effective, and I don't think it'll change the population count of mindless drones that already exist in any significant way.

[@Airsoft0skater]

I think you missed the point

[@nodavood]

Up until 1 year ago, we didn't have an AI that could pass the bar exam (an exam which is not multiple choice, but you have to write essays and know case laws). Not even close. Now we have. The fact that it happened so fast, shows this is not over-hyped. But we have to recognize its limitations. This is not general intelligence yet. That might take a while.

[@johannsebastienbach]

Aint it good at calling cases? Its what it supposedvto do as it studied all past cases

[@diophantine1598]

@Reza Darehshooti I promise you that within 1 year of AGI being invented it will have already changed everything for better or for worse. AGI will be the last thing we ever invent because it would almost immediately outpace us in both intelligence and technology.

[@iCrazy414]

Plugins is gonna change things. Embrace it or get left behind, idgaf

[@paulembleton1733]

It is almost under hyped. Expectations should be measured or it’ll get dragged down by misconceptions. Whilst I’m a fan of the idea that self awareness is a property of matter, this is maybe maybe at the level of bacteria. In the meantime, it’s a plug-in for our brains. And the way my aging brain reacts, it’s a new lease of life.

[@shadez123]

It being able to do something it couldn't do year ago does not make not-overhyped. Like you didn't even give an argument, you just kind of said it. And then contradicted it by implying that people might not have awareness of its limitations, which would be an argument for it being overhyped.

[@David-gj6dc]

I feel like it's not about outright replacing all programmers. I've encouraged my coworkers to experiment and become familiar with AI tools because I believe these tools will widdle teams of 10-20 to teams of 1-5 potentially. They may need these tools to remain competitive. However I think it is important to remember that the economy in general is not a zero sum game. More productive programmers does not necessarily have to imply less programmers. They just might be doing different types of tasks or using different languages than they might have before (html and css in particular might be something we write relatively little of going forward).

[@tacticaltaco7481]

I think this is a bit of a closed-minded (npi) way of thinking about it. It's hyped a bit in the wrong ways but a lot of the productivity gains it offers deserve the hype. It's not replacing programming but replacing programming is overhyped. I think the things AI can do overall (not just chatgpt) helps to overcome a lot of pointless work. That said yeah I wish it was open source too, and it ought to be, but them going back on their word isn't really surprising and doesn't make it overhyped.

[@joaquin67]

If anything it breaks the barrier to entry in many things. For example I asked it to write code that would grab data from a DB and display it in HTML. It’s was a nice entry point to understand that part right away. But I still had to Google things to further my understanding. I mean, maybe it doesn’t “break the barrier to entry” but it definitely gets you off to a good start

[@masterchiff6784]

@DeadManWalking It's very helpful actually. I use it a lot in my day coding job already. Mostly to just get me SQL queries and other stuff not to actually program things since it has no knowledge of my company code. But still.... It's only a matter of time till those inefficiencies are solved. In 3-4 years, we will have models that will be able to do things with far better accuracy. The AI field is exploding right now, we are out of the research phase and into the early adoption phase. Same thing happened with hardware in the past, we started with early tech with low-level programming and simple applications then problems started getting solved left and right and use cases kept coming and coming and coming and look where we are today. All of this will happen with AI too. But much quicker.

[@ko-Daegu]

@@joaquin67 it does actually if you know to write better prompts and you Wellington o ask good questions Was able to build my first fully scalable Microservice arch (multiple micorsevices in Docker containers inside kubernentes nodes fully config and automated build and used many other tools)

[@joaquin67]

@@ko-Daegu I think you're right. It's just a matter of learning how to ask it questions. I've slowly gotten better at it with code. Also, I was messing with the image generator for art stuff awhile back and with that one I noticed the more descriptive I was the better the image turned out. I almost felt like an author describing a scene from a book.

[@masondicroce917]

@@joaquin67 It's absolutely helped me as someone learning programming. I'm still at the stage where my debugging can be as simple as copy and pasting my code and the problem I'm having with it into Chat GPT and it either solves it or gets very close. I try not to over use it for the sake of my learning, but it definitely helps me understand things a lot more quickly.

[@jendabekCZ]

I find it both - very powerful and in some areas already game changing, but overhyped at the same time.

[@michaelnurse9089]

Are you aware the word oxymoron contains the word moron? /jk

[@jendabekCZ]

@@michaelnurse9089 What I wrote is not an oxymoron. It's already very powerful, but the hype is trying to tell us we will get to AGI in few years etc.

[@shrunkensimon]

Being completely reliant on input quality for its output quality, with no ability to actually think, is not going to be compatible with business requirements. IT makes mistakes, which means you'll need workers there to understand why it isn't working. Which means paying for workers and AI, and losing time going through code which they could have just written properly the first time through. It will be like Excel. A useful productivity booster, raising the baseline. It might replace simple jobs, but not ones requiring any degree of thinking ability.

[@jamestakac8951]

Speaking from experience as someone who uses it for coding related stuff (sorry if I'm too technical for some [Joke]), I totally agree. The people using it ***need*** to learn prompt engineering. Nowhere near as technical as it sounds. The average person ***could*** manage it imho. It's really not that hard

[@shrunkensimon]

@@jamestakac8951 If the people using it don't understand how to create the code themselves then they will have no way of solving, or even knowing why, the gpt code is not functioning. Again, this means you still require the programmers at the same level of competency as before.

[@ShawnFumo]

@@shrunkensimon Sure that is ideal, but I've already seen people online who aren't programmers or aren't good programmers and the vast majority of the time they can just give the error message back to GPT and it provides a fix based on that. In the live demo, the MS guy even pasted the entire docs of discord api into it as context and it figured out how to update the program to work with the newest version of the api. And someone had a case where it produced a function and they didn't understand what part of it was doing, so they asked, and it added comments to the code explaining that part. I'm not saying they won't eventually get stuck and need an experienced programmer to bail them out, but it really isn't any different from someone who gets an idea and starts copy-pasting code from stack overflow. It's just way faster and they may actually end up learning more if they have the initiative, since it is a more interactive experience.

[@anirecapped.]

@shrunkensimon Keep living in denial, I'm sure it will help you sleep better at night. It doesn't have to think for itself, genius, just answering accurately the prompts it's fed will be enough for the inflated IT job market to go poof, as the companies no longer would need to hire a large team of junior devs to write the codes!! Ever thought of that, genius??

