So you want to replace workers with AI? Watch out for retraining fees, they're a killer The lucid ramblings and art synthesized by ChatGPT or Stable Diffusion have captured imaginations and prompted no shortage of controversy over the role generative AI will play in our futures. As we've seen with CNET and Buzzfeed, executives are no less dazzled by AI's creative potential to replace workers with profits. But one …
Contrary to what most think, the Luddites were not anti-technology. When the weavers were put out of work by the new industrial processes it was the Luddites that demanded cross-training to the new technology, they wanted compensaiton schemes, they wanted fair wages for newly trained people on the new machines, they wanted health insurance, pensions for those injured and too old to work. They wanted to work with the new industrial barons but he owners of the "dark satanic mills" refused, spread the idea that the Luddites were anti-technology, that the Luddites were not willing to accept the new order of man working for machines, not the other way around. The mill owners had the power to get the government to hunt down and execute the Luddites for daring to ask for fair treatment for the disposessed.
We've already seen what a 21st Century mill owner looks like, Mr Elon Musk buying up Twitter and pushing everyone out who refuses to do things his way. As AI becomes more useful and starts put of we techies out of jobs, and it will to a greater or lesser extent, how long before we have to form the "Luddites V2" and how long until were brutally suppressed by a system that will always prize profit over human diginity.
The best description I've heard of today's 'AI' is just a jumped up autocomplete. It needs humans to curate all the data, humans to hand fettle, design the structure and build the transform in the network (for that is all it is).
The excitement over GPT3 is just the latest incarnation of people getting excited over implementing tech (and the notion that crude approximations of networked nodes is clever like a brain as it's structured like a brain), as opposed to why the tech solves a given problem. A while ago it was things like Alexa. Before that it was apps on phones, Before that it was Victorian clockwork automaton (as back then the world was deterministic).
The thing about GPT3 and its ilk is that as a thing it is clever, however its output isn't.
We've already seen what a 21st Century mill owner looks like, Mr Elon Musk buying up Twitter and pushing everyone out who refuses to do things his way. .... Plest
What a great load of nonsense that is, Plest, and haven’t you yet noticed the present is brutally suppressed by a system that will always prize profit over human diginity .... and now is being forced to be constantly fighting against the system, wasting and depleting its own blood and treasures, to try to ensure news and evidence of the situation does not escape and become general universal knowledge, although that rear guard action was discovered an impossible task to achieve and master some long time ago.
> the present is brutally suppressed by a system that will always prize profit over human diginity
"the present" ? It has always been that way, even before we were able to control fire. The other way around is actually true: More and more workers successfully stand up and organize for better conditions. Globally it is that way, but of course there are local backlashes.
That's not precisely true. The specific group directed by Ludd weren't specifically against the technology. They were against the technology's effects on them. The new machinery meant that their skilled manual labor was not needed because a person with less skill could get the same thing done. They weren't advocating for a general increase in worker's rights or services; they wanted compensation for themselves because technology made their previously high-paid work obsolete.
We techies, as you call us, are in the business of automating things. Stuff that previously took inefficient processes involving a lot of people can now be done automatically by a computer; there aren't banks of people doing mathematical calculations manually and repeatedly anymore, for instance. People have lost jobs that have been automated, and new jobs have arisen to take their place. That has neither been completely for the better or for the worse, but it has happened and if you work in technology, you have aided it. Now that you think it's likely to affect you, you've changed your tune. Perhaps you're compatible with Ludd's thinking after all, when you could be advocating for everybody.
Techies like you are happy to work hard to create a world where there are no economic opportunities besides crime and menial physical labor that pays unlivable wages.
You know else can be operated by a person with no education and only five minutes of training? The guillotine!
In fact, my point was basically the opposite of what you just said. With technological advancement, there is going to be disruption, and that will harm somebody. There are two ways you can deal with this. Only one way works.
Option 1: Don't allow technological advancement when it will harm someone (almost never, in other words). You can't functionally do this. Things break quickly. Take any era where you thought things were best for the workers and reflect on the fact that, without technological advancement in the previous decades, that era would not have worked that way. If we'd banned the telephone because "Think of what will happen to those telegraphists who spent time learning Morse", we wouldn't have had jobs for telephone operators. If we hadn't allowed cars to be manufactured because of all the damage it would have done to people with jobs related to animal-based transport, no wonderful manufacturing economies. Similarly, when those things become automated, it makes sense to adopt the improvements.
Option 2: when people are harmed, have a plan for how to help them. Instead of trying to hold back progress so everyone can keep doing what they've always done (often without considering whether what people have been doing is something they enjoy doing), provide resources to those whose jobs have been lost. The world changes, and those who live there will have to change as well. The best thing we can do is to make sure people have assistance making necessary changes.
