Earlier quoted context omitted.
Here you go: https://www.biorxiv.org/content/10.1101/2024.11.11.623004v1
I'm not convinced. This is using the tooling and paradigms invented by humans.
Ah yes. I believe this is commonly called "moving the goalposts".
S1: A $6 R1 competitor?
241–250 of 430 posts
Re: S1: A $6 R1 competitor?
#242Earlier quoted context omitted.
You made a joke about questioning reality, I simply entertained it. You can do whatever you want with it, wasn’t a slight at all.
It may have been in the form of a joke, but I certainly wasn't joking. I think it is interesting what actions cannot be done by humans.
I wasn’t joking either. Things are just getting started with this AI stuff, and I feel like programmers will experience that “de ja vu” phenomenon that they talk about in the Matrix, that eerie feeling something isn’t right.
Look, why have game developers spent so much time lazy loading parts of the game world? Very rarely do they just load the whole world, even in 2025. See, the worlds get bigger, so even as the tech gets better, we will always lazy load worlds in.
It’s a context issue right? Developers have just recently been given this thing called “context”.
But yeah man, why do we think just because we walked from our house to the supermarket that this reality didn’t lazy load things. That’s how programmers have been doing it all along …
Anyways
Re: S1: A $6 R1 competitor?
#243This thing that people are calling “reasoning” is more like rendering to me really, or multi pass rendering. We’re just refining the render, there’s no reasoning involved.
How are you defining "reasoning"? Because I see these sorts of gnostic assertion about LLMs all the time about how they "definitely aren't doing " by gesturing at the technical things it's doing, with no attempts to actually justify the negative assertion. It often comes across as privileged reason trying to justify that of course the machine isn't doing some ineffable thing only meat-brains do.
From my other ridiculous comment, as I do entertain simulation theory in my understanding of God:
Reasoning as we know it could just be a mechanism to fill in gaps in obviously sparse data (we absolutely do not have all the data to render reality accurately, you are seeing an illusion). Go reason about it all you want.
The LLM doesn’t know anything. We determine what output is right, even if the LLM swears the output is right. We “reason” about it, I guess? Well in this case the whole “reasoning” process is to simply get an output that looks right, so what is reasoning in our case?
Let me just go one ridiculous level lower. If I measure every frame the Hubble telescope takes, and I measure with a simple ruler the distances between things, frame by frame, I can “reason” out some rules of the universe (planetary orbits). In this “reasoning” process, the very basic question of “well why, and who made this” immediately arises, so reasoning always leads to the fundamental question of God.
So, yeah. We reason to see God, because that’s all we’re seeing, everything else is an illusion. Reasoning is inextricably linked to God, so we have to be very open minded when we ask what is this machine doing.
Re: S1: A $6 R1 competitor?
#244Earlier quoted context omitted.
The hard part is coming up with a good way to grade results. Which you need to update the weights based on the outcome, otherwise the model will not actually learn anything.
For the "looping" I'm talking about you don't need to update the weights. It is simply, old context in, new context out, new context in, new-new context out, etc. Of course, keeping that coherent over numerous loops isn't going to be easy. No doubt there is a chance it goes off the rails. So you might have a section of context that is constantly stable, a section of context that updates each loop, etc. In the other r…
Then I'm not quite sure what benefit you expect to derive from it? Making e.g. QwQ-32 loop isn't hard - it often does it all by itself, even. But it doesn't translate to improvements on every iteration; it just goes in circles.
Re: S1: A $6 R1 competitor?
#245https://imgflip.com/i/9j833q
(ptheven)
Re: S1: A $6 R1 competitor?
#246If chain of thought acts as a scratch buffer by providing the model more temporary "layers" to process the text, I wonder if making this buffer a separate context with its own separate FNN and attention would make sense; in essence, there's a macroprocess of "reasoning" that takes unbounded time to complete, and then there's a microprocess of describing this incomprehensible stream of embedding vectors in natural lan…
The problem is that RL is extremely inefficient. It's one thing to use it for fine tuning an LLM to do the chain of thought trick and quite another to do thinking entirely from scratch. The pretrained LLM does a lot of heavy lifting there. And it would have to be RL for your idea to work since there is no "thinking" dataset for a novel token space. There isn't even one for existing LLM token space, but they have the…
> The problem is that RL is extremely inefficient.
Wait What? That is an odd way of defining it. That's like saying turing machines are inefficient way to solve TSP. You would , at the least, want to define this in terms of complexity or put this into context of domains and observability.
RL's by definition is a field that is about finding efficient problems in the domain of choice[1]. There are likely regimes in LLM/LRM learning where RL can be quite efficient, polynomial time even in the state space, we just need to explore and find them. For example you can use Dynamic Programming as a "more" efficient way to solve MDPs[1] because it is polynomial in the state space X Action space.
[1]https://web.stanford.edu/class/psych209/Readings/SuttonBarto...
Re: S1: A $6 R1 competitor?
#247Earlier quoted context omitted.
The hard part is coming up with a good way to grade results. Which you need to update the weights based on the outcome, otherwise the model will not actually learn anything.
For the "looping" I'm talking about you don't need to update the weights. It is simply, old context in, new context out, new context in, new-new context out, etc. Of course, keeping that coherent over numerous loops isn't going to be easy. No doubt there is a chance it goes off the rails. So you might have a section of context that is constantly stable, a section of context that updates each loop, etc. In the other r…
When you look out your eyes, that rectangular viewport is all your context. For example, we cannot fit Jupiter into your viewport.
So, if we can never fit large concepts like Jupiter into your viewport (your context window), does it follow that you will never understand Jupiter? Or is there no way to take a picture of Jupiter and make it smaller so it fits into your viewport?
See, coherence is just resizing or reimagining things so it fits into context. Context can never hold it all because we have the capacity to always imagine infinitely large things.
So I don’t fully know if it’s old context out, new context in, but could be just the right context, just enough of it, and just the right looking context so we can assemble the right worldview. This process would have to go on infinitely, and that is the world as you see and understand it.
Re: S1: A $6 R1 competitor?
#248In case you’re not sure what S1 is, here is the original paper: https://arxiv.org/html/2501.19393v1
To enforce a minimum, we suppress the generation of the end-of-thinking token delimiter and optionally append the string “Wait” to the model’s current reasoning trace to encourage the model to reflect on its current generation.
EDIT: Never mind, end of thinking is represented with followed by the word 'answer', so the code dynamically adds/removes from the list of stop tokens.
Re: S1: A $6 R1 competitor?
#249I have a bunch of questions, would love for anyone to explain these basics: * The $5M DeepSeek-R1 (and now this cheap $6 R1) are both based on very expensive oracles (if we believe DeepSeek-R1 queried OpenAI's model). If these are improvements on existing models, why is this being reported as decimating training costs? Isn't fine-tuning already a cheap way to optimize? (maybe not as effective, but still) * The R1 pap…
If what you say is true, and distilling LLMs is easy and cheap, and pushing the SOTA without a better model to rely on is dang hard and expensive, then that means the economics of LLM development might not be attractive to investors - spending billions to have your competitors come out with products that are 99% as good, and cost them pennies to train, does not sound like a good business strategy.
What I still don’t understand is how one slurps out an entire model (closed source) though.
Does the deepseek paper actually say what model it’s trained off of, or do they claim the entire thing is from scratch?
Re: S1: A $6 R1 competitor?
#250LLMs still feel so magical. It’s like quantum physics. “I get it” but I don’t. Not really. I don’t think I ever will. Perhaps a human mind can only comprehend so much.