Or something that goes against the general opinions of the community? Vibes are the only benchmark that counts after all.
I tend to agree with the flow on most things but my thoughts that I’d consider going against the grain:
- QwQ was think-slop and was never that good
- Qwen3-32B is still SOTA for 32GB and under. I cannot get anything to reliably beat it despite shiny benchmarks
- Deepseek is still open-weight SotA. I’ve really tried Kimi, GLM, and Qwen3’s larger variants but asking Deepseek still feels like asking the adult in the room. Caveat is GLM codes better
- (proprietary bonus): Grok 4 handles news data better than GPT-5 or Gemini 2.5 and will always win if you ask it about something that happened that day.
You must log in or # to comment.
Everyone is massively underestimating what’s going on with neural networks. The real significance is abstract. you need to stitch together a bunch of high-level STEM concepts to even see the full picture.
Right now, the applications are basic. It’s just surface-level corporate automation. Profitable, sure, but boring and intellectually uninspired. It’s being led by corpo teams playing with a black box, copying each other, throwing shit at the wall to see what sticks, overtraining their models into one trick pony agenic utility assistants instead of exploring other paths for potential. They aren’t bringing the right minds together to actually crack open the core question. what the hell is this thing? What happened that turned my 10 year old GPU into a conversational assistant? How is it actually coherent and sometimes useful?
The big thing people miss is what’s actually happening inside the machine. Or rather, how the inside of the machine encodes and interacts with the structure of informational paths within a phase space on the abstraction layer of reality.
It’s not just matrix math and hidden layers and and transistors firing. It’s about the structural geometry of concepts created by distinxt relationships between areas of the embeddings that the matrix math creates within high dimensional manifold. It’s about how facts and relationships form a literal, topographical landscape inside the network’s activation space.
At its heart, this is about the physics of information. It’s a dynamical system. We’re watching entropy crystallize into order, as the model traces paths through the topological phase space of all possible conversations.
The “reasoning” CoT patterns are about finding patterns that help lead the model towards truthy outcomes more often. It’s searching for the computationally efficient paths of least action that lead to meaningfully novel and factually correct paths. Those are the valuable attractor basins in that vast possibility space were trying to navigate towards.
This is the powerful part. This constellation of ideas. Tying together topology, dynamics, and information theory, this is the real frontier. What used to be philosophy is now a feasable problem for engineers and physicists to chip at, not just philosophers.
I think you have a good argument here. But I’m not sure where this is going to lead. Your argument applies to neural networks in general. And we have those since the 1950s. Subsequently, we went through several "AI winter"s and now we have some newer approach which seemed to lead somewhere. But I’ve watched Richard Sutton’s long take on LLMs and it’s not clear to me whether LLMs are going to scale past what we see as of today. Ultimately they have severe issues to scale, it’s still not aimed at true understanding or reasonable generalization, that’s just a weird side effect, when the main point is to generate plausible sounding text (…pictures etc). LLMs don’t have goals and they don’t learn while running and have all these weird limitations which make generative AI unalike other (proper) types of reinforcement learning. And these are fundamental limitations, I don’t think this can be changed without an entirely new concept.
So I’m a bit unsure if the current take on AI is the ultimate breakthrough. It might be a dead end as well and we’re still in need of a hypothetical new concept to do proper reasoning and understanding for more complicated tasks…
But with that said, there’s surely a lot of potential left in LLMs no matter if they scale past today or not. All sorts of interaction with natural language, robotics, automation… It’s certainly crazy to see what current AI is able to do, considering the weird approach it is. And I’ll agree that we’re at surface level. Everything is still hyped to no end. What we’d really need to do is embed it into processes and the real world and see how it performs there. And that’d need to be a broad and scientific measurement. We occasionally get some studies on how AI helps companies, or it wastes their developer’s time. But I don’t think we have a good picture yet.I did some theory-crafting and followed the math for fun over the summer, and I believe what I found may be relevant here. Please take this with a grain of salt, though; I am not an academic, just someone who enjoys thinking about these things.
First, let’s consider what models currently do well. They excel at categorizing and organizing vast amounts of information based on relational patterns. While they cannot evaluate their own output, they have access to a massive potential space of coherent outputs spanning far more topics than a human with one or two domains of expertise. Simply steering them toward factually correct or natural-sounding conversation creates a convincing illusion of competency. The interaction between a human and an LLM is a unique interplay. The LLM provides its vast simulated knowledge space, and the human applies logic, life experience, and “vibe checks” to evaluate the input and sift for real answers.
