👉 To try everything Brilliant has to offer for free for a full 30 days, visit brilliant.org/CodetotheMoon/ You’ll also get 20% off an annual premium subscription. They have courses on LLMs and Quantum Computing!
You said that even if you are sure that the generated code will have at least one bug in the code, it will be faster to fix it, one is not a static number. Some errors can be nested much deeper than in human-written code, errors that never appear until that one day when they have catastrophic consequences. Humans also make a lot of mistakes, but it seems to me that when you use ai, it's much easier to keep that mistake in the code base. I'm not saying that no one should use AI generation/support, but I think there will be industries that will never use it. AI gives you broader and faster access to knowledge, but it's much more likely to introduce an error that's hard to detect than with classic sources, especially when it comes to information that's not as readily available. I hope this will change, but at this point I don't recommend using AI with niche technologies.
Absolutely, the number of bugs in LLM generated code is not limited to 1. The point I was trying to make there was that folks often use "LLMs write buggy code" as an argument as to why LLM code gen has no utility. I was trying to explain that this is not the dealbreaker many think it is. The bottom line is that to use code gen effectively, you *need to be very good at reviewing the code it wrote*. Any sort of complacency there will lead to big problems. This is what seems to get folks in trouble, and then code gen gets a bad reputation because of it. I think that's part of the mindset shift that is needed.
Two complexifications. 1) Snyk has shown that the code quality goes down in roughly a "both ends of the U" pattern - so "beginning intermediate" devs can use AI tools but not beginners - they will be too gullible or ignorant. 2) "Bugs" are not the only consideration - it depends on what bugs there are, and whether they are eradicable etc.
I tried multiple LLM and running a few locally with ollama. I found it's mostly useful for boilerplate code, generate tests, better code completion in some cases. For complicated business logic, understanding context and constraints, it's not very helpful then code by hand. And it works better for widely adopted languages like JavaScript and python. But not so much for niche/new languages like zig and dart. But it helps sometimes with onboarding tools and languages, instead try to find answers with search/stack overflow and go through docs, it might be faster just prompt it for answer and explanations.
This is generally in line with my experience as well. I've found it to be fairly good at Rust too, which is important to me. It definitely can't pull a rabbit out of a hat so to speak, but for getting the ball rolling on a task it is usually pretty great.
5:39 reading garbage instead of reading actual working examples from docs... yeah 7:19 "one bug" is a wishful thinking. You just found "this one" and was unable to find more because you read unfamiliar code
I think a key component of using AI code gen effectively is thoroughly understanding what the AI wrote. It seems like it is cases where that isn't true where people get themselves into trouble. The point I was trying to make is that some seem to think AI code gen will only be useful when it generates perfect code 100% of the time, which just isn't true.
GPT argued with me for 10 minutes last night that 12-1=9. LLMs are not AGI. They are probability based token connectors. Stupid ones. The more high-level the topic, and the more training data on that high-level topic, the better they do. (Unless its 2nd grade math apparently)
Haha that's pretty funny. Yeah, they suck at math. I still don't fully understand why this is, but it seems like there are ongoing efforts to fix this. I believe there is a standard LLM benchmark suite which focuses solely on math.
Exactly. Math is a solved problem. Language models are not meant for math. Nor are they meant for logic or reasoning. They can explain subtraction to you but not solve a concrete subtraction for you. That's what computation is for, not model inference.
I agree that my productivity has reached an all-time high. I use various language models to assist me in summarizing papers and documentation. I want to continue learning, so I don’t allow the language model to write complete code for me; instead, I use it to improve my work and point out any errors.
You are definitely on to something at 8:30. If there's one thing I'd like to see more of in the programming scene, is modular development and a better focus on architecture and design.
Great discussion! Here is what I believe is missing: - Not all digital execution is code - at least not as we have come to know it over the last 20 years. - My Tesla navigates using a neural network that does not resemble code as you have described. - The value of code in the 'new world' is guaranteed repeatable execution. Said another way, there is no judgement - just script processing. - I believe more and more digital execution will result from the will of neural networks. - Coding will be largely relegated to 'tool scripting' where repeatable execution might be wanted/required. My 2 cents, Chuck Boecking
For me, personally, I only use it to get some idea or need a question answered with a example set of code. I have completely turned off AI code suggestions, as they interfere with my train of thought when I code, so for me generally it has been good for "common" questions without many contextual variables needing to be included. It can help people, and I've definitely seen that, but sometimes it also gets to a point where you have wayyy to much context needed for it to work in a proper manner.
wow, that's interesting! never met one of those. The first I'd heard of this is a recent interview with the creator of the Odin language - apparently he is in that camp as well.