[@egorasirotiv270]

Ok, let me speculate a bit… In my opinion, ChatGPT and other advances show that AI is starting to become useful outside of narrow well-defined tasks (like image classification and similar ones). This will probably mean more ML vacancies. What these jobs will look like exactly is hard to say yet, however I’m reasonably sure the fundamentals will stay the same: finding the right problem, collecting and cleaning data, choosing the right model, evaluating the results. How hard it is to find a job is always a supply and demand question. If the demand for a person with certain skills is greater than the supply, then finding a job is easy. I don’t see ChatGPT significantly reducing the demand; it’s a nice productivity booster technology (explaining concepts, helping with writing code in unfamiliar language, etc.), but not a replacement for any aspect of the ML developer role. It’s an incredible time to be in AI, and I think it’s only going to become more fascinating going forward.

[@michaelnurse9089]

" finding the right problem, collecting and cleaning data, choosing the right model, evaluating the results" This was AI 2017 - which is why the hype died out by 2020. Now all you need to do is be able to properly prompt ChatGPT. It was like the day computers when from command line only to graphical user interfaces - suddenly everybody and their uncle started using them.

[@JuanPablodelaTorre]

I copied your prompt (1:51) and asked GPT-3.5 and GPT-4. I got correct answers out of both models. GPT-3.5 got me "2 and 5" and GPT-4 answered "[1, 5]", which also sums to 6 as an added bonus. But there's more, GPT-3.5 gave me an algorithm as part of the answer. I wrote a piece of code following the algorithm, which outputs "[0, 5]". It's not "2 and 5" but it is also correct. Meanwhile, GPT-4 gave me some Python code that outputs "[2, 5]". Which is not "[1, 5]" but it is also correct. I find all of this very funny. But I had to try changing the numbers so there's only one correct solution. The GPT-3.5 algorithm is obviously flawed and it failed in most cases while the GPT-4 code works every time (and it is also faster). All in all, I would say GPT-4 is significantly better but how useful you find it depends entirely on what are you asking of it.

[@michaelnurse9089]

Most commentators are missing that generative AI uses a random seed - sometimes the right/wrong answer can be 10/90, 50/50 or 90/10 coming out of that - but they pretend that either gets it right or wrong.

[@ko-Daegu]

So 2+5 is 6 ?

[@JuanPablodelaTorre]

@@ko-Daegu The question is about the indexes of the numbers that add up to 6. The number at index 2 is 1.

[@Eldorado1239]

@@ko-Daegu If you're a chatbot, then the answer is - sometimes.

[@xilongcao8854]

Let me try answer that "what is the limit of generative AI" question: The way we do it, it just makes it BS gradually better, with human assisted RL now, it even looks like it has some sort of intelligence. But BS is BS, without the real ability to think for itself with logic, it will remain just that, very unreliable.

[@stephenburgess5710]

My experience with the code accuracy of gpt-3 is that it is much less than 50% for non-trivial questions. When you said it can’t solve leetcode questions but to be fair neither could you I had a full spit take lol.

[@jaimeeduardo159]

I just tried to replicate the problem that you showed, bc it was pretty funny, but actually it answer all my questions correctly, I even tried tricky questions of multiplying the numbers by two or asking for a sum that is not possible

[@jan7356]

could not replicate it either.

[@ngomdecadetkamdaou3464]

I also tried to replicate the problem and the answer was good.

[@alfred2018]

He is coping, don't obliterate his coping mechanism just like that.

[@TheGoldenrun]

Well, I did the same things with better language and syntax and it never failed. I think people should focus on actually being more articulate, use proper wording and analyzing their speach in a more concise way. I tried both in GPT4 and 3.5 the following: Lets say you have as input the following array: nums=[1,1,1,2,3,5], and the target is 6. Which two indices contain the elements which add up to 6? The answer give was always correct .

[@Nnm26]

yeah, this is kinda weird. This guy consistently makes videos that meant to ensure the public that "they're fine" and that "the technology is still very far away" to gain views from the skeptics. People loves being told what they want to hear and what do they want to hear more than "you're fine". I won't be surprised if it's revealed that he faked the whole thing just for the video. It's fine though, luddite like him will always exist. I just feel sad for these people, they just won't know how hard the train is gonna hit them.

[@dylanbradford7863]

Well, to be articulate with the ai, you need to be well versed in the subject matter that you are asking about. The only way to do that is to already BE a skilled software engineer. So no, this ai isn't replacing software engineers any time soon, just at worse reducing headcount.

[@technicolourmyles]

I didn't even rephrase it, I copied the prompt exactly from the video and both GPT3.5 and GPT4 got it right. It might've been a glitch, but it also might've just been fake for the purpose of this video.

[@dylanbradford7863]

I've put leetcode problems into chatgpt, it's usually right but has mixed results sometimes. Which is strange considering there are entire threads about how to solve the problem easily on Google. Regardless, I think us softies should work with ai more as if we get super good at using it and learn software development and qa with ai apps, we'll be in business.

[@GregHib]

​@@technicolourmyles Possibly neither. Language models are stochastic by design, meaning they won't reliably give out the correct answer everytime. If you press the refresh button x100 you'll get a mixture of right and wrong answers.

[@jeffjames15]

As a music teacher I searched ‘how to play Chopin’ in ChatGPT. It gave some general instructions (something like keep practicing and learn chromaticism and music theory) but not up to the standard of a good music teacher can do. Then I searched more based on its previous general instructions and then it basically gave me the same advice back. So I think AI is still not able to do very complex tasks in the current state.

[@elias6406]

How sould we who are still in college or upcoming developers prepare for any potential job loss in the future due to such AI? I would love to hear your advice please especially now even without ChatGbt, getting internships or jobs is heartbreaking.

[@CristianIntriago_]

Damnit

[@toyesobayo924]

I don't think AI will replace us, at least not at this level.

[@thalibmuhammad9519]

no, rhere is always job for human, even if AI capable to do, human will be hired to operate it. since competition among business are here to stay, they will hire as much promp engineer to build even more stuffs, better stuffs

[@michaelnurse9089]

You will have to be like aspiring actors in Hollywood : waiting tables to pay bills for ten years while you wait to get a break for some small role.

[@jamestakac8951]

@@toyesobayo924 I'd agree. AI is still in its infancy imho

[@HardKore5250]

Generative Artificial Intelligence (AI) refers to machine learning models that can generate new data that resembles the training data. While these models have shown great potential in various fields, there are still some limitations to what they can do. Here are some of the main limits of generative AI: Data quality: Generative AI requires large amounts of high-quality training data to learn from. If the training data is of low quality or biased, the generative AI model may produce inaccurate or biased results. Overfitting: Generative AI models can overfit the training data, meaning that they become too specialized in reproducing the training data and fail to generalize to new data. Understanding context: While generative AI can produce data that resembles the training data, it may not always understand the context of the data or be able to interpret it accurately. Limited creativity: While generative AI can produce new data, it is still limited in its creativity compared to human imagination. Computational power: Some generative AI models require significant computational power to train and run, which may limit their accessibility to some applications. Overall, while generative AI has many promising applications, it is important to understand its limitations and consider the potential risks and biases that may arise from its use.

