(This was written about 9 months ago. Its not out of date... yet)
People think that AI is going to DESTROY some jobs and CREATE some jobs. It may be too early to tell if this is true. Even so, here are some thoughts.
1) Who will win? Who will lose?
2) Historically in the long term society was better off after a tech change (e.g., we live longer now than we did in the farm-era). Will that happen here as well?
3) We have some sense of what kinds of jobs will be destroyed. But what kind will be created? Will they be interesting? See later in this blog for a job you might not have thought of.
4) Many jobs will change. If you are over X years old then think about how much technology has changed your job even before the AI revolution. (The value of X may vary depending on how high-tech you are.)
For an intelligent view of the questions above, see here.
For my view of one aspect of this, read on.
There is one job which has been either created or expanded by AI:
Annotator.
(See here for an ARTICLE about these jobs, from which I got most of the rest of this post. The word ARTICLE is in caps so when I refer to it later you'll know what I am referring to.)
The job consists of looking at pictures and labeling things.
Here is a direct quote from the manual:
LABEL real items that can be worn by real people.
Does Lady Gaga count as a real person? She sometimes wears dresses made of meat. Should that count? See here for a real article about her and see here for the Weird Al parody of Born that Way. Note that this is Weird AL, not Weird Artificial Intelligence. (See here for the Weird AI for Weird AL problem.)
They do this to create data for AI.
a) The jobs don't pay well though there are some exceptions.
b) The jobs are boring though there are some exceptions (and that may depend on the worker).
c) This job is needed because AI keeps running into edge cases. This may have happened with AI's attempts to solve my GROUP ONE-GROUP TWO prez-VP problem from a prior blog post here or my baseball-brother-pitchers post here.
c) KEY: People in AI used to think this is a temporary thing and these jobs will soon be automated. This might not be the case. There are SO MANY edge cases that AI encounters. The more we expect from AI the more edge cases there will be.
A quote from page 26 (of the ARTICLE pointed to above) which is informative if you can parse it. I think.
Put another way, ChatGPT seems so human because it was trained by an AI that was mimicking humans who were rating an AI that was mimicking humans who were pretending to be a better version of an AI that was trained on human writing.
As I've said before, what's going to happen is that the world is going to figure out that "advanced model" AI is way too computationally expensive for it to be useful for anything.
ReplyDeleteSmall model AI (models that can run on your cell phone or a desktop) will be plentiful and free and useful for some things. People like me, who want to read the Wiki page or the actual papers that discuss the things we search for will continue to have no use for LLM-based AI, but people who just want to know simple trivial things and don't care about the details or intellectual content (or that it's wrong 10% of the time) will be happy with small-model AI.
But computers are getting faster and AI will get cheaper, you scream. Well, no, they're not. The RTX5080 benchmarks at less than 1.5 times faster than the two generations older RTX3080. Back in the day, every two years, things were twice as fast (from 1985 to 2005). Nowadays, a generational improvement is 20% every two years, if you are lucky. And that will often come with a similar increase in power drawn, i.e. zero improvement in computation per watt.
@David_the_bostonian_in_Tokyo. How has tokyo digested these advancements? As capital of the future that i was back in the 1980s, 1990s, how did the city, capital translate these advancements. I remember the first cab ride, amazed by their built in gps long before smartphones were in existence. Or self-driving shuttle buses on work campuses, in the 1990s (not sure if you recall this). Tokyo was well ahead of everyone else. Did the LLM technology change anything over there? Or did things stagnate more or less since the 1990s? I didn't see much more, substantial change lately.
ReplyDeleteThe answer to your question is: no, Japan doesn't feel like some sci-fi writer's high-tech future.. Sure, everyone's got talking GPS, but they did ages ago, too. (Japan led the world on smart phone adoption, but then everyone else caught up.)
DeleteOn the trains, everyone has their noses glued to their smartphone, doom scolling with drugged-out expressions on their faces.
Background: The boom of the 1980s though very early 1990s fizzled starting in 1992 or so, and _median income_ has been falling ever since. The rich are, however, getting richer, so _average income_ is up. A lot.