LLM = Large *Language* Model. Not Large Computer Science Model, not Large Software Development Model, not Large Software Engineering Model, not Large System Design Model. LLM's do not know how to build applications. They just know how to repeat the language patterns they have been trained on from a large pool of highly relevant examples. When no such examples exists, LLM's will make things up. Then the LLM makes a mistake, it has no way of recognizing it's mistake. When you are building commercially viable software, there are no pre-existing examples to train your LLM on. You are the one inventing the first example.
You raise some great points. I think LLMs are only a part of what AI code generation tools are going to consist of (and to some extent already do). It seems like the best way to improve the quality of generated code right now is via giving them agency and thus a feedback loop, so they can run their code and be given the opportunity to fix any mistakes.
I liked the video. I appreciated your perspective and acknowledgment of biases. Personally didn’t find anything wrong with your line of reasoning and am inclined to agree with many of your points. Great work
4:25 Churn rate of code, bugs introduced, features delivered (without churn), there is visible metrics and papers on the topic, not using as part of the research for this video show the real message behind of it. Just sad to watch it.
thanks for the feedback! how do you define churn rate of code? features delivered is not a good metric, because they are not equal in scope and value. On the surface bugs introduced seems like it would be easier to nail down, but I don't think it is either. For example there could be "bugs" which have negligible or debatable customer impact - bugs that we might willingly accept for an X% increase in velocity. Would love a link to the papers you are referring to!
@@codetothemoon chrun, code removed close after inserted into the codebase, there is a paper on the topic using the public data from Github, look for it. Features delivered "before" and "after", also have a paper on the topic, with a control sizeable control group. Bugs introduced by automated, semi-automated and no AI tools, there is also a paper on the topic. There IS metrics, there is for more than a years a conversation on this topic. So please, don't spread misinformation like this video mostly was. Feels good isn't a measurable metric.
@@ShinSpiegelThat's sadly automatic. Any lengthy nuanced comment is almost guaranteed to be deleted for being "controversial". Just my own observations. I've tested this by posting my own long comments from before RUclips got overzealous with the anti-spam and they get auto deleted.
@@LiveType yeah, already unsub on this channel, tired of all the AI things, and the la k or research from lots of content creator. I’ll be more strict on the creators I follow, even if this mean follow no one.
I think AI is great in terms of refreshing your knowledge. It's like a human that knows everything. I was exploring solana blockchain and boy was it hard, but the chatgpt got me started.
definitely agree! It seems like most folks have a very positive experience using it as a means of learning. Its utility there seems far less disputed than for code generation.
For me AI automates a lot of the busy work with coding, like writing boiler plate code and formatting data. So I would say there is definitely an increase in productivity. I think without knowing how to code in the first place, and knowing it well, it becomes a lot more difficult to use AI to code. It may not produce exactly what you want, will make mistakes, etc. If you do not have the skills to properly debug and fix the code manually then you will spend hours trying to get chat gpt to get it right. I do think relying too much on AI for coding will lead to something like what we see with calculators and arithmetic, where our ability to do the "grunt work" decreases substantially. If AI reaches the point that it can totally replace programmers, I think it will be good enough to replace a lot of other occupations as well. LLMs are not restricted to text anymore, they can parse visual input, audio input, and if you map the sensor data into something the AI can understand it can process that too. We honestly might be seeing "smart" robots in our life times, and then who knows what happens.
The really big question around AI is cost. And it’s a big one that is being ignored. All the the talk about the value of AI assumes that the current cost to use AI will remain the same. But we have no reason to believe that’s true. In fact, we have every reason to believe that eventually AI providers will need to generate profit. OpenAI and google and Amazon can’t Contis to lose billions of dollars per year on their AI services forever
I have a love/hate relationship with AI. I've been using it with Rust to improve my Rust skill and the amount of nonsense it gives is unbelievable, I feel like using it makes me more dumb than not using it. I also encountered a situation where I ask it to convert a float or integer into a comma separated string using regex ( I hate regex). and after more than 30 mins and countless attempts, it keeps looping into giving a regex that either adds a comma including in the decimal part if its more than 3 or works fine in float only but doesn't work with integer. but Id argue its a good tool if you already know what you're doing.
