Hello, I agree with you! Just want to say I work with LLMs like ChatGPT and Llama etc. At their current state, they will not replace you, but they really enhance your learning capability. In the right hands and with domain knowledge, it can be used to speed up your processes. It has failed me a lot of times, and it has also impressed me. They are not very good at solving complex problems they have never encountered.

I had a similar experience, I'm a software engineer with 3 years of experience now. I am working on a rather large side project in mostly Python, and for one part I thought it would be nice to include some machine learning. I know some basics of machine learning, I played with TensorFlow before, but I did not have the level of understanding to solve this. I tried the help of GPT-4, but I could not make it work, I tried at least 5 times with new chats but it could not even come up with a solution which did not crash. I tried helping it with the errors the code produced, the input, the expected output and everything. But it could not make it work, and the domain knowledge required was very limited, I even wrote some crude unit tests with what was expected but no luck.

My golden rule with ChatGPT is "less is more." I get way better responses when I give it smaller requests. I primarily use it as a pocket-tutor - it is phenomenal at explaining stuff. "Hey, can you explain what this LINQ statement is doing? It is a bit complicated at first glance: " "Hi, I have this piece of code that is doing x. Are there any suggestions you have for me? Any glaring optimizations to be done? " It's also really good at adjusting or producing small snippets of code (again, try to keep it short). "Hey, I have List. Can you convert it to List where each Object1 meets condition x? Object1 looks like this: Object2 looks like this: Then create and return a JSON response containing a key "data" and the value as the List. If List is empty, then the value for the "data" key should be an empty array."

I use cgpt allot to save me time creating C# scripts for Unity, but that would be standard enough stuff, character controllers and the likes. I could easily write them myself but it just saves me so much time to just alter what cgpt spits out, rather than starting over from scratch on every new project. A few weeks ago I also got it to give me a python script that crawls the lottery website for historical lottery draws, parse that info into a CSV file, and then I massaged a generator script out of cgpt to predict the next bonus number based on the trained data from the CSV file . . now that was a fun few hours 😄

I must say it's easy when using a new language to get the syntax in how to do stuff and ask questions about specific lines but in general i noticed when the code is too long it just spit out parts glued together and then you get bad results, on the other side when you want a specific thing function/method and stuff like that it can save time and it's a nice addon, also in debugging it is quite handy, you give it the line and class and it will do a decent job from time to time

13:32 This is what IA is good for: meme generator. Totally useful.

Working with any language model teach you the new kind of a frustration feeling, completely different type. When you are getting more and more mad and can give up at any second. It is good when you can have some templates to work with, when you give up and polish it by yourself, but when you don't know the particular technology or programming language or a library it is just better to study it yourself. Especialy when it just completely invents new functions out of nowhere.

great video, I feel like is very relevant to some of the current concerns people have about AI

Great video. In summary in the right hands (devs) can benefit from gpt through creativity, boilerplate code, niche domain knowledge etc.

I think you are hitting the context / descretation limits of ChatGPT more than anything else. In my own code I've had to be very conscious about how much context I expose to ChatGPT for any specific question, even when I was developing simple python code. I think you had the right idea with that text document but there was a shorter path probably: I would have gone "take pages X-Y give me a json representing the class structure of what you find there" (I would re-generate the response until the JSON was close to what I would expect/easy to fix). Then I would re-correct nesting in the JSON (kill off unneeded elements); and feed it back a new prompt (as you did with the text file) but again re-generate the response until it was close to reasonable. I would abstract by file or by function followup questions patching individual elements but I wouldn't rely on ChatGPT to remember / keep in context the full model. I actually had a scenario with a large python module where ChatGPT 4 would just generate three cyclical bugs over and over (where it would forget the previous bugs) which felt it had well under 100 lines of memory (which ended up in me manually dealing with the issue). With actively regenerating responses and more functional questions and abstractions it seems to perform better.

Very nice video. I don't know if someone has already tried making up anything with math especially compound interest. I' ve tried everything and it can't make it right. Always brings something strange. Furthermore if you try difference tabs each will bring something different. Even the results using the same formula will result different

You are missing two important points: 1) "pages" are not a concept when you retrieve from a PDF, because the text is arbitrarily chunked 2) if you need to have GPT-4 "see" a PDF that contains complex layout (i.e. the indentation inside the table), you need to pass screenshots, so that the layout is "seen"

Dalle-3 images have something called a "Gen ID" if you generate and image and say show me the Gen ID for this image, on the next image you generate you can say "... in the style [Gen ID]" Its some random hash. Seems to work OK.

