The End Of Programming
HTML-код
- Опубликовано: 21 дек 2023
- Let's explore how AI is going to completely change programming and software to the point where we won't need programmers anymore. I love coding, so this was a difficult conclusion to come to, but it seems inevitable. Let me know your thoughts in the comments below!
Enjoy :)
Join My Newsletter for Regular AI Updates 👇🏼
www.matthewberman.com
Need AI Consulting? ✅
forwardfuture.ai/
Rent a GPU (MassedCompute) 🚀
bit.ly/matthew-berman-youtube
USE CODE "MatthewBerman" for 50% discount
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
Media/Sponsorship Inquiries 📈
bit.ly/44TC45V
Links:
My Slides - docs.google.com/presentation/...
Matt Welsh's Talk - • Large Language Models ...
Aider - github.com/paul-gauthier/aider
McKay Wrigley Example - • Meet Jarvis: My GPT-4 ... 
Наука
Tell me why you agree or disagree, I'm open to all feedback!
could we again write or iterate on the code to become efficient and write itself in assembly
Also what would you suggest to learn coding by accounting professional so it can be useful and they can augment the difflicultly to ai
I agree overall, but disagree with the timeline. I think it's impossible to predict how fast it will change (especially if we get AGI in 2024, remember exponential technology. Your predictions seem a bit linear to me) I was on board with your 1-2 year predictions since in general there will be a huge boom no matter what, but then when you mentioned 5 years having AI teams and non technical people being able to make software... I think that's going to happen much faster and come within the 1-2 years.
Let's assume we get AGI in 2024, everything you predicted will come within 1-2 years then. It's insane to try to predict 10 years into the future when even 1 year from now isn't clear with exponential technology and potentially singularity can happen at any moment (if it hasn't already)
But I get it that you need to be a bit conservative for most of your viewers.
it's funny I had actually just commented about this on your last video and I 100% agree with you
>English as the hottest programming language.
I tried Mixtral yesterday just a little bit, Greek to Rust works wonderfully. I cannot say the same for GPT. GPT works wonders for English to Rust, but not Greek to Rust. I will try some more the following days in Greek. Rust code however is ASCII only, that is to say english only.
Software Engineering is not just coding.. in fact coding is the last thing to think about while designing software systems…
We know this to be true, but the bean counters value beans above all else. And short-term beans are their favorite kind. If you think a US based company is going to try and see the big picture or play the long game, you don't know them very well.
Now companies need full stack developers which is totally coding, maybe they will converge with devops
Also let me tell you as a Gen AI Engineer myself let me shed some light on the reality. Firstly I have access to SOTA Gpt models and even prodding it multiple times it is not able to figure out or write correct code on very very complex scenarios and keep in mind I already guide towards the correct code but still it is not able to do it. Secondly in one week my highest code check ins was are 3278 lines of code my lowest was 1589 lines and yes people at my company can't fathom to matc my speed. So ya I really think at the current gen AI can write well known snippets very well but not actually full scale production level codes. So it's very very far in future where AI will be able to completely write it all and even then we humans are gonna move on to making even complex and futuristic code because technology never sleeps we keep on evolving. 🤟🏻 So I still believe there will be programmers atleast the good ones.
@@myboringdesktop😅
@@pranitrocki agree. Remember this golden rule : God is a better designer of human than human is a better designer of machine AI.
Thinking back, there have been many points where advancements in programming languages and tools seemed to indicate the “end of programming”. What tends to happen is that the problems get more complex and expand to fill the void. So while the majority of code that gets written by hand today will seem trivial to write in the future, there will still be a need for developer-like roles.. the scope will just expand.
I feel like AI will eventually be able to make full stack websites, but due to companies trying to lazily cash in on making a website with 0 effort, the stuff it'll generate will be the most bland SEO slop websites imaginable and it'll become a massive race to the bottom while actually good websites will be coded by people
That's what I think. The complexity of projects is about to go parabolic. Bigger projects and maybe more custom built to users
Scope expansion has always been part of the process, even if one does not think it will happen. lol
Exactly, things we hardly imagine to develop right now, will be a feature to build in our normal job 10 years from now, and will go beyond software but of course using software as a foundation.Try to imagine developing a solution that folds and organizes your clothes. A solution for a supermarket chain that actually does the groceries and takes the products to a household. Like I said, these are things I can imagine, the things we are going to develop are hard to imagine right now
Appreciate your videos. Little bit of topic but a word of advice:
Do not sign up on Replit with a credit card. There is no way to terminate the subscription and there is a lack of information on how to proceed to cancel the subscription. The only way is to block the card.
I'd like to share my experience. I have been desperately wanting Copilot to write solid, simple Java code. I have spent nearly a year for hours on end in front of Copilot, Bard and ChatGPT. The amount of misleading hallucinations is so great it has become a hindrance. To the point that I am nervous and jittery whenever another hallucination happens. I was desperate because I wanted to automate my processes. But the rapid churning out of correct followed by wrong code has disrupted my productivity so much I have given up and am now much faster writing correct code with IDE plugins, without Copilot.
Just a guess but maybe they didn't have a lot of Java datasets to train them on?
@@kaizen5023 Thanks for that comment.
The web is chock-a-block with Java tutorials. Github has the entire Apache Software Foundation collection of massive repos (mostly Java based) plus all the Spring repos. So not sure why Copilot sucks.
It's not just Java. I've had problems with Kubernetes yaml autocomplete hallucination too (AWS EKS)
It's like a one step forward, two steps back cha cha dance.
Give it a year
@@skoumastv it's just past the one year mark
I noticed coding hallucinations within a single answer. I want to try ChatDev next, have you tried it? It's supposed to be able to run the code it generates.
Learning programming is learning a scientific way of thinking and solving problems. I imagine a future that nobody needs to program, but coding would be a mind development tool to teach how they can think logically.
Again. Dinosaur thinking. This is going to be a train wreck watching people over twenty-five fail to adapt to AI commerce and education et al.
As if thinking scientically and logically will make people more human
@@asggerpatton7169 He has a huge point. Programming a classical, strictly logical von Neumann computer is the only mental activity I can think of that only rewards logical thinking, and gives you instant feedback.
Neuronal AIs don't work in a strictly logical way, they work statistically. Prompting them is not strictly logical either, like having to use capital letters or to tell them to do things step by step to get through to them, which is just try and error.
I see a great danger in abandoning classical, strictly logical computing and programming and switching over to prompting neuronal intelligence alone - the neuronal intelligence we are somewhat familiar with, the human brain, has shown how illogical, malicious and destructive it can be, just look at the planet or the rotten leadership of two out of the three super powers.
Neuronal intelligence has been a dangerous path in biology and I am afraid it will be a dangerous path in technology as well. Biological neuronal intelligence with low emphasis on logic has resulted in huge damage to the planet, technological neural intelligence with low emphasis on logic is growing way faster than biological evolution and may result in incredible damage way faster as well.
The whole point is to scale, having slaves workers called ai why keep being the slave, u can be your own boss...why worry about the boss firing you it should be the other way around
This, but only a few will get it. :)
I've been coding in my job since 1999, I have been trying to move away from it but there always seems to be a need. I have noted a massive decline in coding skills as new developers are added to the market and I think this is exacty what is needed. However, coding is only one part of the development process, achitecure is where the vision is converted from scope of requirements to conceps and designed. AI saying yes to all requirements is not the best approach, so some humans will need to be in the loop for consulting still... phew
I agree that programming will change dramatically, just as we no longer use punch cards. We likely won't have to handcraft code from scratch in the future. However, technical people who can manage coding agents and understand/fix the code they produce will remain critical, if not become even more important. Coding is headed toward becoming an artisanal skill in the next 10 years - it will be impressive if someone can code from scratch, but most people won't need to. It will be similar to how we don't need a carpenter to build a desk anymore when we have options like IKEA.
Yea, of course you need someone to manage. The for sure thing is that the programmers will be replaced by ai coders just like how the mcdonalds cashiers are replaced by kiosks.
If LLMs directly connect with end devices, why do we need software engineers?
@@matthew_berman
Precisely.
Not only will AI generate the code in seconds/minutes that would take a human software engineer weeks, months, or years to accomplish, it will have either a built in or a separate AI that checks for errors, and corrects the code on the fly.