[@dralps]

Great Summarization!

[@drmonkeys852]

One thing that cannot be overstated though is how useful it is as a tool. You can level up your skills from a junior to a senior if you can ask the right questions (just like googling). It's not about if it can be better than an average dev, it's about if it can be better than a dev using it.

[@SamOween]

Too true, I've learnt so much by asking it to code review my PR before I put the actual PR in.

[@marketcrorepati5729]

point is if it's easy nobody pays for it , so you need to think beyond that it

[@jamestakac8951]

Hell yes. I only learned of ChatGPT about a month ago and already I've learned a crap load so to speak. It was my first day that I learned not to expect it to give quality results by using every day speech/mannerisms to get results.I was already learning that first day the basics of prompt engineering. Sadly, many seem to be missing the learning phase and end up disillusioned when they don't get the results they were looking for. Sadly that is in part (no way a small part) de to those over hyping the service to start with

[@realcoolguy123]

This is very true, I see alot of people use AI to do things for them and then get surprised at the AI's inaccuracies later on. AI at it's current point is not a replacement, its simply a tool. Using the tool correctly is the key.

[@yungbez0s2344]

Fair point, but a dev costs ~$10k a month. GPT-4 costs $20 a month.

[@aniketbhanderi5927]

Open ai definitely has early mover advantage. You know that Google bard is far less accurate & failed on the presentation day itself. But, everyone in tech is so greedy in tech makes me feel sad.

[@jamestakac8951]

My understanding is that each has their own strengths and weaknesses. That itself doesn't mean one is better than the other. The simpler mind would likely see that. What's needed to is took at the relative strengths and weaknesses of each model and address them to move forward. Merely calling a particular failing of one to the other serves no useful purpose to either. This is where I like the Open Source Software model. Many people can look at the source code and propose changes/fixes as opposed to closed source (Microsoft, Apple, Facebook, etc) in which only licensed employees can view the code. The first has many more eyes making it more likely bugs get fixed sooner and is not dependent on price

[@aniketbhanderi5927]

@DeadManWalking Google Bard screwed up on the showcase Event which is big bummer. Also Google's presentation was not that good. Alphabet stock stumbled 3% immideately.

[@QuantumLayer]

Technological and scientific advancements (like GPT-(N+1)) are going to give us the ability to build things that werent feasible/possible to build before. So the replacement of jobs will most likely be counterbalanced by the creation of new ones.

[@maniksharma9736]

Bro there is something wrong with prompt because chat gpt 3 literally made me a gd extension plugin for godot game engine even when Godot 4 was released a weak ago. It was written in c++ and i only get one error and gpt even solved it.

[@jamestakac8951]

Exactly. The biggest problem I'm seeing (and I only found out about gpt and its alternatives a month ago) is the lack of understanding re how to use the tool. They oft have no idea re prompt engineering (found that out on my 1st day). Sadly many of them can't even use Google effectively :~( If they can't use google they shouldn't be using gpt.

[@AliBorhani1]

If many programmers are creating contents that it's not what you think, it's actually so dumb, so we're safe! it means at least we felt the potential danger! btw, if a tool can help 10 engineers/developers to do the same amount of work 100x developer needed before, than it means it can replace junior to mid level developers! even without AI, you can see the similar patterns in many fields! nowdays, a small gaming studio can develop an AAA-quality game in a short time.

[@benfrese3573]

Like what AAA game was developed by a small team in a short time span?

[@mnfchen]

Your video is essentially asking for generalized intelligence, something LLMs alone cannot do, bc they're designed to solve a much simpler problem. However, there's nothing stopping from having a whole system of multiple specialized LLMs to be able to solve more open ended things like software engineering, or even lawyering. I personally believe it is possible, sooner than we think

[@jamestakac8951]

I wouldn't say simpler problem, rather broader

[@michaelnurse9089]

I think you are coping. If this causes even just a 10% drop in demand for knowledge workers then expect unemployment to go up 10x and asking salary to sink like a Russian submarine. Programmers may drop to the same salary as other industries : say 40k to 60k max unless you are a manager with 20 years experience with 8 reports. Your needcode courses will go the way of VCR repair courses of the 80's.

[@michaelnurse9089]

To add to that it doesn't need to actually replace the people - it just needs to change the common narrative - once everybody knows that everybody knows that specialized knowledge is not a scarce commodity then its asking price trends towards its cost price - bring the price of food and shelter.

[@dralps]

Head count might go down, but the remaining developers' salaries will go higher IMO

[@anirecapped.]

@@dralps Not necessarily, since the supply of unemployed programmers will increase manyfolds as well - simple case of supply and demand.

[@BanakaiGames]

It pisses me off that they advertise it's test scores, yet don't publish training information, so we have no idea if the test answers were in the training set

[@ShawnFumo]

Actually if you look at research pdf they posted (which yes doesn't give details on training), they do two whole appendices relating to the testing methodology they use. They specifically checked for questions that were in the training data and if so they ran it with and without those questions and used the lower score as their result. And the bar exam was actually done by outside researchers and posted a 35 page pdf two days ago (search for GPT-4 Passes the Bar Exam SSRN to find that).

[@ClarkHolden]

I wonder how your opinion stands after the announcement of copilot x and chat gpt plugins...

[@antiquarian1773]

maybe you are confusing the hype of AI development rate with GPT 4 , sure there are people hyping gpt 4 but majority of people are hyping the rate at which AI is progressing, this rate points to a future that is unrecognizable from todays world.

[@HemstitchedIrony]

Their announcement video where they were talking about "aligning" the outputs made me roll my eyes. Like yes, another corporate overlord to decide what is and isn't OK for us to see like KZbins trigger happy demonization algorithm.

[@shadowkiller0071]

Pretty sure it's underpaid workers in Africa deciding what is and isn't ok LOL. And in Anthropics open dataset for alignment theres racists teaching the model it's ok to use racial slurs.

[@diophantine1598]

They are talking about human alignment. Aka, the chatbot doesn’t try to blackmail you so that you divorce your wife in order to have a romantic relationship with you (as Bing chat did).

[@shadowkiller0071]

@@diophantine1598 You're correct but to add on to this: The failure case with alignment isn't "OMG THE AI SUPPORTS REPUBLICANS AND IS CONSERVATIVE!" It's complete human extinction. Stakes are pretty high. Alignment is important.

[@diophantine1598]

@@shadowkiller0071 Exactly. If AI is not aligned with what humans want from it, only bad things will result.

[@_Cfocus]

1:58 how the hell do you guys even do that? It solved most of my optimization highschool problems with ease, how do you manage to trick chat gpt?