This has also been my own experience while using 4o for Swift. It was great for learning basic Swift/SwiftData/SwiftUI. But as my application got more complex and I wanted to do more with it. AI became useless. I learned more from Apples own projects they release than from what the AI was telling me I should do. I encountered numerous incorrect statements from 4o that made my understanding of Swift also incorrect.
How can we know that AI is going to be GREAT at code in the future? Answer: They have only been CRAZY funded and researched for about 3 years, and they're already good enough to keep moving the goalpost and tech bros has to come back and give new claims about things they will never do... Remember a year ago when everyone said AI will NEVER be able to do math. Boom, half a year later they beat the majority of experts. Just like any other profession this will continue to disrupt the field, there is no point in
Good one.. To me to be honest software developers are not inventing new stuff in daily basis.. most of the things they do is repetitive. For example a web developer can only produce a website or maintain it or upgrade it to facilitate new functionalities.. that can be easily replaceable via AI. Where the AI can not compete is genuine idea.. For instance you have a business idea where you are thinking to bring a completely new tool for UX design out of the todays norm.. AI is pretty much useless there.. But I do agree with you... for everyday task like you said converting to another database.. pretty sure somewhere someone already faced that issue, there fore its not original one.. what you are doing is repeating the steps that other guy have faced thousand miles away from you. In that case AI will replace those human labor once it self correct the buggy stuff and reaches the ability to test out its own stuff..
AI can't generate 100% reliable and high-quality code. That's a fact. If I were a CEO, I would probably only care about saving money, so AI generating acceptable code would be enough for me. However, as a researcher, I want my code to be the best it could possibly be, so I have to write it myself. And yeah, I make mistakes, too. The difference between me and AI is that I can learn from my mistakes, and I don't need a huge data set and a power plant to do that.
Agree with your first point! I do think to get the most out of these tools we need to be almost militant when it comes to reviewing the code they generate. Re: difference between you and AI - I think you may have needed a large data set as well, though maybe not quite as large. Everything you learned prior to writing the code that enabled you to write it - including natural language, programming concepts, etc. But you surely did it with quite a bit less power than LLM training clusters - human brains are notorious for their low power requirements in comparison 😎 That said, totally see your point that for many people and in many domains (maybe research is a canonical example of one) just writing it yourself can make a lot of sense.
You don't need a power plant because you aren't responding to millions of people across the world at the same time. It's not like it takes a power plant to run a single instance of an LLM.
@@codetothemoonThe translation is pretty decent. The problem is the horrible voice generated. Sound too bad... If I didn't know your channel, I just pass it.
Like most humans I am guilty of this. I feel like there is a negative implication here but I am unclear what it pertains to. My creation of this video?
The same way AI can copy art and output a full picture, AI will copy full software, without needing for code, only with the executable. You will prompt apps directly. At that point coders will be fully replaced.
that's an interesting idea - llms writing machine code directly. seems like something they'd be capable of, I guess the problem is it wouldn't then be reviewable by humans. I'm skeptical that software developers will be completely replaced - though I wouldn't be surprised if traditional coding takes a back seat to other skills.
In the short term, yes there will be a shift in how software developers work and there may be some new roles being created. In the longer term all software will be created by artificial intelligence. Including triple-A games. Everything. Writing code may be left as a hobby, but humans will not be required. The comparison to the replacement of physical work does not transfer over to the replacement of intellectual work. There is nowhere to go where the AI can't follow, or even take the lead.
👉 To try everything Brilliant has to offer for free for a full 30 days, visit brilliant.org/CodetotheMoon/ You’ll also get 20% off an annual premium subscription. They have courses on LLMs and Quantum Computing!
The video looked interesting, but the flashes of light combined with the sound of drawers closing, just to keep the attention, got the better of me.
thanks for the feedback, it really helps!
Less flashing bro.
thanks for the feedback, definitely won't be using the glitch effect as much anymore!
You said that even if you are sure that the generated code will have at least one bug in the code, it will be faster to fix it, one is not a static number. Some errors can be nested much deeper than in human-written code, errors that never appear until that one day when they have catastrophic consequences. Humans also make a lot of mistakes, but it seems to me that when you use ai, it's much easier to keep that mistake in the code base. I'm not saying that no one should use AI generation/support, but I think there will be industries that will never use it.
