@@SimGunther Literally all the time, partially because it--at least the basic version--can't browse the internet. But like @giarc1988 alludes to above, it's more than that.
I work as a software developer for 20 years. Yes, solving complex problem is main part of the job. AI can help by providing most common solutions and starting point what to google, but that's not always applicable, cause code is specific for applications that are far from something typical. Another problem is that AI is great for creating simple applications without context: here is piece of log file, I want to extract some information such as timestamp, values, confidence factors into CSV - ChatGPT can do it and even write some valid regexps. Then I want to plot these and/or find outliers - again, no problem, another 20 lines long Python script that works. But it really fails in tasks that are dealing with math, kinematics, geometry (my current area) and in general problems that have wide context. But knowing limitations means that AI can be still very useful for some auxiliary function, simple python code snippets (i'm not familiar with using pandas, sympy etc.), for rubber ducking or giving overview about some topics, typical solutions, algorithms or libraries. Bad stuff: it takes away simple, yet rewarding tasks that are nice to have to relax for half a hour.
In agent based systems you can assign different roles and your AI is checking (and testing) the code. The output is a lot better than just plain use of LLM.
AI is only as smart as its training data. It can’t actually think or reason. If it encounters a situation that it wasn’t trained to handle, it will hallucinate. AI is an amazing copilot, but it can’t replace humans ability to deal with unknown situations and use common sense.
On our interviews at my current job, we give people 4 coding questions that tend to use a specified input and output but leave the how completely open. They are designed to take an hour or two, but we let candidates who pass the first interview take them home, solve them in their own time with internet enabled. The second and final interview consists partly out of a code revoew to see how they take criticism
AI replacing large part of devs is just a matter of when, not if. Devs are too expensive compared to a high quality AI system which can be used as a service.
So far it's a good tool, but I find it often uses sub-optimal solutions and sometimes simply forgets to include certain requirements nor can it handle complex requirements. So no where near ready to replace developers, but certainly can be a huge efficiency boost for already experienced developers. Also it's pretty good to use in place of third party existing documentation or code reviewing.
AI for me is a great tool. My productivity increased exponentially. But i use it mainly as an assistant, not something i trust. I will never trust AI. Never. It's an algorithm, running on a machine, you never know when it can fail due to various reasons, and if it does the responsibility for the bad code is on me. I will use only the code i can review and test in my projects.
It's no longer worth it. But not because of ai. The market is flooded with developers and there are hardly any jobs. Companies startup, get funding, hire tons of developers, and then lay them all of and outsource for maximum funds. There are hardly any jobs and seniors are taking mid level and junior roles because of this.
I think there is a big problem with how that research paper is being interpreted. If you pay attention, the trends that are discussed do not start in 2023, as you'd expect for something released in late 2022! You can see the shift ongoing in 2021, 2022. You don't even need to know that there was a massive boom in hiring of junior software engineers in that period to understand there is a confounding factor here. I am willing to bet that the main driver behind those trends is in fact the massive increase in the size of software engineering teams we saw during the pandemic, which often meant hiring people who were very junior, often switching careers, with not a lot of training especially in terms of best practices. That said, I agree that AI is best used today as a semi-trusted colleague. I've been using it to research errors I get when writing Rust, and it's been a massive help. I am now able to learn new concepts and understand things in detail much quicker than I would before having to dig through documentation and forums. The main thing is not to copy/paste blindly, but make sure you understood what the code does.
Say it gets to point where you give it a 200 page specification to build a non-trivial application and it spits out a full application. Then 6 months later it needs a change. You update the specification and feed it to the AI and it spits out a new version but since the AI has also probably been updated there is almost no chance the new application can just be dropped in as a replacement.
EXACTLY CORRECT! Generative systems give you ONE shot to provide "requirements" for them to spit out some arbitrary but COMPLETE solution. If you need to alter or add 5% of the requirements later, these systems will design an entirely NEW solution with zero likelihood its data schema and services will be backward compatible.
The specification that perfectly describes a non-trivial application matches the complexity of the application that it should lead to. So it is programming in a language that's shit for programming.
I could go on for days listing all my issues with AI but the biggest problem by far is that it will never just respond with, "I don't know how to do that" it will just keep spitting out code telling you that it is what you need, and in many cases you can see the glaring mistakes in the code as its writing it.