I don't disagree completely but I find reducing the amount of information you are giving it and then effectively spoon feeding more and more information as you go is effective. Sometimes having to resummarise. Note: the interfaces to reduce the labour intensiveness are coming that will follow prompt flows with agents that can try and check, improve, cross-reference existing code or fetch data, etc, etc.. with semi or near full autonomy. I find that most users that can't get their desired output are not strategic enough in their prompts or they don't give enough good examples or ask too much in one go. I am a very average JS developer but It has 3x my productivity and I can easily see that being 5 by next year with more tools. Thanks for the vid

Do you see the same legal issues with using code produced by chat GPT as you do with github co-pilot? Both are LLMs trained on large sets of data that they never give proper attribution for.

True. I have written epochs of text from activity to activity, functionality to states in a nested descriptions way until I realized the hard truth - I need to learn to code. Which I did. And the speed at which I'm developing my app is now waay faster because i don't have to rely on the answers it produces. I now say "I check for human solutions first" then and only if nothing else helps I go and slap a prompt.. to be honest the main functionality gpt could finish in my alpha version. The v alpha made by gpt worked! But the code was ugly and I had no idea wtf is going on there...and if it works after an update, and if it didn't...

I still don’t think this is fair, because the client wouldn’t send a pdf to a non-developer. They sent you a pdf, because you are. Instead, they would use some kind of project-management/devOps plugin that is built for purpose and that makes sure the requirements/story format is such that the AI will produce a good result. However, having said that, this will still require a developer reviewing the AI’s pull request. So even when those interfaces/standards are adopted, there will still be a demand for developers. Although much smaller. However, having said THAT, I am highly doubtful that AI will become reliable enough such that companies are willing to make that leap. Also I cannot imagine a company offering a product or service, without understanding it’s underlying codebase, because it was ai-generated. AND having said that, AI works on text and images, because humans are error-correcting sense-making machines and are used to working with ambiguous information. In that sense, any of those AI generated cat images, essays, code can be correct or good enough. The inconvenient truth nobody talks about is that deep learning models cannot by definition produce deterministic, highly accurate results. There are also limits in computation, due NP-hard problems, irreducibility, Goedels theorem etc. People forget that ChatGPT was developed and fine-tuned for close to a decade by swarms of classification farms all over africa and south-america. Just to produce the model we currently have. These version bumps have not brought about any improvement. All they have done is hook it up to api’s to allow it to do speech and image generation etc. Finally: AI cannot surpass humans or yield anything new, because that would break the fundamental laws of physics. It might show us something that was obvious, but we didn’t notice. But it cannot by definition uncover anything new.

Wooooo! New vid!

The title of the video is silly because AI is already replacing developers, junior devs to be specific. AI is currently a talent multiplier allowing experienced devs to build code much faster than just manually typing it. In order to get the most out of AI you have to have a fundamental understanding of what needs to be done and what good code looks like, something most junior devs have not developed yet.

It's funny. This is exactly what I experienced whenever I spent time with it. There is no way, Yet, that an ai model is going to have a complete scope in specific situations and relates to a general overview. Just like you said it requires someone with the actual knowledge to know for example where it went wrong and tell it so. Nobody starting out is going to get past that. There is a reason why after ALL these years ... there is STILL a pilot in the front of a plane. So, exactly that's what I found. ChatGPT is terrible at conversation memory and reference. It's gotten better in 4. The same with Dall-e for example. No matter how many times you tell it there is only one it spits out three. It gets it right only one and takes out correct info from previous correct outputs. It is good for example doing or leading you to saying doing specific calculations, algebra, functions etc by itself. I'll give you a good example not code related. If you give it: a city in the U.S., the square footage of a roof, the angle, the material the roof is, things like that. It can calculate how much rain water you can collect including amount for loss etc. Which you could say turn into a program. Or say how much electricity you can expect from 5 100watt solar panels and exactly a brand of deep cycle batteries including loss etc. It is good for calculations. It's just not good so far taking direct steps and building upon it. Many times, it's common it cuts the complete output off and only gives part of it which is frustrating. So the question is like many things - is it good starting off by creating your own base and writing it yourself and maybe adding little snippets it can do? Or do you try to procedurally work with and constantly go back and forth building up time and frustration back and forth trying to get it to do what you need to the point where you are saying (like outsourcing a bad programmer for help :-) I should have just done this myself. 🙂But. It does spit out something useful in seconds sometimes and it is getting better. In the meantime, people need real solutions programmed now not 2025. And even then. After you got it done.... people always want changes. (oh we forgot to add?.. wait somethings wrong... how do we do?) In a way, and maybe I'm wrong, besides "Learn to make $10,000 a month with ChatGPT" on KZfaq :-) the only thing I see is: A need for more people who know what they are doing especially if you understand this stuff and are even better fixing or debugging things. At least for a year or two. // Prompt: I want to create a website for a Web Developer. ChatGPT: (generated image of a guy over a keyboard with three fingers on one hand.) hahahahah. sorry for the ramble thanks for the video