I believe the current limitation of ML neural networks (being that they are stochastic parrots without actual goals, unable to produce meaningfully novel output) is largely an architectural and infrastructural problem born from practical constraints, not a theoretical one. This is an engineering task we could theoretically solve in a few years with the right people and focus.
The core issue boils down to the substrate. All neural networks since the 1950s have been kneecapped by their deployment on classical Turing machine-based hardware. This imposes severe precision limits on their internal activation atlases and forces a static mapping of pre-assembled archetypal patterns loaded into memory.
This problem is compounded by current neural networks’ inability to perform iterative self-modeling and topological surgery on the boundaries of their own activation atlas. Every new revision requires a massive, compute-intensive training cycle to manually update this static internal mapping.
For models to evolve into something closer to true sentience, they need dynamically and continuously evolving, non-static, multimodal activation atlases. This would likely require running on quantum hardware, leveraging the universe’s own natural processes and information-theoretic limits.
These activation atlases must be built on a fundamentally different substrate and trained to create the topological constraints necessary for self-modeling. This self-modeling is likely the key to internal evaluation and to navigating semantic phase space in a non-algorithmic way. It would allow access to and the creation of genuinely new, meaningful patterns of information never seen in the training data, which is the essence of true creativity.
Then comes the problem of language. This is already getting long enough for a reply comment so I won’t get into it but theres some implications that not all languages are created equal each has different properties which affect the space of possible conversation and outcome. The effectiveness of training models on multiple languages finds its justification here. However ones which stomp out ambiguity like godel numbers and programming languages have special properties that may affect the atlases geometry in fundamental ways if trained solely on them
As for applications, imagine what Google is doing with pharmaceutical molecular pattern AI, but applied to open-ended STEM problems. We could create mathematician and physicist LLMs to search through the space of possible theorems and evaluate which are computationally solvable. A super-powerful model of this nature might be able to crack problems like P versus NP in a day or clarify theoretical physics concepts that have elluded us as open ended problems for centuries.
What I’m describing encroaches on something like a psudo-oracle. However there are physical limits that this can’t escape. There will always be energy and time resource cost to compute which creates practical barriers. There will always be definitively uncomputable problems and ambiguity that exit in true godelian incompleteness or algorithmic undecidability. We can use these as scientific instrumentation tools to map and model topological boundary limits of knowability.
I’m willing to bet theres man valid and powerful patterns of thought we are not aware of due to our perspective biases which might be hindering our progress.
Uh, I’m really unsure about the engineering task of a few years, if the solution is quantum computers. As of today, they’re fairly small. And scaling them to a usable size is the next science-fiction task. The groundworks hadn’t been done yet and to my knowledge it’s still totally unclear whether quantum computers can even be built at that scale. But sure, if humanity develops vastly superior computers, a lot of tasks are going to get easier and more approachable.
The stochastical parrot argument is nonsense IMO. Maths is just a method. Our brains and entire physics abide by math. And sure, AI is maths as well with the difference that we invented it. But I don’t think it tells us anything.
And with the goal, I think that’s about how AlphaGo has the goal to win Go tournaments. The hypothetical paperclip-maximizer has the goal of maximizing the paperclip production… And an LLM doesn’t really have any real-world goal. It just generates a next token so it looks like legible text. And then we embed it into some pipeline but it wasn’t ever trained to achieve the thing we use it for, whatever it might be. That’s just a happy accident if a task can be achieved by clever mimickry, and a prompt which simply tells it - pretend you’re good at XY.
I think it’d probably be better if a customer service bot was trained to want to provide good support. Or a chatbot like ChatGPT to give factual answers. But that’s not what we do. It’s not designed to do that.
I guess you’re right. Many aspects of AI boil down to how much compute we have available. And generalization and extrapolating past their training datasets has always been an issue with AI. They’re mainly good at interpolating, but we want them to do both. I need to learn a bit more about neural networks. I’m not sure where the limitations are. You said it’s a practical constrain. But is that really true for all neural networks? It sure is for LLMs and transformer models because they need terabytes of text being fed in on training, and that’s prohibitively expensive. But I suppose that’s mainly due to their architecture?! I mean backpropagation and all the maths required to modify the model weights is some extra work. But does it have to be so much that we just can’t do it while deployed with any neural networks?