Just a while ago I asked ChatGPT to write an encryption function for me. About 10 prompts later I knew that ChatGPT can't code, it can just mash together what it has already seen somewhere. And that was one function. Until "AI" will be able to read und understand thousands of lines of code (that includes how the UI will look and respond) and develop it further according to the users wishes, I will be retired.
For me, AI is some sort of search engine. It saves me time googling around and I take everything from AI with a huuuuge grain of salt. Some of the generated code flashes some cool ideas - and some suggestions are straight wrong. By now, AI is a tool. And I think we should use it this way.
I used to look at Stack Overflow solutions and then implemented this into my code. However I analyzed the code and improved it and make it fit into my existing code. I am doing the same thing with AI code, I look at it, then manually modify it and make it even better. But I can only make it better because I have a history of creating code out of nothing. My brain is trained to create code since the 1980's :-)
Well, we’re all discussing “what’s happening today,” but these technologies are steadily improving. it means our fears may become reality in the future. currently, these technologies struggle with reasoning and logic in conversations, but a lot of work is happening behind the scenes. even in the case of "Devin," the company is diligently working to fix issues and release an improved version.
I feel like im tuning into a lot channels where I'm more experienced than the author, sorry to say but I feel like its the case , does anyone know where can I get some senior content 😅
No your not because there is no such thing as a professional. There is no bar no test nothing to verify you are indeed professional. You are just saying you've been paid to write code....so you're opinion is no different than anyone else...sucks huh
Not everyone of course but I think a lot of people want it so badly to be something it isn't and in the meantime aren't using it for what in my opinion it's good at which are things like handling unstructured data. I think no matter how good it becomes though the show is still going to be run by actual people that know how to utilize whatever tools they have available better than others.
Ask chat gpt for two large primes, tell it to multiply them, tell it to divide the result by one of them. It's always wrong and it always pretends to be right. Useless...
I just asked ChatGPT if Hannahanah is a palindrome. It correctly answered: No. Then I asked if it could turn it into a palindrome by adding one character. It said: Yes, and proceeded to explain to me how adding an n in the middle would produce this palindrome: Hannahananah. Because of course it did.
You are skipping the sheer amount of CO2 that is emitted when training LLMs, no need to address any form of frontend tool since the critical failure is unrelated to development but why this trend is not sustainable. You can research this, there are GOOD scientific papers on this, besides, ever heard of living human brain organoids wired into silicon chips? Yes, as result of the climate concerns resulting from training LLMs they are now focusing on this and it is dark. I cannot post links since they will delete this comment. Besides, use your skills and go into security as in OSINT. If you want to be good at OSINT/OWASP/etc. you need to know programming.
I disagree about AI making code itself worse... to start with, yes, it will dissuade in depth learning, and later, the pro coders and code still being around, AI will learn to condense code in super optimal, minimal instruction, hyper optimized, well annotated, readable stuff, with visual systematic flow representations of what the code elements do, loops and geometry representied in 3d or multidimensional diagrams so even the human gets to have a multi modal view of an ai code brain.
AI is a tool, the results depend on how you use it. The efficiency of coding is going up tremendously. Don‘t use it and you are not competitive anymore. If you have a team of 10 engineers, you might be able to reduce the team to 5-6 and keep same output.
We keep moving the goal posts on AI, so many software engineers in particular seem to think that an AI must be god-like and must be 99.99% correct on 0-shot responses to think their job might be threatened. Now answer me this, As a software engineer, how many bad human software engineers do you know? Now this is where the fun begins…. Now ask yourself this, honestly, what percentage of software engineers may think YOU are a bad software engineer? Devin, ChatGPT, Claude, etc. are as dumb as they are going to be today as they ever will be, they will only get better, faster, and cheaper, and most likely at an accelerated rate. Will you?
Like stackoverflow. Okay. So at some point, you will ask it a question and it will link you the answer it gave another person 10 years ago which does not work anymore but yet it is mad at you and will end the conversation.
The reason why humans disagree about religion etc is hallucination, they are dead convinced of themeselves. flame wars online, disagreements, it's all becauise people are convinced they are right and they may have just m issed a word of context.
