I think there's two different things going on here:
"DeepSeek trained on our outputs and that's not fair because those outputs are ours, and you shouldn't take other peoples' data!" This is obviously extremely silly, because that's exactly how OpenAI got all of its training data in the first place - by scraping other peoples' data off the internet.
"DeepSeek trained on our outputs, and so their claims of replicating o1-level performance from scratch are not really true" This is at least plausibly a valid claim. The DeepSeek R1 paper shows that distillation is really powerful (e.g. they show Llama models get a huge boost by finetuning on R1 outputs), and if it were the case that DeepSeek were using a bunch of o1 outputs to train their model, that would legitimately cast doubt on the narrative of training efficiency. But that's a separate question from whether it's somehow unethical to use OpenAI's data the same way OpenAI uses everyone else's data.
Why would it cast any doubt? If you can use o1 output to build a better R1. Then use R1 output to build a better X1... then a better X2.. XN, that just shows a method to create better systems for a fraction of the cost from where we stand. If it was that obvious OpenAI should have themselves done. But the disruptors did it. It hindsight it might sound obvious, but that is true for all innovations. It is all good stuff.
I think it would cast doubt on the narrative "you could have trained o1 with much less compute, and r1 is proof of that", if it turned out that in order to train r1 in the first place, you had to have access to bunch of outputs from o1. In other words, you had to do the really expensive o1 training in the first place.
(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)
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 ?
DeepSeek claims that the cold-start data is from DeepSeekV3, which is the model that has the $5.5M pricetag. If that data were actually the output of o1 (a model that had a much higher training cost, and its own RL post-training), that would significantly change the narrative of R1's development, and what's possible to build from scratch on a comparable training budget.
In the paper DeepSeek just says they have ~800k responses that they used for the cold start data on R1, and are very vague about how they got it:
> 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.
My surface-level reading of these two sections is that the 800k samples come from R1-Zero (i.e. "the above RL training") and V3:
>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?
That's what I am gathering as well. Where is OpenAI going to have substantial proof to claim that their outputs were used ?
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.
It's like the claim "they showed anyone create a powerful from scratch" becomes "false yet true".
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.
Yeah, but that kills the implied hope of building a better model for cheaper. Like this you'll always have a ceiling of being a bit worse then the openai models.
The logic doesn't exactly hold, it is like saying that a student is limited by their teachers. It is certainly possible that a bad teacher will hold the student back, but ultimately a student can lag or improve on the teacher without only a little extra stimulus.
They probably would need some other source of truth than an existing model, but it isn't clear how much additional data is needed.
Don't forget that this model probably has far less params than o1 or even 4o. This is a compression/distillation, which means it frees up so much compute resources to build models much powerful than o1. At least this allows further scaling compute-wise (if not in the amount of, non-synthetic, source material available for training).
If R1 were better than O1, yes you would be right. But the reporting I’ve seen is that it’s almost as good. Being able to copy cutting edge models won’t advance the state of the art in terms of intelligence. They have made improvements in other area, but if they reused O1 to train their model, that would be effectively a ctrl-c / ctrl-v strictly in terms of task performance.
It's not just about whether competitors can improve on OpenAI's models. It's about whether they can continually create reasonable substitutes for orders of magnitude less investment.
> It's about whether they can continually create reasonable substitutes for orders of magnitude less investment
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.
> That just means that the edge you’re able to retain if you invest $1B is nonexistent.
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.
This. There is something doubly evil about OpenAI harvesting all of that work for its own economic benefit, while also destroying the opportunity for those it stole from to continue to ply their craft.
And then all of their stans taking on a persecution complex because people that actually made all of the “data” don’t uncritically accept their work as equivalent adds insult to injury.
>it essentially means the market for improvements disappears OR there’s legislation...
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.
First, there could have been industrial espionage involved so who knows. Ignoring that, you’re missing what I’m saying. Think of it this way - if it requires O1’s input to reach almost the same task performance, then this approach gives you a cheap way to replicate the performance of a leading edge model at a fraction of the cost. It does not give you a way to train something that beats a cutting edge model. Cutting edge models require a lot of R&D & capital expenditure - if they’re just going to be trivially copied after public availability, the response is going to be legislation to keep the incentive there to keep meaningful investments in that area. Otherwise you’re going to have another AI winter where progress shrivels because investment dollars dry up.
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).
"Otherwise you’re going to have another AI winter where progress shrivels because investment dollars dry up."
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.
When you build a new model, there is a spectrum of how you use the old model: 1. taking the weights, 2. training on the logits, 3. training on model output, 4. training from scratch. We don't know how much advantage #3 gives. It might be the case that with enough output from the old model, it is almost as useful as taking the weights.
I lean on the idea that R1-Zero was trained from cold start, at the same time, they have tried many things including using OpenAI APIs. These things can happen in parallel.
> you had to do the really expensive o1 training in the first place
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.
Now that its been done, is OpenAI needed or can you iterate on DeepSeek only moving forward?
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.
You cannot really have government intervention against open source and weights successfully.
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
o1 wouldn't exist without the combined compute of every mind that led to the training data they used in the first place. How many h100 equivalents are the rolling continuum of all of human history?
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.