@@JROD082384 thats a little too far in the future tho, as it is you might be able to tell it to generate a website (frontend) but it would look generic, the code isn't the best and overall not that great yet, and if were talking highly technical and complex architecture, then AI as of rn is nowhere near that level, IMO Ai's biggest weakness is context, it will understand what you tell it but it wont understand where you are using it, thus creating solutions not the best for your situation
Ikea gave you the lowest available quality in a modular format. it works for cheap students, but it will never survive generations like a proper whole wood piece of furniture nor will it ever retain or increase in value. all ikea gave you is the lowest quality at the lowest price possible in a modular format. it has it's place, but if you know what is good, you d rather take old shabby whole wood furniture over ikea crap. you will live better with less expenses. But if you want your apartement to look like any only fans influencer, ikea is the way to go
dude at this point i don't wanna live anymore.
dont say that! it'll be exciting times
Hey, theres always UBI to look forward to lol.
@@sanseverything900after all the chaos this world will experience yes. governments aren’t in favor of ubi just yet
I am actually looking forwards to this. My favourite job ever was as a Product Manager for SAP Commerce, an ERP-like Java-based webshop / e-Commerce suite of software. Do you know why? Because I enjoyed the orchestration between technical and functional and Marketing and Roadmapping. I think product managers are going to be in huge demand soon, if companies are going to remain competitive. The ones who can spec AND orchestrate well. The ones who have a foot in every camp. We’re heading back into a renaissance for the polyglot person. Specialists beware!
Compared to AI, you're the specialist my guy. At the pace it's moving, heed your own warning.
Your job is easily automated with AGI
Next in the chain to replace, don't worry
You meant "polymath person"?
@SimonHuggins you should give detailed information on what are looking forward to. From your post you appear to understand CICD pipelines and possibly the technology stack etc based on your domain. Address the claim in that context in terms of say complex distributed cloud based systems and ROI. I am disappointed in your response as from your comments you appear to be someone who actually knows what he is talking about. @matthew_berman hope you are tracking as this is a tesosour what is to come if you are up for the debate.
I studied computer science in college and became quite proficient in programming. However, back then, in 2007, programmers didn't have many opportunities, especially in developing countries in Southeast Asia, so I chose not to pursue a career in programming. Last October, out of curiosity, I stumbled upon AI and got intrigued, which led me to start coding again. I agree that AI has helped me catch up, and coding is challenging since there have been many changes over the years.
Opo. Salamat sa inpormasyon.
Appreicate you sharing your story!
@@matthew_berman I never imagined that we could build an app by ourselves in just a single day.
What career did you pursue?
@@b3at1 IT support, haha. Just a simple one.
You don't need to know the arcane syntax for code as much with AI, but it still helps to think like a programmer so you know what to tell the AI to do. I find that due to prompting challenges and limited context, it helps a lot to structure projects for AI using a component architecture. AI is much better at writing small to medium sized functions with explicit inputs and outputs where it doesn't have to understand the scope of the entire project all at once.
Just like humans
yup! U'll carry a reference!
The inability of current AI models to plan and create large full software is something that is obviously going to change very soon.
I'm saying that as a software engineer.
It scares me to think what will happen when this will be available on a large scale.
this rings very true to me (without being competent to judge myself), as it sounds extremely similar to my experience with using ai to write fiction.
Sure, if we make the time frame long enough, then AI will probably be able to guess any kind of program anyone could possibly ever want before we even become aware this might be something we want.
Projecting too long into the future doesn't make sense at the current rate of change, but for at time small enough so that it still does make sense, I imagine that certain very fundamental forms of logical reasoning and programming skills will still be highly beneficial to learn for some humans.
Yup
I think all Programmers will soon turn into Requirements Engineers. Even today 50% of my work as a Software Engineer is to iterate over User-Stories with Product-Owners and Project-Managers to refine their idea and make sure no corner case will break the existing product.
Just think about how many possible solutions exist to a simple prompt like "build the game snake". Will the snake wrap around the boarder? should the snake grow when eating? when will the snake get faster? What about an end screen showing the points before starting a new game? should there be a way to pause the game? etc.
This. Most doomers does not understand coding is actually servicing another people and solving their problems.
And what will happen for sw-project prices in future? They will crash, and every time in history when this was happened, demand of devs has been increased.
it's incredible how the already good videos on this channel become even more interesting when Mattew gives more space to general topics. Compliments
Totally agree, i happen to be one of those non-technical people and i've been able to build stuff i never imagined would be possible, the funny thing is, because AI still falls a little short, you're still going to have to learn a few things while you're at it, you're really learning how to code without it being the main objective.
@chandamubanga your statement should have ended at "non-technical people" and leave it at that as the claim made is actually highly highly highly loaded from a technical perspective spanning an industry that dates back to the late 1950s. @matthew_berman hope you are tracking as this is a tesosour what is to come if you are up for the debate.
I tried using it to create a project, but it kept telling me that it was only a AI tool and that it was not able to create what I wanted. However; it gave me instructions for certain parts of the code and also helped me with using different languages to achieve the same result. Maybe if I simply allow it to build things for me regardless of language I’ll get better results. I also ran into a messages that said that I have to implement my own security methods. I haven’t tested it with databases and serves, but those have been the only restrictions I’ve came across. Other than that it’s an a amazing tool.
2012: In five years there will be no truck drivers
2023: In ten years there will be no truck drivers or programers.
I know, this time it's different.
Another great video, thanks! Been a big fan of you and the channel since you started, really appreciate the insights. For the past year, I pictured a 'black box' with dual 4090's or an H100 on my desk that just 'knew' every line of our code base and could write, build, test, iterate and optimize our code base, even using visual modality to assess the designs, etc. And I think we're heading there. But one thing I learned/realized ~6mo back is that as the context window grows, it's not linear but exponential.
So as your code base gets bigger and bigger, it takes exponentially more resources to "hold" that code base in active memory for inference processing and the resource cost just starts shooting for the moon. Not to say there aren't current and future strategies and solves for this to be sure - there are big developments in optimized training to minimize parameters, fine tuning and even specializing the processing units for LLM's and matrix calculations. So exciting times to be sure 💪
I think a possible solution to codebase/context size is breaking the application into domain based components and having specialist agents working on each domain. They'd need a way of coordinating the end solution, so some architectural model would be needed, but it could be specialised in that, rather than detailed technical delivery.
Kind of like how the real world is supposed to work 😂
I disagree, I used the new chat gpt to write a simple probability analysis script and even though it got it right from a technical perspective it was wrong from a logical perspective and it’s implementation made it difficult to reuse and thus have tiny to scrap it and do it myself. It’s good for programmers but definitely not viable for a professional setting where you can’t just toss some code in you don’t understand and hope it hold up let alone debugging.
I can confirm your experience, senior software engineer here (BSc, MSc software engineering) with close to two decades industry experience. I am playing around with GPT 3.5 and also the paid one (4.0), it can produce some generic code with the proper instructions, but it's a lot of work. I could spend an hour commanding it to create something that any dev with 2-3 years experience will create manually in less than 15 minutes. also forget about anything complex, it's not capable of writing complex inter-connected source code unless you spend 6-7 hours for an output that would take an average developer just 3 hours to write manually. it's nice for data analysis of text though
lets come back to this comment in 2025
@@ginebro1930I've heard the same thing in 2023, you're on
@@ginebro1930 Professional programmer here, I'm spending my time learning and improving my skills in development, I pass on going out, drinking alcohol somewhere or playing games too much, I live in the virtual world as closely as an organic being can, I spend my time researching until my brain can't process anything else for the day, and I challenge the "idea" that you will ever be able to create software remotely as efficiently as me, grab your best AI, see if it holds a candle to my custom combination of agents and the way I use them, being that in 2025, 2030 or whatever year you wish to challenge me.
Coding Project Issues
Code Reusability:
The AI-generated code might have been too specific or not modular, making it difficult to adapt or extend for other uses.
Lack of functions or classes that encapsulate behavior for reuse.
Readability and Maintainability:
The code might lack comments, clear variable names, or a logical structure, making it hard to understand and maintain.
Use of complex or convoluted logic that is difficult to follow.
Hardcoding Values:
The AI might have hardcoded values that should have been variables or parameters, reducing flexibility.
Lack of Error Handling:
Insufficient error checking and handling, making the code fragile in the face of unexpected inputs or situations.
Performance Issues:
Inefficient algorithms or data structures might have been used, leading to performance bottlenecks.
Security Vulnerabilities:
Overlooking security best practices, potentially introducing vulnerabilities.
Solutions and Best Practices
Code Review: Conduct thorough code reviews to identify logical flaws and areas of improvement in readability, maintainability, and reusability.
Testing: Implement comprehensive testing (unit tests, integration tests) to uncover logical errors and ensure the code behaves as expected in various scenarios.