You can learn a lot from it but use it to guide you to the correct path not beliving everything it says. Use little bit of logic, usually the stupid answers it gives are plain stupid for real.
this is just a basic versionof AI in 3 years with all the money invested will improve dramatically, im machine learning engineer and this will be fixable when we have more gpus and data from code repositories. stop focusing con coding and more on solutioning. please do not try to fight on code and focus on solutioning and new technologies.
Industry goes in waves, CEOs fan the hype trains for capital, the reason why A.I is hot is because companies using A.I are getting rewarded cause VCs are convinced it's the future which makes CEO's recalibrate their business targets. A.I is great for narrow tasks. PWA, Alexa, Cross-Platform, IoT ... when it's been milked it will still have a place, but they wont be hiring ML eningeers to clean CSVs anymore and adjust hyperparms on GCP..
AI can read and summarize things, but if I need to write code AI only slows me down. I have to fix everything it spits out. Additionally, everything it gives me I must independently. If AI gives you a perfect answer, it's most likely plagiarized.
![PySpark Full Course [2024] | Learn PySpark | PySpark Tutorial | Edureka](http://i.ytimg.com/vi/sSkAuTqfBA8/mqdefault.jpg)
I would rather see an AI say, "I don't know." rather than a confident wrong answer! We get that enough from the management types!
Yeah I actually wish it could be more honest and maybe be more explicit on where the information is actually pulled from
@@codingwithdeeAnd if you ask it for sources, it gives you fake links. 🙄
@@LouisianishSadly, it has given sources that don't even exist 😢
Except it can't do that because it doesn't know it is "wrong". If it could know that, it would be far better than it currently is
@@SimGunther Literally all the time, partially because it--at least the basic version--can't browse the internet. But like @giarc1988 alludes to above, it's more than that.
I work as a software developer for 20 years. Yes, solving complex problem is main part of the job. AI can help by providing most common solutions and starting point what to google, but that's not always applicable, cause code is specific for applications that are far from something typical. Another problem is that AI is great for creating simple applications without context: here is piece of log file, I want to extract some information such as timestamp, values, confidence factors into CSV - ChatGPT can do it and even write some valid regexps. Then I want to plot these and/or find outliers - again, no problem, another 20 lines long Python script that works. But it really fails in tasks that are dealing with math, kinematics, geometry (my current area) and in general problems that have wide context. But knowing limitations means that AI can be still very useful for some auxiliary function, simple python code snippets (i'm not familiar with using pandas, sympy etc.), for rubber ducking or giving overview about some topics, typical solutions, algorithms or libraries. Bad stuff: it takes away simple, yet rewarding tasks that are nice to have to relax for half a hour.
Please your WhatsApp number 🙏
In agent based systems you can assign different roles and your AI is checking (and testing) the code. The output is a lot better than just plain use of LLM.
AI is only as smart as its training data. It can’t actually think or reason. If it encounters a situation that it wasn’t trained to handle, it will hallucinate. AI is an amazing copilot, but it can’t replace humans ability to deal with unknown situations and use common sense.
On our interviews at my current job, we give people 4 coding questions that tend to use a specified input and output but leave the how completely open. They are designed to take an hour or two, but we let candidates who pass the first interview take them home, solve them in their own time with internet enabled. The second and final interview consists partly out of a code revoew to see how they take criticism
AI replacing large part of devs is just a matter of when, not if. Devs are too expensive compared to a high quality AI system which can be used as a service.
So far it's a good tool, but I find it often uses sub-optimal solutions and sometimes simply forgets to include certain requirements nor can it handle complex requirements. So no where near ready to replace developers, but certainly can be a huge efficiency boost for already experienced developers. Also it's pretty good to use in place of third party existing documentation or code reviewing.
AI for me is a great tool. My productivity increased exponentially. But i use it mainly as an assistant, not something i trust. I will never trust AI. Never. It's an algorithm, running on a machine, you never know when it can fail due to various reasons, and if it does the responsibility for the bad code is on me. I will use only the code i can review and test in my projects.
This mirrors my experiences ar work. Great video!
It's no longer worth it. But not because of ai. The market is flooded with developers and there are hardly any jobs. Companies startup, get funding, hire tons of developers, and then lay them all of and outsource for maximum funds. There are hardly any jobs and seniors are taking mid level and junior roles because of this.
