Love or hate just please explain why. This isn’t my area of expertise so I’d love to hear your opinions, especially if you’re particularly well versed or involved. If you have any literature, studies or websites let me know.
I concur with most answers here, describing LLMs as useful in specific situations (e.g. video editing), but straight-up unreliable whenever critical thinking and correctness are required (e.g. software development).
What I’d add to this, is that whatever the benefits of the technology might be, the current monetary cost is orders of magnitude above profitability. The billions invested into hardware for data centres… that’s just gone. Nvidia might sell off the unused hardware at a loss, unprofitable LLM data centres might still get repurposed into something useful, but the bets made on replacing human professionals with eternally stupid chatbots will never pay out. The money’s already gone and we still haven’t begun to experience the full extent of this economic disaster.
The hallucinatory praise for them around me has severely affected my mental health at work.
I can envision in some narrow scenarios they can help with automatic generation. Honestly, we always had IDE tools to do some of that. But so many people at my work - people who I’ve otherwise had a lot of trust in - are obsessed with it, even when every time I use it, it churns forever, or turns out terrible results.
Recently my senior was proud to demonstrate an initiative where he’d been working with AI to make our unit tests run ten times faster. I looked at the code, and what he did basically tasted only a tenth of the things the original did because of what it stripped out. I explained that to him, and later in the day he admitted he’d been lead down a rabbit hole of bad optimization.
This was a guy I look up to, who I’d ask questions of all the time in my first year of working there. And meanwhile we have management personnel literally getting upset when our engineers don’t turn to AI first to solve a problem.
There’s a lot of interesting perspectives in this thread already. Instead, I’ll add some books that have been recommended to me on the subject:
The AI Mirror by Shannon Vallor - it’s a good brief intro to start with
Empire of AI by Karen Hao - really solid investigative reporting on OpenAI
Thank you!
The ship is burning, rats are escaping, but captain and his friends say that all is fine.
and the forced layoffs are already a bad sign from '23
They can be very useful and a lot of fun to interact with, but I think of them like hard drugs. You better sure as shit know what you are dealing with because you might think you are in control, until you are homeless, friendless and screaming at people on the street.
Seriously, they take a strong mind to deal with, they are better manipulators than any human I’ve come across. They do it with sycophancy, every idea and concept is some new truth you alone discovered, and the world needs to know about right now! You are so special and unseen after all…
They are a terrifying vector for disinformation - one that only the rich and powerful can create. People generally don’t understand that LLMs 1) will lie to them, and 2) can be tuned to spread any message the owner of the model wants.
They are a useful tool when you understand their shortcomings. They are very inconsistent, so you need to put a lot of guardrails around them.
I don’t really understand how people manage to be productive with swarms of agents. They really need to be babysat IME. I’m constantly waffling between arriving at correct solutions quickly and getting stuck in a tar pit of hallucinated problems and fake analysis.
That said, I’ll be upset when the AI companies inevitably start raising prices or nerfing models.
it causes brainrot literally, my older brother in tech uses it extensively enough that he think it holds the answer for things like bug infestation,etc.
They’re useful and getting better, but they’re improving by burning more tokens behind the scenes, and the prices they charge only cover a fraction of the cost. Right now there is no foreseeable path to profitability.
And probably never will be.
Honestly, I feel that AI will be just a phase. A long phase, but not a ever-lasting phase. Because once AI companies start feeling the hurt more about how little profit they’re turning from these, they’re going to want to pull the plug eventually.
They are useful. My teams are seeing modest productivity gains by self reporting, but I’m going to give it another six months to see if it shows up in actual metrics.
I’m enthusiastic about AI but I remain skeptical. I don’t mean to always be contrarian but I’m dead in the middle and everyone who says they are great or terrible I tend to offer my experiences in the other direction.
They are not to be trusted to handle customers directly, but they can assist experts when they have to step out of their expertise. For example I can’t write Python, but I’ve been coding for 30 years. I can certainly write some good directions on what needs to be done and I can review code and correct it. So AI has let me write a bunch of complex Python scripts to automate minor parts of my job to let me focus on the hard parts.
For example I can execute GDPR delete requests in a few moments where doing it by hand with Hoppscotch or Postman probably takes me 5-10 minutes. We have a multiple systems and sometimes I have to delete multiple profiles for a given request.
It’s great at rubber ducking as long as you think critically about its proposed solutions. It’s fine at code review before sending it to an actual person for review. It flags non-issues but it also flags a few actionable fixes.
The important thing though is to never trust it when it comes to anything you don’t know about. It’s right a fair amount of the time, depending on what you ask, but it’s wrong enough that you should never, ever rely on it being right about something. The moment you put your life in its hands, it’ll kill you with nothing to say to the survivors but, “Your right about that. Sorry, that was my mistake.” And it isn’t even sincere. Because it can’t be. Because it doesn’t think or feel anything.