Refactoring: Refactor the code to improve modularity, readability, and reusability. This includes using functions, classes, and clear naming conventions.
Documentation: Ensure the code is well-documented, explaining the logic, usage, and any assumptions made.
Security Audit: Perform a security audit to identify and fix potential vulnerabilities.
Performance Optimization: Analyze the code for performance bottlenecks and optimize using better algorithms or data structures where necessary.
Visual and Interactive Elements
Flowcharts: Create flowcharts to visualize the logic and flow of the program. Tools like Lucidchart or Draw.io can be used for this purpose.
Code Walkthroughs: Use interactive tools like Jupyter Notebooks for step-by-step code walkthroughs, demonstrating how each part of the code works and interacts.
Refactoring Tools: Utilize IDE refactoring tools to restructure the code for better modularity and reusability.
By addressing these areas, you can enhance the quality and robustness of the code, making it more suitable for professional settings.
This is a fabulous summary of the state of LLMs and the impact over the next few years. I think you are being quite circumspect with yoir timescales too, but they seem good when you bear in mind that we always think things will take less time than they really do (think ‘self driving cars’ and ‘genomic medicine’. Whatever the timescale, your message is timely and I think, bang on. And I had no idea about the Sonar thing! 😊Something I use in every project. Appreciate all your hard work - both with Sonar and this channel which has been a huge inspiration over the last year while I have been expanding my know-how in this area. Thanks Matt!
I totally agree and ..at its core what you say is indisputable .. its starting right now with the nc/low code exposition and the empowerment of the citizen developer .. this will address the niche space and problems experienced with small and medium size business.. with convergence of technologies such as blockchain 6G - IoT... an issue to think about is regulation legislature and access and the monopoly that open ai threatens to achieve through its API platform and store ..alternative such as opensource innovators may help keep the transformative impact alive for those with tools , drive and a vision
You might get some basic to intermediate code out of an AI, but Will it be supportable? Will you be able to deploy it to a production system? Will you be able to maintain it? Will you be able to patch the underlying systems that run the code? I think if you look at the entire team that it takes to run major production level software systems, you might be able to replace the junior programmers. But we've still got a ways to go before we can start declaring programmer's obsolete, and even further in order to get rid of the operations(SRE, Devops, Sysadmin) side of the equation.
5 years before jr programmers could replaced on a team, 10 years before lead engineers, 15+ before a product manager could manage all the roles of a software team without others to write, deploy and maintain a software system.
In all likelihood, engineers aren't going anywhere anytime soon. We will use AI to accelerate the building, deployment, and maintenance of software. But I think you'll still see the combination of AI + smart engineers, working together to produce more competitive outcomes than just AI.
I expect there to be a hot industry of consultants that have to come in and unfuck software systems that were built and productionized by AI in concert with people who didn't know what they were doing.
Dinosaur. In ten years it will be all gone. Everything will be different and all the hours, coding, education, career, all for not. It will be extinct. But, people keep rationalizing. The tsunami is coming. Its tough to have everything you ever knew and have been become obsolete almost over night. Coding is dead. People are gonna lose their minds. Start practicing meditation now.
You don't understand that llms are just basic Ai's they are ANI when AGI appears then it's game over, Artificial narrow intelligence are an infinitesimal percentage of what a AGI can do and AGI are an infinitesimal of a ASI and AGI it's pretty much a fact that will happen taking into account the progress of ai in general Which is exponential o even faster at least for the last year.
@LuisBustos-jq8sz I understand what LLMs are, but it seems like you don't understand what LLMs are. LLMs are just next word prediction. AGI has to be able to reason about the world and we've seen nothing close to that yet. We aren't really any closer to AGI than we were at the beginning of the year. We took a huge leap forward at the end of 2022 with chatgpt, but since then all the improvements have been incremental in performance. The gains have been in speed and size of models. Maybe we'll make another leap forward in the next year, but there's no evidence of that yet.
@@durden0 Quality changes are not a gradual process, stuff changes from quantity to quality spontaneously although you are able to predict when a quality change is going to happen you are only able to do so with the critical constant of whatever you are measuring, again AGI is not going to be "oh yeah we were working on the AGI for the last 10 years and we made 10% advancements towards it, each year" it is an unpredictable quality change thats going to happen spontaneously somewhere down the ai development line, The huamn brain is prediction machine based on pattern recognition just as LLMS are LLMS are only able to predict language in this case when human brains can predict other stuff too. So again, everything that humans do (intellectually) will become eventually obsolete as ANI get better and AGI appears. The solution is to use AGI and AI in general to better our "hardware" in order to don't get depressed because we are now "useless". But AI will without a doubt replace human coding at all levels known to man.
@@LuisBustos-jq8sz This discussion is about when AI will overtake human coders and replace them in the work place. If you boil down my comments, i'm essentially saying, "it's gonna be a while before that happens, cause there's not good evidence that we're close." And you're response appears to be (though abstractly) "yeah but AGI will happen suddenly and without warning".... But you don't back up that assertion with evidence or data.
Honestly I'd be curious to see this new AI workflow. Everytime I consider using AI, it just doesn't seem worth the trouble except for some of the most basic tasks. The code it produces is usually flawed or it just can't help meaningfully because it doesn't know anything about the project as a whole to tie things together. And I still get stuck testing and looking over the AI's work to make sure it all actually works like it's supposed to. And if it doesn't... well then I have to debug it anyway. And if you can't program, I don't see how you can debug the AI's work properly.
Thus far the good use cases are fairly limited, like a strictly pure function that does some kind of relatively simple transformation, it's good at. It seems like a useful tool a programmer can use in some spots, but I haven't seen anything that seems to be be making the entirety of programming obsolete anytime soon. I might change my mind if I see some amazing workflow someone is doing with AI that's producing this perfect flawless code, but thus far my own experiences haven't been all that great. Once you get past the honeymoon phase of "OMG I typed a request and the computer is coding it", it just doesn't seem to be all that amazing at coding.
But if I had to pay a guy to write code for a diy drone and now if I can get it to fly with basic knowledge of just English and lot of GitHub and ChatGPT, that’s what’s gonna make “programmers” and not programming obsolete.
I think that the project context problem will be solved fairly soon. I'm working on a solution myself
I strongly disagree. In day to day programming my productivity has skyrocketed. As described in the video , I've become more of an orchestrator. It's indeed crucial to have coding experience, but I'm basically just scanning the code now as it presents itself (including comments) , do some tests on the side. I'm not losing time anymore with code syntax I forgot , but I can instantly recognize it and correct on the fly if necessary. AI code will only get better and better. You should embrace it if you want to be one of the very few programmers left in 5 or 10 years.
I do low level work. It only helped me when I want to do some mundane task (generate a shuffled array of indices, print some tensor in a formatted way...). Other than that, it never seems to get anything right and the autocomplete starts getting annoying after a certain point. So far, I've only seen it being effective in web development, where I believe there is a huge amount of code available online on which it can train on, and even then, it's only generating basic web pages that anyone can learn to do in a week.
AI helps but is not even close to be able to write larger systems consisting of 500k lines of C# code. I use gpt4 and copilot for small, well defined tasks. I'm mentally prepared that this could change any day, and I could become obsolete. I just try to enjoy the time I have left, including my generous pay check. We may be the last generation that will do this kind of low level hardcore programming.
i really appreciate how open and honest you are
I've just lost a lot of respect for you as this video shows how little you get into the real details, while claiming to have been "thinking about it a lot". Your level of undertstanding is at around 20% of my journey of discovery into "why programming is dead". At that point, I would have agreed with you.
For context, i am Lead SWE with 17 years experience, I read 3 AI papers each day, 5 years experience in AI.
What you missed:
- Complexity resolution scales exponentially. AI finds a problem twice as complex four times as difficult to solve.
- Overall iterative coding requires human coder input, and always will.
- Coders are not paid for hourly services but for expertise.
- Humans prefer talking to humans instead of robots.
- Most companies aren't cost optimisers. Many still use 90s software.
- Aggregate demand is infinite. When a project is finished, company wants to code something new.
- Look at production difficulties with RAGs. Read some experiences. AI stumbles on complexity of detail, and human input is required at every step, negating most of utility. These things aren't as easily solvable as you think.
- Closed source code is plentiful. Debugging embedded C++ software on a mobile device or microcontroller is distant future. You need physical hands to interact with devices.
- Early adoption of fully automated systems will be sensitive. A manager sees a bad result, and he gets pissed off much more. Psychological feeling of helplessness vs if there was a human to talk to. Will hinder early adoption.
