They learn between model iterations. You're right, it isn't the same thing as Junior developers' competence improving with experience - the current model's weaknesses are locked in. But it does mean that much of the Junior level thinking and mistakes will be outgrown by successor models.
But they don't retain anything from your on-the-job training. The next model iteration is yet another junior fresh out of college, and knows nothing about the painful training procedures its predecessor put you through.
Nothing prevents an LLM agent from writing a bunch of "notes to self" and using that. And the next model from picking those notes up and using them. Coding agents already do some of that natively.
Hell, we might eventually get an LLM to say "wow the old AI was an incompetent idiot" after reviewing all the notes and session logs. That's how we know we reached human parity!
Yes... but the next session with the same model is yet another junior fresh out of college that knows nothing about the painful lessons the last session put you through ten minutes ago, either.
> All of that said, AI is going to directly cause job loss, I’m calling it now. Not as much as the doomsayers predict, but more than most people expect.
Unless there is some limit to model development we can't currently foresee, plain economics will see to it that white collar job losses will be close to total. Likewise blue collar if we don't find a limit to spatial AI and robotics development.
The problem with all these discussions is that no-one rubbishing the job-apocalypse forecasts can say why or how progress will peter out - beyond pointing to economic limits ("it's a bubble") which won't apply over longer terms. Given the pace of progress the last few years, and this inability to say why job losses won't scale with the tech, anyone ruling them out is either wish thinking, or showing a staggering failure of imagination.
If there's a reason the losses will be "Not as much as the doomsayers predict", say what it is.
The limit to job loss is completely unrelated to AI capabilities. Rather it is social.
There is a breaking point where if enough people end up jobless it will lead to genuine bloody uprisings. I won't pretend to know where exactly that point is, but I am more than happy to state that it is before "nobody has a job anymore" is reached.
I'm sure someone is thinking that job loss needs to be gradual enough that they can get technology to the point of having killer drones ready to take out any individual instantly, before any uprising threshold is crossed. If the abundance of drones keeps rising and surveillance continues toward "total", then we are headed toward this possibility.
Who wants to uprise if it means instant death for the uprisers and everyone they care about?
And if things move gradually enough we are like frogs in boiling water. Think about how if many of the things openly happening today were to happen 50-100 years ago how much resistance there would have been.
The thing about the whole “frogs in boiling water” thing is that it doesn’t work. They’ll jump when it gets too hot.
To counter it with another idiom, consider the concept of having nothing to lose. Remember I’m not claiming it’ll be fun, easy, or anything like that. What I’m saying is that when push comes to shove and enough people genuinely have nothing to lose, it will not be pretty, and I’m not willing to bet on the rich and powerful coming out on top, regardless of how slowly and gradually they try to make it happen.
I think it’ll suck a lot for everyone, and specifically I’d be willing to put money on the rich and powerful wishing they’d had a little more empathy and foresight.
I was thinking this months ago and asked AI what would be done.
IT quickly spit out a half dozen things it expected to be issues and the counter measures that would be deployed instantly and likely proactively.
It showed me videos of things like these 24/7 drones for people detected and less lethal things like soundwaves as deterrents.
It showed me videos of cables hardening and other systems being used to prevent cutting cables and stealing electricity.
Discussed the size of the jails that will be built, an expected number of people that move into a few groups of cybercrime (and which would continue to thrive with that) -
Had numbers of number of people that will be eating mealworms for lunch, the political and cult shifts that will occur, all sorts of interesting things.
I'll be putting out a movie about all this that AI already knows and expects to happen. And this is all with currently known and in use technologies.
Millions will be leaving many cities, and direct to needy from farms systems may keep people alive. Office buildings being converted to mealworm and similar farming may happen.
I do not see the millions of people who make a living via call centers to be able to find similar paying jobs. Most of them, and the people who currently make a living supporting them (sandwich shops, cleaning, etc) - will be competing for delivery jobs, which will depress wages, and they will be competing with robots and x-tunnels, etc.
I need to get a working title for all this info I gathered and come back and edit.
> If there's a reason the losses will be "Not as much as the doomsayers predict", say what it is.
OK, I'll make an attempt:
1. AI capabilities have obviously exploded at an amazing rate over the past few years, but I think most people in the field view a lot of the "Bobby grew a ton by age 13, he'll obviously be 100 feet tall in a few years"-type analysis of a few years ago to be wrong. Or, at least, people see limits to current AI tech, and that completely new/unknown approaches will eventually be needed. Of course, AI never really gets worse, and I can easily see a lot of problems (e.g. hallucination rates) being greatly improved even with just existing tech.
2. I think tons of jobs will get obliterated. I think you'd have to be insane to go into radiology as a med student right now. Tons of people currently make their living driving, and robots can already do a lot of that. More broadly, there are already lots of jobs that are basically "data in, one unambiguously correct answer out" that AI will excel at. Creative jobs will also be affected. I read a report recently about how AI dramas are all the rage in China, and they're already displacing jobs for actors.
3. But I disagree that losses will be "close to total". There will still be a strong desire for humans to actually decide on the "what do we make?", even if it's mostly made by the AI/robots. For a particular depressing and macabre analogy, think of the American South during slavery. Even though most of the actual labor was done by slaves (in the analogous case AI/robots), there were still jobs directing the work to be done.
