S1: A $6 R1 competitor?

If we had, there would be no reason to train a model with more parameters than are strictly necessary to represent the space's semantic structure. But then it should be impossible for distilled models with less parameters to come close to the performance of the original model.

Yet this is what happens - the distilled or quantized models often come very close to the original model.

So I think there are still many low-hanging fruits to pick.

What I am more surprised about is why models actually seem to have to produce "internal thoughts" instead of random tokens. Maybe during training having completely random tokens in thinking section derailed the model's thought process in a same way background noise can derail ours?

It should be considered a distinct field. At some level there is overlap (information theory, Kolmogorov complexity, etc.), but prompt optimization and model distillation is far removed from computability, formal language theory, etc. The analytical methods, the techniques to create new architectures, etc. are very different beasts.

If I don’t need a huge model and GPU, then AI is little more than an open source program running on an idle PC.

I feel like AI was NVIDIA’s lifeboat as GPU mining waned. Don’t see anything after that in the near future.

The above 'hack' is one of many realizations of the above differences.

(1) The abstract changed how I thought about this domain (advanced reasoning models). The only other paper that did that for me was the "Memory Resource Management in VMware ESX Server". And that paper got published 23 years ago.

(2) The model, data, and code are open source at https://github.com/simplescaling/s1. With this, you can start training your own advanced reasoning models. All you need is a thousand well-curated questions with reasoning steps.

(3) More than half the references in the paper are from 2024 and Jan 2025. Just look at the paper's first page. https://arxiv.org/pdf/2501.19393 In which other field do you see this?

Not sure what caused the above but In my opinion not only is the training affected by the date of training data (ie it refuses to answer properly because every year of the training data there was fewer or lower quality examples at the end of the year), or whether it's a cultural impression of humans talking about going on holiday/having a break etc in the training data at certain times and the model associating this with the meaning of "having a break".

I still wonder if we're building models wrong by training them on a huge amount of data from the Internet, then fine tuning for instruct where the model learns to make certain logical associations inherent or similar to the training data (which seems to introduce a myriad of issues like the strawberry problem or is x less than y being incorrect).

I feel like these models would have a lot more success if we trained a model to learn logic/problem solving separately without the core data set or to restrict the instruct fine tuning in some way so that we reduce the amount of "culture" it gleans from the data.

There's so much that we don't know about this stuff yet and it's so interesting to see something new in this field every day. All because of a wee paper on attention.

Though it's poetic justice that OpenAI is complaining about someone else playing fast and loose with copyright rules.

I think if OpenAI (or any other company) are paid for their compute time/access as anybody would, then using content generated by other models is fair game. Because it's an active/ongoing cost and not a passive one.

Whereas if someone trained on my dumb Tweets or HN posts then so be it; it's a passive cost for me - I paid my time to say x thing for my own benefits (tribal monk-e social interaction) therefore I have already gotten the value out of it.

I’d go so far as to add one more layer to monitor this one and stop adding layers. My thinking is that this meta awareness is all you need.

No data to back my hypothesis up. So take it for what it’s worth.

this is just standard decoding, the stream of vectors is called the k/v cache

It's also so not far from Meta's large concept model idea.

Seems like similar concepts. I think there is some potential to improving how LLMs improve and further their own reasoning lines, but I'm no AI mage.

I reflected on the pop-psychology idea of consciousness and subconsciousness. I thought of each as an independent stream of tokens, like stream of consciousness poetry. But along the stream there were joining points between these two streams, points where the conscious stream was edited by the subconscious stream. You could think of the subconscious stream as performing CRUD like operations on the conscious stream. The conscious stream would act like a buffer of short-term memory while the subconscious stream would act like a buffer of long-term memory. Like, the subconscious has instructions related to long-term goals and the conscious stream has instructions related to short-term goals.

You can imagine perception as input being fed into the conscious stream and then edited by the subconscious stream before execution.

It seems entirely possible to actually implement this idea in this current day and age. I mean, it was a fever dream as a kid, but now it could be an experiment!

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 base model to work off of. When the thought is expressed in English, the model already knows the relationships between the tokens in the thought, it's merely repurposing it for a "thinking" application.

We need to cluster the AI's insights on a spatial grid hash, give it a minimap with the ability to zoom in and out, and give it the agency to try and find its way to an answer and build up confidence and tests for that answer.

coarse -> fine, refine, test, loop.