- Self driving cars work just fine 99% of the time. It's the 1% stopping adoption. Programming will be the same. The 1% bug will happen 30 days after product launch, everything will stop, and a human won't be able to fix it in 5 minutes since they didn't code it.
- Telling an AI what to do, you need a pro coder. Using natural language and arguing with AI does not exactly save time.
I agree programming will change in 5-10 years, but your estimate of 1-2 years is really laughable. The skill upgrade you expect from AI will require extremely expensive and clean data sets, and new training techniques. People are talking about this, but this will take 5 years to do. Many companies will invest and fail.
> programming is difficult
Practice, practice, practice...
> Copilot
For me it is just annoying, it just made for simple pieces. Good for documentation research
> AI better at coding
Well... ask it to do a rope algorithm or optimized DAG lmao
So far, no AI were able to do both
Still good enough for regex but even there, i have to double check and fix it
> just give prompts
What if your prompts are fucked? Your requirements are everything
> everyone will be programmers
Without knowing data structures, how to store data properly, etc All those programs will be just small tools, not products that work at scale ¯\_(ツ)_/¯
I'm not even talking about low level programming
So far, it's impressive for many things, it's a good enough tool
AI is a wet dream for mid programmers
AI is just a tool for software developers/engineers
I feel empowered and augmented but I don't feel replaced AT ALL
I think juniors and mids will be replaced for sure, it is going to be hard
I'm not scared for people like me, I feel for the people that need to learn everything to reach a good level of programming
That's what I see every day with juniors, they are a bit lost with all the tooling and languages they need to know
An AI is as good as an intern, maybe one day it will be as good a junior. For now I doubt it will be as good as an experienced software developer.
Also it doesn't have intuition when debugging or visualizing the flow of data in their mind
It's going to be nice for many small iterations, it is going to be nice to fix a small problem quickly and iterate... but to make a real scalable system... or to make a very secretive system, i doubt
I'm happy that people can have a taste of programming and do things without being restricted!
I doubt good software developers/engineers will be replaced, too much valuable! AI will be super fancy tools to work EVEN FASTER than before.
I will use an AI like an exoskeleton !
To some degree, I agree with your comment (certainly at this very moment).
However, something you should try, that might give you a new(ish) perspective:
Over comment, and name your functions/methods/variables with names intended to be understood by an LLM.
Comment first, before you write any code.
Do your function definitions, all of them, before you begin programming.
If you are using an arcane or recent library/module, drop documentation directly into the code as a comment.
Personally, I use Sublime + the LSP plugin and the code happens as tabbed out completions.
What this does is it provides the "prompt" for the LLM, so that it can provide useful completions.
Then, as I begin each line of code, the completions are MUCH more accurate and useful.
Often, if the completions are failing some specific way, I will literally drop a comment in, and "talk" to the LLM, providing it context, or even direct 'orders'. Right there at the problematic line.
In this way, the code becomes almost a conversation between the developer and the LLM which is running completions. It WILL listen to directives in comments.
Then, after I'm finished, I'll go back and delete the useless comments.
Give that a shot, and see if you find the completions a bit more helpful.
For many projects I've been working on, with a bit of terrain preparation (documentation comments, etc.), I'm catching myself *tab*, *tab*, *tab*, *tab*, *tab*, *tab*, *tab*.......
Well, it just wrote the entire method.....
It isn't that I couldn't write the same code myself, it is simply easier and quicker to *tab* completion it out, especially for rote stuff.
Cheers!
@@_SimpleSam I will try for fun
I use AI code gen mainly to script small scripts when i'm busy writing code on something else in the same time, it's like having two other hands to me ¯\_(ツ)_/¯
> write a bash script that use X to do Y with those requirements, and do it with those steps, etc
Also it is useful for documentations while i'm working
I will try your method of code gen do see what it can do, thanks
There are broader implications with the upcoming AI explosion, not just programming we will need to re-imagine human job roles pretty much accoss the board. Especially once AI/ML + robotics + 3d printing becomes common place. Excited to see where those all goes however we do need to have more frank discussions on the implications of where AI/ML can lead us. Great work.
I think there will be two types of countries in the future, the first will convert into chaos, people become homeless and looking for food.
the second will use AGI robotics for their people advantage, like make Farmer Robot, Worker Robot, ... etc., there where people will live in paradise.
Great video bro! You were just a few years older than me it seems and your track and history goes along with mine not in doing but in remembering lol
Exelant video, thanks man, really puts into perspective. Much appreciated!
7:04 "And, granted, I'm not building production-level applications anymore." Well, there ya go. Just getting a small piece of functionality working, which I agree, AI can be quite helpful with, is massively different from getting many small (and large) pieces all working together cooperatively and not collapsing under their own weight.
his naivety and lack of programming knowledge is quite evident in every video he puts out.
AI can help get some boilerplate code working for you, it's not going to write a project and maintain it for you.
@@brianclifton9399 Exactly, I am sorry to say but he is out of touch with wtf a good programmer does. His last video was a promo and in that promo he used AI to create a password setup using raw text input no encryption nothing... and said "This looks great".... Even the basics...
I disagree. This sounds like people 3 years ago saying that ai images will never be as good as human art. I think of it as some sort of fallacy in understanding exponential improvements. Ai in my opinion will shine in exactly the management of large systems that human programmers can't keep in their mind. I can see it vastly optimizing, debugging, and keeping up to date enormous codebases in a way human teams will never be able to do. You guys are looking at a first order Taylor expansion of an exponential change
@@cyberpunkspike And not even that: Because it doesn't really understand it can only adopt some code-snippets it has already seen before and is not able to struct the boilerplate code better then you yourself could google it in a few minutes. - It can't correctly structure code because it simply doesn't know what you want to achieve at all in the end.
I agree that most junior level programming will disappear in the next 5-10 years, specially in business specific applications (aka "VBA crimes"). However, as you can learn from the adoption rate of no-code/low-code, the industry is very slow to change and adopt. The main point is: it is very uncertain if AI can anytime soon can create higher level functions. What you can see today is replicating some "hello world" stuff, very unspecific, generic copy&paste from the net. It is amazing and will change how code is written - however, i am uncertain it can replace a senior developer who works on a production system with sometimes hundreds of complicated business rules on a single function. As IA does not understand anything, i am afraid it might get stuck at the 70% line - it can produce something that is almost good enough, but getting the next 10-15% to make it useable for production might take an insane amount of effort. I am also unsure if AI generated code will every be good enough for critical functionality, as LLMs just replicate what they have as in input.
We use LLMs *a lot* every day, and we have a term for the typical LLM results: "word vomit". As you get more experiences with generated content, you can spot "word vomit" or "pixel vomit" very quickly: AI generated images are off in many ways and so is much of the code it generates today.
To be clear: it will become much better and an below average developer will be in deep trouble - as the job will become much harder for a developer in the future. I also assume that developers will get paid MORE and will be in much higher demand. Why? When i created my first applications in 1995 most developers where fullstack - and the stack was very thin. I remember when PHP was introduced and the Java Server Pages guys made fun of it. Java Server Pages died out soon after, and the PHP ecosystem exploded. Then no-code showed up 20 years later and people predicted the death of all programming languages. Look inside a typical company, you might still find JSP applications and even worse: some niche programming language that was all the rage 20 years ago and is long forgotten - but still used by this one business app, that is kind of critical - not critical enough to be rewritten in an current language, of course.
My prediction is therefore:
#1 some teams will go full AI and will experience crazy productivity gains - and most of the developers will continue doing what they do today for the next 10-15 years.
#2 Yes, some copilot based technology will be used and we will see a big increase in productivity. Some companies already scale down teams by 2-3 developers (from 8-11) - and this seems to be highly effective for multiple reasons.
#3 It will take a VERY long time until LLMs will be adopted for on-premise production solutions. I would not be surprised if this is 10+ years
#4 we might soon enter a "AI cold winter" due to legal issues - blocking any investments for years to come. Expect companies to ban google copilot and similar tools due to legal concerns
#5 i am uncertain if there is a big first mover advantage in the application of AI at this point - companies seemed to rather wait for somebody else to take the risk and then move in quickly as soon as regulations and technology has become production ready.
#6 There are still a significant number of development teams who have not adopted even the basic practices: some teams even work without version control, no standard build pipelines or any testing. You can find projects that got started in 2023 with ZERO automated testing, no QM in the team, no coding guidelines etc. - in companies with 1K developers.