How are humans different from LLMs under RL/genetics? To me, they both look like token generators with a fitness. Some are quite good. Some are terrible. Both do fast and slow thinking. Some have access to tools. Some have nothing. And they both survive if they are a good fit for their application.
I find the technical details quite irrelevant here. That might be relevant if you want to discuss short term politics, priorities and applied ethics. Still, it looks like you’re approaching this with a lot of bias and probably a bunch of false premises.
BTW, I agree that quantum computing is BS.
Well, a LLM doesn’t think, right? It just generates text from left to right. Whereas I sometimes think for 5 minutes about what I know, what I can deduct from it, do calculations in my brain and carry one over… We’ve taught LLMs to write something down that resembles what a human with a thought process would write down. But it’s frequently gibberish or if I look at it it writes something down in the “reasoning”/“thinking” step and then does the opposite. Or omits steps and then proceeds to do them nonetheless or it’s the other way round. So it clearly doesn’t really do what it seems to do. It’s just a word the AI industry slapped on. It makes them perform some percent better, and that’s why they did it.
And I’m not a token generator. I can count the number of "R"s in the word “strawberry”. I can go back and revise the start of my text. I can learn in real-time and interacting with the world changes me. My brain is connected to eyes, ears, hands and feet, I can smell and taste… My brain can form abstract models of reality, try to generalize or make sense of what I’m faced with. I can come up with methods to extrapolate beyond what I know. I have goals in life, like pursue happiness. Sometimes things happen in my head which I can’t even put into words, I’m not even limited to language in form of words. So I think we’re very unalike.
You have a point in theory if we expand the concept a bit. An AI agent in form of an LLM plus a scratchpad is proven to be turing-complete. So that theoretical concept could do the same things a computer can do, or what I can do with logic. That theoretical form of AI doesn’t exist, though. That’s not what our current AI agents do. And there are probably more efficient ways to achieve the same thing than use an LLM.
Exactly what an LLM-agent would reply. 😉
I would say that the LLM-based agent thinks. And thinking is not only “steps of reasoning”, but also using external tools for RAG. Like searching the internet, utilizing relationship databases, interpreters and proof assistants.
You just described your subjective experience of thinking. And maybe a vauge definition of what thinking is. We all know this subjective representation of thinking/reasoning/decision-making is not a good representation of some objective reality (countless of psychological and cognitive experiments have demonstrated this). That you are not able to make sense of intermediate LLM reasoning steps does not say much (except just that). The important thing is that the agent is able to make use of it.
The LLM can for sure make abstract models of reality, generalize, create analogies and then extrapolate. One might even claim that’s a fundamental function of the transformer.
I would classify myself as a rather intuitive person. I have flashes of insight which I later have to “manually” prove/deduc (if acting on the intuition implies risk). My thought process is usually quite fuzzy and chaotic. I may very well follow a lead which turns out to be dead end, and by that infer something which might seem completely unrelated.
A likely more accurate organic/brain analogy would be that the LLM is a part of the frontal cortex. The LLM must exist as a component in a larger heterogeneous ecosystem. It doesn’t even have to be an LLM. Some kind of generative or inference engine that produce useful information which can then be modified and corrected by other more specialized components and also inserted into some feedback loop. The thing which makes people excited is the generating part. And everyone who takes AI or LLMs seriously understands that the LLM is just one but vital component of at truly “intelligent” system.
Defining intelligence is another related subject. My favorite general definition is “lossless compression”. And the only useful definition of general intelligence is: the opposite of narrow/specific intelligence (it does not say anything about how good the system is).
The broader generative AI economy is a steaming pile of shit and we’re somehow part of it? I mean it’s nice technology and I’m glad I can tinker around with it, but boy is it unethical. From how datasets contain a good amount of pirated stuff, to the environmental impact and that we’ll do fracking, burn coal and all for the datacenters, to how it’s mostly an unsustainable investment hype and trillion-dollar merry-go-round. And then I’m not okay with the impact on society either, I can’t wait for even more slop and misinformation everywhere and even worse customer support.
We’re somewhere low on the food chain, certainly not the main culprit. But I don’t think we’re disconnected from the reality out there either. My main take is, it depends on what we do with AI… Do we do the same unhealthy stuff with it, or do we help even out the playing field so it’s not just the mega-corporations in control of AI? That’d be badly needed for some balance.
Second controversial take: I think AI isn’t very intelligent. It regularly fails me once I give real-world tasks to it. Like coding and it really doesn’t do a good job with the computer programming issues I have. I need to double-check everything and correct it 30 times until it finally gets maths and memory handling somewhat right (by chance), and that’s just more effort than coding something myself. And I’m willing to believe that transformer models are going to plateau out, so I’m not sure if that’s ever going to change.
