There are three parts of this talk, the last part (~last 15mins) bluntly address the question "how human roles shift in a world with advanced AI?"
My main takeaways from the speaker's position are: purely technical skills will get devalued because they tend to be verifiable tasks and AI will get better. Also effort will shift from building to evaluation of AI and systems. Speaker also mentioned that more of our time will go to decision making while AI takes over delivering on decisions.
I found many arguments to be at odds with one another bc the speaker does not address how someone can gain understanding without accumulating purely technical skills. It simply does not add up that someone who lacks purely technical skills in an area can make good decisions. Experience is often a by product of accumulating technical acumine in a particular problem area.
"It simply does not add up that someone who lacks purely technical skills in an area can make good decisions."
I've been thinking about this and your related points, and I've been (with growing concern) instinctively in agreement.
Another analogy has been coming up. I started programming (as a kid) in the late 1970s, first with BASIC (of course) but pretty quickly moved to machine language for obvious performance reasons, and then to C, followed by many other languages. In the late 1980s, at university, I learned proper assembly language, and used it quite a bit.
Back then, I was pretty good with assembly/machine language. Not super star or expert level, but comfortably solid. Since then though I've spent very little time at that low level.
Granted, such early exposure and experience has been a pretty big fairly indirect benefit. But over the years, I've met and worked with many really smart, productive people who have never had any experience at that low level, and that lack was rarely noticeable.
For any given 'layer of abstraction', there is some use in being familiar with the lower layers. I think, in general, the closer the lower layer is, the more useful the knowledge is. For example, how useful is deep knowledge of CPU microcode to a Java programmer? Perhaps some, but not a lot.
In the tech world, GenAI is likely to push most everyone quite a ways up the tower of abstractions. We're in the middle of that right now, and it certainly feels pretty alarming to me.
The biggest uncertainty that comes into my mind is how insanely fast this change is coming upon us. My crawl up the abstraction tower happened over the course of about 12 or so years, from machine language/assembly to C then to Perl then to many other things. In retrospect, that felt pretty fast.
GenAI is seemingly pushing us in that direction many times more quickly.
> Nobody has good answers for any of these things. It's a folly to think someone else has it figured out.
We don’t know that. Only history will tell us who figured it out and why.
What I don’t understand is why there is a clear decision between decisions and delivering on decision, other than the artificial need to retain control.
I don’t see a real roadblock that will prevent AI making better devisions than humans.
I always thought that the Gettier problem was an academic curiosity.
Someone who justifiably holds beliefs about the eventual impact of AI that later turn out to be true---can they be declared to have been knowledgeable about the subject matter?
It could be that the ratio of the amount of context required for a good decision to happen to the amount of context an entity is willing or able to provide is not high enough. E.g. the AI might dictate my life better than me if I upload every single thing about myself to it, but for privacy reasons I only tell it a small amount about me.
And maybe corporations make similar decisions.
(You might counter that there will be local AI. This is true but it’s still risky to let an autonomous system have all your sensitive information, especially in the AI hacking age.)
I was wondering that too. I suppose if we take it to the extreme end of the spectrum, "end world hunger" is a great decision but how to actually realize it is another problem.
The pipeline problem is already well-discussed. Just learn the stuff even though the AI can do it, so you can acquire that supposedly valuable judgment. My concern is that if the only things left are the subjective and the social, only people who talk a big game are going to be valued.
I have a question about “domain expertise” as a component of future knowledge worker requirements. How does one gain such expertise in a context where thinking is expected to be delegated to AI (shifting from problem solving to question asking, as noted in this paper)?
How does one learn to pose the right questions when basic ones are rarely “manually” answered? that is, without an AI assistant’s help
This pattern appears in schools, where AI interferes with human development that typically demands long and difficult effort of actually answering questions
For certain functionality, we import libraries. For others, we've handwritten solutions. Even so — sometimes we have to crack open the repo for a library to really dig in to what's going on. That seems to go for domain experts as well.
I think we've always been working with the scenarios of what's solved for us and what we need to solve manually. AI is certainly not an abstraction layer, but there are ways to use it where you are still driving, learning, and shipping.
AI won’t function unless you give it a purpose (at least for now). Isn’t it crucial, then, that we define that purpose as specifically as possible?
Both AI and humans make mistakes. What they have in common is that review is necessary.
2) The system has determined that you need 10% calories if you sit on a bench all day instead of playing football.
3) The system has determined you don't need to watch TV/YouTube/Netflix, so that energy will be redirected for more food production. [You still get 90% anyway.]
4) Distribute (step 6) is not necessary. Please go to the queue for the two daily meals. [Estimated time, 6 hours each.]