#7 decision makers do not like to take risks. Generative AI is currently the biggest liability and risk for a company. Yes, it is an amazing opportunity for a startup and a few early adopters - but most people have way to much to loose and nothing to gain from investing into it. So why should it take off quickly?
Just because we know how to build professional software does not mean that people actually do it.
Just because we have the power of generative AI for development does not mean that it will be adopted anytime soon - even IF it works "good enough".
Hello,
I started to code at around 16, now 53, now going more to management and a bit tired of coding, still love it.
I'm now more manager of a small team, developing mobile app, that is quite popular, we have like 20 app on the same code base, iOS and Android with now many part in common using KMP.
This is quite scary an exciting at the same time, still, I don't believe the end is that near from my point of view, I saw a lot of math problem solving, or new code generated to solve problem that do not require interaction with a framework. Nor doing reusable code of some sort. What would it be if the code is written by AI that have a bug cause it did not thing about code reuse, or how it will find a fix of a crash cause because an old version of and android OS don't have a library in it!
What about fixing current legacy code, do we just ask it to rewrite it all? and will it make it bug free, he is still trained on code that may have some bug in it. So I am right if I say he will introduce bug that we, human, did not have fix yet or thing about it!
And, what about human, yes, us, human, when we're gonna realize we're loosing our job. So many technology have been delayed just because of if we go too fast, we loose money, and how about people that will just stop right away working before it's too late and paralyze a complete sector. Like in Calgary at the port, they go on strike because of bad condition AND automation that will replace them! Human factor will slow down all this, an AI may look good at initial writing from scratch a little feature, but I believe we're still far from AI writing RUclips or Facebook. And I don't see any AI writing an OS like, Windows for ARM or next gen of smart phone OS.
I home I'm not naïve!
Bless you, thank you for all your videos to educate us and inspire us
Huge issue currently is exceptionally poor spatial reasoning. Ie AI coders, in my experience, have no idea what a good UX is nor how to present it without very detailed design guidance.
I started programming in BASIC on a Commodore Vic-20 in 1981. I can already hear the "OK Boomer" now :) . In my 20s and 30s I got so immersed in programming that I almost became one with the code.
Over the years I moved into systems and architecture and spent the last 14 years at Stanford. I was working on AI when I left in August.
I no longer need an income. I have lots of ideas but no longer have the mental ability to manage 100s of functions and dozens of libraries in my head. LLMs have entirely changed my future.
Thanks to AI agents I am creating a complex app for a global restaurant chain all by myself. Even in my intellectual prime I could not come close to my productivity now. I spend my time doing product definition now instead of searching how to's all day.
Pretty similar to my situation, i am building financial app all alone when it couldn’t be done with 10 programmers before
Now imagine if a precocious 17 YO uses chatGPT to write this complex app for a global restaurant chain. They will probably be able to get something working, but it will probably come crashing down and they teen will have no ability to recover. I agree with everything in this video but think there is still room for experienced developers. Imagine if programming tasks as earth moving. I used to have a shovel and was an expert at using it. Everyone now has a backhoe. On simple tasks like digging a ditch, the 17 YO will be just as good and maybe better than me. But on a complex task like building a highway tunnel, I have the experience to avoid disaster
@@asmithmd1 That is true now. ChatGPT often gets the code wrong even when I give very detailed requests. But in a few more years AI will have learned from all of us how to do everything we do and will do it better. My decades of experience as well as the experience of everyone else will get absorbed into AI. We will have to re-evaluate our purpose and value before too long.
@@billcollins6894 How can you no longer have as much mental ability as back then, isn't the brain "use it OR lose it", not "use it AND lose it"? What about all of the cognitive reserve and rewiring promises of the current mainstream neurology - all overhyped and wrong?
Or could it be other health issues that are slowing you down, like say your hipocampus has gotten less active due to something in your body from environment or different diet habits or whatever? Or gradual loss in myelination from avoiding fats in your diet? I hope you don’t do that.
Please answer, like how has it been with your intellectual abilities over the decades, how has the experience felt, I wanna know.
What ai agents are you using?
I truly value your candor. Reflecting on history, when Gutenberg invented the printing press, many monks who previously hand-copied texts were initially upset, fearing it rendered their skills redundant. However, this innovation significantly broadened the availability of written material, contributing to a more informed and educated society. A similar situation is unfolding now. It's a choice between embracing this new technology and progressing with it, or resisting it and risking being left behind.
very well done video, the best argument I have seen so far on how ai will replace coders though I am curious on how ai is going to handle the non-coding parts of a developers job like gathering detailed requirements from stakeholders, feature alignment with diff teams and handling tickets/issues from customers/clients
Excuse me... Lot's of the tested models were not even able to initiate variables correctly although they have seen probably 100 different examples of snake game code in different languages. The actual AIs simply don't know what they are doing! They just combine at the moment.
I don't think it will change it all that much. Co-piolot/gpt are pretty good but they constantly get things wrong and I often find I was better off just doing the thing 100% myself. They are just reproducing code examples from other humans but seems limited on its ability to innovate. Which is good for us. Every now and then it will get me out of a pinch though. They are great tools and I hope we keep them as that.
The best use case for AI ime is solving really small problems one at a time. e.g. when I forget specific syntax or how to do a particular math function. I've never once had success with it doing something on a larger scale so its not really worth trying until it gets better.
This will change. It's just the beginning.
@@EllysEscape Yea its great for iterations and someone to rubber duck with. But hallucinations are real, and the knowledge pool it's trained off is often average at best so it often does things in an less than optimal way. But it's still impressive.
@@djRichyRichhNo it won't, unless most companies are okay with patched up messy code blocks.
Hi Matt. Not so sure I agree. I think it’ll help make super coders, but I don’t see a LLM coding an app anytime soon. Any app I’ve ever worked on I’ve spent nearly as much time testing and bug fixing. Now those things are more important. You’ll be running an app that you just have to trust. And how do you make fixes? Trust it to correct itself? Enhancements? I’m just not sure we are that close yet.
Why can't you explain what the bug is and tell it to fix it? Obviously there will need to be human QA people still.
@@allanshpeley4284 Because often it either hallucinates what the solution to said bug is and confidently responds with another (wrong) solution (or sometimes the same previous solution not even changed) or just flatout refuses to acknowledge there is a bug because it has no ability to test or gain context for what it's written. Perhaps one day it will be able to test its own code in any language or setting, but if it was today I'd be using it. (it's not)
@@allanshpeley4284 bc AI doesn't have brain to think. It's not intelligent xd
Thank you for the video I was glued the entire time. Thank you for the valuable sources from others, it made everything easy to follow. I just started my CS Degree online and seeing this is slightly discouraging. Would you have any tips for someone who is still learning to code and what I can do to secure my future in Computer Science?
Thank you for confirming all the struggles with programming that you have. I am a business professional who has been coding for years to make my tasks more productive, and never knew what was normal for others. :D
Having a tool that does a good job of roughing out basic programming tasks doesn't replace programmers any more than having a lint checker replaces programmers. Programmers will just be more efficient.
I personally think that AI will will stagnate on a few key areas that prevents it from replacing programmers altogether. This is specifically refuting the "PM -> coding AI -> QA" loop mentioned in the video.
1. Debugging large projects - AI is good at building individual components, but to integrate multiple components together AI will need to fully understand the previous code outputs to the very last detail or it will write the wrong design that breaks in rare and strange ways that do not happen in code written by a human. Even if the AI is very advanced and can fix these problems when prompted, non-technical human QA will have trouble understanding and prompting it back to the AI. This problem can only be solved by replacing the QA with AI as well - and hence creating an AI that never makes any mistakes through solving any it finds internally. That is an extremely tall ask compared to where AI is currently.
2. Performance - nontechnical people may have an idea of the functionality of a product, but ultimately code (even AI-generated) still runs on computers and are subject to computational complexity. Performance optimizations are highly context-dependent and generally online training data tells AI the most general way to solve a problem, which will be slower in compute. This implies that a technical person with knowledge of programming will still be required to create any software that gets value in being performant.
3. Training data - I am pessimistic in general of whether LLMs can ever get out of the training data problem. Inherently, even the best LLMs only work as long as one similar example is present on the internet. They stop working if asked to do something that the internet has no example for, and thus the LLM is unable to train to understand. Even if LLM compute power and quality improves exponentially, the quantity of meaningful training data will not, so LLMs will struggle with any niche subjects which are common in programming in the cutting-edge.
This is not to say that AI to replace programmers will never come, but I don't think the path of improving LLM technology will get us there. I think we truly need something that can think and solve problems like humans and therefore solve the data problem with meaningful AI-generated data.