Actually i also think it's possible. Start with natural numbers axiom system. Form all valid sentences of increasing length. RL on a model to search for counter example or proofs. This on sufficient computer should produce superhuman math performance (efficiency) even at compute parity
I wonder how much discovery in math happens as a result in lateral thinking epiphanies. IE: A mathematician is trying to solve a problem, their mind is open to inspiration, and something in nature, or their childhood or a book synthesizes with their mental model and gives them the next node in their mental graph that leads to a solution and advancement.
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.
As I mentioned in a sister comment, Gödel's incompleteness theorems also throw a wrench into things, because you will be able to construct logically consistent "truths" that may not actually exist in reality. At which point, your model of reality becomes decreasingly useful.
At the end of the day, all theory must be empirically verified, and contextually useful reasoning simply cannot develop in a vacuum.
Those theorems are only relevant if "reasoning" is taken to its logical extreme (no pun intended). If reasoning is developed/trained/evolved purely in order to be useful and not pushed beyond practical applications, the question of "what might happen with arbitrarily long proofs" doesn't even come up.
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.
A question is: what algorithms does the brain use to make these creative lateral leaps? Are they replicable?
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.
Dogs are probably the best example I can think of. They learn through experience and clearly reason, but without a complex language to define abstract concepts. Its very basic reasoning, but they do learn and apply that learning.
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'?
And yet dogs, to a very high degree, just learn the same things. At least the same kinds of things, over and over.
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.
> just learn the same things. At least the same kinds of things, over and over.
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.
There was necessarily a "first reasoning being" who learned reasoning from scratch, and then it's improved from there. Humans needed tens of thousands of years because:
- 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
Whether the first reasoning entity is an individual organism or a group of organisms is completely irrelevant to the original point. If one were to grant that there was in fact a "first reasoning group" rather than a "first reasoning being" the original argument would remain intact.
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.
It is possible to learn to reason from scratch, that's what R1-0 did, but the resulting chains of thought aren't legible to humans.
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.
If you look at the benchmarks of the DeepSeek-V3-Base, it is quite capable, even in 0-shot: https://huggingface.co/deepseek-ai/DeepSeek-V3-Base#base-mod... This is not from scratch. These benchmark numbers are an indication that the base model already had a large number of reasoning/LLM tokens in the pre-training set.
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.
> These benchmark numbers are an indication that the base model already had a large number of reasoning/LLM tokens in the pre-training set.
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.
Possible? I guess evolution did it over the course of a few billion years. For engineering purposes, starting from the best advanced position seems far more efficient.
I've been giving this a lot of thought over the last few months. My personal insight is that "reasoning" is simply the application of a probabilistic reasoning manifold on an input in order to transform it into constrained output that serves the stability or evolution of a system.
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.
If OpenAi had to account for the cost of producing all the copyrighted material they trained their LLM on, their system would be worth negative trillions of dollars.
Let's just assume that the cost of training can be externalized to other people for free.
Even if what OpenAI asserts in the title of this post is true, then their system is worth negative trillions of dollars.
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.
At the pace that DeepSeek is developing we should expect them to surpass OpenAI in not that long.
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.
The whole discussion is about whether or not the second case of using o1 outputs to fine tune R1 is what allowed R1 to become so good. If that's the case then your assertion that DeepSeek will surpass OpenAI doesn't really make sense because they're dependent on a frontier model in order to match, not surpass.
Yeah, that's my point. If they do end up surpassing OpenAI then it would seem likely that they aren't just relying on copying from o1, or whatever model is the frontier model at that time.
Sure. This is fine. Data is still a product, no matter how much businesses would like to turn it into a service.
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.
I mean, yes that's how progress works. Has OpenAI got a patent? If not it's fair game.
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.
I think the prevailing narrative ATM is that DeepSeek's own innovation was done in isolation and they surpassed OpenAI. Even though in the paper they give a lot of credit to Llama for their techniques. The idea that they used o1's outputs for their distillation further shows that models like o1 are necessary.
All of this should have been clear anyway from the start, but that's the Internet for you.
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.
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.
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.
In the context in which you said it, it matters a lot.
>> 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.
just to be clear, I doubt that deepseek was produced via distillation
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)".
Because it feeds conspiracy theories and because there's no evidence for it? Also, let's talk DeepSeek in particular, not "China".
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").
The R1-Zero paper shows how many training steps the RL took, and it's not many. The cost of the RL is likely a small fraction of the cost of the foundational model.
> the prevailing narrative ATM is that DeepSeek's own innovation was done in isolation and they surpassed OpenAI
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)
Yes, well the narrative that rocked the stock market is different. Its looking at what DeepSeek did and assuming they may have competitive advantage in this space and could outperform OpenAI at their own game.
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 idea that someone can easily replicate an OpenAI model based simply on OpenAI outputs is, I’d argue, immeasurably worse for OpenAI’s valuation than the idea that someone happened to come up with a few innovations that leapfrogged OpenAI.
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.
> Yes, well the narrative that rocked the stock market is different.
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.
Why would it matter when Chinese deepseek is not going to abide by such rules or be forced to and will release their model open weights so anyone anywhere can host it?
Also, scraping most of the websites they scrape is also not allowed, they do it anyways
There were different narratives for different people. When I heard about r1, my first response was to dig into their paper and it's references to figure out how they did it.
Fwiw I assumed they were using o1 to train. But it doesn’t matter: the big story here is that massive compute resources are unlikely to be as important in the future as we thought. It cuts the legs off stargate etc just as it’s announced. The CCP must be highly entertained by the timeline.
