(with the caveat that all we have right now are accusations that DeepSeek made use of OpenAI data - it might just as well turn out that DeepSeek really did work independently, and you really could have gotten o1-like performance with much less compute)
Human reasoning, as it exists today, is the result of tens of thousands of years of intuition slowly distilled down to efficient abstract concepts like "numbers", "zero", "angles", "cause", "effect", "energy", "true", "false", ...
I don't know what reasoning from scratch would look like without training on examples from other reasoning beings. As human children do.
In an axiomatic system, those solutions are checkable, but how discoverable are they when your search space starts from infinity? How much do you lose by disregarding the gritty reality and foam of human experience? It provides inspirational texture that helps mathematicians in the search at least.
Reality is a massive corpus of cause and effect that can be modeled mathematically. I think you're throwing the baby out with the bathwater if you even want to be able to math in a vacuum. Maybe there is a self optimization spider that can crawl up the axioms and solve all of math. I think you'll find that you can generate new math infinitely, and reality grounds it and provides the gravity to direct efforts towards things that are useful, meaningful and interesting to us.
At the end of the day, all theory must be empirically verified, and contextually useful reasoning simply cannot develop in a vacuum.
On the contrary, when reasoning about the real world, one must reason starting from assumptions that are uncertain (at best) or even "clearly wrong but still probably useful for this particular question" (at worst). Any long and logic-heavy proof would make the results highly dubious.
Unless the brain is using physics that we don’t understand or can’t replicate, it seems that, at least theoretically, there should be a way to model what it’s doing with silicon and code.
States like inspiration and creativity seem to correlate in an interesting way with ‘temperature’, ‘top p’, and other LLM inputs. By turning up the randomness and accepting a wider range of output, you get more nonsense, but you also potentially get more novel insights and connections. Human creativity seems to work in a somewhat similar way.
Emergent tool use from multi-agent interaction is a good example - https://openai.com/index/emergent-tool-use/
In your particular case the prompt would look something like: <pubmed dump> what are the plants that aren't poisonous to most people?
A general reasoner would recover language and relevant world model from pubmed dump. And then would proceed to reason about it, to perform the task.
It doesn't look like a particularly efficient process.
- humans experience reality at a slower pace than AI could theoretically experience a simulated reality
- humans have to transfer knowledge to the next generation every 80 years (in a manner that's very lossy), and around half of each human lifespan is spent learning things that the previous generation already knew
Reasoning could very well have originally been an emergent property of a group of beings.
The animal kingdom is full of examples of groups being more intelligent than individuals, including in human animals as of today.
It’s entirely possible that reasoning emerged as a property of a group before it emerged in any individual first.
What I wonder instead is whether reasoning is a property that is either there or not there, with a sharp boundary of existence.
To your point, experience is the training. Without language/data to represent human experience and knowledge to train a model, how would you give it 'experience'?
They were pre-designed to learn what they always learn. Their minds structured to readily make the same connections as puppies, that dogs have always needed to survive.
Not for real reasoning, which by its nature, does not have a limit.
Its easy to train the same things to a degree, but its amazing to watch different dogs individually learn and reason through things completely differently, even within a breed or even a litter.
Reasoning ability is always limited by the capacity of the thinker to frame the concepts and interactions. Its always limited by definition, we only push that limit farther than other species, and AGI may eventually push it past our abilities.
Do this continually through generations until you arrive at modern society.
First you must invent the universe.
That was the easy part though, figuring out how to handle all the unintended side effects it generated is still an ongoing process. Please sit and relax while we are solving the few incidentals events occurring here and there, rest assured we are putting our best effort to their resolution.
This manifold is constructed via learning a decontextualized pattern space on a given set of inputs. Given the inherent probabilistic nature of sampling, true reasoning is expressed in terms of probabilities, not axioms. It may be possible to discover axioms by locating fixed points or attractors on the manifold, but ultimately you're looking at a probabilistic manifold constructed from your input set.
But I don't think you can untie this "reasoning" from your input data. It's possible you will find "meta-reasoning", or similar structures found in any sufficiently advanced reasoning manifold, but these highly decontextualized structures might be entirely useless without proper recontextualization, necessitating that a reasoning manifold is trained on input whose patterns follow learnable underlying rules, if the manifold is to be useful for processing input of that kind.
Decontextualization is learning, decomposing aspects of an input into context-agnostic relationships. But recontextualization is the other half of that, knowing how to take highly abstract, sometimes inexpressible, context-agnostic relationships and transform them into useful analysis in novel domains.
This doesn't mean a well-trained model can't reason about input it hasn't encountered before, just that the input needs to be in some way causally connected to the same laws which governed the input the manifold was trained on.
