I’m not sure how to phrase this so it is bit of a ramble…
In my work, a big push/play into AI, LLMs and Informatics is underway. My experience is that it can be really helpful but ultimately it misses the ability to make a tight call. In data rich areas, it can be helpful but in data poor areas it struggles and I’m not sure what it can bring (yet/currently). For people that have a limited view/experience of the field it can facilitate an advanced conversation w/o really knowing the ins and outs but as a start point it can be useful.
As a parallel, my son had a task from school - read (story) book about a journey to escape to the US over the Pyrenes during WWII. The idea being they then discuss the ins and outs of it in class. My son read it but a friend of his got his AI agent to summarise it. For some reason my mind went to - well is this a bad thing, i.e. not reading it but getting the distillation instead? Is this just a different way to learn?
Then of course there was the recent booing at a lecture given at Oxford (or some such) by an ex Google exec when he mentioned AI… the irony was not lost - happy to use AI to write up coursework but if it will take the job I’m paying to study for, well that is just wrong!
In a way this whole thing will pass me by as I, at the most, will have one more job in me, and have (hopefully) gained enough experience to ride out the next 5-10 years before AI really bites.
I can summarise for you
Immigrants=bad/too many, you out of touch lefty vs stop reading that online shit, you stupid right winger
You can thank me later
Yeah it is a meaningfully worse way to learn. And I think this where we need to be keep front of mind the purpose of getting the information being sought. Chris Hayes (an MSNBC anchor) has been doing a lot of good work on this stuff (less from a tech perspective and more an a meaning perspective) and points out that there are tasks where sometimes the struggle is the point. This is obviously true for something like exercise, but is also true in a lot of professional work cases and undoubtedly true when it comes to education or learning in general. There is a significant difference in what your brain goes through in reading a book and coming up with a one paragraph analysis vs simply being able to parrot one you got from an LLM. Likewise with writing. Lots of people think that good writing comes from a mind with a clear perspective, but in most cases the act of writing reveals the flaws of your thinking and it is the exercise in itself that clarifies your thinking allowing your 10th draft of something to appear as the product of a clear mind. You may get there quicker using an LLM to help you structure your unstructured thoughts and find the flaws in your position, but what is that doing to your thinkin? How different is how you are able to think about the issue different at the end? And that is before you even get into issues about the mind as a muscle.
So I guess for me the main consideration is what is the point of the task. If you are at work and have a comparable situation, simply surfacing the relevant information in summarizable form is likely sufficient for your needs and it doesnt so much matter if you’ve internalized the lessons from that exercise. In school it does though as that is the point of school. I think there are likely a lot of ways in work where this will matter as well where it is important that moving forward you are the owner of the insight and if you outsourced the generation of that to an LLM your ability to correctly apply that over your domain will be lacking.
I personally get a lot of use of the simple “explain this to me like I’m an idiot” aspect of it being a super charged personalized google search. I work in tech and am not techy so find myself in a lot of conversations where I need that sort of quick explanation for WTF these people are even talking about. Far better than a google search ever did, allowing you ask “why” 3 or 4 times really effectively gets to the point that you understand it to the level you need to better than a Wiki ever could. But my best use of it is in running what are essentially red team exercises. A lot of my work is in identifying and evaluating trade offs and that is work were its hugely useful to have people to run your work by. In my past I’ve worked in structures where other colleagues had to review and pressure test my work. It was really productive and valuable, but I dont have that in my current role. Copilot has filled in the gap really nicely. You have to be careful because its default is agreeableness to the point of sycophancy so my prompt has to actively redirect it to a perspective of skepticism. And there are now countless cases where its just gone insane and answered completely different questions, sometimes cross pollinating from other conversations when it isnt supposed to be able to. But honestly that is no different than working with a human.