I lied. There's a third option: the Luddite option. That's basically where you take all the improvements from technology that affects others without thinking of them, but as soon as something looks like it will affect you, you demand special treatment. In the specific case of Ludd and his compatriots, it was well-paid production work that was automated, but they couldn't accept that anyone else could produce fabric and thought that they were more important than the people getting jobs operating the new machines. This is the hypocrisy demonstrated by the original post: "As AI becomes more useful and starts put of we techies out of jobs". Now I've seen the code that GPT spits out, and I don't have that much fear of that, but if it happened, it's the way improvements change the world. Technology has done it to others, and it may do it to us. It is our job to prepare and adapt, and if we're going to ask for more resources if it happens, we should ask for them for others as well.
This! Elon, along with daddy Bezos have introduced the work culture where individual employees are every bit as expendable as worker ants in ant hill.
At this point, it's going to take nothing short of a repeat of a French or Soviet revolution to change things for the better.
You get ChatGPT to read what it produces, of course.
Conceptually, the system should disappear up its own rear with a 'pop'.
Seriously, its like any other technology. It could be useful but you just know that 90%+ of its output is going to be the press release equivalent of cat videos. But considering what exudes from the typical marketing department is mostly worthless pap anyway nobody will notice. Or care.
(Note how Tesla disbanded its marketing / communications departments and seems to be doing just fine. Currently reports are that something over 18% of new car sales in California are electric cars with the bulk being Teslas.)
Self service, although I now use them and find them useful when in a rush, are also annoying because they are designed to make the company more money but the customer gets no discount for essentially doing the companies job for them. Not to mention the "You are a thief unless you can prove you're not" cameras that record you as you use the machine. We all should be sending in Right to be Forgotten requests for those machines and those recordings. They can try and claim they are entitled to keep them due to "security" but if no theft has happened, they have no right and I've never seen any GDPR warning for them in any store.
ChatGPT, although good does give out wrong answers and knows it. I asked it a powershell question the other day and it gave me the code. Which I tried, the commandlet didn't work. I told ChatGPT it didn't work as the commandlet didn't exist. It then confirmed "You're right, the commandlet was discontinued on blah blah blah", I gave up asking it after that as it gave another command that it knew was no longer valid. (I've already forgotten what the commands were, stupid brain).
The discount you get is your time, normally you can be queuing for ten minutes, removing everything from your trolley and putting it back into bags in your trolley in a rushed way. Self scan lets you walk up to the payment device, scan a barcode, tap a couple of times, pay and off to the car.
> The discount you get is your time
Not really.
I'm old enough to remember when supermarkets used to boast that if there were more than X (typically 2) people ahead of you then they would open another checkout.
Of course that never helped you personally because the entire population of the store would magically migrate to the newly opened till before you'd even noticed. Meanwhile the wanker in front would spend half an hour trying to pay with the coins from their piggy bank.
-A.
> that'd run you in the neighborhood of $4.28 million just to train it.
But that isn't how people learn. That one day someone decides that something you know is obsolete, so you have to return to primary school and start learning The cat sat on the mat and everything else you've ever know, all over again.
People learn incrementally: building their knowledge on what the already knew. Sure, there are times when a paradigm shift comes along and people have to relearn a skill or set of rules (such as when a country changes the side of the road its drivers drive on). But those a few and far between. And usually highly compartmentalised. When the UK when from pounds-shillings and pence to decimalised currency, that did not mean anyone had to relearn how to read or write.
So it should be with training AIs.
" Sure, there are times when a paradigm shift comes along and people have to relearn a skill or set of rules (such as when a country changes the side of the road its drivers drive on)"
That's not how paradigm shifts really work.
A shift in paradigm is best desribed as like "taking one set of glasses off, then putting on a different set wich give you a very different view of the same world". A very good analogy (or example) is the Rubin Vase optical illusion. Viewing it through one paradigm, you see faces. Exactly the same image viewed through another paradigm, you see a vase. It is very hard to see both at the same time, it is one or the other, yet always the same image you are looking at. All you are changing is the way you interpret it and it is a radical change.
That is pretty much how Thomas Khun defined it when he coined the term.
Lets suppose the Rejoin EU Party wins the next election, and they decide that we are all going to drive on the right side of the road.
Yes, I need to relearn which side of the road to drive on, and I need to relearn which way to drive round roundabouts, but all the stuff like knowing to stop when I see a red traffic light, selecting the correct gear, knowing what the foot pedals do, etc; that's all going to stay the same, so I would likely be at an advantage over a learner-driver attending their first lesson. And of the course both me and the learner driver have some prior knowledge, like being able to read which helps interpret street signs, and having previous experience on the roads as a non-driver.
> So it should be with training AIs.