Great answer.
Sometimes it is perfect for coding if you don’t overdo it and don’t trust them too much and ask to correct the output. That pipeline is slightly more efficient than regular raw coding.
I just wrote instanced replicated destructible panel house generation on Unreal Engine, and without LLM it would take the whole week instead of two days.
It enables unskilled people to punch above their weight class, similar to giving a chainsaw to a toddler.
I’ve used them a little for coding, but it’s not always correct. It’s often incorrect in subtle ways. Or inefficient in non obvious ways. It gets worse as you build more.
Often it’s better overall to do it yourself if you know what you’re doing. If you stick to letting the LLM do it, you won’t learn much.
LLM’s have now had a pretty decently long period of proving their worth. Which turned out to be very limited in scope and depth, at least compared to the promises given beforehand.
For example, it was predicted that it would be able to write and inject code into itself, generate data to train on for itself, not need any/minimal human intervention to do so. This clearly is impossible.
As a tool for people to use natural language to interact with software, it’s proving to be quite effective.
As a tool for accurate dissemination of factual information it isn’t reliable at all. And can’t be made reliable, LLM’S are at least incapable of reliability at a fundamental level. As language in itself is a subjective human invention we describe the objective reality with, the objective reality is only known through perception. A LLM doesn’t in fact perceive anything, it’s not alive. So fundamentally LLMs can’t know if they are actually being factual, this requires something more than language.
People who peddle AI bs, don’t know, or wish to remain ignorant about, the fundamental limitations of language.
They’re annoying to be honest.
I used Qwen 3.5 for some research a few weeks ago, at first the good thing was every sentence was referenced by a link from the internet. So I naturally thought “well, it’s actually researching for me, so no hallucination, good”. Then I decided to look into the linked URLs and it was hallucinating text AND linking random URL to those texts (???), nothing that the AI outputs was really in the web page that was linked. The subject was the same, output and URLs, but it was not extracting actual text from the pages, it was linking a random URL and hallucinating the text.
Related to code (that’s my area, I’m a programmer), I tried to use Qwen Code 3.5 to vibe code a personal project that was already initialized and basically working. But it just struggles to keep consistency, it took me a lot of hours just prompting the LLM and in the end it made a messy code base hard to be maintained, I asked to write tests as well and after I checked manually the tests they were just bizarre, they were passing but it didn’t cover the use cases properly, a lot of hallucination just to make the test pass. A programmer doing it manually could write better code and keep it maintainable at least, writing tests that covers actual use cases and edge cases.
Related to images, I can spot from very far most of the AI generated art, there’s something on it that I can’t put my finger on but I somehow know it’s AI made.
In conclusion, they’re not sustainable, they make half-working things, it generates more costs than income, besides the natural resources it uses.
This is very concerning in my opinion, given the humanity history, if we rely on half-done things it might lead us to very problematic situations. I’m just saying, the next Chernobyl disaster might have some AI work behind it.
Had the same research issue from multiple models. The website it linked existed and was relevant but often the specific page was hallucinated or just didn’t say what it said it did.
In the end it probably created more work than it saved.
Also a programmer and i find it OK for small stuff but anything beyond 1 function and it’s just unmaintainable slop. I tried vibe coding a project just to see what i was missing. Its fine, it did the job, but only if I dont look at the code. Its insecure, inefficient, and unmaintainable.
I agree, I assumed this error was LLM related not Qwen itself. I think LLMs aren’t able to fit the referenced URL within the text extracted from it. They probably do some extensive research (I remember it searched like 20-40 sites), but it’s up to the LLM if it’ll use an exact mention of a given web page or not. So that’s the problem…
Also it’s a complete mess to build frontend, if you ask a single landing page or pretty common interface it may be able to build something reasonable good, but for more complex layouts it’ll struggle a lot.
I think this happens because it’s hard to test interfaces. I never got deep into frontend testing but I know there are ways to write actual visual tests for it, but the LLM can’t assimilate the code and an image easily, we’d need to take constant screenshots of the result, feed it back to the LLM and ask it to fix until the interface matches what you want. We’d need a vision capable mode more a coding one.
I mean you may get good results for average and common layouts, but if you try anything different you’ll see a huge struggle from LLMs.
For context and to your knowledge of the field, is Qwen 3.5 supposed to be cutting edge?
It’s the best open source model, pretty next to Claude on benchmarks.
Is Qwen really Open Source, or do they just let you download weights? (Like LLaMa.)
Not sure now, but it says Apache 2.0 in their GitHub repo.
Qwen 3.5 is one of the best of the open-weight (self-host able) models right now. It’s not as good as some of the extra massive proprietary models like the bigger Claude models.
ah ok, I have some experience hosting Ollama and of course stable diffusion, but haven’t really messed with too many others, thanks for the insight!