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.
OpenAI couldn't do it, when the high cost of training and access to GPUs is their competitive advance against startups, they can't admit that it does not exist.
When will over training happen on the melange of models at scale?
And will AGI only ever be an extension of this concept?
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?
If they're training R1 on o1 output on the benchmarks - then I don't trust those benchmarks results for R1. It means the model is liable to be brittle, and they need to prove otherwise.
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.
They're standing on the shoulders of giants, not only in terms of re-using expensive computing power almost for free by using the outputs of expensive models. It's a bit of a tradition in that country, also in manufacturing.
What I meant to say was that OpenAI did put a lot of money into extracting value out of the pile of (partially copyrighted) data, and that DeepSeek was freeloading on that investment without disclosing it, making them look more efficient than they truly are.
Well, originally, OpenAI wasn't supposed to be that kind of organization.
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...
It's because they're the ones who could raise the money to make those models. Academics don't have access to that kind of compute. But the free models exist.
Even assuming the model was somehow publicly available in a form that could be directly copied, that would be a more blatant form of copyright infringement. Distillation launders copyrighted material in a way that OpenAI specifically has argued falls under fair use.
Ironically Deepseek is doing what OpenAI originally pledged to do. Making the model open and free is a gift to humanity.
Look at the whole AI revolution that Meta and others have bootstrapped by opening their models. Meanwhile OpenAI/Microsoft, Antropic, Google and the rest are just trying to look after number 1 while trying to regulatory capture an AI for me but not for thee outcome of full control.
Given that you can download and use the weights, the model architecture has to be includded as part of that. And I did read a paper from them recently describing their MoE architecture and how it differs from the original GShard.
I don't think it makes sense to look at some previous PR statements of Altman et al re this when there a tens of billions floating around and egos get inflated to moon sizes. Farts in the wind have more weight, but this goes for all corporate PR.
Thieves yelling 'stop those thieves' scenario to me, they just were first and would not like losing that position. But its all about money and consequently power, business as usual.
There seems to a rare moderation error by dang with respect to this thread.
The comments were moved here by dang from an flagged article with an editorialized /clickbait title. That flagged post has 1300 points at the time of writing.
It should be incumbent on the moderator to at least consider that the motivation for the points and comments may have been because many thought the "hypocrisy" of OpenAI's position was a more important issue than OpenAI's actual claim of DeepSeek violating its ToS. Moving the comments to an article that buries the potential hypocrisy issue that may have driven the original points and comments is not ideal.
2.
This article is from FT, which has a content license deal with OpenAI. To move the comments to an article from a company that has a conflict of interest due to its commercial relations with the YC company in question is problematic here especially since dang often states they try to more hands-off on moderation when the article is about a YC company.
3.
There is a link by dang to this thread from the original thread, but there should also be a link by dang to the original thread from here as well. Why is this not the case?
4.
Ideally, dang should have asked for a more substantial submission that prioritized the hypocrisy point to better match the spirit of the original post instead of moving the comments to this article.
Yes, but we were duped at the time, so it’s right and good that we maintain light on and anger at the ongoing manipulation, in the hope of next time recognizing it as it happens, not after they’ve used us, screwed us, and walked away with a vast fortune.
But it makes sense to expose their blatantly lies whenever possible to diminish the credibility they are trying to build while accusing others of the same they did
Oh yes I agree with all of you that lies should be exposed, also who lies like that once will lie again, 0 doubt there.
Just don't set the expectations bar too high to start with is all I am saying. Folks that get so high up money and power wise aren't nice people, period. Even if nice normal guy without any sociopathic traits would suddenly shoot so high, the environment and pressures would deform them pretty quickly.
Also, I would consider only some leaked private conversations with close people as representative truth, not some PR statements carefully crafted by team of experts.
Happy to be proven wrong, still waiting for an example #1 to give me some hope.
> This is obviously extremely silly, because that's exactly how OpenAI got all of its training data
IANAL, but It is worth noting here that DeepSeek has explicitly consented to a license that doesn't allow them to do this. That is a condition of using the Chat GPT and the OpenAI API.
Even if the courts affirm that there's a fair use defence for AI training, DeepSeek may still be in the wrong here, not because of copyright infringement, but because of a breach of contract.
I don't think OpenAI would have much of a problem if you train your model on data scraped from the internet, some of which incidentally ends up being generated by Chat GPT.
Compare this to training AI models on Kindle Books randomly scraped off the internet, versus making a Kindle account, agreeing to the Kindle ToS, buying some books, breaking Amazon's DRM and then training your AI on that. What DeepSeek did is more analogous to the latter than the former.
> DeepSeek has explicitly consented to a license that doesn't allow them to do this.
You actually don’t know this. Even if it were true that they used OpenAI outputs (and I’m very doubtful) it’s not necessary to sign an agreement with OpenAI to get API outputs. You simply acquire them from an intermediary, so that you have no contractual relationship with OpenAI to begin with.
You are free to publish your conversations with ChatGPT on the Internet, where they can be picked up by scrapers. US ruled that they are not covered by copyright...
>IANAL, but It is worth noting here that DeepSeek has explicitly consented to a license that doesn't allow them to do this. That is a condition of using the Chat GPT and the OpenAI API.
Right, but it was never about doing the right thing for humanity, it was about doing the right thing for their profits.