I think there is a big problem with how that research paper is being interpreted. If you pay attention, the trends that are discussed do not start in 2023, as you'd expect for something released in late 2022! You can see the shift ongoing in 2021, 2022. You don't even need to know that there was a massive boom in hiring of junior software engineers in that period to understand there is a confounding factor here. I am willing to bet that the main driver behind those trends is in fact the massive increase in the size of software engineering teams we saw during the pandemic, which often meant hiring people who were very junior, often switching careers, with not a lot of training especially in terms of best practices.
That said, I agree that AI is best used today as a semi-trusted colleague. I've been using it to research errors I get when writing Rust, and it's been a massive help. I am now able to learn new concepts and understand things in detail much quicker than I would before having to dig through documentation and forums. The main thing is not to copy/paste blindly, but make sure you understood what the code does.
Say it gets to point where you give it a 200 page specification to build a non-trivial application and it spits out a full application.
Then 6 months later it needs a change. You update the specification and feed it to the AI and it spits out a new version but since the AI has also probably been updated there is almost no chance the new application can just be dropped in as a replacement.
EXACTLY CORRECT! Generative systems give you ONE shot to provide "requirements" for them to spit out some arbitrary but COMPLETE solution. If you need to alter or add 5% of the requirements later, these systems will design an entirely NEW solution with zero likelihood its data schema and services will be backward compatible.
The specification that perfectly describes a non-trivial application matches the complexity of the application that it should lead to. So it is programming in a language that's shit for programming.
I could go on for days listing all my issues with AI but the biggest problem by far is that it will never just respond with, "I don't know how to do that" it will just keep spitting out code telling you that it is what you need, and in many cases you can see the glaring mistakes in the code as its writing it.
Always love your top notch analysis Dee. Amazing
Overall i would say yes its getting sloppy. Im newly graduated and wow ChatGPT was a mess as a teachers aid.
Thank you for the analysis Dee. I agree with every part.
Just a while ago I asked ChatGPT to write an encryption function for me. About 10 prompts later I knew that ChatGPT can't code, it can just mash together what it has already seen somewhere.
And that was one function. Until "AI" will be able to read und understand thousands of lines of code (that includes how the UI will look and respond) and develop it further according to the users wishes, I will be retired.
For me, AI is some sort of search engine. It saves me time googling around and I take everything from AI with a huuuuge grain of salt. Some of the generated code flashes some cool ideas - and some suggestions are straight wrong. By now, AI is a tool. And I think we should use it this way.
You’re too reasonable
I used to look at Stack Overflow solutions and then implemented this into my code.
However I analyzed the code and improved it and make it fit into my existing code.
I am doing the same thing with AI code, I look at it, then manually modify it and make it even better.
But I can only make it better because I have a history of creating code out of nothing.
My brain is trained to create code since the 1980's :-)
Well, we’re all discussing “what’s happening today,” but these technologies are steadily improving. it means our fears may become reality in the future. currently, these technologies struggle with reasoning and logic in conversations, but a lot of work is happening behind the scenes. even in the case of "Devin," the company is diligently working to fix issues and release an improved version.
I feel like im tuning into a lot channels where I'm more experienced than the author, sorry to say but I feel like its the case , does anyone know where can I get some senior content 😅
I use AI for coding every day. I am a trained professional. For everyone else, please refrain from using AI.
No your not because there is no such thing as a professional. There is no bar no test nothing to verify you are indeed professional. You are just saying you've been paid to write code....so you're opinion is no different than anyone else...sucks huh
How do you know they don't have actual training? Which would add weight to their opinion. @@troy5370
@@troy5370the difference between a hobbyist photographer and a professional photographer is the professional makes a living on it
I need help
Not everyone of course but I think a lot of people want it so badly to be something it isn't and in the meantime aren't using it for what in my opinion it's good at which are things like handling unstructured data. I think no matter how good it becomes though the show is still going to be run by actual people that know how to utilize whatever tools they have available better than others.
It wont replace a soft ware engineer today but in 20 years it will. Learning code is not a priority for my 6 yr old.
I think eventually AI will replace coders but will create more demand for software architects.