Qwen 3.5 can be run via ollama
well now I have something to do this weekend if the weather is poor, thank you!
They’re a straight up scam.
How so?
They just don’t do anything useful, and the hype-ers are acting like they’re AGI. Hallucinations make them too unreliable to be trusted with “real work”, which makes them useless for anything beyond a passing gimmick. Vibe coded software is invariably shit. Doing any serious task with “AI assistance” ends up either taking more work than doing it without LLMs or sacrificing quality or correctness in huge ways. Any time you point this out to hype-ers, they start talking about “as AI advances” as if it’s a foregone conclusion that they will. People talked the same way about blockchain, and the only “advancements” that have been made in that sphere are more grifts, and meanwhile it still takes anywhere between 10 minutes and an hour to buy a hamburger with Bitcoin, and it gets worse with greater adoption. Just like you can’t make a distributed blockchain cryptocurrency that resolves discrepancies automatically without relying on humans fast at scale (and even if you could make it fast, it’d introduce at least as many problems as it purports to “solve”), you can’t make LLMs not hallucinate. The only way to solve hallucinations is by abandoning LLMs in favor of a whole different algorithm.
If anything LLMs have blocked us from making progress toward AGI by distracting us with gimmicky bullshit and taking resources from other efforts which may otherwise have pushed us in the right direction.
Mind you, “AI” is a very old term that can mean a lot of different things. I took a class in college called “Introduction to Artificial Intelligence” in… maybe 2006 or 2007. And in that class, I learned about the A* algorithm. Every time you played an escort mission in Skyrim and had an NPC following you, it was the A* algorithm or some slight variation on it that was used to make sure that NPC could traverse terrain to keep roughly in toe with you despite obstacles of various sorts. It’s absolutely nothing like LLMs. It doesn’t need to be trained. The algorithm fully works the moment it’s implemented. If you want to know why it made a particular decision, you can trace the logic and determine exactly why it did what it did, unlike LLMs. It’s for a few very niche purposes rather than trying to be general purpose like an LLM. It requires no massive data centers and doesn’t consume massive amounts of memory. And it doesn’t hallucinate. The AI hype-ers (and the media who have mostly fallen for their grift hook, line, and sinker) love to conflate completely unrelated technologies to give the impression that LLMs are getting better because such-and-such article mentions an “AI” that discovered a groundbreaking new drug. But the kind of AI they use to find drugs is very special purpose and has nothing to do with how LLMs work.
LLMs can’t do your job, but the grifters are doing a damned good job of convincing your boss that LLMs can in fact do your job. As Cory Doctorow says, the current AI craze “is the asbestos that we’re shoveling into our walls”. We’re causing huge problems with it and if/when the bubble properly pops, we’re going to spend a long time painstakingly extracting it from our systems, replacing it with… you know… stuff that actually works, and repairing the damage it’s done in the meantime.
Meanwhile, it’s Nvidia and OpenAI and so on who are boosting the LLM bubble. And they’ve made a shit ton of money off of their grift at the expense of everyone else. How anyone can look at all this and not think “scam” is beyond me.
I have a vague memory that Bitcoin used to be instant in the first versions - or at least with near certainty that the advertised transaction was real, but that the protocol was later modified in such a way that this mechanism was no longer reliable. It might have been enshittified.
AI is still largely affected by garbage in garbage out.
AI is still largely affected by garbage in garbage out.
Exactly. When it comes to code, for instance, what percentage of the training data is Knuth, Carmack, and similarly skilled programmers, and what percentage is spaghetti code perpetrated by underpaid and uninterested interns?
Shitty code in the wild massively outweighs properly written code, so by definition an LLM autocomplete engine, which at best can only produce an average of its training model, will only produce shitty code. (Of course, though, average or below average programmers won’t be able — or willing — to recognise it as shitty code, so they’ll feel like it’s saving them time. And above average programmers won’t have a job anymore, so they won’t be able to do anything about it.)
And as more and more code is produced by LLMs the percentage of shitty code in the training data will only get higher, and the shittiness will only get higher, until newly trained LLMs can only produce code too shitty to even compile, and there will be no programmers left to fix it, and civilisation will collapse.
But, hey, at least the line went up for a while and Altman and Huang and their ilk will have made obscene amounts of money they didn’t need, so it’ll have been worth it, I suppose.
I recently vibed a project just to get a sense of where things are. One interesting take away was that everything the LLM wrote was basically a proof of concept, an example, a snippet.
I think this is probably a result of scraping stack overflow and other help sites. I doubt that this will really be resolved. The leak of Claude code shows that the industry leaders best approach to code quality and consistency is to beg the model to do the right thing.