I'm sure we could create a fully generalized reasoning manifold which could handle anything, but I don't see how we possibly get that without first considering and encountering all possible inputs. But these inputs still have to have some form of constraint governed by laws that must be learned through sampling, otherwise you'd just be training on effectively random data.
The other commenter who suggested simply generating all possible sentences and training on internal consistency should probably consider Gödel's incompleteness theorems, and that internal consistency isn't enough to accurately model and interpret the universe. One could construct a thought experiment about an isolated brain in a jar with effectively unlimited neuronal connections, but no sensory connection to the outside world. It's possible, with enough connections, that the likelihood of the brain conceiving of true events it hasn't actually encountered does increase meaningfully. But the brain still has nothing to validate against, and can't simply assume that because something is internally logically consistent, that it must exist or have existed.
To quote DeepSeek directly:
> DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning. With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors. However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates cold-start data before RL.
On the other hand, my take on it, the ability to do reasoning in a long context is a general capability. And my guess is that it can be bootstrapped from scratch, without having to do training on all of the internet or having to distill models trained on the internet.
But we already know that is the case: the Deepseek v3 paper says it was posttrained partly with an internal version of R1:
> Reasoning Data. For reasoning-related datasets, including those focused on mathematics, code competition problems, and logic puzzles, we generate the data by leveraging an internal DeepSeek-R1 model. Specifically, while the R1-generated data demonstrates strong accuracy, it suffers from issues such as overthinking, poor formatting, and excessive length. Our objective is to balance the high accuracy of R1-generated reasoning data and the clarity and conciseness of regularly formatted reasoning data.
And deepseekmath did a repeated cycle of this kind of thing mixing in 10% of old previously seen data with new generated data from last gen in a continuous bootstrap.
Whether or not you could have, you can now.
In this study, we demonstrate that reasoning capabilities can be significantly improved through large-scale reinforcement learning (RL), even without using supervised fine-tuning (SFT) as a cold start. Furthermore, performance can be further enhanced with the inclusion of a small amount of cold-start data
Is this cold start data what OpenAI is claiming their output ? If so what's the big deal ?
> To collect such data, we have explored several approaches: using few-shot prompting with a long CoT as an example, directly prompting models to generate detailed answers with reflection and verification, gathering DeepSeek-R1-Zero outputs in a readable format, and refining the results through post-processing by human annotators.
>We curate reasoning prompts and generate reasoning trajectories by performing rejection sampling from the checkpoint from the above RL training. In the previous stage, we only included data that could be evaluated using rule-based rewards. However, in this stage, we expand the dataset by incorporating additional data, some of which use a generative reward model by feeding the ground-truth and model predictions into DeepSeek-V3 for judgment.
>For non-reasoning data, such as writing, factual QA, self-cognition, and translation, we adopt the DeepSeek-V3 pipeline and reuse portions of the SFT dataset of DeepSeek-V3. For certain non-reasoning tasks, we call DeepSeek-V3 to generate a potential chain-of-thought before answering the question by prompting.
The non-reasoning portion of the DeepSeek-V3 dataset is described as:
>For non-reasoning data, such as creative writing, role-play, and simple question answering, we utilize DeepSeek-V2.5 to generate responses and enlist human annotators to verify the accuracy and correctness of the data.
I think if we were to take them at their word on all this, it would imply there is no specific OpenAI data in their pipeline (other than perhaps their pretraining corpus containing some incidental ChatGPT outputs that are posted on the web). I guess it's unclear where they got the "reasoning prompts" and corresponding answers, so you could sneak in some OpenAI data there?
The reasoning prompts and answers for SFT from V3 you mean ? No idea. For that matter you have no idea where OpenAI got this data from either. If they open this can of worms, their can of worms will be opened as well.
I assume in their API logs.
Maybe they needed OpenAI for their process. But now that their model is open source, anyone can use that as their cold start and spend the same amount.
"From scratch" is a moving target. No one who makes their model with massive data from the net is really doing anything from scratch.
They probably would need some other source of truth than an existing model, but it isn't clear how much additional data is needed.
They are using the current SOTA tools and models to build new models for cheaper.
That’s why it’s so hard to understand the true cost of training Deepseek whereas it’s a little bit easier for cutting edge models (& even then still difficult).
“Hey OpenAI, if you had to make a clone of yourself again how would you do it and for a lot cheaper?”
Nice move.
Tbh a lot of people in the world would love this outcome. They will use AI because not using it puts them at a comparative disadvantage - but would rather AI doesn't develop further or didn't develop at all (i.e. they don't value the absolute advantage/value). There's both good and bad reasons for this.
