What will be left for us to work on
icml.cc
icml.cc
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.
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.
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.)
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?
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.
Let me introduce you the the executive team of every company.
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
I think the answer lies in baking the good old fashioned way and ignoring the factories.
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.
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.
I'm not saying this as in I love it but I just don't see how we'd go back
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.
The utilitarian view of mental 'work' might shift in the same way. Deliberately opting for mental excersize not because of nescessity but for mental health and enjoyment.
Just like in the physical case, there will be widespread mental obesity because many will not.
Thoughts?
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.]
No idea how it got it so wrong.