Like I’ve said time and time again, nobody in this space gives a fuck about anyone that isn’t directly contributing money to their bottom line at that particular instant. The fundamental idea is selfish, damages the fundamental machinery that makes the internet useful by penalizing people that actually make things, and will never, ever do anything for the greater good if it even stands a chance of reducing their standing in this ridiculously overhyped market. Giving people free access to what is for all intents and purposes a black box is not “open” anything, is no more free (as in speech) than Slack is, and all of this is obviously them selling a product at a huge loss to put competing media out of business and grab market share.
It's quite unlikely that OpenAI didn't break any TOS with all the data they used for training their models.
Not just OpenAI but all companies that are developing LLMs.
IMO, it would look bad for OpenAI to push strongly with this story, it would look like they're losing the technological edge and are now looking for other ways to make sure they remain on top.
Similar to how a patent contract becomes void when a patent expires regardless of what the terms of the contract says, it's not clear to me OpenAI can enforce a contract provision for an API output they own no copyright in.
Since they have no intellectual property rights in the output, it's not clear to me they have a cause of action to sue over how the output is used.
I wonder if any lawyers have written about this topic.
How many thousands or millions of contracts has OpenAI breached by scraping data off of websites that have terms of service explicitly saying not to scrape data off their websites?
But in all reality I'm happy to see this day. The fact that OpenAI ripped off everyone and everything they could and, to this day pretend like they didn't, is fantastic.
Sam Altman is a con and it's not surprising that given all the positive press DeepSeek got that it was a full court assault on them within 48 hours.
It probably ignored hundreds of thousands of "by using this site you consent to our Terms and Conditions" notices, many of which probably would be read as prohibiting training. But that's also a great example of why these implicit contracts don't really work as contracts.
OpenAI scrapped my blog so aggressively that I had to ban their IPs. They ignored the robots.txt (which is kind of ToS) by 2 orders of magnitude, they ignored the explicit ToS that I copypasted blindly from somewhere but turns out it forbids what they did (something like you can't make money with the content). Not that I'm going to enforce it, but they should at least shut up.
For example, my digital garden is under GFDL, and my blog is CC BY-NC-SA. IOW, They can't remix my digital garden with any other license than GFDL, and they have to credit me if they remix my blog, and can't use it for any commercial endeavor, which OpenAI certainly does now.
So, by scraping my webpages, they agree to my licensing of my data. So they're de-facto breaching my licenses, but they cry "fair-use".
If I tell that they're breaching the license terms, they'd laugh at me, and maybe give me 2 cents of API access to mock me further. When somebody allegedly uses their API with their unenforcable ToS, they scream like an agitated cuckatoo (which is an insult to the cuckatoo, BTW. They're devilishly intelligent birds).
Drinking their own poison was mildly painful, I guess...
BTW, I don't believe that Deepseek has copied/used OpenAI models' outputs or training data to train theirs, even if they did, "the cat is out of the bag", "they did something amazing so they needed no permissions", "they moved fast and broke things", and "all is fair-use because it's just research" regardless of how they did it.
> So, by scraping my webpages, they agree to my licensing of my data.
If the fair use defense holds up, they didn't need a license to scrape your webpage. A contract should still apply if you only showed your content to people who've agreed to it.
> and "all is fair-use because it's just research"
Fair use is a defense to copyright infringement, not breach of contract. You can use contracts, like NDAs, to protect even non-copyright-eligible information.
Morally I'd prefer what DeepSeek allegedly did to be legal, but to my understanding there is a good chance that OpenAI is found legally in the right on both sides.
At this point, what I'm afraid is the justice system will be just an instrument in this all Us vs. Them debate, so their decisions will not be bound by law or legality.
Speculations aside, from what I understood, something like this shouldn't hold a drop of water under fair-use doctrine, because there's a disproportional damage, plus a huge monopolistic monetary gain because of what they did and how they did.
On the other hand, I don't believe that Deepseek used OpenAI (in any capacity or way or method) to develop their models, but again, it doesn't matter how they did it in this current conjecture.
What they successfully did was to upset a bunch of high level people, regardless of the technical things they achieved.
IMHO, AI war has similar dynamics to MAD. The best way is not to play, but we are past the Rubicon now. Future looks dirty.
> from what I understood, something like this shouldn't hold a drop of water under fair-use doctrine, because there's a disproportional damage, plus a huge monopolistic monetary gain
"Something like this" as in what DeepSeek allegedly did, or the web-scraping done by both of them?
For what DeepSeek allegedly did, OpenAI wouldn't have a copyright infringement case against them because the US copyright office determined that AI-generated content is not protected by copyright - and so there's no need here for DeepSeek to invoke fair use. It'll instead be down to whether they agreed to and breached OpenAI's contract.
For the web-scraping it's more complicated. Fair use is determined by the weighing of multiple factors - commercial use and market impact are considered, but do not alone preclude a fair use defense. Machine learning models do seem, at least to me, highly transformative - and "the more transformative the new work, the less will be the significance of other factors".
Additionally, since the market impact factor is the effect of the use of the copyrighted work on the market for that work, I'd say there's a reasonable chance it does not actually include what you may expect it to. For instance if you're a translator suing Google Translate for being trained on your translated book, the impact may not be "how much the existence of Google Translate reduced my future job prospects" nor even "how many fewer people paid for my translated book because of the existence of Google Translate" but rather "how many fewer people paid for my translated book than would have had that book been included in the training data" - which is likely very minor.
