I wouldn’t assume a person is wrong about AI when they say it lies 80% of the time. Nor if someone else had the converse opinion.
If you are using it for topics that are well represented in the training process it’ll probably be much better than for topics which aren’t well represented.
If a person happens to be looking for help in a topic they know well that isn’t well represented, they might form a different opinion than someone looking for help in a topic that they don’t know well that is well represented.
At least that’s my theory, for why I find it so #$%@ing useless for my xen based homelab config.
I think there’s only so far we can go from our singular personal experiences.
I like to look at the studies coming out on AI effects in the workplace, on productivity and profitability, on learning outcomes, etc. We have “science” and empiricism as very helpful and successful tools in helping us manage this world, no need to listen to Sam Altman, we can observe the world and see what is happening.
I can ask AI to give me a line of code to search a Mongo database and it works 90% of the time. Great time saver.
Co-worker use AI to fill out a form so I could do work based on the information on the form. What would normally take me 15 minutes to setup, has now taken me 4 hours because of all the mistakes on the form. This co-worker just got an award for using AI to automate their workflow.
Because my work insist we use AI for everything, I now take a 2 sentence email and put it in AI to fluff it out and “make it look professional”. I send that to my boss who takes that email and ask AI to summarize it.
The pushback to AI is because people are using it for stupid stuff and turning their brain off.
Same. I’m mainly working with inherited course content at this point but have some assignment redesign coming up. Part of that’s assuming LLM use and arranging assignments to foster AI literacy, both in understanding places where you can use it and places where it’s going to attempt the useless like making line breaks part of variable names. I hope to have time to keep a gradebook for the big four (Gemini, Claude, Copilot, ChatGPT) all the way through though, not being in a position to do local inference, preventing the LLMs from serving up the assignment development in response to student queries is a concern.
If you’re referring to K-12 or continuing direct to undergrad because of parental expectations, I’d say the kids’ appraisal of the education system’s pretty accurate. Given a gig to do assigned things to get points to get a grade, copy/pasting AI output to turn in is a logical optimization. It’s been my experience engineering students are particularly rational about this.
I mostly teach folks who’ve worked for a bit and are coming back to classes to further their careers or to shift career positioning. So, as an instructor, if you show how course material facilitates solving on the job problems it connects a lot more than with folks at a life stage where they’ve only had the chance to work summer jobs.
The big big thing for LLMs in the next few years that I just realised this week is that those who are early adopters will have a year or more of context about them stored in the providers and get much better answers out of AI than those who don’t.
As you say, relevant context matters a heap and this is where google may catch up with the normies who have a google mail account.
I think there is a little bit of a bias from those who work in the industry when it comes to inevitability and perplexity.
I have used AI, both various frontier models, and my own local inference models (previously llama 3.3:70B and right now Gemma4:26b and Laguna S 2.1, and while they have been extremely helpful at times, they definitely have their limits.
I primarily use them for web assisted research, IT/homelab troubleshooting and some low key python scripting for home lab automation.
Sometimes they are brilliant and get me to the answer quickly.
Other times it takes several trial and error cycles but we eventually get there.
In some cases they have given me the wrong answer so many times in a row that I just give up, do a traditional search and find the answer almost immediately, after having wasted an excessive amount of time doing it the “AI way”.
They are certainly a helpful tool to have in my back pocket, but in my experience the error/hallucination rate is real. What this means - at least for me - is that I don’t trust them, and never use them for anything that I would consider important, or if the risks are high if I get it wrong. For making trivial shit easier though, sure. I’ll use them. Worst that will happen is that they waste my time.
I don’t think I’ll ever get to the point where I trust letting agents loose on either my digital life or in “meat space” though, which is why I have stayed with the chat oriented use cases.
I have seen how fantastically wrong they can get things on occasion, and I don’t want to set that loose on my files, my calendar, my work, or even my social commitments. I just don’t trust it for that kind of stuff.
That, and I consider the prospect of losing competence because I become increasingly reliant on AI to the point where I either forget how to do things myself, or never learn how to do new things because I am constantly using AI as a crutch, to be a really scary prospect, so I try to not over-use it.
But there are some really helpful use cases where I will try to take full advantage of the local models I have though.
For instance, I am in the process of creating a photo processing script with the help of my local models.
I have for a long time disabled anything and everything cloud on my devices because I hate the modern constant corporate surveillance reality, so I have been missing out on such features as backing up or being able to search my photos.