It is not good at building individual components, at times it can manage to cobble together a working module of spaghetti code after multiple iterations.
@@brianclifton9399 Completely true. All AI code I've seen so far was extremely basic and had no security related measures AT ALL in it.
The actual state of AI codding can only impress absolute beginners.
As we see again and again it does not even understand that one must/should define variables before usage. - It simply does not know what it's doing.
3 years
@@GodbornNovenI'm sorry, what do you mean with "3 years"? - Until what exactly?
Your so right. Thank you for this video.
Dude, thanks for it!
As long as I have my fellow humans along with me for this AI future. I think we'll be fine because going in alone sounds scary.
This may be true for mainstream and standard applications like web development or database apps. However, I am having hard time to make gpt4 to code engineering programs like multi domain systems modeling. Even after repeatedly propmting with corrective statements, i couldnt make it write me a code for a very simple engineering problem. Not sure how fast these gpt models will continue to improve, but this is the case for now.
I think GPT-6 will handle it
Although eventually this will happen.
You really underestimate how much production level apps run of complex legacy code and also what dependencies they might have.
The examples shown are great but a lot of details are left out.
Also there were no mentions of hallucinations, which are a huge hurdle to overcome and will not be accepted with industries that require certain levels of compliance.
This timeline is more of a guess than anything…
Thank you for wisdom
I look forward to the buggy mess of untested auto generated software we have to suffer through for the next 10 years. :) Make sure you keep a paper/personal record of anything important, and hopefully it fails in my favour.
And this is different how? I personally am awful at language in school. But when I used chat GPT4 to make GPT 3.5 chat bot w mobile and Stripe accept. So I am able to create and build AI tools, auto gen AI agents Etc.
@@timmay6634 The issue I have is that from my experience the AI really doesn't do a good job of assessing the priority of where to emphasize specifications. It easily drops parts of the specifications, also it doesn't appear to apply the same logic from the specifications to all parts of the code it generates (even within a single prompt, and it gets worse when you try to span generation of consistent code between multiple prompts). Basically it generates very complex sophisticated output for a variety of requests yet it makes the most trivial mistakes in others. This is why I think we are going to suffer through auto generated code, because people assume that because it's able to generate sophisticated code without error it can do simple stuff as well, and based on what I've seen so far, this assumption isn't justified or correct. I expect with time it will get better but in the interim we are going to get some jank errors. I hope I'm wrong though and tools emerge that allow it to process more complex specifications, processes and business constraints.
15:27 that's literally me. I was non-technical and was able to hang in a hackathon because of ChatGPT. I knew some basic Python but used ChatGPT like copilot. And I was shocked some of the more experienced participants were unaware of how good ChatGPT is.
Exactly!
Exactly this! There are so many people with their heads in the sand on the pace of advancement with this technology.
Yeah, a lot of denial in the comments. Even some experienced guys don’t get it, or don’t want to get it.
@@darylallen2485 i can't tell you how many software engineers I talk to that are all either super skeptical about the tech and don't use it, or are like completely unaware.
I went to a GENERATIVE AI hackathon, and I was the only person in my team that even had an OpenAI API key. Team had 2 20+ year vets in OpenAI and 2 students in MS program for Machine Learning. As soon as we used GPT endpoint, our results improved like 1000%.
I think it's because a lot of SWE's don't stay up to date and see it as a job so when they come home they don't care much about new tech.
Sounds to me they are not experienced... Most programmers use GPTs.. Why is it that you assume that all programmers hate AI? So many people here lack what it takes to be a programmer. blog.boot.dev/computer-science/ai-taking-programming-jobs/ You might learn something.
Would love to hear your thoughts about the followings:
1. Both in your example (eight gpt pilot) and the example with Jarvis you asked the so to create a log in site. I'm guessing that there are more then 100,000 repos that already done that and gpt had acesses to learn them. Meaning it recycled an existing code. Whould it be able to create an end to end app that uses micro services, work loads, docker and k8s integrations and deploy to aws for example? An app that has complicated logic and complicated ui? How many iteration will that app require for gpt-pilot to complete? Can it go into a loop that he can't solve?(happened to me on gpt 4 were he given me over and over wrong answers) of course I know to correct what i asked but it required me to understand what is wrong...
2. What do you think will happen to restricted code base such as military systems or enterprise companies such as nvidia, space x or Tesla? Do you think they will share their code base with external ai service that doesn't have an open source and can scan their code bases and use its knowledge for their competitors? What the us military will say if a drone that was programmed by ai malfunction and target civilians or the forces that lunch him?
3. How will the AI be able to handle niche areas of programming, such as adobe extensions for example(I am developing extensions for living and while gpt and copilot help me develop extension, they do a lot of errors and I assume it has something to do with the fact that there are not that many of repos on GitHub for this set of problems I.e. automating adobe tools)
And how will they test the extension if it requires to run in a closed environment?
Thanks a lot for your videos! They are interesting and insightful!
Super interesting. Sharing on x
Anecdote: back in about 1990 the office next to mine which had 30+ COBOL developers suddenly emptied .. they had been laid off and been replaced by a software tool running on a minicomputer. The surviving team was about 4 business analysts who took customer requirements and entered them into the computer which then created/configured the mega app. There were also 2 nerdy techies locked in a little room. They maintained the minicomputer and coded odd fragments of custom code.
I expect the general sw development world to soon proceed in the same way : large teams of coders will be replaced by a handful of business analysts plus a handful of AI technical gurus.
One point to note : those still employed will be the best-of-the best. The mediocre and lazy will have no place in all this. Only a handful of the best newbies will be hired each year - and they will be apprentices to the techie and business gurus, with the intent for them to step into their shoes a year or two later.
Mediocre and lazy ones will go pump their muscles to rob the best of the best; this is natural selection. Many of us are not born geniuses, and believe me, we will find a way to interact with you because there are more of us than you might think. This will be a global collapse, as it will happen simultaneously in many professions: music AI, leading AI, voice AI, actor AI, scriptwriter AI. Optimistic scenarios about basic income are fairy tales for suckling children
I am a fullstack developer myself, and I definitely agree with this - the timelines being the hardest to predict accurately, but no doubt programming as we know it will change drastically, and rapidly.
Really intrigued by the "Jarvis" project and its capabilities in automating software creation with AI-assisted dialogue. I'm working on a hobby project of my own with a similar concept. I'm building a full-stack template, focusing on a robust, modular monolithic architecture (the most pragmatic for 95% of cases, imo) as its foundation. The goal is to train AI using a lot of examples based on this foundation, allowing it to autonomously develop and deploy enterprise-level software (again, built within the robust 'framework').
Seeing this "Jarvis" only makes it more obvious that further developing that concept in general has potential to significantly redefine the programming landscape. Genuinely looking forward to not having a job (even though I love programming) sometime in the future!
As a "full stack developer", have you actually used AI to actually write coherent code at any scale beyond a snake game? If so, I'd question your credentials as a developer.
@@brianclifton9399 as a "full stack developer", I have not tried writing a snake game at all. I'm using AI to generate bits and pieces of a larger system where I am still responsible for the overall architecture.
Giving it enough examples of different bits and pieces, it's quite capable of writing coherent code, in my experience. Not saying there's no challenges here though.
Question what you want, I'm still going to try.
True, all the things around the source code, packages, versions, libraries, plugins, framework configurations and on and on. Large tech companies can have whole teams just focusing on that, but to do it oneself, I recently found Maven complicated enough.
I remember trying to talk to our family's Apple II+ computer in the early eighties. Pretty much the only thing I could get it to say was "Syntax Error." If we can just get machines to do what we want by communicating with them in natural language, that will be much more... natural.
I'm also working with a team of AI agents for dev coding. I totally see the human in the loop, no question about that, because:
a) it's impossible that an initial prompt covers all the challenges that appear along the way,
b) while building and researching you're exposed to several scenarios that will change you human vision about what or how to build a project, and
c) AI will always need alignment as long as we humans coexist with them.
always and never are very risky words in nowadays ;/ we will see what ai can do in next year, but hope you are right!
I can't code without AI - for me this current sitaution is the beginning. I have medical background and now thanks to generative AI I may do data analysis and build apps:)
Thank you for the oldster shout out! I grew up programming just after lunch cards, but assembler and c were unforgiving mistresses. Memory management was the bane of my existence for many years.
Excellent and illuminating discussion, Matthew.
Appreciate the talk, some interesting points I agree with, but probably just as many I don't, read below if you are interested.