A lot of my work requires synthesizing large volumes of information and drawing insights from it. That is one of the commonly stated use cases for LLMs but I strictly avoid that for most of my work. LLM’s are not good at reading the subtext of things people say unless it has access to a broad range of ancillary data and conversations where someone has pointed out “when x says y he is really talking about z” and that is really important if your insight is coming from human feedback. But the bigger issue for me is why I need that info synthesized in the first place. If that is the source of all the recommendations I make then I need to do the work to understand it. There are going to be connections I miss downstream if Ive outsourced this work of understanding to the LLM. Moving forward I think this is going to be one of the key drivers of who is good at this job (producing good outcomes) vs who can move quickly. Unfortunately I think a lot of corporate decision makers will favour the latter which will be shit for the job market but also shit for consumers of anything built in this way.
Dawkins was a grifter, he used highly ‘constructed’ phrases from the Bible to debase established philosophical issues, which he hardly ever went into.
As well, he published at a time when the paradigm was strongly veering into nihilism and postmodernity, and as well, widening liberties, all which made his work seem timely and correct. Its was actually a trashy philosophy.
As to AI, I rest with Prof Penrose, computation is not consciousness.
I saw an article yesterday which suggested LLMs are simply going to replicate the underlying philosophical conjectures we already have. How else can they escape that trapping?
AI bubble is going to crash badly and lots of professional investors are already giving warnings, and linking this to other large corporations engaged with Ai going for public share sales this summer. All which indicates the sector knows this is maximum recovery time. It will begin to crash after those peaks.
The point about reading the book and reading the shortened summary was to me an interesting point - is the extracting the information important or the reading or the reading and the extraction. I for one would be for the struggle - but that speaks more to my underlying mindset of the struggle being the point of life. A 14 year old may see reading as a unneeded struggle - time away from gaming or footy or maths or sleeping. Will this particular struggle be of use in 25/30 years down the road for the individual? The risk is that we do everything like that - get AI to “help”… can’t escape the feeling that eventually the problems will catch up.
May companies will use AI/LLM to optimise (i.e. cut costs) but it feels like it will come at the expense of the core trade by degrading the essential and core expertise of said business.
p.s. agree that writing is the best way to actually explore what one actually means… I obviously struggle
I had a strange disagreement with a poster on some other forum on a related subject. The premise was around Neumann probes - self-replicating space probes built for what ever purpose (mining, warring, exploring, watching etc..). I ventured that if such machines could be created would they not eventually evolve? The other poster was very much in the “No” camp and me in the “Possibly yes” camps. Anyway human consciousness “emerged” at some point in the last 0.5 M years and so why could machine consciousness not similarly emerge at some point? It’s only been around for the last 50 odd years after all.
That is the point. The struggle will always be of use because that is how you actually come to understand things, or in the case of skills be able to do something. The incentive is going to be huge for those in education to bypass that and get straight to the output that approximates what someone who understands could produce, so I’m not arguing about what a 14 year may perceive as valuable. Only about the implications of people doing that on things where the friction is the point, where the point is to learn something. Or especially in the case of kids and education, in learning how to learn.
Is there a future where knowing anything is made redundant by everyone outsourcing that to personal LLMs and agents? I think that is very unlikely, not just from a technological perspective but because even if that was possible eventually there would be push back at how grim an existence that would be. So what that makes a very real possibility is a large percentage of the population who can approximate knowledge but actually know nothing. And that is similarly grim (or, the UK politics thread on TAN)
in construction, in Australia, id argue that is already the case, not that its particularily relevant to AI just yet, but yes, we now have an industry where its more important to know how to look something up, than to actually perform the task
Interesting paper here. One of the common job related concerns of AI is the jobs it will be best positioned to replace are entry level jobs, the ones people need to learn enough to do the jobs that will still exist alongside AI. There has been a slowdown in hiring of these roles and the conventional wisdom is this the AI effect already hitting
This paper suggests this isn’t being driven by outsourcing to AI but attitudes related to work from home - the perspective junior workers require supervision that is too difficult to provide remotely and so it’s the work from home junior jobs that have disappeared largely accounting for the slow down.