Well, yes - and then we'd be getting back to "actual" AIs, with internal models that are properly explained/explanatory. Expert Systems, purely as an example, can have their rule set modified piecemeal, without having to start the whole process again.
But these massive nets being sold as "AI" - there is no way to point to any of the numbers in the model and say "we can tweak that to get this result", or "removing this and adding that to the training set will change this set of numbers, and nothing else" so no piecemeal updates for you.
And depending on how you've built your AI, you might be able to do that. Many systems that are in active use aren't built that way. You can only add so many layers on top of one, most of which simply filter the results that no longer apply and make the model try again, before you start inhibiting the performance you're looking for. A model based on millions of words can't just read a few million more and insert them in because it's already weighted things based on the old words and those weights will effect any future data.
If we end up using lots of massive models in the future, machine learning researchers may start pivoting to methods that make it easier to perform incremental training. So far, that hasn't been necessary so people haven't bothered doing it.
People learn incrementally: building their knowledge on what the[y] already knew. [... So it should be with training AIs.
But that's the whole point, right? Knowledge is not a characteristic of "generative AIs", as far as I understand, nor is it the basis of their operation, so they cannot be trained that way.
As demonstrated by the well known self-training bail-application-recommendation AI of a few years ago. It trained on feedback about it's own output. It recommended that people with a high chance of re-offending not be granted bail, and those recommendations were approved, retraining it gradually to reject all bail applications from black applicants.
There is one fundamental concept all organisations looking to replace customer facing people with AI "chat bots" need to take on board.
If I am a current or prospective customer of yours and I contact your company, I want to talk to a REAL LIVE PERSON. Not some effing bot.
If all I get is an effing bot or you make it difficult to get through to a person, I'll take my business elsewhere.
It really is as simple as that.
It is bad enough phoning up some organisation for assistance with something and having to wade through shit loads of shite "automated assistants", which 99% of the time are able to do nothing I can't do myself online - ERM THAT'S WHY I'M PHONING UP because your automated systems aren't helping!!!! - without getting put through to "a plastic personality".
I hate to say this, because I share your sentiment - but it really doesn't matter that you want to talk to a real live person. You say: 'ok, I'll walk away'. And I tell you: sure, walk away from tesco self-service check out and use sainsbury's self-service checkout. Walk out of your current bank's online banking system, sign up to use another online banking system. Once every business adopt this or that model, there is no other, and your freedom is either to use the system or not use it at all. And they WILL all adopt it, because it's less cost, more profit.
I was making a comment along the lines of "Eliza was a small language model. These are large language models." and it occurred to me that modern children have never seen Eliza.
If they had, they would have the proper skepticism, and would actually understand the situation, instead of regarding it as "AI magic"
Ah well. They'll learn. School of hard knocks and all that.
Don't expect those hundreds of thousands of tech jobs losses to not be replaced by overseas labour. Between Blockchain and AI, and amid the hundreds of thousands of on going tech layoffs. I would recommend out of work tech workers pursue in demand careers not in the tech sector. Chat GTP was trained using Kenyan labour at about $2/hr, if you think you can compete with that, good luck. In the meantime the excessive hiring within the tech bubble has led to a severe shortage of skilled workers, I might recommend pursuing jobs in the real economy for those going through layoffs. Perhaps in the construction industry with historic job vacancies would be a good fit.
I've had a play with ChatGTP.
I asked it some questions about something I was researching at work. It helped but more than half the links were wrong.
I also asked it about a very difficult to achieve programming problem with very tight performance criteria, for which I had a working solution for. It failed to provide a similar solution even with coaching. In fact none of the proposals would have met the performance required. Many of the proposed solutions were naive at best and would have easily been discarded by someone knowledgeable in the field.
AI has its uses but people need to understand and accept it's limitations too.
The AI version costs tens of millions to train to a point. It becomes less valuable day by day after that.
OK, high capital cost, v low marginal cost of production. We know about things like that. What we end up with is some, few, of such expensive to produce Big Things producing output for many, many customers. You hope that the difference between marginal price collected and marginal cost to produce pays back that capital cost. Or, average production cost is well above marginal until some massive scale (average will, asymptotically, approach marginal as overheads are paid for as scale increases), you need to sell a lot to make back the investment.
This isn't that much different from the economics of search engines. Or even social media. You get very little from each user of Google or Twitter (G gets maybe $10 a year per user?) and you've costs in the billions of providing the product at any level to anyone.
To make it work you've got to sell the shit out of millions of pieces of output that is. And, as with social media, advertising might well be the way to pay for it. Because that way you can give it away "free" and thereby gain the scale of pennies and pennies that pay back the billions.
Interesting that lately people seem to talk about technology as it is something that has existed by itself and now has been generously "given" to us. We forgot how technology advanced from us and not by it's own.