So I don't see much of a high-tech boom, and really never have. There was some talk of self-driving light rail, but that's not been heard from recently. (Tokyo only has two light rail lines, and they're gloriously low-tech and have been around since like forever. (Setagaya Line: 1923, Arakawa Line: 1913)
The thing I do see is amazingly long-term monster construction projects to remove grade crossings on commuter lines, so that said commuter lines don't disrupt the neighborhoods they pass though any more. Seriously kewl, but not "high tech" (even though it's seriously kewl construction tech). Also an amazing amount of new really big office (and also condo) building construction largely funded by Chinese money (according to a native informant who works for the World Bank). Condo prices are inflating insanely: I suspect that's mostly foreign speculative investment. Real people can't afford them. I read a novel by a popular Japanese author on the social woes of families living in a 50-story condo complex ("tower mansions", although a "mansion" in Japanese just means a unit in a reinfoced concrete building.)
So Japan is functioning fairly well without much high-tech BS, in my opinion. Lots of talk about AI in the newspapers, but I think it's mostly talk with not much actual use.
@David_the_bostonian_in_Tokyo. Wow, some insight here. Thanks for the take. I think I can appreciate no self-driving taxis in Tokyo more now, having real humans with white gloves who use the somewhat antiquated inbuilt GPS devices drive me around. Also, happy to see the ridiculous QR code culture not be part of society and daily life. Yes Chinese money has entered in more ways than was beneficial, having brands like LAOX,now a chinese majority owned company, advertised on GINZA billboards.
DeleteTo say something somewhat vaguely close to being on-topic here, the September 3, 2026 issue of Science (the AAAS rag) has a discussion of LLM use in conference paper reviewing. One thing it said was that while LLMs reduce the time time for basic fact checking, that's largely made up for by the time it takes reviewers to check the LLM review itself for accuracy, so there's no statistically significant time or effort saving. And there was no discussion of the costs involved. It's a must-read for folks here.
ReplyDeleteAnother older thing, was that LLMs (due to their 10% error rate) are already being seen as problematic in the context of an existing relatively large business, but for an individual or small group of folks doing a startup, being able to prototype things quickly (where the user is taking responsibility for fixing LLM mess-ups) is real useful.
So that means that while actually useful, they ain't going to be doing away with very many jobs where it's important to get things reasonably correct the first time. Again, as people filing court briefs are still finding out the hard way.
FWIW, the article you recommend says the following.
ReplyDelete"Rather than seeing a highly paid physician, I could have most of my health care needs met with this technology and a reasonably compensated medical professional, who may get less than one-third of the pay of a doctor."
In my opinion, this is stupid beyond words. First of all, the rest of the industrialized world manages to provide medical care to all of its legal residents at reasonable total costs and affordable charges to patients. It's only in the US that medical costs (and doctor incomes) are insanely high and huge swaths of the population are left without any access to care whatsoever.
But even worse, it represents a horrific level of misunderstanding of what's involved in good medical care, and a sky-high level of intellectual hutzpah. We in AI have a long history of insisting that we're going to replace doctors, and it's been incredibly stupid (and has failed) every time. We're complete idiots here, and need to be told so. (Also, there's this really ugly money-grubbing aspect to it: we think doctors are overpaid, and so want a cut of that money.)
LLMs have proved really really bad at giving medical advice, and despite there being a zillion startups thinking they're going to replace radiologists, there are more radiologists today than ever before. (There are some fundamental problems with "neural net" based image recognition stuff, including the (recently discovered) point that mammalian visual systems use high-level information to improve low-level visual performance. Sure, there's some low-level processing along the tract from the retina to visual cortex, but once at the VC, high-level information becomes available.) Also, the "neural net" stuff is completely opaque: the image recognition stuff passed all the tests. Until, that is, someone showed it a picture of a cow on a beach. It turned out that the "rule" it had inferred from the training was "if it's got green pixels, it's a cow.".
AI is still really really dumb.