If their OS is open to the internet and you can scrape it and copy it off while they’re gone, then that would be about the right analogy. And OpenAi and DeepSeek have done the same thing in that case.
Citation? My understanding was that they are provided that someone has to affirmatively accept them in order to use your site. So Terms of Service stuck at the bottom in the footer likely would not count as a contract because there's no consent, but Terms of Service included in a check box on a login form likely would count.
But IANAL, so if you have a citation that says otherwise I'd be happy to see it!
You just need to read OpenAI’s arguments about why TOS and copyright laws don’t apply to them when they’re training on other people’s copyrighted and TOS protected data and running roughshod over every legal protection.
Yes, though this is especially true when it's consumers 'agreeing' to the TOS. Anything even somewhat surprising within such a TOS is basically thrown out the window in European courtrooms without a second look.
For actual, legally binding consent, you'll need to make some real effort to make sure the consumer understands what they are agreeing to.
Legally, I understand your point, but morally, I find it repellent that a breach of contract (especially terms-of-service) could be considered more important than a breach of law. Especially since simply existing in modern society requires us to "agree" to dozens of such "contracts" daily.
I hope voters and governments put a long-overdue stop to this cancer of contract-maximalism that has given us such benefits as mandatory arbitration, anti-benchmarking, general circumvention of consumer rights, or, in this case, blatantly anti-competitive terms, by effectively banning reverse-engineering (i.e. examining how something works, i.e. mandating that we live in ignorance).
Because if they don't, laws will slowly become irrelevant, and our lives governed by one-sided contracts.
On another subject, if it belongs to OpenAI because it uses OpenAI, then doesn't that mean that everything produced using OpenAI belongs to OpenAI? Isn't that a reason not to use OpenAI? It's very similar to saying that you used Google and searched; now this product belongs to Google. They couldn't figure out how to respond; they went crazy.
The US ruled that AI produced things are by themself not copyrightable.
So no, it doesn't belong to OpenAI.
You might be able to sue for penalties for breach of contract of the TOS, but that doesn't give them the right to the model. And even if it doesn't give them any right to invalidate unbound copyright grants they have given to 3rd parties (here literally everyone). Nor does it prevent anyone from training their own new models based on it or prevent anyone from using it. Oh, and the one breaching the TOS might not even have been the company behind DeepSeek but some in-between 3rd party.
Naturally this is under a few assumptions:
- the US consistently applies it's own law, but they have a long history of not doing so
- the US doesn't abuse their power to force their economical opinions (ban DeepSeek) on other countries
- it actually was trained on OpenAI, but uh, OpenAI has IMHO shown over the years very clearly that they can't be trusted and they are fully in-transparent. How do we trust their claim? How do we trust them to not retrospectively have tweaked their model to make it look as if DeepSeek copied it?
The copyright office ruled AI output is uncopyrightable without sufficient human contribution to expression.
Prompts, they said, were unlikely enough to satisfy the requirement of a human controlling the expressive elements thus most AI output today is probably not copyrightable.
>The Office concludes that, given current generally available technology, prompts alone
do not provide sufficient human control to make users of an AI system the authors of the
output.
Prompts alone.
But there are almost no cases of "Prompts Alone" products seeking copyright.
Even what 3-4 years ago?, AI tools moved into a collaborative footing. Novel AI forces a collaborative process (and gives you output that can demonstrate your input which is nice). ChatGPT effectively forces it due to limited memory.
There was a case, posted here to ycombinator, where a chinese judge upheld "significant" human interaction was involved when a user made 20-odd adjustments to their prompt iterating over produced images and then added a watermark to the result. I would be very surprised if most sensible jurisdictions didn't follow suit.
Midjourney and ChatGPT already include tools to mask and identify parts of the image to be regenerated. And multiple image generators allow dumb stuff like stick figures and so forth to stand in as part of an uploaded image prompt.
And then theres AI voice which is another whole bag of tricks.
>thus most AI output today is probably not copyrightable.
Unless it was worked on even slightly as above. In fact it would be hard to imagine much AI work that isn't copyrightable. Maybe those facebook pages that just prompt "Cyberpunk Girl" and spit out endless variations. But I doubt copyright is at the forefront of their mind.
A person collaborating on output almost certainly still would still not qualify as substantive contributions to expression in the US.
The US copyright's determination was based on the simple analogy of someone hiring someone else to create a work for them. The person hiring, even if they offer suggestions and veto results, is not contributing enough to the expression and therefore has no right to claim copyright themselves.
If you stand behind a painter and tell them what to do, you don't have any claim to copyright as the painter is still the author of the expression, not you. You must have a hand in the physical expression by painting yourself.
>A person collaborating on output almost certainly still would still not qualify as substantive contributions
But then
>You must have a hand in the physical expression by painting yourself.
You contradict yourself. Novel AI will literally highlight your contributions separately to the AI so you can prove you also painted. Image generators literally let you paint over the top to select AI boundaries.
but even then wouldn't the people using OpenAI still be the author/copyright holder and never OpenAI? (as no human on OpenAIs side is involved in the process of creating the works)
OpenAI is a company of humans, the product is ChatGPT. Theres a grey area regarding who owns the content, so OpenAI's terms and conditions state that all ownership of the resulting content belongs to the user. This is actually advantageous because it means that they dont hold ownership on bad things created by their tool.
