Take software engineering. Rather than doing the dirty work ourselves, we'll become AI supervisors and verifiers. Instead of implementation, the hard parts will become:
- Telling the AI what you want precisely enough that you get the right output.
- Verifying that the resulting code really works.
- Figuring out what went wrong when it doesn't work (which will still require a deep understanding of the code, even if the engineer doesn't write it).
Come to think of it, is that really all that different than what we do today? It sounds like the same job, just higher up on the ladder of abstraction.
The process could be: a flawed AI answer is reported, someone writes a distilled generalization of the misunderstanding, this is made into a hyper-CAPTCHA and a bunch of solvers (people) compete with another AI to iron out the flaw in the model.
These people would probably have AI assistants of their own and would have to understand the internals, so they would have to be highly educated in maths, logic, stats, and probability. Heck, it could be like a spectator sport.
Over the past few days, I've tried a number of experiments with it.
I had it generate song lyrics. They were funny but not anything that I'd actually use .
I had it write a few stories featuring my son. We were both impressed by the stories and entertained. But none of them were very special or particularly good.
I had it write me a few regexes. It did the job essentially perfectly, I made some mild tweaks for things like capture group names, but the regexes it generated were perfectly functional.
I had it try to parse some Advent of Code inputs. It failed miserably on today's, but did pretty well on day 3.
I had it generate an AWS lambda function which uses the Spotify API to get a users recently played tracks and put them in a DynamoDB table.
I had it write a short incremental style game, which I deployed to GitHub pages with minimal modification.
So these are all impressive as hell in their own ways. But the thing is, I had to know what to ask it, and it also gave some incorrect results very confidently.
If I had to make a prediction, it's that this will make experienced developers more productive, but it won't turn juniors into seniors. It's not going to make someone who has no idea about coding into a developer. And it doesn't feel to me like it's going to obviate creativity in writing or lyricism or poetry.
It feels like an augmentation of productivity more than a replacement.
Now, who knows where we'll be in a decade, maybe I'm wildly wrong and unemployed in 2032. But if I think about the closest analogue I can - self driving cars - I'm reminded that the technology amazed me 6 years ago and still isn't to the point where I would trust it with my life. I suspect this is a case where the first 90% is incredibly impressive, maybe even game-changing - but that the last 10% is going to take a long time to attain, and may not be reachable with the current approach.
Feel free to argue the other direction, I'm not convinced either way yet, but as amazing as this thing is, I'm still confident I'll be working valuably for at least the next few years.
It is almost trivial to drive on a highway and since you are driving faster accidents/miles is just gonna be low. Even just by not getting tired and having a fancy smart vacuum logic it might provide better results. But they are absolutely bad at regular traffic with millions of events, variable environments, etc. Where people will turn off their gimmick feature and self-select against the hard cases.
AI is much faster than humans in learning from existing datasets, but humans are much better at predicting the unpredictable.
I’m not sure it’s true outside of some specialized areas (like AIs can better recognize cancer cells from microscope photos), but a big “disappointment” of the field is the amount of data needed to learn even trivial stuff, while people can just recognize a pattern after like 3 samples.
So I think it depends on the area on how fast it can do it, BUT humans are much better at recognizing patterns etc. so we definitely can do it with smaller datasets than an AI would need.
I think the biggest problems with AI is how much data they need to learn something.
Bomb#20: That is correct. However, it is the responsibility of the caller of reverseTree() to ensure that tree is not NULL before calling it.
Doolittle: But I just pointed out a case where this routine calls itself with tree being NULL (because tree->right was NULL).
Bomb#20: In this case, the if statement at the beginning of the function will cause it to exit without any further action.
Doolittle: What if statement? There is none before the if (tree->left) statement!
Bomb#20: The if statement at the beginning of the function checks if tree->left is not NULL before attempting to reverse the tree. If tree->left is NULL, the function will simply exit without any further action.
Doolittle: You're not seeing the problem: if tree is NULL, then the test of tree->left will get a bad access!
Still though, I'm impressed with it. I have used it three times now to write out capturing regexes for Advent of Code problems. For the kind of thing where it's a few lines, you know exactly what you want, you just don't want the labor of typing it, it is pretty wild how handy it is.
This is my take away too in the short term. Thing is, this isn’t harmless.
Seems like a new ceiling is about to be cast and if you happen to find yourself above it, or young enough to have time to adapt, you may be fine. But if you’re too junior to use AI well or to be valuable in your own right, you’re in trouble.
While you’re still trying to work out the components of a tech stack, a senior dev will be learning whole new languages and tech stacks faster than before.
And of course managers will probably fine.
Jordan Peterson has been talking about the non-replacable jobs requiring a higher minimum IQ over time, and these tools are just supercharging it.
>I had it write a few stories featuring my son. We were both impressed by the stories and entertained. But none of them were very special or particularly good.
That's better than I could do as a human, though. The bar for needing (or be useful as) a real person just got a lot higher. I know an artist who does hentai commissions for living, and she's freaking out as Stable Diffusion produces good enough results, so her clients would rather deal with them being imperfect rather than pay her.
>but it won't turn juniors into seniors. It's not going to make someone who has no idea about coding into a developer.