Maybe a parallel model that handles the visualization stuff. I imagine its training would look more like computer vision. Mind palace generation.

If you're stuck or your confidence is low, wander the palace and see what questions bubble up.

Bringing my current context back through the web is how I think deeply about things. The context has the authority to reorder the web if it's "epiphany grade".

I wonder if the final epiphany at the end of what we're creating is closer to "compassion for self and others" or "eat everything."

My literal first thought about AI was wondering why we couldn't just put it in a loop. Heck, one update per day, or one update per hour would even be a start. You have a running "context", the output is the next context (or a set of transformations on a context that is a bit larger than the output window). Then ramp that up ... one loop per minute, one per second, millisecond, microsecond.

That’s pottery, not sculpture. Traditionally in sculpture you start from a block of marble or wood, but you can also make sculptures of cast bronze or welded steel (or clay, but you don’t use a spinning plate).

Remember that the Lisp AI-labs people were working on non-solved problems on absolute potatoes of computers back in the day, we have a semblance of progress solution but so much of it has been brute-force (even if there has been improvements in the field).

The big question is if these insane spendings has pulled the rug on real progress if we head into another AI winter of disillusionment or if there is enough real progress just around the corner to show that there is hope for investors in a post-deepseek valuation hangover.

And thus, by this logic, every company in the world should just be buying as many servers as they can get their hands on, because More Servers = More Value.

Obviously, this is not happening. It doesn't take much analysis to start listing the many and manifold reasons why. Many of those reasons will apply to GPUs as well. Just as if everything in AWS got 10x faster, overnight, this would not create a situation where everyone suddenly starts grabbing more servers in AWS. Obviously everyone would start trimming down, even if perhaps in a few years time they'd find some way to use this burst of power such that they can use more later. This can't happen overnight, though. It would take time, and not "weeks" or "months" but "years" at scale.

Incorporating the important variable of time in the analysis, if AIs become literally hundreds of times cheaper to run, today, then it is perfectly logical that the near-term demand for the hardware to run them is also going to go way, way down. However much potential AI may have, it is fairly clear looking out at the AI landscape right now that there isn't really anyone out there unlocking vast amounts of value and sitting there wringing their hands because they just can't get more GPU compute. The GPU rush has been from fear that someone will figure out how to "really" unlock AI and then they'll be stuck without the hardware to compete.

It may be the case that vastly cheaper AI will in fact be part of unlocking that value, and that as the AI industry grows it will grow faster as a result... but that's still going to be on a multi-year time frame, not a tomorrow time frame. And all those GPUs and all those valuations are still broadly based on them being valuable real soon now, not in a few years, and all those GPU purchases are on the assumption they need them now, or on a timeframe where we can't be waiting around, rather than waiting for some rounds of exponential doublings to bring price down. The hardware curve in 5 years may be higher but the curve in the next year would be lower, and by a lot.

And, you know, who's to say we're done? I doubt there's another 100x in there, but is someone going to eke out another 2x improvement? Or a 10x improvement? Making it easier to run lots of experiments makes it much more likely for that to happen. I'm skeptical of another 10x general improvement but 10x improvements for specific, important use cases I can't rule out.

Edit: I should also point out this is an extremely common pattern in technology in general. Often the very hardest part is producing a thing that does a particular task at all. Once we have it in hand, once we can use it and learn how it operates and what its characteristic operating modes are, once we can try modifications to it in the real world and see what happens, optimizing it becomes much easier, sometimes explosively so by comparison. Taking any first iteration of a tech that is practical and then trying to straight-line demand based on it is silly, in all sorts of ways and all directions. The internal combustion engine, for example, has had a myriad of impacts on the world and certainly after various improvements many, many millions if not billions of them have been made... but any company that reacted to the first couple of cars and just went ballistic buying those first-generation internal combustion engines would have lost everything, and rather quickly.

  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 <|im_start|> followed by the word 'answer', so the code dynamically adds/removes <|im_start|> from the list of stop tokens.

So how to externally prod it to think more when it does terminate? Replacing thought termination with a linguistic signifier of continued reasoning plus novel realization seems like a charmingly simple, principled, and general approach to continue to traverse concept space.

The fact that we can compress a GPT-3 sized model into an o1 competitor is only the beginning. Maybe there is even more juice to squeeze there?

But even more, how much performance will we get out of o3 sized models? That is what is exciting since they are already performing near Phd levels on most evals.