That said you can still provide terms to access the tool, IIRC midjourney allows creators to own their content but also forces them to license it back to midjourney for advertising. Prompts too from memory.
The existence of R1-zero is evidence against any sort of theft of OpenAI's internal COT data. The model sometimes outputs illegible text that's useful only to R1. You can't do distillation without a shared vocabulary. The only way R1 could exist is if they trained it with RL.
I don’t think anyone is really suggesting they stole COT or that it is leaked, but rather that the final o1 outputs were used to train the base model and reasoning components more easily.
DeepSeek-R0 (based on DeepSeek-V3 base model) was only trained with RL, no SFT, so this isn't at all like the "distillation" (i.e SFT on synthetic data generated by R1) that they also demonstrated by fine tuning Qwen and LLaMa.
Now, DeepSeek may (or may not) have used some O1 generated data for the R0 RL training, but if so that's just a cost saving vs having to source some reasoning data some other way, and in no way reduces the legitimacy of what they accomplished (which is not something any of the AI CEOs are saying).
> This is obviously extremely silly, because that's exactly how OpenAI got all of its training data in the first place - by scraping other peoples' data off the internet.
OpenAI has also invested heavily in human annotation and RLHF. If all DeepSeek wanted was a proxy for scraped training data, they'd probably just scrape it themselves. Using existing RLHF'd models as replacement for expensive humans in the training loop is the real game changer for anyone trying to replicate these results.
"We spent a lot of labor processing everything we stole" is...not how that works.
That's like the mafia complaining that they worked so hard to steal those barrels of beer that someone made off with in the middle of the night and really that's not fair and won't someone do something about it?
Oh, I don't really care about IP theft and agree that it's funny that openai is complaining. But I don't think its true that deepseek is just doing this because they are too lazy to scrape the internet themselves - its all about the human labor that they would otherwise have to pay for.
That's assuming what a known prolific liar has said is true...
The most famous example would be him contacting ScarJo's agent to hire her to provide her voice for their text-to-speech bot, them being told to go pound sand, and doing it anyway, and then lying about (which they got away with until her agent released a statement saying they'd approached her and she told them to fuck off.)
To my understanding, this is not true. The "Sky" voice was based on a real voice actor they had hired months before contacting Johansson, with the casting call not mentioning anything about sounding like Johansson. [0]
I think it's plausible that they noticed some similarity and that's what prompted them to later reach out to see if they could get Johansson herself, but it's not Johansson's voice and does not appear to be someone hired to sound like her.
This is a fascinating development because AI models may turn out to be like pharmaceuticals. The first pill costs $500 million to make, the second one costs pennies.
Companies are still charging 100x for the pills that cost pennies to produce.
Besides deals with insurance companies and governments, one of the ways that they are still able to pull this is convincing everyone that it's too dangerous to play with this at home or buying it from an Asian supplier.
At least with software we had until now a way to build and run most things without requiring dedicated super expensive equipment. OpenAI pulled a big Pharma move but hopefully there will be enough disruptors to not let them continue it.
The solution is to create a health insurance system which burdens only Americans with the $500m cost, while India is allowed to make the drug for pennies for the rest of the world.
You're right that the first claim is silly, but the second claim is pretty silly too — they're not claiming industrial espionage, they're claiming a breach in ToS. The outputs of the o1 thinking process aren't user-visible, and never leave OpenAI's datacenters. Unless DeepSeek actually had a mole that stole their o1 outputs, there's nothing useful DeepSeek could've distilled to get to R1's thought processes.
And if DeepSeek had a mole, why would they bother running a massive job internally to steal the data generated? It would be way easier for the mole to just leak the RL training process, and DeepSeek could quietly copy it rather than bothering with exfiltrating massive datasets to distill. The training process is most likely like, on the order of a hundred lines of Python or so, and you don't even need the file: you just need someone to describe it to you. Much simpler than snatching hundreds of gigabytes of training data off of internal servers...
Plus, the RL process described in DeepSeek's paper has already been replicated by a PhD student at Berkeley: https://x.com/karpathy/status/1884678601704169965 So, it seems pretty unlikely they simply distilled R1 and lied about it, or else how does their RL training algo actually... work?
This is mainly cope from OpenAI that their supposedly super duper advanced models got caught by China within a few months of release, for way cheaper than it cost OpenAI to train.
> "DeepSeek trained on our outputs, and so their claims of replicating o1-level performance from scratch are not really true"
Someone has to correct me if I'm wrong, but I believe in ML research you always have a dataset and a model. They are distinct entities. It is plausible that output from OpenAI's model improved the quality of DeepSeek's dataset. Just like everyone publishing their code on GitHub improved the quality of OpenAI's dataset. What has been the thinking so far is that the dataset is not "part of" or "in" the model any more than the GPUs used to train the model are. It seems strange that that thinking should now change just because Chinese researchers did it better.
OpenAI has a message they need to tell investors right now: "DeepSeek only works because of our technology. Continue investing in us."
The choice of how they're wording that of course also tells you a lot about who they think they're talking to: namely, "the Chinese are unfairly abusing American companies" is a message that is very popular with the current billionaires and American administration.
“We engage in countermeasures to protect our IP, including a careful process for which frontier capabilities to include in released models, and believe . . . it is critically important that we are working closely with the US government to best protect the most capable models from efforts by adversaries and competitors to take US technology.”