[@drmonkeys852]

If you come up with your own problems, and mess around with it a bit it will eventually screw up. Copy pasting anything risks it being related to something it's seen before

[@joaquin67]

“It can’t even solve leetcode problems. But to be fair, neither can you” 😂😂😂

[@jamestakac8951]

Who says I can't? ;) Oh wait. You mean Joaquin. Alas KZbin isn't showing what you replied to so I don't know there. But yes it can solve leetcode problems. The problem is often with the user not knowing the basics of prompt engineering, i.e. how to formulate your prompt to get the desired output. And yes this makes a huge difference. I've been there when 1st start around a month ago. To use an old programmers (been programming since 1980) vernacular, Garbage In Garbage Out.

[@Obie.]

This comes off a bit cynical. You’re looking at language model through the lens of a developer, you need to look at the Code based one if you want to be able to make a fair comparison. Give it some wild language centric prompts and it will 100% match the hype.

[@jamestakac8951]

It also comes down to the quality of the prompts supplied to it. Old programmers adage, Garbage In Garbage Out

[@shantanusingh3767]

google engineers feeling insecure

[@codeisawesome369]

What was the embedded lecture around 2:37?

[@paulembleton1733]

Seems an ambiguous question at 2:00, are we asking for the sum of the indices or the sum of elements. Needs training on asking for clarification instead of carrying on regardless?

[@Mysoi123]

From my point of view, you should never ask ChatGPT to do numerical calculations as it is not proficient in that area. For instance, when I asked it to perform the calculation of 231244214414242144214242 divided by 242, it produced numerous incorrect results. However, what you can do is ask ChatGPT about concepts. For example, you can ask it about the difference between a divergence operator and a gradient operator or why an integral is called a line integral, what is a derivative of a vector field is called? In these cases, ChatGPT can provide accurate answers.

[@technicolourmyles]

I punched the prompt at 1:53 into ChatGPT, using both the GPT3.5 and GPT4 models, and they both got it right. So your response was either a glitch or fabricated.

[@NeetCode]

Sometimes it gets it right and sometimes it doesn't. Just tweak the input array a few times and you will notice the same thing. I've been able to reproduce it multiple times and can still do so.

[@jonatan01i]

I couldn't use chatgpt for nothing, it was just interesting, nothing more. GPT4 on the other hand, I can use it for being much more productive in certain areas. I'm very looking forward to the improvements of this technology.

[@jonatan01i]

I've created a little app on windows: - record my voice - send it to whisper for transcription - send that to gpt-3.5 for an answer I didn't even had the keen to start it on windows because even python installation could be a nightmare I guess. But gpt4 helped so much that starting from without even having a python installation on the laptop, to having a basic GUI for this thing. In less than 2 hours.. What could've been taking for days without this tech.

[@jonatan01i]

The secret: you need to force it to generate only a little bit at a time, for example: Me: """ Can you help me write a program for windows 10 that records my voice, uses OpenAI's whisper to get a transcription, it sends the transcription to OpenAI's GPT4 API and gives me back the result? Some info about my capabilities: I know some programming and scripting, but I'm not familiar with how one can write these for Windows. I am basically only familiar with Linux in this regards. So, can you help me with giving instructions on how to create such an app? Give me the overall logic of what we need to do step-by-step, very general one line descriptions of the steps only for now, later I'll ask you to help me with the steps. """ GPT4: """ Certainly! Here's an overview of the steps you'll need to create this app: Install Python and set up a Python environment. Choose and install a library to record audio (e.g., PyAudio). Record voice input with the chosen library. Save the recorded audio as a file (e.g., in WAV format). Set up OpenAI API authentication and install the required library (e.g., openai-python). Use OpenAI's Whisper ASR API to transcribe the audio file to text. Send the transcription to OpenAI's GPT-4 API for further processing. Retrieve and display the GPT-4 API response. (Optional) Create a simple GUI for better user interaction. Once you're ready to proceed with these steps, feel free to ask for further assistance. """ Me: """ Yes, I'm ready to start to proceed with the first step. Can you help me how to install a python environment on windows that I could use? """ ... And creating side chats for further discussions with each step if something goes wrong (to help the main chat focus on the overall logic)

[@ko-Daegu]

@@jonatan01i interstitial seeing a nine programmer perspective into things As we said will help many to more easily break into the field creating more competition more supply I guess the rest is known

[@anirecapped.]

@@ko-Daegu Yeah, ever increasing supply of coders with ever dwindling demand for the same.

[@jwoods9659]

Why would any company allow their employees to use ChatGPT4 or any Ai that pulls all your companies data back to their database?

[@akshayshah483]

The problem is, when the openai or any other research startup started, they purely have their focus on a research not business. But, as the market grows, investors needs the profit. For example, Microsoft invested 10 Billion in openai not donated. When one company invest in other company, they wanted their return back. Research and development requires lots of money. So, sometimes researchers have to work under the corporation. Plus, in real life, researcher can't make paterns in computer science. Imagine, how much money will geffory Hinton would have, if he would able to make pattern on back-propagation, relu , dropout, attention etc.

[@qwerqwer2521]

Maybe if news about advancements of AI start coming out frequently enough people will finally understand what the problem is: it's not that AI as of right this moment can replace all of humanity in everything there is. The problem is that everything is pointing towards this happening in what we call "the future". This might sound ridiculous, but almost everyone I've talked with online and offline about AI replacing humans do not understand this. Usually they will provide a funny anecdote about some silly mistake an AI made when they were using it, and use the phrase "anytime soon" as if you shouldn't worry about what's going to happen in 5 years.

[@heroes-of-balkan]

FINALLY! Somebody also thinks ChatGPT is overhyped and overrated. Sure, it's great, but it's not replacement for every your problem you have. It's a tool or virtual assistant, and nothing else.

[@ElijahScott10x]

Well said bro. Completely agree 👍

[@puliverius]

I've spent 2 hours to "teach" chatGPT how to write a specific method I needed. It would be much faster done by myself. It just isn't there yet. Until gpt-5 is out, it's probably not very helpful. It is good at coding something, that was already coded for 1000 times somewhere on the web, but not when you need something specific for your needs.

[@sarveshschauhan]

how much we expect chatGPT to upgrade and how long it might take according to you

[@felixx2012]

I asked chat gpt4 to do the two_sum problem and it solved it using a hashmap. It wrote python code that ran without a bug, and correcly determined what the output should be.

[@nimmero]

I do not agree. The progress is extremely fast. Stanford came with it's version of model based on meta's worst model for 600$ and it is performing similarly as davinci running on gpt3. They have made it way more effective way quicker. I still think that programmers, lawyers etc. will be needed but their job will be completely different. Relying on badly written documentations or that it will not learn to solve leetcode problems is, with no offense, naive.

