New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Comments
It isn't really a positive thing that it will replace a lot of jobs, high unemployment brings boredom and boredom brings crime. But it probably won't happen that soon, hallucination in LLMs is mathematically impossible to solve and it will never be reliable enough. Other kinds of AI such as for medical use are useful, but LLMs will fall out of favour with time and likely be replaced in future as we understand emulated conciousness more.
Build more RAM manufacturing factories.
The AI market is yet to be saturated.
Just wait a year or more. After that It (RAM) will be cheaper than ever. You can change product but the cycle will remain same because basics will never change.
Find the root cause of this ridiculous price hikes, I don't think we are short on supply as long as you can pay for it to acquire regardless of origin. So that is more of price war and holding demand by suppliers & factories.
Demands for AI is just an excuse.
That's how we end up with unaffordable houses. People investing for the sake of investing (not even as a landlord in the case of housing).
But, I think LLM market will grew even when it crashed, as the problem lies on the supply and demand mismatch, resulting in overinvestment of greatly depreciating assets.
For example, let's assume currently (2026) there's one rack producing 1M tokens/s. Because of bandwidth improvement alone, we could assume the rack installed in 2029 will be capable of supplying 2M tokens/s even with larger model.
Yes, it should burst because paying $135 for 16GB DDR4 doesn't make any sense.
The bubble will burst because of malinvestment. There's no reason to overbuild the capacity of an underutilized asset that will soon become obsolete due to the hardware generation cycle,
[citation needed]
I hate LLMs as much as the next guy, but that's a pretty strong claim.
OpenAI has admitted this themselves. https://arxiv.org/pdf/2509.04664
It boils down to an inherant flaw where if the LLM is not certain, it makes a guess, but presents that guess as if it is fact.
It may change in future of course if they make some advancements, but I doubt it any time before the crash happens.
There are still a lot of use cases for LLMs as simply an interface between the user and complex data. They don't need to be completely infallible when answering questions. If you provide them access to a dataset, definitions about how the data is structured, and rules about how to query it, they can turn a normal high school level user into a data analyst. Like everything else, AI is super polarizing. It seems there are two camps, one super pessimistic and one super optimistic. The truth is somewhere in between.
There's nothing you can do really, you just have to wait for the AI bubble to pop.
After that, as the AI corporations fail and get bailed out by governments using your money maybe you'll be able to get some of the old hardware they no longer want but not the new kit, the new kit is needed to feed the next bubble.
That paper only explains why hallucinations are common. Where does that claim mathematical impossibility? In fact, it states the opposite, saying that the issue is in training incentives and is (theoretically) solvable:
And, for a trivial (but impractical) counterexample to any mathematical proof:
Either you didn't read that paper, or you linked the wrong one by mistake.
They just need to say, "I'm not sure, I'm only X% confident in this answer.". Solved.
US government already said they won't bail anyone out. But knowing Trump, he'll just wait for bankruptcy and then change his mind.
Sadly that's harder than it sounds. Maybe one day, but for now, LLMs are unreliable and garbage.
Optimize everything again so all can run smoothly in 2GB ram ddr2.
It will. People are on denial if they think otherwise.
need some cheap chinese ram so I can upgrade my dated 5950x 64gb ddr4 system, I guess I could sell my 64gb ddr4 ram for 64gb ddr5 ram, to help cushion blow...
That's what Math does; strong claims. Venture into Math and you can be sure you will find strong claims all over the place, many strong claims. That's one of the beauties of Math.
"It may change in future" you say. Yes, that's true, it will change when the model changes but then it will not be LLM anymore it will be something better than LLM.
EDIT: Perhaps a world model will not suffer from the same flaws, "large language models" seem to be prone to some serious flaws that cannot be solved.
Exactly, but the posted research paper does not make any such claims, and in fact claims the opposite with a counterexample. I deal with cryptography, so I'm well aware of both the powers and limitations of mathematical proofs.
While LLMs are prone to serious flaws and I suspect it would be impractical to solve all of their flaws reasonably, it's still wrong to state that there exist mathematical proofs that it is impossible to solve since, after all, making such a claim necessarily summons a computationally-unbounded adversary agent.
LLMs evolved not much recently, the recent improvements in output we're seeing are based on better/more training data and more computational power, much more computational power.
In short: BRUTE FORCE efforts, as we know such efforts never scale well.
If the models are not improving much beyond BRUTE FORCE effort then what the future holds is merely more of the same and little to no innovation.
EDIT: OpenAI claiming DeepSeek distilled its models to gain an edge (https://www.bloomberg.com/news/articles/2026-02-12/openai-accuses-deepseek-of-distilling-us-models-to-gain-an-edge)
There's not much innovation going on in the LLM field as LLM models are already feeding on each other, this field has hit a snag.
EDIT2: LLMs killing activity on stackoverflow and similar sites, major sources of training data, is bad news for AI coding assistants, LLMs are killing the geese that lay its golden eggs.
Tl;dr no math proofs.