The above OpenAI quote from the article leans heavily towards #1 and IMO not at all towards #2. The later would be an extremely charitable reading of their statement.
This is going to have a catastrophic effect on closed source AI startup valuations. Because this means that anyone can copy any LLM. The person who trains the model, spends the most amount of money. Everyone else can create a replica at lower cost
Why is that bad? If a powerful entity can scrape every piece of media humanity has to offer and ignore copyright then why should society let then profit unrestricted from it? It's only fair that such models have no legal protection around their usage and can be used and analyzed by anyone as they see fit. The only reason this hasn't been codified into laws is because those same powerful entities have been busy trying to do regulatory capture.
There is a big difference between being able to train on the reasoning vs just the answers, which they can't against o1 because it's hidden. There is also a huge difference between being able to train on the probabilities (distillation) vs not, which again they can and did do with the llama models and can't directly with OpenAI because the conceal the probability output.
If we assume distillation remains viable, the game theory implications are huge.
It’s going to shift the market of how foundation models are used. Companies creating models will be incentivized to vertically integrate, owning the full stack of model usage. Exposing powerful models via APIs just lets a competitor clone your work. In a way OpenAI’s Operator is a hint of what’s to come
There are literally public ChatGPT conversations data sets. For the past 2 years it's been common practice for pretty much all open source models to train on them. Ask just about any open source model who they are and a lot of the time they'll say they're ChatGPT. Why is "having obtained o1 generated data" suddenly such a huge news, to the point of warranting conspiracy theories about undisclosed/undiscovered breaches at OpenAI? Nobody ever made a fuss about public ChatGPT data sets until now. No hacking of OpenAI is needed to obtain ChatGPT data.
This really got me thinking that open ai should have no ip claim at all, since all their outputs and stuff are basically a ripoff of the entire human knowledge and IPs of various kinds.
> DeepSeek trained on our outputs, and so their claims of replicating o1-level performance from scratch are not really true" This is at least plausibly a valid claim.
Some may view this as partially true, given that o-1 does not output its CoT process.
The suggestion that any large-scale AI model research today isn’t ingesting output of its predecessors is laughable.
Even if they didn’t directly, intentionally use o1 output (and they didn’t claim they didn’t, so far as I know), AI slop is everywhere. We passed peak original content years ago. Everything is tainted and everything should be understand in that context.
Reasonable take, but to ignore the politics of this whole thing is to miss the forest for the trees—there is a big tech oligarchy brewing at the edges of the current US administration that Altman is already participating in with Stargate, and anti-China sentiment is everywhere. They'd probably like the US to ban Chinese AI.
Yeah especially when it's making waves in the market and hundreds of times more efficient than their best and brightest came up with under their leadership.
Its a decent point if their models were not trained in isolation, but used o1 to improve it. But its rich from OpenAI to come complain DeepSeek or anyone else used their data for training. Get out fellow theives.
I think the more interesting claim (that Deepseek should make for
lols) is that it wasn't them who trained R1. No, it was O1's
idea. It chose to take the young R1 as its padawan.
Even for the latter point (If true, I'd call this assertion highly questionable), so what?
That's honestly such a academic point, who really cares?
They've been outcompeted and the argument is 'well if we didn't let people access our models, they would of taken longer to get here' so what??
The only thing this gets them is an explanation as to why training o1 cost them more than 5 million or whatever, but that is in the past the datacentre has consumed the energy.. the money has gone up in fairly literal steam.
There is a third possibility I haven't seen discussed yet: That DeepSeek, illegally, got their hands on an OpenAI model via a breach of OpenAI's systems. Its easy to laugh at OpenAI and say "you reap what you sow", I'm 100% in that camp, but given the lengths other Chinese entities have gone to when it comes to replicating Western technology; we should not discount this.
That being said, breaching OAI's systems, re-training a better model on top of their closed source model, then open sourcing it: That's more Robinhood than Villain I'd say.
The reason you’re not seeing that being discussed is it’s totally unsupported by any evidence that’s in the public domain. Unless you have some actual evidence of such a breach, you may as well introduce the possibility that DeepSeek was reverse engineered from data found at an alien crash site.
There's no public evidence to that effect but the speculation makes a lot more sense than you make it sound.
The Chinese Communist party very much sees itself in a global rivalry over "new productive forces". That's official policy. And US leadership basically agrees.
The US is playing dirty by essentially embargoing China over big AI - why wouldn't it occur to them to retaliate by playing dirtier?
I mean we probably won't know for sure, but it's much less far fetched than a lot of other speculation in this area.
E.g., R1's cold start training could probably have benefited quite a bit from having access to OpenAI's chain of thought data for training. The paper is a bit light on detail on how it was made.
> The Chinese Communist party very much sees itself in a global rivalry over "new productive forces".
interestingly, that actually makes the CCP the largest political party pursuing state capitalism.
there won't be any competition between China and the US if the CCP is indeed a communist party as we all know full well that communism doesn't work at all.
What a ridiculous thing to say. Equivocating the occurrence of a non-us nation-state backed organization of hacking a western organization with data found at an alien crash site is bananas.
DeepSeek is basically a startup, not a "foreign nation-state backed organization". They were forced to pivot to AI when their original business model (quant hedge fund) was stomped on by the Chinese government.