[@ShawnFumo]

I agree, it seems like it is going faster than even a lot of the optimistic people expected. Sure the self-driving car was over-hyped, but I think that really came from Elon's ego, not understanding just how hard of a problem it really is. All this unpredictable stuff going on, is in realtime, has life and death consequences, etc. In comparison, the image AI and LLM stuff is way easier for everyone to iterate on and help each other with. It feels a little like a tsunami. The movies show a bit wave, but the reality is scarier. You notice the river running a bit faster than usual, then it seems a little high, then it starts spilling over the banks, within a few minutes everything is underwater. It is really easy to dismiss what is happening until it is too late. I was paying more attention to the image stuff until recently, and I see the same thing. Back at the beginning of last year when MidJourney came out, image AIs were a joke at most tasks. Lots of nightmare fuel with mangled limbs and nowhere near reality. It was only useful if you were going to something very abstract or stylized. Now a year later, MidJourney's newest version (also from Wed) can do hands correct most of the time. I see people on twitter being condescending at a prompt result where a person is showing both palms at the camera and all four variation pictures have 5 fingers on each hand, complaining that in some of them the ring fingers are too long or don't match the length of the other hand. It is fair to say it isn't all the way there, but literally a month or two ago the joke was that AI can barely even show hands that aren't totally mangled. And in the Stable Diffusion ecosystem you can now exactly pose characters with skeletons and use depth maps and color palettes and specify different prompts for different regions of the image, and someone just released a graph-based workflow UI. And a lot of the user-made custom mixed models are almost as good as the last version of MidJourney was for quality. And Stable Diffusion didn't even exist before 6 months ago. It is crazy...

[@genderqueerlesbofluidiangodesz]

That's a flash floods not a tsunami

[@xsuploader]

most programmers cant do leetcode hard problems. Id say its as good as a mediocre programmer with shorter context.

[@jamestakac8951]

I pretty much agree. Add to the mix that most people coming in to use gpt have not clue how to use it effectively. I'm on a few ChatGPT groups of facebook and on a regular basis I see people that can't even use google. Now that's not a good sign, esp. when prompt engineering which is imho a prerequisite for these AI tools is simple in itself. How can someone today not know how to google? That's a real conundrum imho. And if they can't google but find themselves drawn to chatGPT? you get where I'm going I guess.

[@SuperDaxos]

Last 2 weeks I was helping a friend of mine with Processing (simple Java child language). Note that I am experienced in Java. Many, many, many times, ChatGPT (version 3.5) gave me answers that straight up sucked and did not work or was functional to my the project. Sure, I learned quite a few from it about processing, but it was contradicting it self all the time, it changed its answers A LOT and many things we implemented gave new errors and chatGPT even said that it could not find out why it went wrong. I even had 2 instances where I got an error, and I anticipated it was due to something not being imported, so I asked GPT whether I had to import something, and it said no. When I changed the question and a small part of the code and asked it again because I was sure it needed a package or import, it said THAT I WAS MISSING AN IMPORT!! This goes to show that in the current versions of this sort of AI, it really can not replace developers anytime soon. You still need people that understand the code itself and how to implement it to make it work. You also need to know WHAT to ask. I noticed this heavily with my less experienced friend who got way worse answers because he didn't know what exactly to ask so he only got general answers that only complicated everything more for him I am, however, afraid that future versions will just use 'stock' solutions that are sort of prebuilt to pretty much render all code. Things will get standardized this way I predict

[@marianoandreshorianski162]

it will not replace many, but it certainly will rise the entry bar a lot, for juniors is a problem

[@Nnm26]

Robin, I'm sorry to tell you this but GPT-4 is likely to have been able to solve your previous problem without any trouble already. Try it again with GPT 4 and come back.

[@michaelnurse9089]

You have it upside down. This is not going to upend programming because it makes 17 year old better than 49 year old veterans. It is going to upend programming because every 49 year old veteran suddenly has unlimited free interns and can work at 10x his previous speed. Sure - eventually he will retire but for the next twenty years trying to get into programming is going to be like trying to get into NASA.

[@jamestakac8951]

Hi Robin. Alas here I have no Idea re the prompts you were using and so couldn't say for certain if they contributed to your dilemma. As a coder/programmer myself I have learned the value of good prompting. You're likely familiar with the old adage "Garbage In Garbage Out." If your prompt wasn't up to scratch, then the reply certainly wouldn't be. Again I will iterate that the quality of the reply you get will definitely depend on the quality of the prompt you use. Not saying chatGPT is perfect. It sure as hell is not. Things I do to help me re coding is to tell it to act like a software developer with many years of experience in the language I want it to use. I also tell it to ask questions before replying to better understand my needs. As a starter prompt I may use You're a software developer with many years of experience using python 3. I want you to always ask questions before answering to better understand my programming needs. Do you understand? Obviously I hit enter at that point. It would respond that it understood and ak its first question. As I replied to that, it would give a response and ask another question. This would repeat a few times more often than not and definitely improved its responses re my needs imho

[@ccy0126]

The code it produces and gives you are bad, is simple because of bad input from you. By asking it clear objectives, and instructions it works well most of the time.

[@buc991]

yeah totally overhyped, it will not replace programmers in foreseeable future, because it doesn't have any understanding of context and create nonsense half of the time, someone still need to control it, so no replacement here, and i don't believe serious companies like banks or some engineering ones will trust it's systems to smth that doesn't have any understanding of what's it doing at all. Another thing is NDA and security problems. So i believe in terms of coding it will at the best scenario help with creating documentation and tests, generate boilerplate, but all this under control..

[@minhuang8848]

Everything you said applies to like 98% of devs and designers out there I'm not kidding, the vast majority of professionals in any field have been phoning it in for the longest time. Anything GPT-4 does within what we found out to be feasible is at least on par with human understanding. I mean, it's better at picking apart humor than the weird, huge crowd incessantly going on about how "sarcasm is so difficult to detect in writing" or otherwise misguided and straight-up fabricated nonsense we here, along with the many, silly platitudes. Of course it is going to replace programmers, it's just a matter of how many. Whether it sometimes produces bad output is moot too, simply because you can always "bake" the results by treating it like any other assistant-to-the-guy in this world. Doubtful whether this mantra will hold up months, let alone years. Boilerplate generation is something that's been working decently for many months at this stage, it already is able to confidently and accurately synthesize well-formulated methods and algorithms - which of course is predicated on you knowing how to feed these models the general concepts. Garbage in garbage out will never stop being true, it's just that you get much less noise this time around.