As reasoning improves the models could start with a basic set of principles and build from there. Of course for facts grounded in reality RAG would still likely be the best, but maybe with enough “reasoning” a model could simulate an approximation of the universe well enough to get to an answer.

Before LLMs we had N-Gram language models. Many tasks like speech recognition worked as beach search in the graph defined by the ngram language model. You could easily get huge accuracy gains simply by pruning your beam less.

s1 reminds of this. You can always trade off latency for accuracy. Given these LLMs are much more complex than good old N-Grams, we're just discovering how to do this trade.

Because I see these sorts of gnostic assertion about LLMs all the time about how they "definitely aren't doing <thing we normally apply to meat-brains>" 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.

The models' latent spaces are insanely large. The vast, vast majority pretty much has to be irrelevant and useless, it's just that the training commandeers random fragments of that space to link up the logic they need and it's really hard to know which of the weights are useless, which are useful but interchangeable with other weights, and which are truly load-bearing. You could probably find out easily by testing the model against every possible thing you ever might want it to do, just as soon as someone gets around to enumerating that non-enumerable collection of tasks.

These bogus <wait> tokens kind of demonstrate that the models are sort of desperate to escape the limitations imposed by the limited processing they're allowed to do -- they'll take advantage of thinking time even when it's provided in the silliest manner possible. It's amazing what you can live if it's all you have!

(Apologies for the extended anthropomorphizing.)

Because that's what gets the clicks...

Saying they spent a boatload of money on the initial training + iteration + final fine-tuning isn't as headline grabbing as "$5 million trained AI beats the pants off the 'mericans".

That sounds very useful, albeit a bit different than how current "chat" implementations would work, as in you could control both ways of the conversation.

Don’t be, that’s pretty common.

https://en.wikipedia.org/wiki/Multiple_discovery

It sounds like you need to play with some of the existing 32B models with better documentation on how to run them if you're having trouble, but it is entirely plausible to run this on a laptop.

I can run Qwen2.5-Instruct-32B-q4_K_M at 22 tokens per second on just an RTX 3090.

And I think the second point is due to The Market thinking there is no need to spend ever increasing amounts of compute to get to the next level of AI overlordship.

Of course Jevon's paradox is also all in the news these days..

What about AI enabled drones and guided missiles/rockets? The case for their effectiveness is relatively simple in terms of jamming resistance.

Indeed. People are welcome to go "all in" on whatever nonsense gambling they want to do with their personal investments, but national security demands actually thinking about things - adversarially. Because the enemy will as well.

It's perfectly possible to lose a war by investing in expensive superweapons that under deliver. The Nazis were particularly bad at this.

If we're not talking about cyber war exclusively, such as finding and exploiting vulnerabilities, for the time being national security will still be based on traditional army.

Just a few weeks ago, italy announced a 16bln€ plan to buy >1000 rheinmetall ifv vehicles. That alone would make italy's army one of the most equipped in Europe. I can't imagine what would happen with a 500$bln investment in defense,lol. I don't agree with what Meloni's government is doing, but one of the ministers I agree more with is the defense minister Crosetto

Furthermore, what is being shown, at least for the time being, is that open source can be and is crucial in aiding developing better models. This collides with the idea of big, single "one winner takes it all" VC mentality (because let's be honest, these defense pitches are still made by startup/VC bros)

Yeah, I was thinking about this while trying to figure out author affiliations.

There was a Stanford paper a few years ago that dusted off some old intelligence concepts and the authors seemed excited about it.

But given the pace of AI, it's difficult to look in new directions. It will probably take an AI winter and some unbridled enthusiasm immune to burnout to make some real progress outside of feed forward neural networks.

Is (human-based) general intelligence not one of the fundamental enabling elements of literally every human activity throughout history, regardless of how many layers of automation and technology one has to peel back to get to it?

Can you maybe imagine how the ability to create arbitrary amounts of general intelligence, completely divorced from the normal lengthy biological process, could upend that foundation of human activity?

> They have the vision of the ‘AI’ they’ve seen in movies in mind, see the current ‘AI’ being used and to them it’s basically almost the same, their brain is mental bridging the concepts and saying it’s only a matter of time.

I've found that most AI-related movies exclusively focus on "quality ASI" scenarios, which are mostly irrelevant to our current state of the world, as an immense amount of danger/value/disruption will arrive with AGI. People who are seriously reasoning about the impacts of AGI are not using movies as references. "Those stupid movie watching idiots" is just a crutch you are using to avoid thinking about something that you disagree with.