AI gives you broader and faster access to knowledge, but it's much more likely to introduce an error that's hard to detect than with classic sources, especially when it comes to information that's not as readily available. I hope this will change, but at this point I don't recommend using AI with niche technologies.
Absolutely, the number of bugs in LLM generated code is not limited to 1. The point I was trying to make there was that folks often use "LLMs write buggy code" as an argument as to why LLM code gen has no utility. I was trying to explain that this is not the dealbreaker many think it is.
The bottom line is that to use code gen effectively, you *need to be very good at reviewing the code it wrote*. Any sort of complacency there will lead to big problems. This is what seems to get folks in trouble, and then code gen gets a bad reputation because of it. I think that's part of the mindset shift that is needed.
Two complexifications. 1) Snyk has shown that the code quality goes down in roughly a "both ends of the U" pattern - so "beginning intermediate" devs can use AI tools but not beginners - they will be too gullible or ignorant. 2) "Bugs" are not the only consideration - it depends on what bugs there are, and whether they are eradicable etc.
I tried multiple LLM and running a few locally with ollama. I found it's mostly useful for boilerplate code, generate tests, better code completion in some cases. For complicated business logic, understanding context and constraints, it's not very helpful then code by hand. And it works better for widely adopted languages like JavaScript and python. But not so much for niche/new languages like zig and dart. But it helps sometimes with onboarding tools and languages, instead try to find answers with search/stack overflow and go through docs, it might be faster just prompt it for answer and explanations.
This is generally in line with my experience as well. I've found it to be fairly good at Rust too, which is important to me. It definitely can't pull a rabbit out of a hat so to speak, but for getting the ball rolling on a task it is usually pretty great.
5:39 reading garbage instead of reading actual working examples from docs... yeah
7:19 "one bug" is a wishful thinking. You just found "this one" and was unable to find more because you read unfamiliar code
I think a key component of using AI code gen effectively is thoroughly understanding what the AI wrote. It seems like it is cases where that isn't true where people get themselves into trouble. The point I was trying to make is that some seem to think AI code gen will only be useful when it generates perfect code 100% of the time, which just isn't true.
GPT argued with me for 10 minutes last night that 12-1=9. LLMs are not AGI. They are probability based token connectors. Stupid ones. The more high-level the topic, and the more training data on that high-level topic, the better they do. (Unless its 2nd grade math apparently)
Haha that's pretty funny. Yeah, they suck at math. I still don't fully understand why this is, but it seems like there are ongoing efforts to fix this. I believe there is a standard LLM benchmark suite which focuses solely on math.
Exactly. Math is a solved problem. Language models are not meant for math. Nor are they meant for logic or reasoning. They can explain subtraction to you but not solve a concrete subtraction for you. That's what computation is for, not model inference.
I agree that my productivity has reached an all-time high. I use various language models to assist me in summarizing papers and documentation. I want to continue learning, so I don’t allow the language model to write complete code for me; instead, I use it to improve my work and point out any errors.
You are definitely on to something at 8:30. If there's one thing I'd like to see more of in the programming scene, is modular development and a better focus on architecture and design.
Great discussion! Here is what I believe is missing:
- Not all digital execution is code - at least not as we have come to know it over the last 20 years.
- My Tesla navigates using a neural network that does not resemble code as you have described.
- The value of code in the 'new world' is guaranteed repeatable execution. Said another way, there is no judgement - just script processing.
- I believe more and more digital execution will result from the will of neural networks.
- Coding will be largely relegated to 'tool scripting' where repeatable execution might be wanted/required.
My 2 cents,
Chuck Boecking
For me, personally, I only use it to get some idea or need a question answered with a example set of code.
I have completely turned off AI code suggestions, as they interfere with my train of thought when I code, so for me generally it has been good for "common" questions without many contextual variables needing to be included.
It can help people, and I've definitely seen that, but sometimes it also gets to a point where you have wayyy to much context needed for it to work in a proper manner.
Some of the best programmers that I know don't even use lsp.
wow, that's interesting! never met one of those. The first I'd heard of this is a recent interview with the creator of the Odin language - apparently he is in that camp as well.
