Assuming its current most familiar and popular form - the LLM?
I think we may have already reached the point of marginal returns. Not enough available energy to drive useful output that can effectively replace the best people within each relevant domain and industry. This post below articulates my views on the subject better than I can.
What are you guys observing out there in the real world?
My job has tried to get us to use an inhouse “AI” that is supposed to help people troubleshoot machine issues. Only the most incompetent people seem to use it and it functions largely as a glorified search engine of past failures. At best this is a crutch for people who will never be good at the job. In the meeting where they first began encouraging us to use the AI, I told my manager that we should try using regular intelligence first.
Every one seems desperate to leverage what they are told is a revolutionary technology. I get middle manager types sending teams message attachments that are Copilot-created excel spreadsheets interpreting big chunks of data. These will sometimes point to interesting trends. I get others sending messages that are clearly written by AI which is nauseating. At a managerial level, it seems to be largely a tool for people who do not understand how to use existing purpose-built tools that do exactly what they are after.
For my part I have used the google AI at work when devices are failing in apparently novel ways that the manual does not provide for. It is useful for this but is mostly just a shortcut for the same information I could get using a regular search engine, if search engines hadn’t been completely pwned by SEO.
I have also used it in my personal life to help in circuit design but one must be very very careful, limit the scope of what you asking of it and triple check its reasoning and math. It did help with some tasks that would have taken a bunch of trial and error to get right.
I’m not really sold on it, to be honest.
AI reminds me a lot of self driving cars; they’ve been “almost there” for like 15 years now yet any time I order a taxi there’s still a driver at the steering wheel. That being said I use Claude every day, primarily to help me write. It’s interesting to see what it’s good at, which is to interpret text and point out it’s flaws, but ask it to write something original and it’s laughably bad. Like Bardmau said it seems to be pretty good at certain low level work, but if you’re actually good at what you do there is no need to panic about it taking your job.
This is very telling. It reminds me so much of the work I was doing in my previous life at corpoglobohomo world. We were migrating all of our on-prem infrastructure and software over to the cloud.
And of course, back then, the suite of cloud software and tools just didn’t work all that well. The company was hemorrhaging money executing migrations that took months longer than what was suggested by the cloud experts. And even in the event of successful migrations, the existing features never worked as promised. I’m not sure if things ever improved since I departed before I could see whatever the end result was. But everything took months & years longer to complete and whatever the finished product was ended up costing more and working less than whatever they had originally implemented (I’m sure that this wasn’t the case with some components - I can only speak to huge chunks I was personally privy to).
I just also wonder if these LLMs are already somewhat at the max of what they’re capable of doing. It seems like it tops out at just automating workflows and being a more efficient personal assistant to you than you can be for yourself. At least as far as what would be useful to your average consumer.
I actually don’t think Claude is that good at interpreting text or pointing out its flaws. Probably better than the average person, who although they can read, is functionally illiterate and possesses minimal readhing comprehension.
What I have seen it be useful for in regards to writing is as a dictionary and thesauras. But it can’t evaluate sentiment or quality of writing any more than it can do that with anything that requires sentimental judgement (which is to say it can’t do these things at all, it just spits back out what you tell it).
I think yes, AI is in a bubble that will burst and crush so many brown 3rd worlders’ dreams. I support it, not in the gay “AI art le bad” or “AI consumes water?? whaaat” way. I don’t like AI because it’s not real innovation. The concept itself is a scam, too much exposure with too little output. AI is here to stay, yes. But it will not be as strong as people think it is.
AI is both proloteriat and elite at the same time, that I respect. It made self expression easy and freed it from the shackles of artists, but it also made self expression worthless. The people with worthy ideas abstained from using AI because it got associated with “slop”.
Furthermore, AI can’t sustain itself via profits. As mentioned in this thread, there isn’t enough demand for AI. The complex grew too big too fast, and it will burst at some point. China’s already bleeding goypoints in its stock exchange, I give it until 2030 the western world follows suit.
In conclusion, yes, AI is in a bubble that will burst badly, but what remains from the wreckage is here to stay.
I think the AI phenomenon and its ongoing success, has much more to do with the anxiety that is inherent with being near the top of the S-curve and desperately needing the sort of gains you were getting further down than with any of its actual merits. I think this is also why the Altman’s and musks are so correspondingly desperate to convince every one how revolutionary this technology is. Did the bipolar junction transistor need this level of hype?
The other end of it is that is more or less necessary that search engines did put AI assistants at the top of their searches because of how shitty search results have gotten. Like a lot of things this whole thing seems to exist at an intersection of what the redactors call enshittification, diminishing returns, the laws of low hanging fruit etc. Not to get any one’s hopes up but the ridiculous amount of money being spent on AI does make me think of old Tainter and his whole ‘the decline of marginal returns on investments in complexity is what leads to collapse’ idea.
I might have posted this before, but the most feasible-sounding AI hype isn’t about doing new things, it is all about shedding social complexity in routine operations. Think 1-click lawsuits and automating away bureaucracy.
Of course, the former can’t work unless the court system can adapt to that strain, and for the latter DOGE was an absolute failure. But we do get solutions that create their very own new problems and leave them half-solved.
They might not just be responding to the waning ROI half of Tainter’s thesis. They might be praying for a technology to decalcify our institutions, give society the surplus required to reorganize without totally disenfranchising too many people at one time. This would allow them to avoid total collapse.
This is arguably the thing that AI is worst at. It doesn’t understand social complexity or context in the same way that we do. It seems to not be able to hold onto that much context in working memory while it completes whatever operation its assigned.
Therefore the output you get is something approximating what you asked for, but to any well versed person within the domain related to the operation it comes off as someone trying to create something that LOOKS like it makes sense on the surface, but lacks any substance underneath that veneer.
This is also why mathematicians and engineers constantly double check and triple check their AI driven work as it just drops important variables and numbers all of the time without reason out of nowhere, or forgets what it’s already done or defined previously so then further wastes compute resources redoing old work. Now if it’s dropping empirical quantitative context, just imagine how far removed it must be from completely qualitative and conceptualy abstract context.