> I argue that there will be plenty for us to work on, grounded in the “AI as normal technology” thesis
There is no doubt that AI is creating more work than it produces. Every single line it writes has to be reviewed by someone because AI cannot take responsibility. And it can spew out decades worth of written code every day. Every line is untrustworthy and must be manually verified.
The problem is I don't want to do this work. You don't either. Everyone I've seen using AI to write code does not do this work. Anyone who says they do this work is lying, to you or themselves or a little of both. People just YOLO and ship it.
If that is the only work left to do when using AI, I'll just do it the old way. If that's not good enough to stay employeed, then I'll just find a new career.
This is exactly correct. The whole point of software is to precisely build what is intended.
The perception of the quality of the result is relative to the original requirements and expectations. It shouldn't be surprising that people who are ignorant of industry specs and who have very low standards are impressed. They've always been impressed by any software at all if it scratches their immediate itch.
The bigger threat to employment is the same as it's always been: the employer's wishes. Many of those wishes are not out of the kindness of their heart. They are forced by regulations and contractual obligation.
What we actually saw was an initial dismissal of how deep that specification can/should go. I'm expecting there to be a wave of people pushing for reform instead when they realize LLMs aren't moving the needle much. That's good. That's where all this energy belongs. We always want better ideas. It doesn't change the fact that humans still need to implement the new specs when they're new. Ah, you say, then the bots take over? No. As ever, that's just packaged as a standard library. There's always another new layer to think about integrating it with which again requires a human in the loop.
What's most surprising isn't the AI, but all the profound ignorance that exists out there, even from big shot execs! All work works like this. If it didn't, the economy would have collapsed during every technological milestone in human history.
It seems likely there will be much work in the area of fixing the slop code that proves to be unmaintainable. Companies right now are racing to push the lowest quality code into prod with no way to even QA it. If you have the skills, it'll be top dollar to bail these people out.
There are three parts of this talk, the last part (~last 15mins) bluntly address the question "how human roles shift in a world with advanced AI?"
My main takeaways from the speaker's position are: purely technical skills will get devalued because they tend to be verifiable tasks and AI will get better. Also effort will shift from building to evaluation of AI and systems. Speaker also mentioned that more of our time will go to decision making while AI takes over delivering on decisions.
I found many arguments to be at odds with one another bc the speaker does not address how someone can gain understanding without accumulating purely technical skills. It simply does not add up that someone who lacks purely technical skills in an area can make good decisions. Experience is often a by product of accumulating technical acumine in a particular problem area.
"It simply does not add up that someone who lacks purely technical skills in an area can make good decisions."
I've been thinking about this and your related points, and I've been (with growing concern) instinctively in agreement.
Another analogy has been coming up. I started programming (as a kid) in the late 1970s, first with BASIC (of course) but pretty quickly moved to machine language for obvious performance reasons, and then to C, followed by many other languages. In the late 1980s, at university, I learned proper assembly language, and used it quite a bit.
Back then, I was pretty good with assembly/machine language. Not super star or expert level, but comfortably solid. Since then though I've spent very little time at that low level.
Granted, such early exposure and experience has been a pretty big fairly indirect benefit. But over the years, I've met and worked with many really smart, productive people who have never had any experience at that low level, and that lack was rarely noticeable.
For any given 'layer of abstraction', there is some use in being familiar with the lower layers. I think, in general, the closer the lower layer is, the more useful the knowledge is. For example, how useful is deep knowledge of CPU microcode to a Java programmer? Perhaps some, but not a lot.
In the tech world, GenAI is likely to push most everyone quite a ways up the tower of abstractions. We're in the middle of that right now, and it certainly feels pretty alarming to me.
The biggest uncertainty that comes into my mind is how insanely fast this change is coming upon us. My crawl up the abstraction tower happened over the course of about 12 or so years, from machine language/assembly to C then to Perl then to many other things. In retrospect, that felt pretty fast.
GenAI is seemingly pushing us in that direction many times more quickly.
Nobody has good answers for any of these things. It's a folly to think someone else has it figured out.
> Nobody has good answers for any of these things. It's a folly to think someone else has it figured out.
We don’t know that. Only history will tell us who figured it out and why.
What I don’t understand is why there is a clear decision between decisions and delivering on decision, other than the artificial need to retain control.
I don’t see a real roadblock that will prevent AI making better devisions than humans.
I always thought that the Gettier problem was an academic curiosity.
Someone who justifiably holds beliefs about the eventual impact of AI that later turn out to be true---can they be declared to have been knowledgeable about the subject matter?
It could be that the ratio of the amount of context required for a good decision to happen to the amount of context an entity is willing or able to provide is not high enough. E.g. the AI might dictate my life better than me if I upload every single thing about myself to it, but for privacy reasons I only tell it a small amount about me.
And maybe corporations make similar decisions.
(You might counter that there will be local AI. This is true but it’s still risky to let an autonomous system have all your sensitive information, especially in the AI hacking age.)