LLM = Large *Language* Model. Not Large Computer Science Model, not Large Software Development Model, not Large Software Engineering Model, not Large System Design Model. LLM's do not know how to build applications. They just know how to repeat the language patterns they have been trained on from a large pool of highly relevant examples. When no such examples exists, LLM's will make things up. Then the LLM makes a mistake, it has no way of recognizing it's mistake. When you are building commercially viable software, there are no pre-existing examples to train your LLM on. You are the one inventing the first example.
You raise some great points. I think LLMs are only a part of what AI code generation tools are going to consist of (and to some extent already do). It seems like the best way to improve the quality of generated code right now is via giving them agency and thus a feedback loop, so they can run their code and be given the opportunity to fix any mistakes.
Developmer :) 0:02
Wow, how embarrassing. Thanks for pointing it out 😞
Sigh, found another mistake at 0:24 too
I liked the video. I appreciated your perspective and acknowledgment of biases. Personally didn’t find anything wrong with your line of reasoning and am inclined to agree with many of your points. Great work
Thanks Nathan, really appreciate the kind words!
AI (GitLab Duo) helped me learn Rust and have a lot of fun in the process.
nice! Had a pretty solid experience with GitLab Duo as well.
the flashing transition was very hard on me when watching.
thanks for the feedback, seems like many others agree! definitely won't be using the glitch effect like this anymore.
4:25 Churn rate of code, bugs introduced, features delivered (without churn), there is visible metrics and papers on the topic, not using as part of the research for this video show the real message behind of it. Just sad to watch it.
thanks for the feedback! how do you define churn rate of code? features delivered is not a good metric, because they are not equal in scope and value. On the surface bugs introduced seems like it would be easier to nail down, but I don't think it is either. For example there could be "bugs" which have negligible or debatable customer impact - bugs that we might willingly accept for an X% increase in velocity. Would love a link to the papers you are referring to!
@@codetothemoon chrun, code removed close after inserted into the codebase, there is a paper on the topic using the public data from Github, look for it. Features delivered "before" and "after", also have a paper on the topic, with a control sizeable control group. Bugs introduced by automated, semi-automated and no AI tools, there is also a paper on the topic. There IS metrics, there is for more than a years a conversation on this topic. So please, don't spread misinformation like this video mostly was. Feels good isn't a measurable metric.
@@codetothemoon Looks like my comment explaining was deleted. I won't comment again.
@@ShinSpiegelThat's sadly automatic. Any lengthy nuanced comment is almost guaranteed to be deleted for being "controversial". Just my own observations. I've tested this by posting my own long comments from before RUclips got overzealous with the anti-spam and they get auto deleted.
@@LiveType yeah, already unsub on this channel, tired of all the AI things, and the la k or research from lots of content creator. I’ll be more strict on the creators I follow, even if this mean follow no one.
I think AI is great in terms of refreshing your knowledge. It's like a human that knows everything. I was exploring solana blockchain and boy was it hard, but the chatgpt got me started.
definitely agree! It seems like most folks have a very positive experience using it as a means of learning. Its utility there seems far less disputed than for code generation.
For me AI automates a lot of the busy work with coding, like writing boiler plate code and formatting data. So I would say there is definitely an increase in productivity. I think without knowing how to code in the first place, and knowing it well, it becomes a lot more difficult to use AI to code. It may not produce exactly what you want, will make mistakes, etc. If you do not have the skills to properly debug and fix the code manually then you will spend hours trying to get chat gpt to get it right. I do think relying too much on AI for coding will lead to something like what we see with calculators and arithmetic, where our ability to do the "grunt work" decreases substantially. If AI reaches the point that it can totally replace programmers, I think it will be good enough to replace a lot of other occupations as well. LLMs are not restricted to text anymore, they can parse visual input, audio input, and if you map the sensor data into something the AI can understand it can process that too. We honestly might be seeing "smart" robots in our life times, and then who knows what happens.
The really big question around AI is cost. And it’s a big one that is being ignored. All the the talk about the value of AI assumes that the current cost to use AI will remain the same. But we have no reason to believe that’s true. In fact, we have every reason to believe that eventually AI providers will need to generate profit. OpenAI and google and Amazon can’t Contis to lose billions of dollars per year on their AI services forever
I have a love/hate relationship with AI. I've been using it with Rust to improve my Rust skill and the amount of nonsense it gives is unbelievable, I feel like using it makes me more dumb than not using it. I also encountered a situation where I ask it to convert a float or integer into a comma separated string using regex ( I hate regex). and after more than 30 mins and countless attempts, it keeps looping into giving a regex that either adds a comma including in the decimal part if its more than 3 or works fine in float only but doesn't work with integer.
but Id argue its a good tool if you already know what you're doing.
