“Which is probably why Academics are reacting so desperately and so foolishly to the existence of chatGPT and other LLMs. They’re desperately trying to prevent people from using the tools in the hope that this will keep up their social status. But this is a doomed enterprise. The mere fact that the statistical text generator can get excellent grades means that the grades are no longer worth more than the statistical text generator.”
To change the example a bit, the entry level – and bread and butter scut work when you were on your uppers – in financial journalism was company reports. “SpudIndustries showed revenue growth from YouTube but analysts worry about credibility” etc. 300 words and there are 4,000 companies on the US markets. 10 per working day -ish. 30 to 50 outlets needed that sort of scut work done. In modern money you might get £30 per piece. You know, -ish, -ish. It was a vast sea of boring low paid work.
Doesn’t exist now, not in the slightest. You can’t get £1 for such a piece. Even if you set up an AI to write it for you then publish it on the net you’ll not gain £1 in ad revenue from it. It’s so cheap to produce it’s not worth even the £1.
You do still get paid for opinion, for analysis (actual analysis that is, here’s why I don’t believe the statistical numbers sort of analysis). Because that’s something different. Can’t, for the life of me, think why academia should be different.
If AI can destroy the universities then that’s at least one thing in its favour…….
This comes down to the problem with LLMs. They’re low value. You can listen to The Velvet Sundown, who are an AI band, and it’s not unpleasant. It’s a lot like The Eagles, Neil Young, Tom Petty. But you know, it’s like mediocre tracks that got put on a b-side of a single late in their career. It doesn’t have that magic of Learning to Fly or Take it Easy.
You could probably get an LLM to write Eastenders scripts. It’s going to notice “‘ere Ian, look after the caff for 10 minutes will ya?” and use it. It’s not writing the St Crispin’s Day speech from Henry V. It’ll design a boring logo for a delivery company, rather than the genius of the FedEx logo. And the more that gets pumped into buying NVidia chips, the cheaper it will get, which will just lead to a sea of worthless mediocrity.
It’ll be the death of a lot of journalism because so much of that is just plausible-sounding bullshit.
CoyoteBlog had an interesting take comparing AI to the industrial revolution, except this time, the jobs that are going to be replaced are those of the elites.
What If The Hand Loom Weavers Were Children of the Nobility?
WB
I think you’ve rather got it there.
At the moment we pay a lot for worthless mediocrity ( architecture, lawyers, doctors, design, academia, politicians etc etc )
AI will expose its true worth.
Bless you, Gunker, I’d no idea that he’d made a comeback.
“ It’ll be the death of a lot of journalism because so much of that is just plausible-sounding bullshit”
Grauniad and the Beeb might be okay then, given they don’t do plausible.
It seems to me that AI will take all the bullshit jobs. The ones that actually do very little, and can’t cause people to die if you f*ck up. LLM AI is great at being close enough for pennies. What it can’t do is be right 100% of the time. So any job that requires precision answers/output will be AI proof. Any job where close enough will do is vulnerable.
Jim @ 9:06
Nor can people for that matter!
The first attempts at ‘AI’ focussed on logical reasoning, the sort of thing that counts as intelligent because it is a hard and unnatural thing for people to do. “How we would like to think we think” But that is just the sort of thing that is easy for machines to do. The hard things are easy and the easy things are hard.
Now, at enormous expense, we have machines that can make a plausible effort at the sort of thing —generating bullshit— that comes quite naturally to people. The AI-proof jobs will not be those requiring 100% precision (we got machines to do that) but the ones that best judge the difference between 40% and 60% nonsense.
Jim,
“So any job that requires precision answers/output will be AI proof.”
That won’t worry our foolish leaders who will still use it for mission-critical tasks because it saves them from having to think.
One of the most interesting pieces I’ve read on AI suggested that in many firms the senior executives will sack lots of workers and so drive up their own bonuses. Then they’ll depart before the downsides appear.
When those do appear the firms will start hiring people again. I dare say, though, they’ll often be younger, cheaper people. In which case lots of forty- and fifty-somethings could be in for a hard time.
Is it worse to be sacked because AI can do your job or because your bosses preferred to pretend that AI could do your job? I must ask ChatNBG.