That just means that the edge you’re able to retain if you invest $1B is nonexistent. It also means there’s a huge disincentive to invest $1B if your reward instantly evaporates. That would normally be fine if the competitor is otherwise able to get to that new level without the $1B. But if it relies on your $1B to then be able to put in $100M in the first place to replicate your investment, it essentially means the market for improvements disappears OR there’s legislation written to ensure competitors aren’t allowed to do that.
This is a tragedy of the commons and we already have historical example for how humans tried to deal with it and all the problems that come with it. The cost of producing a book requires substantial capital but the cost of copying it requires a lot less. Copyright law, however flawed and imperfect, tries to protect the incentive to create in the face of that.
This is possibly true, though with billions already invested I'm not sure that OpenAI would just...stop absent legislation. And, there may be technical or other solutions beyond legislation. [0]
But, really, your comment here considers what might come next. OTOH, I was replying to your prior comment that seemed to imply that DeepSeek's achievement was of little consequence if they weren't improving on OpenAI's work. My reply was that simply approximating OpenAI's performance at much lower cost could still be extraordinarily consequential, if for no other reason than the challenges you subsequently outlined in this comment's parent.
[0] On that note, I'm not sure (and admittedly haven't yet researched) how DeepSeek just wholesale ingested ChatGPT's "output" to be used for its own model's training, so not sure what technical measures might be available to prevent this going forward.
Jeez. Must be really tough to have some comparatively small group of people financially destroy your industry with your own mechanically-harvested professional output while dubiously claiming to be be better than you when in reality it’s just a lot cheaper. Must be tough.
Maybe they should take some time to self-reflect and make some art and writing about it using the products they make that mechanically harvest the work of millions of people, and have already screwed up the commercial art and writing marketplaces pretty throughly. Maybe tell DeepSeek it’s their therapist and get some emotional support and guidance.
Let's just assume that the cost of training can be externalized to other people for free.
If other players can access that data with relatively less effort, then it's futile trying to train your models and improve upon them, as clearly you don't have an architectural moat, just a training moat.
Kind of like an office scene where an introverted hardworker does all the tedious work, while his extroverted colleague promotes it as his and gains credit.
We don't make people figure out how to domesticate a cow every time they want a hamburger. Or test hundreds of thousands of filaments before they can have a lightbulb. Inventions, once invented, exist as giants to stand upon. The inventor can either choose to disclose the invention and earn a patent for exclusive rights, or they can try to keep it a secret and hope nobody reverse engineers it.
The model already embodies the "total sum of a massive amount of compute" used to create it; if it's possible to reuse that embodied compute to create a better model, that's good for the world. Forcing everyone to redo all that compute for themselves is, conversely, bad for the world.
The big question really is, are we doing it wrong, could we have created o1 for a fraction of the price. Will o4 cost less to train than o1 did?
The second question is naturally. If we create a smarter LLM, can we use it to create another LLM that is even smarter?
It would have been fantastic if DeepSeek could have come out with an o3 competitor before o3 even became publicly available. That way we would have known for sure that we’re doing it wrong. Cause then either we could have used o1 to train a better AI or we could have just trained in a smarter and cheaper way.
It is no better for OpenAI in this scenario either, any competitor can easily copy their expensive training without spending the same, i.e. there is a second mover advantage and no economic incentive to be the first one.
To put it another way, the $500 Billion Stargate investment will be worth just $5Billion once the models become available for consumption, because it only will take that much to replicate the same outcomes with new techniques even if the cold start needed o1 output for RL.
My understanding is this effectively builds on OpenAI's very expensive initial work, provides a "nearly as good as" model for orders of magnitude cheaper to train and run, that also provides a basis to continue building on and improving without openAI, and without human bottlenecks.
That cuts OAI off at the knees in terms of market viability after billions have been spent. If DS can iterate and match the capabilities of the current in-development OAI models in the next year, it may come down to regulatory capture and government intervention to ensure its viability as a company.
the attempt in cryptography with PGP and export controls made that clear.
Even if DS specifically is banned (and even effectively), a dozen other clean room replications following their published methods will become available.
It is possible this government will ban all “unapproved” LLMs not running at authorized provider[1], saying it is weapon and AGI or skynet or whatever makes powers that sound important, thus establishing the need for control [2], the rest of the world will keep innovating.
—-
[1] Bans just need to work only economically, not at information level i.e organization with liability considerations will not use “unapproved” ones and they are ones who will bulk of the money and that what they need to protect.
[2] if they were smart they could do this positively without the backlash bans would have. By giving protections to compliant models like legal indemnity for for model companies and users without necessarily blocking others
All of this should have been clear anyway from the start, but that's the Internet for you.