Instead I use a program to automatically sync/back up my photos to my NAS, and whike this ended the cycle of me losing all of my photos every time I had a phone die, it did not solve the aspect of making my photos searchable.
So I brainstormed ways of scripting this with Gemma4:26b (which runs really fast on my GPU) and then had Laguna S 2.1 (a 118B parameter reasoning model) do code review and code augmentation to create a python script I can ad to cron on my server in order to every night pass new photos to my vision model and create searchable text from them.
This is a pretty low risk application, and one I am OK with, because I can review the code myself for any catastrophic problems (I am not a programmer, but I think I can pick out things that would result in data loss, and besides I have my photo folder snapshotted every night. All the model would do is generate searchable text based on the images, which would then be written to a database. Worst case outcome I get some bad descriptions and my search is imperfect. Very low risk.
That’s the extent I’ll use the models for though. I have found that while they are still right most of the time, for anything that is important, even a single digit percentage failure rate is just too damn high for me…
…and I’m not sure if anything could ever change my mind on this topic.
Que? Every single person I know is paying for it and using it personal admin, home office and professionally.
And I don’t mean coding or it at all, I mean regular people, Im a production manager in a hemp lab, it gets used for Word, Excel, PDF formatting, memos, reports, all the things. It’s mostly a mix of Chatgpt and Claude.
Chatgpt has never been useful to me personally but most regular people millennial and late Gen z are using it’s paid service by now.
I guess our circles of friends and acquaintances differ greatly.
I don’t think I’ve ever met anyone who has used a paid AI service. (in person, I’ve certainly talked to people online who do) At least not anyone who has said they do.
Very few people I know use AI at all, even just for basic chat, and those who do have no idea what an “agent” is or how to have it interact with documents. They just use them as basic chat bots.
Some probably use the “assistants” built into their smartphones, not realizing that they are essentially using an LLM, or not considering it as such, but that’s about it.
I’m certainly the only person I know who has gotten into local inference setups.
The only person I know who talks about using AI, is this excellent programmer I know, and he talks about how he is more or less being forced to use it at work, and does it only because he doesn’t want to lose his job, and avoids it everywhere else.
Maybe its a generational thing? I’m late Gen X / Early Millennial, or “Xennial” if you prefer, if that makes a difference.
Or maybe a geographical thing? I’m in New England.
Northern Colorado area so yea might make some difference for sure. We have Microsoft and meta datacenters in Cheyenne, WY. Google not too far behind em.
There’s alot of pushback for sure, especially Cheyenne, for reasons.
But yes, everyone I know is definitely using paid services.
You hit the paywall, you just need answer, take my 20 bucks.
And it’s that perception is why I responded, that people aren’t using AI, it’ll go bankrupt, bubble will burst. Yea maybe a little but not to 90s for com level even. It’s too useful.
Nobody’s saying AI will go away; that’s not a necessary consequence of the bubble bursting. Hell, Nvidia will probably do quite well out of it (eventually). OpenAI, Anthropic, all the companies holding GPUs for DCs etc…they’re gonna go, and their DCs will effectively flood the market by being sold off for pennies on the dollar.
If you do the sums, to make a profit OpenAI needs every working-age adult on the planet to have a ChatGPT account, and it needs some 15% of them to be paying for it. That’s an unheard-of conversion rate, and it also assumes that they have no competitors. It will not happen.
But…if you want evidence that OpenAI and Anthropic are either going to go bankrupt or massively shrink their operations and then drop out of the market due to not being able to train new models at scale, you only have to look at Nvidia: over the last six months, Jensen’s been pushing self-hosting hard. There’s a reason for that - he wants every business in the developed world to be paying for their own private cloud instances, and he doesn’t care whether that’s using Chinese models or US models. That’s why he’s gone ahead and bought HuggingFace.
Yeah, thats what I don`t get. One editor of a german technical magazine I respect also told about his AI journey and it basically boiled down to that: “I installed this and that, gave it internet access and access to my files, it fixed my linux configuration for problem y”. I am thinking “dude, you are the first to cry wolf if you hit any telemetry, but you let a blackbox fiddle with your OS and personal files”. Because you are self hosting doesn`t mean its automatically safe or without errors. I have worked in support and I know that even small, stupid things can cause problems and nowadays most software is even more bloated and unreliable. What if the LLM decides tool x from AUR is the best for this job and downloads malware. While new tech is exiting and I would like a personal computer janitor myself I would have liked at least some warnings in his article.
Making images searchable: isn’t that something imich could do? Imich is a selfhosted photo library that does exactly that plus there are connectors and apps to sync your phones fotos to it. It can leverage AI for object and person recognition.