This has also been my own experience while using 4o for Swift. It was great for learning basic Swift/SwiftData/SwiftUI. But as my application got more complex and I wanted to do more with it. AI became useless. I learned more from Apples own projects they release than from what the AI was telling me I should do. I encountered numerous incorrect statements from 4o that made my understanding of Swift also incorrect.
im getting a lot more work done as a game dev
nice!
How can we know that AI is going to be GREAT at code in the future?
Answer: They have only been CRAZY funded and researched for about 3 years, and they're already good enough to keep moving the goalpost and tech bros has to come back and give new claims about things they will never do...
Remember a year ago when everyone said AI will NEVER be able to do math. Boom, half a year later they beat the majority of experts.
Just like any other profession this will continue to disrupt the field, there is no point in
agree 💯 - I think many are making the mistake of casting long term judgement on these tools without considering how much better they are going to get!
Good one.. To me to be honest software developers are not inventing new stuff in daily basis.. most of the things they do is repetitive. For example a web developer can only produce a website or maintain it or upgrade it to facilitate new functionalities.. that can be easily replaceable via AI. Where the AI can not compete is genuine idea.. For instance you have a business idea where you are thinking to bring a completely new tool for UX design out of the todays norm.. AI is pretty much useless there.. But I do agree with you... for everyday task like you said converting to another database.. pretty sure somewhere someone already faced that issue, there fore its not original one.. what you are doing is repeating the steps that other guy have faced thousand miles away from you. In that case AI will replace those human labor once it self correct the buggy stuff and reaches the ability to test out its own stuff..
AI can't generate 100% reliable and high-quality code. That's a fact.
If I were a CEO, I would probably only care about saving money, so AI generating acceptable code would be enough for me.
However, as a researcher, I want my code to be the best it could possibly be, so I have to write it myself.
And yeah, I make mistakes, too. The difference between me and AI is that I can learn from my mistakes, and I don't need a huge data set and a power plant to do that.
Agree with your first point! I do think to get the most out of these tools we need to be almost militant when it comes to reviewing the code they generate.
Re: difference between you and AI - I think you may have needed a large data set as well, though maybe not quite as large. Everything you learned prior to writing the code that enabled you to write it - including natural language, programming concepts, etc. But you surely did it with quite a bit less power than LLM training clusters - human brains are notorious for their low power requirements in comparison 😎
That said, totally see your point that for many people and in many domains (maybe research is a canonical example of one) just writing it yourself can make a lot of sense.
You don't need a power plant because you aren't responding to millions of people across the world at the same time. It's not like it takes a power plant to run a single instance of an LLM.
Aaaand I'm officially not getting anything out of this channel anymore lmao, sorry
thanks for the feedback - what are you looking to get out of it?
Just an ai doomer that can't see past their biases.
See ya 👋
Interesting perspective. I look forward to the opportunities created by AI.
thanks and I do as well!
haha please at last, remove the french translation. It's horrible. switch back to original track faster than light
thanks for letting me know about this - is it super inaccurate? maybe ai is good for coding but not for automated translations 🙃
@@codetothemoonThe translation is pretty decent. The problem is the horrible voice generated. Sound too bad... If I didn't know your channel, I just pass it.
You just want to continue to make money.
Like most humans I am guilty of this. I feel like there is a negative implication here but I am unclear what it pertains to. My creation of this video?
developmer
yeah someone else pointed this out already, quite sad about it :( somehow got missed by multiple sets of eyes.
The same way AI can copy art and output a full picture, AI will copy full software, without needing for code, only with the executable. You will prompt apps directly. At that point coders will be fully replaced.
that's an interesting idea - llms writing machine code directly. seems like something they'd be capable of, I guess the problem is it wouldn't then be reviewable by humans. I'm skeptical that software developers will be completely replaced - though I wouldn't be surprised if traditional coding takes a back seat to other skills.
In the short term, yes there will be a shift in how software developers work and there may be some new roles being created. In the longer term all software will be created by artificial intelligence. Including triple-A games. Everything. Writing code may be left as a hobby, but humans will not be required. The comparison to the replacement of physical work does not transfer over to the replacement of intellectual work. There is nowhere to go where the AI can't follow, or even take the lead.