So I guess I'm in the "it will be a shit show of epic proportions" camp, but that's still not as bad as some of the worst doomsaying I've seen.
3 seems the strongest of these arguments. The 'other techs plateaued' argument ignores that this is the first tech ever to convert electricity into thought and agency. There isn't a precedent for AI, and until intelligence stops scaling with compute, any assumption of a limit - that may not even exist - being reached in the few years left before jobs are wiped out is arbitrary faith.
I agree though, that business leadership roles will still survive - with some industries, wherever some principle or vision needs to be maintained - with the normal little adjustments humans might prefer to feel out for themselves. Perhaps also politicians, sportsmen, escorts, priests, anyone involved in spiritual and new age therapy. But this is still close to total. And aside from ownership/leadership which can earn in power and influence, it isn't clear how any of these jobs would be paid.
Hinton was probably right, even in 2016. When a med student chooses their residency, they want to choose a career that will be around in 40 years. The tech obviously wasn't there in 2016, but it is tantalizingly close today. I have a family member who is a radiologist who works for a group that deploys AI tools as an adjunct, and is was pretty eye opening the first time that tool caught a critical finding he missed.
Interestingly, there is currently a huge shortage of radiologists because the tech (but, more importantly, the regulatory framework) isn't quite there yet, but again people choosing a medical specialty aren't looking at today or a year or two out, they want a career that will sustain them into old age after investing years and hundreds of thousands in training. People are worried at what the landscape will look like in 5 years, let alone 20, 30 or 40.
Job loss can only be measured in relation to time. If it's gradual enough, it's not job "loss" it's job transition.
My point isn't so much about how many jobs will be replaced by AI, but how quickly. In my mind the doomsayers are predicting 30-70% job loss in the next 3-5 years. I'm not saying the job loss won't be that high; I'm saying it won't be that soon. If I had to guess, I'd predict 5-15 years, which won't be a party for anyone, but it's not immediate devastation.
And of course previous job loss situations were supported by other jobs people could migrate to. If this situation doesn't -- if AI swallows a substantial portion of the entire job market -- then eventually it will come down to whether we have an acceptable way to share the abundance that results.
I also think there will be significant displacement and change, but the size of the pie will grow tremendously, and there will be many, many jobs people haven't thought of to address the bottlenecks.
The pace of progress is precisely why many people qualitatively assume the curve will flatten soon: J curves are generally (obviously not always) unsustainable.
You're arguing that the limits will appear because they usually do. (Correct my paraphrase if this is unfair.) Apart from being blind faith, this argument is oblivious to the fact that capability so far has scaled directly with compute and that the experts developing the models expect that to continue.
That is the lowest effort of lame responses. Look at the actual authors on the paper then. Or, I don't know, actually make a substantive comment about the research in the paper beyond your 8th grade redditor "Ha ha Siri sucks" response.
If it’s a monopoly yes. But there are massive profits in full automation. I’m not expecting costs to go to zero but it’s the only pathway to things getting cheaper by a lot.
I think we get hung up on definitions and end conditions when we are more likely to feel the effects of the asymptotic regions ala AGI or Full Automation
The exact situation I laid out:
“If it takes 10 construction workers at $60,000/annum to build one home, I can forsee the descendants of current AI tech enabling 10 construction workers at $150,000/annum building 5 homes in the same time with an even larger profit margin for the corporation involved.”
Is the most likely condition if we let this technology grow healthily with the exact “full automation” end condition being beyond the point of diminishing returns.
This is a lot of very wealthy workers building products for a lot of people with revenue growing but margins plateauing and therefore absolute profits growing as well. This is an exact repeat of the Industrial Revolution situation,
I think you underestimate just how competitive obscure fields and crafts can be. The world's second-best basketweaver is likely to be painfully aware of their superior rival, and push themselves hard to catch up and surpass them.
What you're really arguing is that SWEs are superior to basketweavers. But I wouldn't be so sure. That basket might well be around and admired long after the software's obsolete and gone.
This is a simple economic fallacy. Your idea holds true if your business has room for growth, but you always need a market for those multiplied outputs. In an economy where every business has multiplied outputs, but demand doesn't scale with those outputs, what do you think happens?
The problem for high-performing developers (even the elitest) is that over a long enough time scale, they won't be remotely competitive with high performing AI. At the moment we need human operators to guide and supervise these AIs, but unless the tech's progress grinds to a halt for some unforeseeable reason, it won't be this way for long.
Assumption, sure. Lacking evidence, though? I see tons of evidence. Circumstantial sure, but it's not exactly a claim out of the blue. In 5 years these things have gone from economically worthless toys to very competent replacements for juniors. Maybe the improvement stops, but that would be an assumption without evidence.
This is the central problem with the dismissals of the tech's capability. Public discourse needs to shift to planning for the economic impact in particular, but the kind of High Brazilism from the naysayers who insist it's a proof of psychosis to even mention AI's potential, makes the inertia in policymakers much easier for them to maintain. Waiting for the financial effects to arrive and then improvising policy is the stupidest way of handling an upheaval on this scale - even if the precise form of those shocks can't be anticipated.
It makes tremendous sense - when understand as reflexive straw-clutching and wish-thinking aimed at reducing the frequency of the poster's nightmares and reducing their diaper expense.
reply