While 'programming' is the shorthand for writing code, I usually think of it as 'getting computers to do what we want' and that can involve a lot of minor detail that can start to get difficult to express in natural language, at least efficiently. "No code"/"Low code" solutions have been around for a long time, this use of LLMs does feel different, but I think the same limitations exist.
LLMs will get better at working out the intention of a human for common use cases, but that vagueness in natural language will bite. You rightfully pointed out how LLMs struggle with 'lost in the middle', which is what happens as soon as a project grows to beyond a toy project. Currently I believe people focusing on LLMs have 'Transformer Blindness'. Yes, the Transformer architecture is amazing, but this 'lost in the middle' problem I believe is a pretty fundamental limitation given how next tokens are predicted. I still agree this will be overcome, but not with just bigger and bigger Transformer models. Active Inference solves the problems we face with Agents and 'exploration vs exploitation' problem a lot better than just throwing more parameters at fixed models and IMO will be the next important leap in development of Agents. It still won't end the need for all manual programming, but it will drastically reduce the need for human programmers yes.
Knowing what you need when building a system is going to be a larger and larger chunk of the problem, which in itself will benefit from a programming background. While a vast amount of software can be built with a SQL database, backend API and web front end, these are not the only systems getting built.
Timelines are impossible to predict, but to me I think there is a fundamental limitation of natural language when combined with the specificity required by computers and building solutions. This makes me think that the "Human in the loop" of programming will persist for much longer than 10 years, but the impact on the contraction of the required software developer workforce is a big unknown, and very much worth highlighting which this video did well. Thanks again for the video, much appreciated!
This. I keep saying it: Computer programming IS "requirements documentation."
Not just programming but a lot of things are going to become obsolete
preferably humans become obsolete
Like in any era, nothing new
even the predictions of what will be obsolete will be obsolete when AI is making them for us... lol
Man, I am following the space for a while now, not just programming, automation will be in a every aspect of our life, my prediction in approximately two years or three people won't need to consult an expert for some tasks performed by expert today. The worst thing we are not prepared. Only the big tech are preparing by doing the massive layoff we are seeing now
We're already obsolete@@Danuxsy
Amazing video, very impactful with some key hooks such as 'this is the worst its going to get' and 'a billion people writing code'. That said, I do disagree, there is nuance here that is missed: such as in the demo where he said 'make it responsive', it's those sub-domain knowledge that will still be required. Another way of putting it is, we don't know what we don't know. PMs would need to specialise in order to know the right dials and permutations to ask of.
Its like when you ask an AI to redo an image, and it never seems to come out right - but if you watch an artist make the prompts, they get so much better results!
the most important questions I have as someone with an ok BG in Python are whether I still need to keep doing traditional programming studies, OR I should only be studying the best use of this technology?
Probably true in the long term, but programming will be one of the LAST skills to become obsolete. There will be a massive graveyard of other white-collar skills (writing, graphic artists, accounting, lawyers, etc.) that will either become obsolete or see a reduction in demand before programming faces a similar fate.
they said tha about artists and today the market for them crashed, you can even ask bing to create a logo free
Bro if I follow the same logic as coding with AI all industries will become more productive cheaper. That's a lot of money earned and saved wich will be used to expand the existing programms that we already have to make more benefits. So we will ironically need more white-collar skills to manage, correct and improve the AI of each sector you've mentionned.
If you were to ask me what would be the fastest, most efficient way to implement AI in a way that would allow non programmers to build complex programs, it would be a node based system, with each nodes connection being controlled by a separate agent. You click on the node, write in natural language what you want the node to do, then connect any nodes too it that would contain information you want that nodes function to use. This would allow the user to guide the AI to only the relevant parts of the code it needs to achieve a function so it doesn't get confused by the overwhelming amount of data. It would also allow the user to troubleshoot each function in their code by checking the input and output values of the nodes.
this is killer app, hope you or someone else will make it soon, very good idea!
@@MartinDlabaja thanks, it's one of a few ideas i have, unfortunately im just a guy i need to make a whole lot of money before I have the resourced to do that
@@homewardboundphotos yeah, and consequent stress load coming with that ...
Dude, fully agree. Love your content...
Incredible video! And 100% agree
Hi Matthew. First of all thank you for your videos. I learn a lot.
I know you make such a title only for a hype. But if not, do you really think that AI is able to understand requirements written by human beings for a more or less big project (have you ever read such a crap? :) If the most advanced LLM has only 25K tokens context how it will help with a codebase of 100MB or 1GB? I work on a project with more than 200 DB-entities and uncountable number of relations between. It's so funny to hear how AI will leave me without a job in close time, but when I ask gpt-4 some questions about what I'm working on it's not able to answer at all. AI-copilots trained on existing code, it's not able to generate something new, like a good architectural solution (i.e. relations between hundreds of classes working well) in a big project. At this moment AI is good only for one-page-DIYs to impress people who know nothing about development.
Coding is not the biggest part of development. And what is AI copilots really good in, is just autocomplete. It really helps to save time to remind syntax or similar, but obviously it can not create something working in real life without an educated human. So, let's talk about that in 5 years, not now :)
Finally: All thumbs up for your comment, that's exactly my opinion! Why can I only click the thumbs up button once?
And even the most average IDEs are more useful for code completion then actual AI models that also offer a lot of bs.
And if you really mean 200 DB-entries, that's nothing in the industry environment. - Even 200000 DB entries are not many.
I watched you using a tool the other day ans I realized how much your knowledge was necessey to wrangle the AI model. I came away from that demo with the impression not that human programmers are going away , but that they will transition into managers of AI assistants.
Who said that software and especially AI development was easy and for everyone? - All the GUI tools you watch youtube introductions about are not tools for productive software used in companies and the industry. These are almost all just "only" tools for testing and maybe for use at home.
@@henrischomacker6097 I certainly did not imply that in the least.
Your best video yet. Thanks
Matt you nailed it again. This video is a reflection of my thoughts. I am so glad to know you and your channel.
Mattew, please name a single intellectual activity that is NOT going to become obsolete when AGI come (I am seriously wondering what my next job could be)
fair point
those with very high degree of responsability, because politics.
What makes you think there will be AGI in your life time?
Agi is here, lmm s are general, the quality just needs to be improved@@dievas1
@@dievas1 how can you be sure you won't see AGI in your lifetime? are you very old?
Not finished with the video but I always thought AI would just enhance programmers and not completely end it
It will enhance us greatly...and then they won't need us anymore
@@matthew_bermanI won't be so pessimistic. I think you overestimate the basic computer literacy of the general population.
But I agree, having another degree next to IT would be probably the best.
BTW, that's the general "prompting proficiency ' ruclips.net/video/BKorP55Aqvg/видео.htmlsi=RxbKTJlcoIxdekEZ
Thats cause you listen to spokespeople no offense. It’s pretty obvious coding will transition to qa .. for geniuses
Matthew - I have enjoyed your AI insights and completely agree with your review. The AI trends you discuss put specific focus on a couple of levels I'd like to share from my experience with English language no-code tools from the late1970's to the present day. Entry level skills that need to be built from elementary school levels are critical thinking, and operational system analysis ( use case domain driven ) . Then learning to build prompt syntax excellence within the boundaries of specific LMs that are trained on data aligned with specific use cases. Example - we are experts in the SME manufacturing space. The manufacturing domain results in manufacturing an English language domain variant that can be easily used to train an LM. Today - this challenge is the easy part. The hard part today is getting code delivered with UI-UX that is more sophisticated delivering what you need to see and do "in 3 clicks or less" on multiple devices. And yes even that standard of UI-UX design can be taught as patterns to an LM. Rock on!
18:40 - Back in 2012, IBM demonstrated similar technology to a select group of people (I was there), asking AI to create an interface where people could post, share and like photos. Within seconds, AI created something very similar to Instagram. What is amazing to people "NOW" is a decade old technology, who knows what they can access now?
100% agree. Never learned to code - until chatGPT came out. That was the moment where I didnt need to learn every detail and could still create working code. Now im getting very fast with it and im learning the details as I build. Its truly a huge game changer.
My goal is to build my version of what you showed where I just assist and guide the AI while it does the work!
Funny how tables have turned. Programmers are now becoming QA Engineers. "write me some code..." now i test it.
That's exactly what I do
That is my workflow now.
@KirkKohler you are incorrect in your observation. I have put money on the table and also reqeusted an open debate on subject. If you are up to I challenge i.e back your subjective observation with tangle exemples beyond the one liner hyperbole I keen to take you on. @matthew_berman hope you are tracking as this is a tesosour what is to come if you are up for the debate.