I did not think this, nor did I think this was what others assumed. The narrative, I thought, was that there is little point in paying OpenAI for LLM usage when a much cheaper, similar / better version can be made and used for a fraction of the cost (whether it's on the back of existing LLM research doesn't factor in)
If the narrative is actually that DeepSeek can only reach whatever heights OpenAI has already gotten to with some new tricks, then markets will probably refocus on OpenAI's innovations and price things accordingly, even if the initial cost is huge. It also means OpenAI probably needs a better moat to protect its interests.
I'm not sure where the reality is exactly, but market reactions so far have basically followed that initial narrative and now the rebuttal.
The latter could be a one time thing, and/or OpenAi Could still use their financial might to leverage those innovations and get even better with them.
However, the former destroys their business model and no amount of intelligence and innovation from OpenAI protects them from being copied at a fraction of the cost.
How do you know this?
> If the narrative is actually that DeepSeek can only reach whatever heights OpenAI has already gotten to with some new tricks, then markets will probably refocus on OpenAI's innovations and price things accordingly
Why? If every innovation OpenAI is trying to keep as secret sauce becomes commoditized quickly and cheaply, then why would markets care about any innovations they have? They will be unable to monetize them.
Also, scraping most of the websites they scrape is also not allowed, they do it anyways
That's what I thought and assumed. This is the narrative that's been running through all the major news outlets.
It didn't even occur to me that DeepSeek could have been training their models using the output of other models until reading this article.
Hmm, I think the narrative of the rise of LLMs is that once the output of humans has been distilled by the model, the human isn't necessary.
As far as I know, DeepSeek adds only a little to the transformers model while o1/o3 added a special "reasoning component" - if DeepSeek is as good as o1/o3, even taking data from it, then it seems the reasoning component isn't needed.
Distillation is a term of art in AI and it is fundamentally incorrect to talk about distilling human-created data. Only an AI model can be distilled.
https://en.m.wikipedia.org/wiki/Knowledge_distillation#Metho...
It seems clear that the term can be used informally to denote the boiling down of human knowledge, indeed it was used that way before AI appeared in the popular imagination.
>> The idea that they used o1's outputs for their distillation further shows that models like o1 are necessary.
> Hmm, I think the narrative of the rise of LLMs is that once the output of humans has been distilled by the model, the human isn't necessary.
If deepseek was produced through the distillation (term of art) of o1, then the cost of producing deepseek is strictly higher than the cost of producing o1, and can't be avoided.
Continuing this argument, if the premise is true then deepseek can't be significantly improved without first producing a very expensive hypothetical o1-next model from which to distill better knowledge.
That is the argument that is being made. Please avoid shallow dismissals.
Edit: just to be clear, I doubt that deepseek was produced via distillation (term of art) of o1, since that would require access to o1's weights. It may have used some of o1's outputs to fine tune the model, which still would mean that the cost of training deepseek is strictly higher than training o1.
Yeah, your technical point is kind of ridiculous here that in all my uses of distillation (and in the comment I quoted), distillation is used in informal sense and there's no allegation that DeepSeek could have been in possession of OpenAI's model weights, which is what's needed for your "Distillation (term of Art)".
Seems reasonable they would be investing heavily in plaing state assets within OpenAI so they can copy the models.
Looking back on the article, it is indeed using "distillation" as a special/"term of art" but not using it correctly. IE, it's not actually speculating that DeepSeek obtained OpenAI's weights and distilled them down but rather that it used OpenAI's answers/output as a starting point (which there is a different method/"term of art").
- v2/v3 (not r1) seem to be cloned from o1/4o output, and perform worse (this cost the oft-repeated 5ish mm USD)
- r1 is specifically a reasoning step (using RL) _on top of_ v2/v3 and performs similarly to o1 (the cost of this is _not reported anywhere_)
- In the o1 blog post, they specifically say they use RL to add reasoning to LLMs: https://openai.com/index/learning-to-reason-with-llms/
But HOW they are necessary is the change. They went from building blocks to stepping stones. From a business standpoint that's very damaging to OAI and other players.
And is this related to the lottery ticket hypothesis?
I have a question (disclaimer: reinforcement learning noob here):
Is there a risk of broken telephone with this?
Kinda like repeatedly compressing an already compressed image eventually leads to a fuzzy blur.
If that is the case then I’m curious how this is monitored and / or mitigated.
That is where artificial intelligence is going. Copy things from other things. Will there be a AI Eureka moment where it deviates and knows where and why the reason it is wrong?
It seems like if they in fact distilled then what we have found is that you can create a worse copy of the model for ~5m dollars in compute by training on its outputs.
Everyone is standing on the shoulders of giants.
Better benchmark scores can be cooked
But if you leave someone in the tech industry of SV/SF long enough, they'll start to get high on their own supply and think they're entitled to insane amounts of value, so...