Well, I mean that is actual fact right now. They are selling subscriptions for a fraction of the cost it takes them to provide the services, and thus far businesses are the biggest users, and they are starting to massively balk at the cost, even at these below cost prices.
Every business I am familiar with that utilizes AI in their processes has started looking at their AI/Token bills and started questioning if it is worth it.
If they were charged an actual amount of money that would give the frontiers a profit, I’m not convinced anyone would still be paying.
Even big finance companies I am aware of who try to use AI for coding high frequency trading applications are looking at their AI bills and thinking, maybe we should just buy everyone Mac workstations and have them run open weights local models instead, because this pricing is insane. The take seems to be, “sure we can get away with less headcount by using AI, but damn, at these prices we are paying more than it would just cost us to hire more people…” And those are the below cost prices. Imagine what happens when the providers need to start breaking even or making a profit.
There is no doubt that AI is a transformative technology, but right now the big frontiers are selling it at a massive loss, and even then it is too expensive for what the market is willing to pay.
Sooner or later this reality is going to hit them, and hard. Either the costs need to drop significantly, or the capabilities need to improve significantly to where the pricing sufficient for them to make a profit is actually worth it to the marketplace, and both of those seem like a huge stretch right now.
According to the analysis by major sophisticated investment banks (Goldman Sachs, Deutschebank, etc.), as it stands right now, more money has been invested in AI than - when considering the time value of money in an NPV-style analysis* - can ever be made back from selling it.
*I don’t know what your familiarity with capital budgeting/finance type calculations is, but when investors invest in something they need to consider what the money they are investing is costing them. In other words, what is the opportunity cost.
How much is it costing me to borrow this money that I am investing, or how much could I be earning by investing this money somewhere else, maybe even in safe treasury bonds.
Every investment organization has an understanding of what their cost of capital (expressed as an annual percentage) is, and they use this to discount cash flows.
Lets say the cost of capital is 8%, then any predicted return cash flow a year from now is worth 92% of its face value. ($1000(100%-8%)). Any predicted return cash flow two years from now is worth ~85% ($1000*(100%-8%)^2). Any predicted return cash flow in three years is worth ~78%, etc. etc.*
Since this is compounding in nature, it doesn’t take many years until future returns are worth very little in terms of present value
In fact with constant cash flows (where you get a fixed amount of money every period in perpetuity the future cash flows - despite being eternal - can be expressed with a fixed present value, because returns many periods out are worth next to nothing, so continuously adding them winds up having near zero value.
This is what they are saying. If they add up all future cash flows over time, discounting them appropriately for the cost of capital, more has already been spent in AI than can ever be earned back.
In other words, what this means is that for any of the frontiers to actually wind up being profitable long term, most of their competitors need to go belly up or otherwise exit the market, with all of their investors losing their investments. And that should be enough to scare markets.
Right now, the only thing keeping them holding on and not running for the lifeboats is their irrational belief that they are somehow exceptional in their investing choices, and that their chosen investment is going to be the one that wins where everyone else fails.
…and with how fast this technology changes I don’t think even the smartest and most knowledgeable person in the world can confidently say that.
Let me provide my own little factoid to this financial analysis.
I have first hand involvement with 2027 business planning for F200’s.
The number 1-2 initiative for everybody I have seen is something-ai. Some number of agents-per-employee, some % cost savings due to ai automation, etc.
Ngreedia is going to be 100% sold out of products for at least the next 18 months if this pattern holds.
But. If 95% of those 2027 initiatives fail to deliver. Thats when we will see people start yanking their cash out. Unfortunately I dont think we will know for another year which direction it will go and how fast.
And with the market and technology changing by the hour, its too soon to say anything will or wont materialize.
Their is an ocean between what the would-be-techno-feudal-lords say about “AI” and what the real world uses of LLM tools are for the 97% of the world who doesn’t care about coding. There is no “AI,” we have no thinking machines, we have some cool language based tools that can infer from available inputs based on finite information.
Those tools are akin to the early hammers and stone tools- I look at them, and then I look at my Makita 80v XGT Demolition hammer. Sam Altman wants me to believe the former performs same as the latter. The tools are amazing because they are new, not because they are good- they very much aren’t.
If the LLM companies and leaders were even somewhat less of transparently ghoulish, power hungry, dead eyed sociopaths, they would have waltzed in as the promethean oracles of an exciting new future. Probably because of American politics, and culture, this emerging technology got some of the worst examples of humanity for it’s evangelists.