I was wondering that too. I suppose if we take it to the extreme end of the spectrum, "end world hunger" is a great decision but how to actually realize it is another problem.
"better" is a qualifier
The pipeline problem is already well-discussed. Just learn the stuff even though the AI can do it, so you can acquire that supposedly valuable judgment. My concern is that if the only things left are the subjective and the social, only people who talk a big game are going to be valued.
There's a reasonable argument to be made that people who talk big game already dominate.
I have a question about “domain expertise” as a component of future knowledge worker requirements. How does one gain such expertise in a context where thinking is expected to be delegated to AI (shifting from problem solving to question asking, as noted in this paper)?
How does one learn to pose the right questions when basic ones are rarely “manually” answered? that is, without an AI assistant’s help
This pattern appears in schools, where AI interferes with human development that typically demands long and difficult effort of actually answering questions
How do you become a baker, in a world of bread factories.
I think the answer lies in baking the good old fashioned way and ignoring the factories.
Fine for bread, but I don't think most customers are interested in the artisanal aspects of business operations.
For certain functionality, we import libraries. For others, we've handwritten solutions. Even so — sometimes we have to crack open the repo for a library to really dig in to what's going on. That seems to go for domain experts as well.
I think we've always been working with the scenarios of what's solved for us and what we need to solve manually. AI is certainly not an abstraction layer, but there are ways to use it where you are still driving, learning, and shipping.
The author has a blog post mirroring their talk in writing: https://www.normaltech.ai/p/what-will-be-left-for-us-to-work
AI won’t function unless you give it a purpose (at least for now). Isn’t it crucial, then, that we define that purpose as specifically as possible? Both AI and humans make mistakes. What they have in common is that review is necessary.
Video recording from ICML up from Arvind Narayanan
https://autonoma.ca/model.png
Thoughts?
1) I guess the robots extract the metal too?
2) The system has determined that you need 10% calories if you sit on a bench all day instead of playing football.
3) The system has determined you don't need to watch TV/YouTube/Netflix, so that energy will be redirected for more food production. [You still get 90% anyway.]
4) Distribute (step 6) is not necessary. Please go to the queue for the two daily meals. [Estimated time, 6 hours each.]
[dead]
Here’s an idea: reversing climate change. That should keep you busy for the rest of your life.
[dead]
> I argue that there will be plenty for us to work on, grounded in the “AI as normal technology” thesis
There is no doubt that AI is creating more work than it produces. Every single line it writes has to be reviewed by someone because AI cannot take responsibility. And it can spew out decades worth of written code every day. Every line is untrustworthy and must be manually verified.
The problem is I don't want to do this work. You don't either. Everyone I've seen using AI to write code does not do this work. Anyone who says they do this work is lying, to you or themselves or a little of both. People just YOLO and ship it.
If that is the only work left to do when using AI, I'll just do it the old way. If that's not good enough to stay employeed, then I'll just find a new career.
Programming is a vocation. It became a career because of sheer luck.
what makes you think companies will do it the "old way" when the new way is much more economically viable?
I'm not saying this as in I love it but I just don't see how we'd go back
This is exactly correct. The whole point of software is to precisely build what is intended.
The perception of the quality of the result is relative to the original requirements and expectations. It shouldn't be surprising that people who are ignorant of industry specs and who have very low standards are impressed. They've always been impressed by any software at all if it scratches their immediate itch.
The bigger threat to employment is the same as it's always been: the employer's wishes. Many of those wishes are not out of the kindness of their heart. They are forced by regulations and contractual obligation.
What we actually saw was an initial dismissal of how deep that specification can/should go. I'm expecting there to be a wave of people pushing for reform instead when they realize LLMs aren't moving the needle much. That's good. That's where all this energy belongs. We always want better ideas. It doesn't change the fact that humans still need to implement the new specs when they're new. Ah, you say, then the bots take over? No. As ever, that's just packaged as a standard library. There's always another new layer to think about integrating it with which again requires a human in the loop.
What's most surprising isn't the AI, but all the profound ignorance that exists out there, even from big shot execs! All work works like this. If it didn't, the economy would have collapsed during every technological milestone in human history.
It seems likely there will be much work in the area of fixing the slop code that proves to be unmaintainable. Companies right now are racing to push the lowest quality code into prod with no way to even QA it. If you have the skills, it'll be top dollar to bail these people out.
I can't even trust a chatbot with internet search. You guys use it for research?
Then don't trust a chatbot for internet search. There's much more to llms than chat bots
Yes you can. This comment might have made sense in early 2025.
Lol no you can't. I had one give me blatantly incorrect information just a bit ago. The first result after the AI version had the correct info.
No idea how it got it so wrong.
Maybe work on a larger font size, pal.