Ask chat gpt for two large primes, tell it to multiply them, tell it to divide the result by one of them. It's always wrong and it always pretends to be right. Useless...
Sometimes its very helpful. And sometimes it is so close to being helpful you want to punch your screen.
I am still interested in software engineering
1:19 made me chuckle
I just asked ChatGPT if Hannahanah is a palindrome. It correctly answered: No. Then I asked if it could turn it into a palindrome by adding one character. It said: Yes, and proceeded to explain to me how adding an n in the middle would produce this palindrome: Hannahananah. Because of course it did.
You are skipping the sheer amount of CO2 that is emitted when training LLMs, no need to address any form of frontend tool since the critical failure is unrelated to development but why this trend is not sustainable.
You can research this, there are GOOD scientific papers on this, besides, ever heard of living human brain organoids wired into silicon chips? Yes, as result of the climate concerns resulting from training LLMs they are now focusing on this and it is dark.
I cannot post links since they will delete this comment.
Besides, use your skills and go into security as in OSINT. If you want to be good at OSINT/OWASP/etc. you need to know programming.
I disagree about AI making code itself worse... to start with, yes, it will dissuade in depth learning, and later, the pro coders and code still being around, AI will learn to condense code in super optimal, minimal instruction, hyper optimized, well annotated, readable stuff, with visual systematic flow representations of what the code elements do, loops and geometry representied in 3d or multidimensional diagrams so even the human gets to have a multi modal view of an ai code brain.
AI is a tool, the results depend on how you use it. The efficiency of coding is going up tremendously. Don‘t use it and you are not competitive anymore.
If you have a team of 10 engineers, you might be able to reduce the team to 5-6 and keep same output.
We keep moving the goal posts on AI, so many software engineers in particular seem to think that an AI must be god-like and must be 99.99% correct on 0-shot responses to think their job might be threatened. Now answer me this, As a software engineer, how many bad human software engineers do you know? Now this is where the fun begins…. Now ask yourself this, honestly, what percentage of software engineers may think YOU are a bad software engineer?
Devin, ChatGPT, Claude, etc. are as dumb as they are going to be today as they ever will be, they will only get better, faster, and cheaper, and most likely at an accelerated rate. Will you?
AI is an improved version of StackOverflow. There is so much time until Skynet takes over
Like stackoverflow. Okay. So at some point, you will ask it a question and it will link you the answer it gave another person 10 years ago which does not work anymore but yet it is mad at you and will end the conversation.
14:47
AI will demonetize software production, lowering the bar for enthusiast coders and boosting open source software production.
I lilke your channel it'sinteresting you sound south African and Indian, although i dont know you coul dbe sri lankan or etheopian dutch lol.
The reason why humans disagree about religion etc is hallucination, they are dead convinced of themeselves. flame wars online, disagreements, it's all becauise people are convinced they are right and they may have just m issed a word of context.
The answer is NO. I can't tell you how many times its wrong doing simple C coding. Maybe in like 5 years it might be worth using.
1,5 Zhlédnutí Zabezpečení Žabka
Not yet.
You can learn a lot from it but use it to guide you to the correct path not beliving everything it says. Use little bit of logic, usually the stupid answers it gives are plain stupid for real.
Sometimes I think when trump blabbers on he must in fact be an AI, if you are offended, replace variable name with biden!
this is just a basic versionof AI in 3 years with all the money invested will improve dramatically, im machine learning engineer and this will be fixable when we have more gpus and data from code repositories. stop focusing con coding and more on solutioning. please do not try to fight on code and focus on solutioning and new technologies.
Industry goes in waves, CEOs fan the hype trains for capital, the reason why A.I is hot is because companies using A.I are getting rewarded cause VCs are convinced it's the future which makes CEO's recalibrate their business targets. A.I is great for narrow tasks. PWA, Alexa, Cross-Platform, IoT ... when it's been milked it will still have a place, but they wont be hiring ML eningeers to clean CSVs anymore and adjust hyperparms on GCP..
Glue aint hallucination it only copied answer of a playful redditor
O2
AI can read and summarize things, but if I need to write code AI only slows me down. I have to fix everything it spits out. Additionally, everything it gives me I must independently. If AI gives you a perfect answer, it's most likely plagiarized.
14:50