Great video Matt!
i wonder how well the software will stand up to incremental tweaks; or if we get stuck making prototypes for which we can't keep them going. it may still matter if a human can read any of the code. would it be ok if LLMs just produced binaries?
Excellent video as always. It might affect individual lives negatively in the short term, but maybe if it pushes humanity as a whole forward without tying it to economic or political interests then there might be a net benefit to people as a whole.
Thank you and agreed!
I wish RUclips had a laughing react.
Domain knowledge will become the most important skill for companies. Even today, it is easier to teach and civil architect to use and build domain specific software than to hire a bunch of C++ programmers and teach them the domain of civil architecture
Agreed domain knowledge will become even more important!
But that’s the role of a PM
Well if AI gets so advanced to the point where it can replace software engineers I don't see why it can't replace other engineering disciplines especially since simulation software exists.
Domain knowledge is something that could also asked by an AI. So I still think good programmer is better than good domain knowledge.
@@matthew_berman maybe it’s just semantics but i think you mean the BA (business analyst) or business prime, and not PM are the SME (subject matter expert) who will validate and and define software requirements
As a junior level programmer planning to start my programming career I have to say your video is kind of scary for me. But even if I work with AI tools on daily basis I don't believe that they will completely replace human programmers. I think that there always have to be specialized people who understand the topic deeply. There are custom frameworks, design patterns in the tech companies as well as complex systems where at least a few specialists will be needed. But the juniors will probably be replaced completely as it is already happening. Quite frustrating for me but we have to face it. Let's see where it goes... Anyway, thanks for really great content.
I'm a generalist indie dev making a game in UE5, and as a generalist I have to fulfil all development roles, from the design, to the code, to the modelling, and even the 2D graphics. AI tools are absolutely going to be a game changer for solo devs and small studios with limited resources. I have a feeling AI 3D modelling is going to be the next big thing, because Luma AI's new Genie tool reminds me of midjourney's early days. It's already remarkably good, and of course this is the worse it will ever be.
I agree. I do software development, and can't imagine my job existing much longer. If there are software developers, we're going to be working a significantly different workflow than what we do now.
AI’s will have the best companies, the best lawyers, And when we go into negotiations? They’ll leave us laughing! And we’ll be thinking “what a bunch of great guys!”
The world is about to get bizarre.
Another par excellence research video brother! A+
Buddy this is not how this works😂
Thank you!
Swap “AI” with “wealthy”. Nothing has changed
@@lonewitness I understand but you clearly don’t think in exponential terms or in long term investment terms. If you think this train is going to end here? You obviously missed the run up on Nvidia. And why? A lack of research and imagination.
Now obviously the absolutely frustrating idiot savant stage of LLMs are like the first ten years of basic and pc write.
And nobody thought the king of search, Yahoo, would be knocked off.
Ten years from now is going to be very weird and even more weirder. Great time for long term investors though.
It’s like you didn’t watch the video brother. This was super informative. Really well put together.
@@thomassynths i think so. And why not? They are going to be the kings and barons. We will be the serfs in the fields.
I agree. Been working in automation and AI for several years and yes, AI will take over most if not all the programmers work in a very near future.
I love your videos. Thank you. I want to add a little spice to your analysis. Yes, like you I want to do everything myself, but it's not really possible. So I want to thank all the programmers who have made the components at the abstraction level where I do my stuff. BTW, I see I made comments here three weeks ago, so I must be rewatching. :)
Don't worry, this guy is just selling his AI courses. If you like programming, don't be discouraged from learning it, it is a very cool hobby, and a way to earn money. And it will be for the foreseeable future. Until AI can write and maintain the whole project for you, people will need programmers. And with its current capabilities, AI is not even close
Agree. Studying ML now. Don’t know what this guy is smoking. Will take a long time for AI to replace programmers. Maybe Copilot will litter code bases all over the world with weird errors that takes millions of human programmers to fix. 😅
I mean first of all, that is exactly what he said in the video.
And second, saying "AI is not even close currently" fails to take into account that it is a million times closer than it was one year ago. So what does that tell us about 1 year in the future?
I do agree that programming will become more and more accessible to those whom had never been exposed to the field, but to say that programming wont't exists isn't really something I would say is a complete inevitability. Like you said, it's a valuable skill to have and on top of that I'd even go so far as to say it requires an entirely different way of thinking, but there will always be that gap between people who "use" the technology and people who "create" the technology, which is probably a more distinct point to think about. Although, yes it will be much easier for people to get into building software only using natural language and A.I. there will always be people that exist on the bleeding edge aka those that are building the software using natural language, A.I. in combination with their knowledge of the underlying technology that builds the software for people that can only use natural language and A.I. You did mention that there will always be humans in the loop, but that there one day won't be and perhaps that's something I'd need to think about a little more, because personally I love programming and since the advent of these new A.I. tools, I haven't felt more alive and don't really see an end to a willingness to "program" something. I'm super excited to for what's to come next, especially when AGI and ASI become a reality. Combine that with brain computer interfaces and quantum technology and we have a recipe for a future in which we will able to compute and "program" the very fabric of reality itself.
Yes the education component at the entry and possibly intermidate level is reduced but that same information can be obtained from books and online courses.
If you can narrow your problem to a few statemennts (lines of code) that is the best approach however that still reqiures knowledge which you still need to obtain.
Finally, if you look it from a code completion or help system perspective then you be able to make the adjustments to the response.
I agree. Long ago I succeeded at customizing a snake game after watching one of your videos. Before that I had the experience of working with developers to write proprietary software and apps. In my experience, even when working with expert programmers, you have to be incredibly precise in your elucidation of the desired end result. The results are only as good as accuracy of the words chosen to define the project goals. This applies to AI devs, just like it applies to human devs.
In general and regardless of the domain, the better the prompt (whether to a fellow human or to an AI) the better the outputs.
New subscriber always watching your videos very helpful❤
if Ai gets so good that it replaces programmers, it's going to be changing how the world works basically replacing everyone so it's not just going to make programming obsolete if it get's that advanced and that's a big if
Even RUclips Channels like this will be gone. AI will be watching AI made content.
A big if? You're talking mad shit for someone who has no idea how fast ai is progressing. Give it 3 years and come back and reply to this comment. If I'm wrong. Mock me. If I ain't. Laugh at your own words
I'm studying MSc in Ai, i know how fast it's moving, the big question is if next token prediction can surpass or even reach human level which even ilya can't quite answer it yet @@GodbornNoven
I decided to make a career change for coding at 40 years old, went back to college and now I'm like 3 weeks from graduation and yeah I agree programming by humans will die. I wish I had known that a year ago but whatever that's life, at least I learned a bunch of machine learning skills which are more likely to last a couple more years.
Good luck, it's tough enough getting older and changing careers as it is, but with what's been happening this last year and knowing what's coming?
Crazy times are ahead.
I'm in my early 50's and have played around with the idea of getting into IT, coding, etc and am now glad that I'm such a procrastinator 😂
I look forward to seeing what technology brings us, I just fear that, at least in America, we aren't ready as a society for so many people to be out of work in the coming decade.
Yep, I've been coding for 20 years too.
An OG of a unicorn.
I totally agree. I was gonna try chase the opportunities but it's moving too fast. I think I'll just retire early 😮
I was going to try and create a channel showing older folks how to use some of these tools to help build a business, and while I was learning some of them they already became obsolete 😂
Guess I'll sit on the sidelines and cross my fingers we smarten up as a society and don't let too many people suffer needlessly when the jobs start to disappear.
Thx for this info, now I know how to think for the future when creating a new business.
Progamming is about to become a niche skill.
What I fear the most are those improvised programmers who don't really understand what's going on behind the scenes and are going to trigger data breaches and security issues without knowing how to fix them.
But that's not a bad thing ... i'll gladly help them for 3K per 8 hours of my slow human brain time ... XD
Joke aside, we can expect to see less diversity in projects code.
Programming may become reduced, but testing will explode, since ai does not really understand what it writes and how it meets requirements, not to mention its hallucinations
Non-technical person here. I seriously considered getting into programming 25 years ago but took a creative media pathway instead. I'm now looking at how AI can help me make computer games, at least basic ones to start with. Although I know next to nothing about coding, I have a feeling these changes will happen even faster than the predicted timeline in this video. They are already hugely impacting my industry in photography and video.
Do you think when we get to your 10 year estimated natural language to code that we will still have "programming languages" or will we grow enough trust in our AI counterparts that we let them code directly in machine language (either traditional binary or a quantum equivilent by then) thus cutting out the need to compile code at all?