“The AI-proof jobs will not be those requiring 100% precision (we got machines to do that) but the ones that best judge the difference between 40% and 60% nonsense.”
Yes that is probably a better way of putting it.
There’s also the issue of degrees of mistakes. A person will make a mistake thats not blatantly obvious, because they’ll probably catch the really obvious ones. But an LLM will make mistakes from the tiniest detail to the most obviously wrong thing. It doesn’t know the difference between being 0.01% out and being 100% wrong. Saying ‘The moon is 239k miles from earth’ is not completely accurate, but its not far away either. Saying ‘The tensile strength of mild steel is 1000 MPa ‘ is just wrong. But the LLM won’t know the difference. So an LLM will be capable of making the sort of errors that would be sackable (or possibly even criminal) offences for a human.
https://archive.ph/I3Qnk
Robot brickies. Had to happen.
Jim,
“So any job that requires precision answers/output will be AI proof. Any job where close enough will do is vulnerable.”
And earlier AI like police cameras for stolen cars are like that. You put a camera a mile down the road and it reads the plate, checks it against the stolen list and it tells the police car down the road to stop the car.
These get it roughly 90% right. You’ll get a situation where it fails to read the plate. It might get it wrong and tell the police to stop ABC123D. So, when the car arrives the police see it’s actually ABC128D and ignore it. And the reverse – the camera sees it as ABC128D and it was actually ABC123D and should have been stopped.
But it doesn’t matter that it only gets it right 90% of the time. That’s still 90% of all stolen cars picked up. 100% would be nicer, but 90% still has a deterrent effect. If people think they can’t use a stolen car for long, they won’t buy them. It’s cheaper and probably better than having another cop sat there punching numbers into a computer.
One of the reasons I’m a sceptic about the scale of AI is that you need particular sorts of problems where it saves on human costs but still has human checking involved.
I doubt that AI will have that much effect on academics. It will probably make continuous assessment of students unworkable (if it hasn’t already) and encourage a return to the final examination system. (CA favour girls, while FE favours boys; so expect squeals of outrage from the usual suspects.) That AI can write an acceptable essay/dissertation on Gödel’s theorem or the wave–particle duality relation without any understanding of either means that academics will have to return to testing students’ powers of understanding rather than their powers of regurgitation. In turn, this could raise standards and make higher education an elite activity again, because only c.5% of an age cohort will have the understanding as opposed to the partial and limited knowledge of the lower ability ranges.
Theo, if AI brings to an end the Age of Credentialism and a return to assessment, hiring and promotion based on demonstrated ability and capability it’ll be a thoroughly good thing.
If I am right, Norman, AI will make Blair’s notion of 50% of each age cohort attending university redundant and unworkable. It could also make academia more orientated towards research rather than teaching.
It’s too soon to say whether AI will cut a swathe through white-collar jobs, but it strikes me as unlikely. My daughter is a senior City lawyer and she says AI often fails the simplest legal tests – for, basically, the reasons that Western Bloke articulates so well. She demonstrated this by asking Gemini to summarise the Consumer Rights Act 2015 and advise on a problem. The errors were serious and the advice defective: a junior solicitor in a High Street practice would have done far better.
“continuous assessment of students” was always bilge. Everything that can be examined in conventional closed book exams should be. Obvs that can’t include lab practicals or project work but it can include damn near everything else.
Moreover the scripts should be anonymous and the examiners marking the scripts should not be the people who lectured the material. In other words, bumsucking along American lines should be made impossible.
And no bloody multiple choice questions either; make the buggers do serious analyses in the exam hall.
Everything that can be examined in conventional closed book exams should be.
That doesn’t make any sense. The books contain data or the tools for processing data. Why would you want to carry those round in your head? All you need to know is they exist & how to access them.
BiS
If you haven’t memorised the formulae, then you haven’t considered their components; and if you don’t understand the components of the formulae, then you don’t understand the formulae. And if you don’t understand the formulae, then innovation becomes almost impossible. Following a recipe without understanding the ingredients eliminates creativity and prevents the development of new dishes…
To expand on Theo’s remark, creativity can be defined as finding a new way through the schemata, or rearranging things that already exist. For that to work your schemata has to be populated already with things to rearrange.
My stock in trade is actually creative problem-solving. The more of a polymath I am, the better at it I get.