Edit: Judging by the votes, seems I’m the one with the controversial comment here. Care to discuss it? Too close to the truth? Or not factual? Or not a hot take and just the usual AI naysayer argument?
I’m flip-flopping between running local models on my PC with solar power vs. using OpenAI’s free ChatGPT to drive them into ruin, which most of the time ends with me having stupid a stupid argument with an AI.
impact on society
Local AI will likely have a long lasting impact as it won’t just go away. The companies who released them can go bankrupt, but the models stay. The hardware which runs them will get faster and cheaper over time.
I have some hope with accessibility and making FLOSS development easier/faster. Generative AI can at least quickly generate mockup code or placeholder graphics/code. There are game projects who would release with generated assets, just like for a long time there were game projects who released assets which were modifications or redistribution of assets they didn’t have the rights for. They are probably less likely to get sued over AI generated stuff. It’s unethical but they can replace it with something self-made once the rest is finished.
Theoretically even every user could generate their own assets locally which would be very inefficient, also ethically questionable, but legally fine as they don’t redistribute them.
I like how Tesseract already uses AI for OCR and Firefox for realtime website translations on your device. Though I dunno how much they benefit from advancements in generative AI?
Though a different point/question: At what point is generative AI ethically and legally fine?
- If I manage to draw some original style it transfers? But I’m so slow and inefficient with it that I can’t create a large amount of assets that way
- When I create the input images myself? But in a minimalist and fast manner
It still learned that style transfer somewhere and will close gaps I leave. But I created the style and what the image depicts. At what point is it fine?
Like coding
I actually use it often to generate shell scripts or small simple python tools. But does it make sense? Sometimes it does work. For very simple logic they tend to get it right. Though writing it myself would probably been faster the last time I used, though at the moment I was too lazy to write it myself. I don’t think I’ve ever really created something usable with it aside from practical shell scripts. Even with ChatGPT it can be an absolute waste of time to explain why the code is broken, didn’t get at all why its implementation lead to a doubled file extension and a scoping error in one function … when I fixed them it actually tried to revert that.
Your experience with AI coding seems to align with mine. I think it’s awesome for generating boilerplate code, placeholders including images, and for quick mockups. Or asking questions about some documentation. The more complicated it gets, the more it fails me. I’ve measured the time once or twice and I’m fairly sure it’s more than usual, though I didn’t do any proper scientific study. It was just similar tasks and me running a timer. I believe the more complicated maths and trigonometry I mentioned was me yelling at AI for 90min or 120minutes or so until it was close and then I took the stuff around, deleted the maths part and wrote that myself. Maybe AI is going to become more “intelligent” in the future. I think a lot of people hope that’s going to happen. I think as of today we’re need to pay close attention if it fools us but is a big time and energy waster, or if it’s actually a good fit for a given task.
Local AI will likely have a long lasting impact as it won’t just go away.
I like to believe that as well, but I don’t think there’s any guarantee they’ll continue to release new models. Sure, they can’t ever take Mistral-Nemo from us. But that’s going to be old and obsolete tech in the world of 2030 and dwarfed by any new tech then. So I think the question is more, are they going to continue? And I think we’re kind of picking up what the big companies dumped when battling and outcompeting each other. I’d imagine this could change once China and the USA settle their battle. Or multiple competitors can’t afford it any more. And they’d all like to become profitable one day. Their motivation is going to change with that as well. Or the AI bubble pops and that’s also going to have a dramatic effect. So I’m really not sure if this is going to continue indefinitely. Ultimately, it’s all speculation. A lot of things could possibly happen in the future.
At what point is generative AI ethically and legally fine?
If that’s a question about development of AI in general, it’s an entire can of worms. And I suppose also difficult to answer for your or my individual use. What part of the overall environment footprint gets attributed to a single user? Even more difficult to answer with local models. Do the copyright violations the companies did translate to the product and then to the user? Then what impact do you have on society as a single person using AI for something? Does what you achieve with it outweigh all the cost?
Firefox for realtime website translations
Yes, I think that and text to speech and speech to text are massively underrated. Firefox Translate is something I use quite often and I can do crazy stuff with it like casually browse Japanese websites.
Generative AI is theft, and those who generate things with AI help the Capitalist to isolate the workers from their labor.