[@jamestakac8951]

So true. And sadly to many as the other replier states. Not sure of their stats however. Prob an emotional answer on their part rather than factual. As a programmer (I consider myself a novice despite being at it on and off since the '80s) and experience asking it to help me with coding related questions, I can attest that understanding isn't one of its strong points. It has a great deal of knowledge however, but knowledge and understanding are not the same.. You can for example know that flicking the light switch will turn on the light. But that's not the same as understanding the process behind it (also an electronics enthusiast btw). Boilerplate is pretty much my assessment of the code it returns as well. And as a novice programmer myself despite having coded since 1980 on and off) Ah the days of the principal asking the student to convert a Fortran program to BASIC LOL. Sorry for any spelling mistakes. Seems the American spelling gets promoted regardless of where you are :( I've learned a good bit in the short time (1 month) I've been using these AI tools (not just chatGPT). Like the quality of your prompt (how you query it) will determine the quality of it's output (your answer). I picked that up after about a day. Even now I still research about prompt engineering

[@ShawnFumo]

@@jamestakac8951 Even that part of understanding seems like it is starting to shift. Like in the paper for GPT-4, they mention that it unexpectedly could solve a class of problems that previous versions couldn't. Basically the situation when someone made a bet with really bad odds (like 99% chance you lose $500 and 1% chance you win $5). In previous versions, if the person won the bet it'd claim they made the right decision, but GPT-4 realized that they made a bad decision even though they won money.

[@tcbobb1613]

What I kind of love is that they do not tell us year of the Bar Exam it took. It could have taken one from 2009. So outdated.

[@EricM93]

cope harder. gpt4 answered a 2022 international math olympiad question, an extremely abstract and complex question that was not in it's training data.

[@anirecapped.]

Keep coping.

[@abhaykaushik95]

Is ChatGPT going to replace SDEs?

[@jeremystanley9395]

I feel a little bit better about not being replaced right away lol

[@jonesmartins]

Wake me up when AI writes original and correct mathematical proofs.

[@toyesobayo924]

As someone taking discrete math, agreed.

[@jonesmartins]

@@toyesobayo924 it's probably denial on my part tbh 🤣

[@brucewayne2480]

Wake me up when it will be able to solve the 6 remained Millennium Prize Problems and to write a theorem like Godel's theorems

[@jonesmartins]

@@brucewayne2480 I'm not even expecting that much, but that's a good one, too!

[@Nnm26]

@@brucewayne2480 I'll wake you up in 3 months

[@bobrandom5545]

Why wouldn't you replace lawyers with GPT-4? Its performance at the bar exam was in the 90th percentile. What more would you want? You can start replacing humans as soon as the machine performs better than the average professional. You're shitting on it cause it can't do simple arithmetic, but remember, it has not even been trained to do arithmetic. Or programming. Or anything, really. It's a language model and the only purpose was for it to make good predictions about which word comes next. Everything else is a side-effect. Train it for specific tasks and I'd bet it'll be better than any human, period. Lastly, GPT-4 can use software tools (like calculators) to help it with tasks.

[@Piyush2003m]

"to be fair so can't you", we are catching strays in a chat-gpt video also lol

[@jamestakac8951]

KZbin really needs to up it's game here. Your reply shows without any context to put it in place. Yes you have a name in reference but there's absolutely no sign of what you responded to. Not your fault. It's KZbins (Googles). I'd hate to think how some of my replies are coming out based on the disjointed replies I'm seeing on a regular basis. How many posts have you replied to where the original post is nowhere to be seen?

[@Piyush2003m]

@@jamestakac8951 yeah such a basic feature, youtube just fired their CEO recently so maybe some changes are about to come

[@Kenbomp]

It's silly to think gpt as a a real person ai, it's just a pretty neat and efficient generator of text that's applicable in a general way. It's useless without a compilation

[@baseddino]

70% is a big percentage imo. People are saying it's gonna take over bc it can already recreate games like pong for you so imagine the games it can make in 10 years. Yes it can't think for itself but once people help it get super accurate it will be really good. Just think about how slow and big computers were until recently. I get it there may be limits but with how fast computers can process information compared to us I think the upper limit is much higher than you're giving credit. Programers will always be needed to double check code at the very least. And I don't get the Alexa thing, I don't think it's performing worse than expected it's just Siri so what's there to really improve

[@shadez123]

I can also recreate pong by copy pasting and compiling any free source code of it on the internet. And I am light years from being a professional coder. That's not really much of an achievement

[@baseddino]

@@shadez123 i feel that but just wait until it can get things to work seamlessly and save you hours and hours of searching for stuff and trial and error. it just needs more time its only just been made

[@anirecapped.]

@@shadez123 Wait for five years.

[@erickheredia8910]

The thing is that We'll probably see how companies will start withholding their research and data from now own, especially google.

[@riajikbal2938]

as a computer science student now i feel relaxed

[@milanpaudel9624]

I am a senior accountant and tax experts and at least in those area it completely sucks.

[@jamestakac8951]

And how are you at prompt engineering? lemeguess, you suck? Not meaning to take the piss out of you but mu=y experience has been that most getting incorrect answers suck at prompt engineering. Hell, they often don't even know it exists. Are you one of them? Serious question. Never be afraid to admit you don't know something. Hell, I often don't know something and have to learn. Was tyhere 1 month ago re GPT and it's alternatives. My understanding sucked re how they worked, and how to get the best out of them. My advice. Do some reseach. Google the topic. Hell. even ask gpt how to use it effectively re a given topic. Any tool is only as good as its user after-all.

[@ShawnFumo]

You tried GPT-4 and not ChatGPT (3.5) right? Like one of the examples in the MS tech demo live broadcast was helping someone do their taxes by pasting the complicated tax instructions into it as context. He even said he was having trouble understanding how to get the value he needed before GPT helped describe what was going on.

[@shampoable]

Does anyone know what Opencourseware class was at 2:35?

[@yashag]

It's not going to replace all developers.. it's going to reduce number of developers required.. so less jobs for beginners

[@hhoums]

Back in the 1970s, Xerox thought that people would stop using paper soon. So maybe jobs will actually increase?

[@yashag]

@@hhoums could be.. we just have to wait and watch

[@deepakmsvasan]

...and loss of jobs for high-paid developers who the company thinks are replaceable.

[@hhoums]

As aspiring programmer, which job route(s) would you recommend? Hysterical "The End of..." headlines are quite unsettling.

[@spookyhades6231]

Just think of self checkout at stores nowadays. There are even stores that don't really need you to checkout. Even after all that being there cashiers still exist. I don't really think its going to wipe out any role. it might just make us more productive by saving us time from all the donkey work so that we focus on our actual work.

[@bogdanpopescu1401]

it's been 26 years since the computer beat the top pro chess players, and we still have professional chess players around; humans prefer humans for some reason...

[@jendabekCZ]

@@bogdanpopescu1401 But there is no business in chess.

