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.
  • hendrik@palaver.p3x.deEnglish
    1·
    12 days ago

    You just described your subjective experience of thinking.

    Well, I didn’t just do that. We have MRIs and have looked into the brain and we can see how it’s a process. We know how we learn and change by interacting with the world. None of that is subjective.

    I would say that the LLM-based agent thinks. And thinking is not only “steps of reasoning”, but also using external tools for RAG.

    Yes, that’s right. An LLM alone certainly can’t think. It doesn’t have a state of mind, it’s reset a few seconds after it did something and forgets about everything. It’s strictly tokens from left to right And it also doesn’t interact and that’d have an impact on it. That’s just limited to what we bake in in the training process by what’s on Reddit and other sources. So there are many fundamental differences here.

    The rest of it emerges by an LLM being embedded into a system. We provide tools to it, a scratchpad to write something down, we devise a pipeline of agents so it’s able to devise something and later return to it. Something to wrap it up and not just output all the countless steps before. It’s all a bit limited due to the representation and we have to cram everything into a context window, and it’s also a bit limited to concepts it was able to learn during the training process.

    However, those abilities are not in the LLM itself, but in the bigger thing we build around it. And it depends a bit on the performance of the system. As I said, the current “thinking” processes are more a mirage and I’m pretty sure I’ve read papers on how they don’t really use it to think. And that aligns with what I see once I open the “reasoning” texts. Theoretically, the approach surely makes everything possible (with the limitation of how much context we have, and how much computing power we spend. That’s all limited in practice.) But what kind of performance we actually get is an entirely different story. And we’re not anywhere close to proper cognition. We hope we’re eventually going to get there, but there’s no guarantee.

    The LLM can for sure make abstract models of reality, generalize, create analogies and then extrapolate.

    I’m fairly sure extrapolation is generally difficult with machine learning. There’s a lot of research on it and it’s just massively difficult to make machine learning models do it. Interpolation on the other hand is far easier. And I’ll agree. The entire point of LLMs and other types of machine learning is to force them to generalize and form models. That’s what makes them useful in the first place.

    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

    I completely agree with that. LLMs are our current approach. And the best approach we have. They just have a scalability problem (and a few other issues). We don’t have infinite datasets to feed in and infinite compute, and everything seems to grow exponentially more costly, so maybe we can’t make them substantially more intelligent than they are today. We also don’t teach them to stick to the truth or be creative or follow any goals. We just feed in random (curated) text and hope for the best with a bit of fine-tuning and reinforcement learning with human feedback on top. But that doesn’t rule out anything. There are other machine learning architectures with feedback-loops and way more powerful architectures. They’re just too complicated to calculate. We could teach AI about factuality and creativity and expose some control mechanisms to guide it. We could train a model with a different goal than just produce one next token so it looks like text from the dataset. That’s all possible. I just think LLMs are limited in the ways I mentioned and we need one of the hypothetical new approaches to get them anywhere close to a level a human can achieve… I mean I frequently use LLMs. And they all fail spectacularly at computer programming tasks I do in 30 minutes. And I don’t see how they’d ever be able to do it, given the level of improvement we see as of today. I think that needs a radical new approach in AI.