AI companies are valued as if everyone will soon use AI for their daily work. But that only works if people already use modern software that can easily connect to APIs.
Most small/medium sized businesses don’t. They either use basic, rigid software or older custom systems that don’t integrate well with AI tools, no APIS or very limited. Connecting AI to those systems would be expensive, complex, and not worth the return, at least not on the short term, those companies today don't usually rely on software integrators. They do however on system and network integrators just to keep their systems running, and doing what they do.
I think small businesses are actually the main group AI companies hope will adopt these tools, because all i'm seeing is adverts for AI agents making your life easier, but when you dig deeper everything is connected to some sort of saas platform, in which you could easily ask any software integrator to build a custom workflow, or even if you would read the docs and have a little technical expertise, build your own, because the API is right there.
So the assumption that AI will quickly replace everyday work across the economy ignores a big reality: the technical and financial bridge is too wide for most companies to cross.
And students comparing notes with a chatbot doesn’t create real economic value.
At least, this is my point of view.
I believe it's going to burst first because of extreme overvalueing the technology behind 'AI', return is way too little on investment, early investors hyped this for the initial investors who want to see returns, but i don't think that's likely on short term.
Lets put ads in our chatbot and in our start menus.
OpenAI Claims DeepSeek Distilled US Models To Gain an Edge (bloomberg.com)
Open> @TimboJones said:
Perhaps you should read this: https://arxiv.org/pdf/2602.07238 AI isn't getting smarter, it's getting more power hungry. And expensive.
Posted on the arXiv preprint server. MIT research.
Form a global regulatory body to regulate RAM purchases
cant make ai more 'smart' than it is on current tech only 'train' more data
I agree, but the claim was about a mathematical proof of impossibility. One could say colloquially that it's impossible that
cat /dev/urandomwould spit out the cure for cancer, but I'd object to anyone saying that it's mathematically impossible. In fact, any proof that it's impossible would also constitute a break in the ChaCha cipher!Honestly, a lot of it is Stack Overflow's own management's fault. Sure, LLMs have directed a lot of people away, but the management is absolutely insane and refuses to listen to the community. The sharp traffic drop actually started before ChatGPT itself was introduced, but it and its ilk certainly helped accelerate the decline.
What management? What community? The SO community is quite insane, in stead of helping newcomers the userbase is dismissive and condescending, and they get rated for it. You'd look days for a missing semicolon on your code while learning only to receive either RTFM comments or This question has been asked before. Yeah no shit sherlock, takes you less time to point out my semicolon then to point me in the direction of the manuals, the docs i have NEVER SEEN BEFORE WOW THANK YOU FOR YOUR GUIDANCE.
I am however very curious where all the salty neckbeards are clustering together after SO, any tips are greatly appreciated.
Rguarsd
And then when you search for something on Stack Overflow, instead of finding the answer, you find the same question you have, with people saying RTFM, or it's already been asked, or it's not an appropriate question for the forum.
And then when you ask a question on Stack Overflow, it's closed because it's an opinion-based question. Which, okay, it's "not a fit" for their format...but they could have had so, so much more traffic if they allowed people to ask for software recommendations or ask what's the best way to do something. Instead, they stubbornly clung to their Q&A format and became dry and now irrelevant.
And if they had set as their absolute goal, I don't know how they could have torqued off their community more. At a time when they desperately needed them.
And as one member put it: "If I were completely new to the network, what would motivate me to answer a question? The company's mission to help Google and other giant tech corps train their AI with good human-sourced data so they can put it in some developer's IDE and charge for it doesn't inspire me. It also completely bypasses/obsoletes any existing answers on Stack Overflow, so there's no hope of building something of lasting usefulness, especially now that the data dump is sequestered on the company's private servers and access to it has been restricted."
Their traffic is now roughly what is was the month they launched 15 yearsa ago.
Well, as a moderator on Stack Exchange, I can say that I see it from the other side. We get lots of people asking a question where, if they had simply searched their own question title on Google, would see another SE/SO post as the top result that answers their question. So we close it as a duplicate, and then get flamed for being toxic. Or we correct some grammar and formatting in the post and get attacked for micromanaging.
SE/SO have a purpose that is often misunderstood. It's not a generic Q&A site, but rather a repository of knowledge. It's more similar to Wikipedia than Quora in that sense. The primary thought whenever a question is asked should be "will this be useful to others in the future and does this further the goal of building a repository of knowledge?".
An answer that boils down to "you forgot a semicolon" is useless to everyone but the original person asking the question, and so really doesn't fit well. A question about why GCC gives seemingly unrelated or nonsensical errors whenever a semicolon is missing, however, would be a good fit. We spend a lot of time curating information, and people asking questions are expected to ask their questions in such a way that they further that goal.
The community has been constantly asking management for better moderation tools so that new users can be onboarded properly rather than simply getting shut down, but management ignores that in favor of "AI EVERYWHERE! WELCOME INCLUSIVE ENVIRONMENT! ZERO-TRUST QUANTUM BLOCKCHAIN BUZZWORDS!".