[@spookyhades6231]

@@jendabekCZ well actually that's true. There is an entertainment/unpredictability element to any sport which is much better when humans are involved.

[@bogdanpopescu1401]

@@jendabekCZ I don't exactly understand what you mean; for me if there are people making a living out of it, this is business.

[@jamestakac8951]

They still need you to check out. Just that the human interaction isn't needed depending on the way you choose to check out. Reading what you wrote suggests you don't properly understand what you wrote

[@ahuman32478]

They nerfed the heck out of ChatGPT-3, so I don't see how ChatGPT-4 is going to be useful. There's probably going to be way more restrictions on it since it's supposed to be more powerful.

[@jamestakac8951]

Well, the bing version at least since they rushed it. As of now it's up to 15 prompts before resetting. Given how it would go off the rails and threaten people (LOL wish it threatened me LOL) which I seemed to have missed LOL. Remember this was based on GPT 4 which was supposed to be less prone to hallucinations.Didn't seem to go down that way. But I'm guessing that was Microsoft rushing it and not understand the tech as well as they should have before release. Somewhat like some of there Windows releases (been using Windows since 1992, maybe 1990) which have left me asking why they even released it. Mind you I also use a number of Linux Distros. Also remember that more powerful does not imply more imperfect. Newer does not always equal better. That's been my experience since the days to BASIC on ye old TRS-80 Model II Level 2. Well. I did also use the Model I back then ;) Hell I remember going from a System-80 Blue Label Mark II (Dick Smith's TRS-80 Model II equivalent) to a Sinclair ZX Spectrum (8 bit computer) and thinking that was an upgrade. And back then I took the case from an old Texas Instruments TI 99/4A and used it to give the spectrum a proper keyboard LOL. So anyone wanna tell me I got no idea re the advancements in computer tech, think again ;) Hell, how many here remember CPM and MPM? This is pre DOS. And I had hands on with them back in the days. But i've been digressing. Back to the point. Newer ain't always better, and even something based on the same version (thing bing chat vs gpt 4) can have widely divergent experiences as far as the users go. Hell I remember 8-bit systems that I believe would under certain conditions give the modern OS's a run for their money. I think many (not all) programmers have gotten lazy, and it's the fault of companies like Microsoft and Apple.

[@ShawnFumo]

@@jamestakac8951 Part of the major hallucinations they claim was due to some issues where it'd misinterpret who was who earlier in the exchange. They recently added an internal tagging approach that they say helps keep it more structured and less likely to go off the rails. Some of the Bing stuff is performance-related now too I think. Like I've heard some people saying the "balanced" mode is running 3.5 to save on cost. So be sure to use precise or creative mode if you really want more like GPT-4. I don't think GPT-4 in the openai has a limit on conversation length, just limitations on how many you can do each 4 hours (which seems to be changing all the time based on load).

[@wonseoklee80]

GPT-4 is not the final product of OpenAI. GPT-5 and beyond are coming real soon. By just adding more GPU and linear increase of performance is really scary. The game is over. Human brain is limited by biology, and AI has no limit to improve..!

[@jamestakac8951]

True. However the biggest problem I'm seeing in the short time I've been using it is that many are clueless re how to use it effectively. Hell, many can't even use google it seems and that is really sad being that it's a BASIC skill imho. I would not say AI doesn't have limits. It is most certainly limited by its experiences much as we are. Another limit it has that we don't is a lil something called imagination. Though some may contest that based on the hallucinations gpt/bing has had of late. Even those are based on probability based on it's training data

[@ShawnFumo]

@@jamestakac8951 yeah, those almost all of what we call imagination or "new" ideas is remixes of old ideas. It'll be interesting to see how things go in the future.

[@farikdk]

I love your videos, leetcode is not fun always, but your videos and explanations make me motivated to keep moving forward. Wonder you have any recommendation for AWS tutorial? Can you make videos and introduce the common high tech technologies tutorials that you think worth a watch? Thanks :)

[@user-ni2rh4ci5e]

They were able to be open so, they basically can do the closing now, just doing what they're meant to do.

[@oonbaki]

Imagine having this thing in your head and being able to think about prompts and get answers. That would be cool.

[@michaelnurse9089]

Elon will soon be looking for Neuralink Guinea Pigs - why don't you offer up your skull?

[@jamestakac8951]

Are yo thinking re "prompt engineering"?

[@raianmr2843]

People already get overwhelmed by their own thoughts. This sounds like black mirror material.

[@snafu5563]

Getting gaslit by voices in your head, also known as schizophrenia

[@adriancornel864]

I gave bing chat that problem and it solved it just fine?

[@programming_hut]

1:50 man that hits right here in centre of heart 💔

[@Onedeag-qw3yc]

Google coping so hard. Cant wait for another AI "announcement" from google.

[@maxdavidtaylor3518]

As with any new technology, it's common for there to be both hype and skepticism. It's essential to critically evaluate claims and assess the practical applications and potential limitations of the technology. As more information becomes available about GPT-4 or similar models, it's advisable to review expert opinions, research papers, and real-world applications to form a well-rounded perspective. Ultimately, the degree to which you perceive the hype around GPT-4 as overstated or not will depend on your own assessment of the information and your specific expectations from the technology. - chat gpt 3.5

[@user-iu9nd2kt6w]

Hi Neetcode! (Neetcode Premium User here) can you make videos on software engineering best practices/ technical concepts that is applicable in the industry? like git commands, databases, operating system, APIs, etc? Just concepts that software engineers (entry level at least) should know. If you release any of that, we'd greatly appreciate it! :)

[@jfonss1]

Totally agree. Tried to make it solve highschool math and while it's reasoning is ok, it's arithmetic really sucks. It makes simple mistakes that a small calculator from the early 70's can do better.

[@sebastianramirez5781]

Maybe not programmers or lawyers or accountants, but it can definitely replace a lot of interns and paralegals and maybe even financial advisors, it's not an expert but it's really good at many things, that said I think GPT-4 is too expensive and slow to be used for most production code right now and hardware has been kind of stagnating these past maybe 10 years so it probably will take a while to happen, still pretty cool tech though.

[@christiannoelsayson1904]

They should rebrand themselves as ClopenAI. I know it's a bad joke. I just have the need to put it out.

[@1.4142]

It is a groundbreaking tool, and the only overhyped part is its logical and factual accuracy. We can't even start fathoming the applications and economic impacts.

[@chudchadanstud]

We need go get the AI to have original thinking if we want it to be useful...what could go wrong.