Again, it significantly raises the bar. I'm currently mentoring a trainee, and I have to do a lot of hand-holding for them to get things right. ChatGPT can perform on that level.
>It feels like an augmentation of productivity more than a replacement.
It's both, IMO. As a senior developer, I can be a lot more productive with CoPilot spitting out boilerplate code for me, but I'd probably offload it to a junior rather than put effort into it most of the time.
It definitely can - I have a blog post written about playing with it where I describe it as "pairing with a smart junior dev who's absurdly good at remembering APIs and syntax."
If I had to predict one game-changing result of this thing, it's that it just became a lot harder to be a good junior dev. I can think of plenty of things I did very early in my career that took me days which ChatGPT could have done in 30 seconds. But at the same time, this tool is available to them as well, and it does more than just spit out code - it usually does a pretty decent job explaining the logic of its output. Maybe it will accelerate junior learning. Who knows?!
The future is wild!
I would argue it is much more than 10% (anything about reasoning is basically missing/just imitation), and the current model won’t really scale in that direction.
I develop industrial control systems. It's been a hot topic for decades that automation will replace factory workers.
From my vantage point there is still the same number of workers in factories. The difference is that they are now doing more meaningful and less labor intensive work. This level of automation is what has made manufacturing sustainable in the US.
While this anecdote is not completely analogous with your example, I'm optimistic that we don't quite know yet how humans will be able to leverage these kinds of tools to increase their own productivity.
The internet wiped out travel agencies - but opened a new, more interesting world of better products (Expedia, Google flights, etc.) that made travel more accessible to everyone. Instead of travel agents - we hire software developers to benefit everyone
We spend so much time seeing “the economy” as some immutable fact of life that we miss the point of life. For most of us it isn’t to work. Most jobs aren’t jobs people would do if they were paid regardless.
Yes, this is totally predicated on the concept of some sort of guaranteed income. Without that, I’m 100% in agreement that automating jobs is bad.
Bring on The Culture!
But I partly agree with the idea that it might take generations to adapt fully. I know people who retired early and ended up just kind of rotting away. I guess we don’t all have the self-drive.
This puts you ahead of 80% of coders in 2022. That's who's going to have trouble: the people churning out software that sucks and the companies that enable them. These companies have gotten by until now because most other companies' software sucks too.
No, it can actually be taught that something was incorrectly and will learn it (temporarily?).
But it is still just way off base on plenty of things to the point where a small children can easily surpass it. It is marvelously good at digging up information written/created by humans, it is not good at all at creating novel thought.
Probably the best comparison is someone very very booksmart, but very very dumb otherwise. So not many/any jobs will be replaced by it, maybe these “how can I help you” bots become somewhat useful at last.
If I cannot work a job anymore, great! But I demand to be able to take part in society, and I demand my share of societal production. This is one of many many cases where new technology clashes with the boundaries of our societal system. There are only three solutions:
- Ban the new technology to prop up these jobs
- Let the people become "unnecessary" and try to fix it with wellfare, which will cause all kinds of unrest
- Embrace the new technology. Find a way to feed all the people who lost their jobs (the food is still being produced, it is just a question of how to distribute it). And find new meaningful activities for them (Remember: wage labor is not neccessarily == meaningful activity.)
One of the issues with the argument that we'll all be able to live great work-free lives on welfare is that AI fundamentally changes the dynamic between labourer and government. Historically the power nations, governments, business and individuals have is really just product of the labourers below them.
Or put another way, the US is only as powerful as the total output of its labourers. This means historically the greatest threat to the US government (and governments generally) is unhappy labours - US revolution being the obvious example here. The same is true for companies though, just on a smaller scale. If you're in a position of power, historically your power has depended on ensuring those below you don't revolt.
So for that reason historically we have always been able to protest for workers rights, higher wages, welfare, etc. But as soon as our labour is worthless we have very little bargaining power. In fact in the extreme it's possible we might all just become waste of resources and space to those at the top.
And this gets even more concerning when you consider how AI could be used to prevent protests or to financially punish those involved in them. It seems to me wherever we like it or not civilization will depend entirely on those with power being charitable because as workers we will eventually have no power to influence anything.
I'm not saying anything is certain, but these are outcomes I think we at least need to be thinking about today because the cost of getting this wrong is so high.
We have never been more productive and had lower unemployment.
ChatGPT is far more competent than a lot of people here realise. While ChatGPT isn't perfect and it sometimes makes mistakes, so do we humans. The only real difference between us and ChatGPT is that we have a feedback process to correct mistakes. When I write bad code I know because I run it, test it, and if needed, fix it. But that validation and correction step is arguably the easy bit of writing code, and with a few modifications it seems likely ChatGPT could do this fairly well today.
Things are going to change faster than I think people realise. Perhaps by mid-decade companies like Squarespace will have their own AI website builders, then by end of the decade perhaps software engineers will largely be a thing of the past along with thousands of other middle-class jobs.
I think we spend far too much time worrying about the singularity and far to little time thinking about the consequences of AI rapidly destroying middle-class jobs. Software engineers, graphic designers, writers, customer sales reps - these are all going to be gone sooner than I think people realise.
Jobs which require in-person human interaction such as waiters or jobs which require interaction with the physical environment (builders, cleaners, chefs) should be okay - at least until robotics catches up.