If you want deterministic you have to go to the API playground and set the temperature to 0.

ChatGPT is not a worker. it is a tool. but even a shovel can replace the labour of a bunch of guys digging a hole by hand.

Google are bringing their own cloud version of VS code which means if we use it they'll just use all the code to try and replace us and make their database stronger so make sure everyone you don't use it.

By using the word Please, it tends to work better.:D

if you use GPT-4 over the API you can use temperature 0 and edit the assistant's responses half way through which helps with some of these issues

It's fun, i write the comments, then watch what you say, and it's match:)

Ugh, no it's not going to be a direct substitution, but when you have a team of 10 devs and you give them all a really good AI partner like GPT4 now the 10 can do the work of 8, company fires the two. Devs aren't being replaced 1 for 1 by AI, but devs with AI are replacing devs without AI by being more effective. In some cases the company will simply expand the work given to the team of 10, but in some cases they will opt to fire. It is already starting to do this acorss the world.

Seemed like a lot of time and energy spent getting very little return.

If you let ChatGPT drive, it will take you directly into a brick wall. It’s more like cruise control.

Developers using LLMs will replace developers not using LLMs. I use it daily to give me simple scrips or second opinions. Sometimes it's garbage, sometimes pure cold, but it hardly slows me down as I do not wait for it to finish and do something else.

You remind me of pro chess players vs computers in the 80s. I’d agree ChatGPT won’t replace coders today. But you are wrong about “not anytime soon”. I will take your money, thank you

But like help tool, is good.

Chat doesn't help people, in programming, who doesn't know what programming is.

Some developers at home right now are using ChatGPT to build code. That is why their code is shit.

Its getting better quiickly though. And .....its all in the prompts.....if you learn how to write prompts better, you will see drastic improvement in output. Right now though, its a good tool to have in your toolkit, like having google search on steroids....

it wont replace us...but it will allow us to do our job faster

To be honest chat gpt is not replacing any decent proffessional any time soon. At this stage it is just a tool and nothing more, a tool that requires a smart enough human for it to be of any value.

There are many tasks AIs can not do when producing runnable code. Many. Lol.

yeah not any time soon, It has like a good 3 more years or something.

my honest opinion is that islt sucks

Well, how would you explain the massive layoffs by big tech companies? I think it is somewhat ignorant to actually believe that AI systems won't replace developers. By replacing I do not mean 100% but a very significant amount. When you now start a startup or some tech project I think you can architect it in a way you will need MUCH less developers in order to develop and maintain the product because we can leverage AI now. Existing companies are caught in there current practices and these practices won't easily change in just a couple of months. And also just because a developer has his job, it does not mean that he is actually needed or his work could be automated away with AI.

I think that the concern at this point (for me at least) is not necessarily that it will replace developers soon. I think that most people realize that if that were possible, many other jobs would suffer the same fate. The concern is, this tool has 10x'd my productivity in my computer science course work, I imagine it can do the same for people working in the programming field. If it can't now, than that is something that it will be able to do soon. Now, the job pool has shrank quite significantly in the last few years, and with the people that are lucky enough to be working now having access to this, it seems very possible that "less" people are needed to complete the same tasks. Less people required in a job market that is already shrinking, is a scary thing for me to think about. I am not trying to be overly "doomy" but these are just my thoughts. On a different note, thank you for making this video, I really get a lot of value from them.

Second

once chat gp5 5 comes online this video will make no sense as it will kill hundreds of thousands of jobs also engineer man is terrible at using the chat bot i have much better success with it

I agree with you but when we will reach Ai singularity most of us will lose our jobs, from simple assistants to software engineer like us