When I read about "chatgpt that", "chatgpt this" it reminds me when people were being shown fireworks (which were created by humans) and being exceptionally surprised as if it was a deity or something. I guess it's the human nature to seek the "divine" in some way or another. Even if that "divine" thing is a creation by Man itself.
Still, some whackos, for their own reasons, are dreaming of a Singularity or some other twisted and perverted use of technology to surprise (as in controll) the folks.
Makes sense I guess...
A better question is what you are going to train it with?
With no paid journalists and artists, who is going to make the training source?
Human produced work which is still under copyright will not be put up on the web for free use in training models.
Those commercial businesses still paying creatives, are absolutely not going to allow unpaid or low cost use for training purposes. The fair-use loophole won't fly when training has been explicitly prohibited.
All my open hardware and source-code has been moved behind various anti-scrape barriers requiring manual intervention. It's no longer in plaintext on github.
Ts&Cs now explicitly exclude machine learning use.
Intelligence, is knowing what to do, when you do not know what to do
How well does this ChatGPT fare at the novel?
Just remember chaps, when this Ai bubble bursts, and the same with this cloud nonsense that everyone has a boner for, falls flat on its face.
You know what to tell them
Remember these words "Per employee profit", it's what your renumeration should be, anything less and you're being robbed
I can't help but think that this article is perhaps leaning on straw men too much.
In the first instance, I agree that training models will need to be updated (and that itself leads to an interesting problem, since over time, the content being used to train said models is likely to be increasingly AI-sourced, similar to how Youtube content-generator bots were fed off data scrapers which indicated what the current popular trends were...), but do they need to be updated weekly?
Barring significant infrastructural changes in the underlying technology, I would have thought an annual refresh would be more than sufficient for the vast majority of purposes.
Secondly, it doesn't really matter how much it costs to train a given model, because this only needs to be done once. After which, you can sell as many copies of the pre-trained data as you want. In much the same way as how software like Microsoft Windows has traditionally been marketed.
To my mind, there's plenty of other issues with AI output - ChatGPT is very capable of producing incorrect answers which look good at first glance, so still needs human curation and verification. And the quality of StableDiffusion images can wildly vary (how many fingers and sets of teeth do you want?), so again, human intervention is needed to curate it's output.
Admittedly, these are things which are likely to improve as we improve the technologies driving these models, and lots of time and CPU cycles will be burnt in doing this training.
But still, it only needs to be done once.
Time was, you could only make teeny tiny models that fit within the available RAM, not these humungous beasts. Which meant that, when let into the big wide world, dealing outside of the training set the cracks would show up much faster and be more obvious to everyone. Although they were a lot quicker to train (at the easy end of exponential data crunching versus linear Moore's Law) so you caould chuck 'em out faster.
Go back to the old automated-trading scams (not that anyone ever actually did this, no, nope, on my life guv): train up (teeny) models on a stock market feed. Make a fair few of these, with random variations (number of nodes, different data subsets) so that they don't all end up identical but they all re-create the historical data pretty well. Take out a great big ad stating how your amazing system managed to track the market so far (well, duh) and promise Great Riches to anyone buying it to "predict" how the markets will act tomorrow. You can even keep adding to the ad copy "certified satisfied customer reports" from purchasers who "struck gold this very week"!
Of course, everyone got a "customised model", most of the predictions were effectively random (and it has been shown more than once that random stock picking is at least as good as most pundits and traders...) and only the "big wins" are ever glorified in the press (who can be trusted to take on the burden of bolstering your ad copy). After a short time, the models are so far out of their depth that anyone can see that they are generating gibberish and the customers quietly feed their copies into the shredder (one of the great advantages of using floppy discs and CDs when buying software, you at least physical revenge). You've quietly folded your company (or, better yet, flogged the whole thing to an even wider boy - sorry, I mean "a respected trading company") and trotted back to Peckham.
But now it has all gone horribly wrong.
- the costs of the models are obscene, with the obvious repercussions (like face-saving: when you call out the quality of the goods, Rodders will look nervous and Del will give you the gab, but the boys playing with this much dosh have got nice suits with strange bulges; metaphorically, of course).
- the cracks start off tiny, compared to the size of the model and the amount of stuff it can regurgitate, so they are easily brushed off (that 10-line routine forgot to declare "loopCount", you fixed it without even thinking, almost out of habit)
- there are only a few models which everyone is playing with, so as the cracks propagate no-one will be spared from falling into the Underworld of Nightmares.
- the more people get themselves invested into something big and shiny, the less able they are to give it up. And they can't even get catharsis from the shredder (although putting one end of the Ethernet cable in might be fun, it'll still wreak havoc on anyone in the vicinity).
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
