Well. in the case of my friend, they were neither desperate nor vulnerable. They were conflicted, and it was making it harder to be artistically productive than they wanted, so they kept trying coaches and therapists, and none of them got to the heart of the matter.
The GPT asked them several layers of questions, based on ACT, CBT and other well-proven therapies, and led my friend to discover for themselves the root of the problem. When they articulated it to the LLM, it gave my friend a list of exercises from various therapies that might help. It volunteered to custom build a selection of exercises, and my friend agreed and answered a few more questions. That led to an exercise plan that my friend has been using, and they are very happy with the result.
I donât think ChatGPT is a substitute for psychiatry or crisis intervention (yet), but I think it might replace some âlife coachesâ.
This seems to be the preeminent AI thread at the moment, so let me just leave this here to cheer us all up:
With hindsight, perhaps we shouldnât have let a predominantly introverted profession take over the world. They donât fear the bunker.
Note, I have tried quickly to find a source supporting the idea that programming is a predominantly introverted profession without success.
Thereâs plenty of anecdotal support[1] and plenty of people asking why it would be true[2], but not much in the way of actual evidence.
I did learn that there are genuine physical differences between how introvertsâ and extrovertsâ brains work[3] and also found an interesting study saying that there is only the very mildest of correlation suggesting that introverts are actually any better at programming.[4]
Anyway, Iâve gone down a rabbit-hole that Iâve found more interesting than my original point.[5] Dear Internet, I COMMAND YOU to ignore my little diversion and stay on topic! [6]
Seems to me like people here are trying to solve too many problems at once.
In the short term, I think people are going to cotton-on to the ersatz nature of current LLMs. My Facebook feed is now inundated with generated images that look like 30âs movie posters and accompanying text that is blackboard-scratching bland (wait is that possible?). Yes, I am more sensitive to such things than the average person but I think theyâll eventually notice.
As for the reason why LLMs produce ersatz things itâs because there is no real-world knowledge built into them. You can ask it why an apple falls but there is no function in there that actually has seen an apple falling or feels getting bonked on the head by an apple. What it does have is a probabilistic map of what innumerable people have said about apples falling. Not at all the same thing.
Yesterday, I asked Anthropicâs Claude for statistics about the accuracy of the major LLMs on day-to-day queries. Iâll not burden you with the (good though prolix) result. It listed the impressive results for mathematics olympiads and the like and then, at the end, said there was little data on everyday questions. Why? Because, er, from anecdotal and a few more rigorous results theyâre terrible. 70% or thereabouts and perhaps the companies donât want to publicize that. How long before you fire a new hire whoâs only accurate 70% of the time?
I suspect LLMs are going to blow up and take a lot of investors with them because of this fundamental problem.
As for what happens over the next ten years? I donât know. Feeling rather quantum right now.
If theyâre making factual errors 30% of the time then theyâre still better than every university student that didnât get a first class degree.
If theyâre merely making judgement errors 30% of the time, this is probably your best employee ever. (Noise: A Flaw in Human Judgement by Kahneman et al is an interesting if a bit long and repetitive study on this)
Theyâre only accurate 70% of the time, but they also donât learn. Thatâs the fundamental problem: An LLM might be able to replace a brand new intern, but it canât replace someone with even a few years of experience. Except if you donât hire brand new interns, where are the experienced people supposed to come from?
Actually, they do. Thatâs kind of the point of them. They just donât learn from their own mistakes, rather from the collective evolution of human knowledge. In theory at least, that should be much much faster.
Actually, from my observed experience across a very wide market space, individuals all the way up to very senior roles indeed are outsourcing their own responsibilities and judgement.
They are still hiring the interns. Thatâs the group they are maintaining. The cheap ones who can borrow the knowledge, experience and judgement of the amorphous entity.
They arenât. You donât give Usain Bolt a bicycle. AI is to make the interns look experienced. You save money by only having inexperienced staff.
Iâm not saying this is a good strategy, but it is what is happening. Is it going flawlessly? Ask Deloitte.
Not really. To learn from the âcollective evolution of human knowledgeâ requires that the training set be updated, and then that the model be updated from the new training set, and then that the model be asked the question again. For me personally to learn a specific task, I just need a single human to say âthis way works better.â
Theyâre certainly trying to. Whether theyâre doing so successfully is another question. More to my point, the CEOs who think âAI can do everythingâ arenât firing themselves.
On further reflection, it also requires that the particular bit of knowledge be captured in electronic form somewhere that the LLM can access. Even in highly technical fields, âstuff thatâs written downâ is only a small fraction of the collective knowledge of the enterprise.
Indeed. Although in that case itâs probably fair to say add that the law of large organisations says that you have to tell people a lot more than once that âthis way works betterâ for it to start to make any kind of difference. And if your âthis wayâ is changing existing behaviours and processes, a further large multiple of times on that.
Oh, and these days you canât ask people (yes, even interns) to do things; they have to be persuaded. If interns just did what they were asked, AI would never have gotten considered by businesses, not even for a heartbeat.
Mm, where I come from, it has become the exception to find someone who shows up on time in the morning.
That may be a slight exaggeration but AI always works, is never in a bad mood, and doesnât need an empathetic supervisor.
How well AI works is not important. Because all companies will have more or less the same (poor or not) AI. Customers are welcome to switch to another provider, but that wonât change anything. Take it or leave it. Iâm afraid thatâs what the near future looks like
Hah if only. LLMs sometimes work, but also when claiming to be working, it may just be lying to you. Trust is a big part of this. LLMs very convincingly make stuff up.
LLMs are a big part of my daily workflow (software dev) but Iâm extremely wary using them. Just yesterday I had Copilot telling me the way to do something, only for it to say the opposite seconds later, then claim it had updated a file (which is had not actually touched).
They make everything up. EVERYTHING. It just that a lot of what they make up (70%, apparently, according to this thread) is either not fact-based or itâs coincidentally true.
Anyone who has worked in an office knows that three things will get you to the top very very quickly:
Agreeing with the boss
Just repeating what others have said or done and taking credit for it
not caring about the impact of your work, just the letter of what youâve been asked because you donât intend to be there long enough for the implications to bite you personally.
This reminds me of one of my students way back in the midâlate 1980s. He worked for British Steel (when they were still a substantial firm with plants all over the country!), and was a part time student.
One day in a discussion of I no longer know what, he told us that the best tactic in meetings was to listen and say nothing till the end, when you then just said what everybody else had been saying but in your own words! That way you would build a reputation as someone who really had a grasp of things.
but AI always works, is never in a bad mood, and doesnât need an empathetic supervisor.