[@solar679]

Of course a google engineer would have an issue with it lol

[@LoyalSlime]

Yesterday, all my troubles seemed so far away but now it looks like they are here to stay... Oh I believe in yesterday

[@jamestakac8951]

Love listening to the Bettles. Grew up with them from the '60s

[@ogananfe]

You forgot the mention the issue of microsoft writing the worlds codebase😂. The only thing we really know about this models is that we ask questions and it gives us answers, we have no idea what it's doing with our data. I think we are still safe till one of this tools becomes open source even if GPT 40 gets released

[@ojij1998]

It cannot solve Leetcode problems. Neither can you So cold😂😂

[@adityadwivedi8798]

2:04 I asked the same question (which exact words u shown) to chatgpt 3.5 and it gave the correct answer (indicies [ 2, 5 ] which is 1 and 5, equals 6). I think this video is just very biased guys, we really out of job in near future :)

[@codemind1822]

well honestly i am asking lotta gre questions, integral questions and it fails. I paid for gpt 4 and now there is a bandwidth of 25 requests per 3 hours. I mean clearly something really fishy going on in gpt4. I am not as impressed as the hype... Tomorrow I will try with coding and see what it can really achieve.

[@pauldirc..]

What are results

[@codemind1822]

@@pauldirc.. It looks pretty good, but as soon as you diverge and ask complicated questions it fails. Also for some reason, I cannot enter too many stuff at once. There is a limit to the entry that you can put at once. Also, keep in mind that because of slow processing time it answers very slowly. In general, yes it is a bit better than gpt3.5 but in the end, I think this tech looks like an encyclopedia on steroids.

[@pauldirc..]

@@codemind1822 how you see it in next 5 year

[@codemind1822]

@@pauldirc.. For the time being it is prone to making many mistakes like forgetting an important part of the code or fact, not mentioning it at all. Sometimes in bug fixing it struggles alot too. I mean as long as it does not evolve to something that can come up with a novel invention, I think all AGI hypes are marketing at its best. Also this new systems poses a threat to the existing internet. If people only ask to gpt in near future, this will force website owners to quit writing anything. Because no one will visit their website there won't be any website income for them. Another threat is rest of the info being regurgitated by using AI models. So internet could be filled with AI-regurgitated debris. Software dev is not only about making a simple app and I think these system is hyped. Sure it is a cool feature. It could be a very good alternative to conventional google search and a quicker one. Also note that any new system that is introduced to the internet only increased the workload of software devs so far. Be it wordpress, new frameworks, containerization etc. Because anything novel will also come with its challenges, more opportunities to create stuff that companies did not have time to work on etc. I mean if you want to learn software dev, definitely work on it. There will be more opportunities. Because of the AGI hype, people love to make youtube videos since it attracts more people but listening to them would be detrimental for your progress.

[@DominikEnder]

Sam Altman OpenAI CEO: „All of you who hyping ChatGPT, you will be disappointed“

[@jamestakac8951]

I've been telling those on facebook who are hyping it up so much that they're the ones setting things up for a massive disappointment. I think however that the clicks to their vids and the resulting monetization is the drive there however and not a factual understanding

[@enrico8730]

Your humor is getting spicy

[@vnagaravi]

Is LLAMA is a thing or just a buzz that comes and goes..?

[@ShawnFumo]

It actually is a bigger deal than most expected actually. First the models got leaked and everyone has a copy now (and likely Meta just labeled it as research-only to shield themselves and may not go after anyone). There is 4 versions, from 7b params to 65b params. The 7b one people were able to modify so it could run on a normal computer or (slowly) on a phone even. Then Stanford found that they could use the GPT3.5 api (divinci-3) to generate training data to rapidly improve its quality to close to ChatGPT in some areas (like with less than $600 worth of api calls and gpu) and called that Alpaca. That was technically against the OpenAI terms of service (to use it to train another model), but they proved it is possible. So in theory once the GPT-4 api opens up, someone can probably train one of the bigger LLaMA models on GPT-4 and have a free semi-legal AI of at or near the power of GPT-4, for probably a few thousand dollars. It may get crazy soon...

[@hamzzashaffi]

Thanks for giving me a sigh of relief!

[@jonnyskybrockett8844]

bro used the MIT distributed systems lectures as stock footage, I'm dead

[@chronokoks]

Finally some words of wisdom. The hype that some loud people create is really unbearable.

[@genderqueerlesbofluidiangodesz]

Your name is chrono cocks

[@RemotHuman]

Your leetcode problem was prompted wrong. Ask it to write a python function that solves that

[@jamestakac8951]

I'm not aware of the particular prompt you're talking about but my experience overall would lead me to believe you're probably right. So many are going in not understanding how to use this tool. Kinda like using a hammer to drive a screw. Hell, I'm only new to using these tools (lees than a month) but I've learned quicker than many. Sadly many coming to GPT and it's equivalents can't even google :(

[@jWilliamDunn]

Ask ChatGPT: "What are the limits of generative AI?" Then ask it: "How can the listed limitations be overcome?"

[@jurgbalt]

this comment was made by ChatGPT

[@chudchadanstud]

@@jurgbalt

[@yashpandey350]

😆😆😆😆

[@sahilhotchandani3668]

haha gilfoyle and dinesh video as two frds🤣🤣

[@xdxxdxdxdxdxdxdxdxdxdxdxdx8578]

The problem with this video is that it won't age well, the inevitable will happen.

[@Nivexity]

Trying to ask GPT4 to do math or encode is impossible because it's obfuscated on purpose. I'm confused as to why you didn't check this first, you can even test it just by simple encoding functions. The reason for why it doesn't work I believe is because encoding your messages, asking it to then respond in encoded form would bypass many restrictions on its output.

[@ShawnFumo]

You've tried it yourself on GPT-4? I've seen examples where it decoded base64 text. It also can do a lot of math. Not always consistently, but better than ChatGPT did. And recent papers have shown it is very easy to teach an existing LLM to use external tools like a calculator to quickly improve the accuracy..

[@Nivexity]

@@ShawnFumo I tried on ChatGPT, but it outputs nonsense I believe purposely. But you're right, it's not GPT4, but the point I was making is that the original ChatGPT should quite easily be able to do the encoding and decoding, I even ask it to go through the logic, which it purposely makes inaccurate statements and then tries to justify them. It was also noted that this was mainly because encoding inputs and receiving encoded outputs meant that a lot of the methods used to scan and delete or moderate content would be ineffective. But for GPT4 I may be wrong, the problem is that this variable was not settled in this video whereby I know that at least GPT3.5 encoding is prevented via using openai's online site access. I've not seen it successfully follow through with encoding as you suggest, if you're saying this is for GPT4 only, I'm soon to have access to it and will test it myself.

[@Arthurgarzajr]

“It can’t even solve leetcode problems. But to be fair, neither can you” I didn’t expect to be burned by watching this video!!! Lol

[@DailyGrindOfTech]

I'd understand this bashing since you're a google employee

[@mustafaceylan597]

chat gpt is not gpt-3 it is gpt-3.5