> Real AI, artificial intelligence, is a fever dream. This is machine learning except the machines are bigger than ever before. There is no intellect.

Do you have any evidence to support this conclusion? And does it even matter? If "fake intellect" can replace a human, that human still has to deal with the very real issue or not having a job anymore. If "fake intellect" is used to conduct mass surveillance, and direct suppression activities towards divergent individuals, those individuals are still going to have a bad time.

The hard part now is relatively trivial. Does anyone think that there is a fundamental and profound discovery that evolution made purely by selection in the last 200,000 years? I mean a true qualitative difference?

Sure—-We call it language, which is just another part of a fancy animal’s tool kit.

Does anyone think there is an amazing qualitative difference between the brain of a chimp and the brain of a human?

No, not if they know any biology.

(Although that does not stop some scientist from looking for a “language gene” like FOXP2.)

So what did dumb mutations and 200,000 years of selection do that a group of dedicated AI scientists cannot do with their own genuine general intelligence?

Nothing—-nothing other than putting a compact energy efficient LLM with reinforcement learning on a good robotic body and letting it explore and learn like we did as infants, toddlers and teenagers.

Each one of us has experienced becoming a “general intelligence”. I remember it hit me on the head in 6th grade when I dreamed up a different way of doing long division. I remember thinking: “How did I think that?” And each one of us who has watched an infant turn into a toddler has watched it as an observer or teacher. This is what makes babies so fascinating to “play” with.

We have to give our baby AGI a private memory and a layer of meta-attention like we all gain as we mature, love, and struggle.

I read the linked article and as a neuroscientist I realized the “wait” cycles that improved performance so much is roughly equivalent to the prefrontal cortex: the part of the CNS most responsible for enabling us to check our own reasoning recursively. Delay—as in delayed gratification—-is a key attribute of intelligent systems.

We are finally on the door step to Hofstadter’s Strange Loop and Maturana’s and Valera’s “enactive” systems, but now implemented in silicon, metal, and plastic by us rather than dumb but very patient natural selection.

Karl Friston and Demis Hassabis (two very smart neuroscientist) figured this out years ago. And they were preceded by three other world class neuroscientist: Humberto Maturana, Francisco Valera, and Rich Sutton (honorary neuroscientist). And big credit to Terry Winograd for presaging this path forward long ago too.

This is not "just" machine learning because we have never been able to do things which we are today and this is not only the result of better hardware. Better hardware is actually a byproduct. Why build a PFLOPS GPU when there is nothing that can utilize it?

If you spare yourself some time and read through the actual (scientific) papers of multiple generations of LLM models, the first one being from Google ~~not DeepMind~~ in 2017, you might get to understand that this is no fluff.

And I'm speaking this from a position of a software engineer, without bias.

The reason why all this really took off with so much hi-speed is because of the not quite expected results - early LLM experiments have shown that "knowledge" with current transformers architecture can linearly scale with regards to the amount of compute and training time etc. That was very unexpected and to this day scientists do not have an answer why this even works.

So, after reading bunch of material I am inclined to think that this is something different. The future of loading the codebase into the model and asking the model to explain me the code or fix bugs has never been so close and realistic. For the better or worse.

  > Real AI, artificial intelligence, is a fever dream. This is machine learning except the machines are bigger than ever before. There is no intellect.

Even if the machine learning isn't really intelligent, it is still capable of performing IF..THEN..ELSE operations, which could have detrimental effects for [some subset of] humans.

And even if you argue that such a machine _shouldn't_ be used for whatever doomsday scenario would harm us, rest assured that someone, somewhere, who either does not understand what the machines are designed to do or just pretends that they work like magic, will put the machines in a position to make such a decision.

This is ai and trend lines point to an intelligence that matches or barely exceeds human intellect in the future.

You’re part of a trend of people in denial. When LLMs first came out there were hordes of people on HN claiming it was just a stochastic parrot and LLMs displayed zero intellectual ability. It is now abundantly clear that this not true.

We don’t fully understand LLMs. That’s why gains like COT are just black box adjustments that come from changing external configurations. We have no way to read the contents of the black box and make adjustments off of it. Yet idiots like you can make such vast and hard claims when nobody really fully understands these things. You’re delusional.

I agree that LLMs won’t allow us to make some super weapon to give us some edge in national security.

Define "intellect".