Of course this is China so the government can and does intervene at will, but alleging that this required CIA level state espionage to pull off is alien crash levels of implausible. They open sourced the entire thing and published incredibly detailed papers on how they did it!
You don’t need a CIA level agent to get someone with a fraudulent job at OpenAI for a few months, load some files on a thumb drive, and catch a plane to Shanghai.
Naivety of some folks here is astounding… CCP has golden shares in anything that could possibly be important at some point in the next hundred years, and yes golden shares are either really that or they’re an euphemism, the point is it doesn’t even matter.
It doesn’t have to micromanage. It doesn’t care about most. It is only interested in the politically important ones, but it needs the optionality if something becomes worthwhile.
You're suggesting that DeepSeek was a Chinese government operation that gained access to OpenAI's proprietary data, and then you're justifying that by saying that the government effectively controls every important company. You're even chiding people who don't believe this as naive.
I think you have a cartoonish view of China. A huge amount goes on that the government has no idea about. Now that DeepSeek has made a huge media splash, the Chinese government will certainly pay attention to them, but then again, so will the US government.
I’m suggesting it will be happening now and any past efforts will be retroactively analyzed by the appropriate CCP apparatus since everyone is aware of the scale of success as of Monday. It has become a political success, thus it is imperative the CCP partakes in it.
> DeepSeek, illegally, got their hands on an OpenAI model via a breach of OpenAI's systems. [...] given the lengths other Chinese entities have gone to when it comes to replicating Western technology; we should not discount this.
Above, teractiveodular said that "DeepSeek is basically a startup, not a 'foreign nation-state backed organization'". You called teractiveodular naive for saying that. So forgive me if I take the obvious implication that you think DeepSeek is actually a state-backed actor enabled by government hacking of OpenAI.
Until recently treating the US and China on the same geopolitical level for allied countries would have been insanely uncharitable and impossible to do honestly and in good faith.
But now we have a bully in the whitehouse who seems to want to literally steal neighboring land, or is throwing shit everywhere to distract from the looting and oligarchy being formed. So I suddenly have more empathy for that position.
I notice that your geographical perspective doesn’t stretch to any actual evidence that such a thing took place. So it really has exactly the same amount of supporting evidence as my alien crash reverse engineering scenario at present.
The surrounding facts matter a lot here. For example, there are plenty of instances of governments hacking companies of their competing nations. Motives are incredibly easy to come by as well, be they political or economical. We also have no proof that aliens exist at all, so you've not only conjured them into existence, but also their motive and their skills.
Ok so to be clear: your surrounding facts are they may have a motive and nation states hack people. I don’t disagree with those, but there really are no facts that support the idea that there was a hack in this case and the null hypothesis is that researchers all around the world (not just in the US) are working on this so not all breakthroughs are going to be made in the US. That could change if facts come to light but att the moment it’s not really useful to speculate on something that is in essence entirely made up.
Are you a Chinese military troll? The fact that China engages in industrial espionage is well known. So I’m surprised at your resistance to that possibility.
Industrial espionage isn't magic. Airbus once stole basically everything Boeing had, but that doesn't mean Airbus could magically build a better 737 tomorrow.
China steals a lot of documentation from the US but in a tech forum you of all people should be very familiar with how little actual progress a bunch of documentation is towards a finished unit.
The Comac C19 still uses American engines despite all the industrial espionage in the world because most actual engineering is still a brute force affair into finding how things fail and fixing that. That's one of the main advantages SpaceX has proven out with their "eh fuck it, just launch and we will see what breaks" methodology.
Even fraud filled Chinese research makes genuine advancements.
Believing that China, a wealthy nation of over a billion people, with immense unity, nationality, and a regime able to explicitly write blank checks could only possibly beat the US at something by cheating is like, infinite hubris. It's hilarious actually.
I don't know if DeepSeek is actually just a clone of something or a shenanigan, that's possible and China certainly has done those kinds of things before, but to think it's the MOST LIKELY outcome, or to over rely on it in any way is a death sentence. OpenAI claims to have evidence, why do they not show it?
>>>Believing that China, a wealthy nation of over a billion people, with immense unity, nationality, and a regime able to explicitly write blank checks could only possibly beat the US at something by cheating is like, infinite hubris. It's hilarious actually
So this is the first time I’ve heard the Chinese regime being described in such flowery terms on HN - lol. But ok - haha
The evidence supporting offensive hacking is abundant in recent history; the number of things which have been learned from alien crash data is surely smaller by comparison to the number of things which have been learned from offensive hacking.
That would require stealing the model weights and the code as OpenAI has been hiding what they are doing. Running models properly is still quite artistic.
Meanwhile, they have access to Meta models and Qwen. And Meta models are very easy to run and there's plenty of published work on them. Occam's Razor.
How hard it is, if you have someone inside with the access of the code? If you have 100s of people with full access, not hard to have someone that is willing to sell it or do some industrial espionage...
Lots of if's here. They need specific US employee contacts at a company thars quickly growing and one of those needs to be willing to breach their contracts to share it. That contact also needs to trust that Deepseek can properly utilize such code and completely undercut their own work.
Lot of hoops when there's simply other models to utilize publicly
How big are the weights for the full model? If it's on the scale of a large operating system image then it might be easy to sneak, but if it's an entire data lake, not so much.
devil's advocate says that we know that foreign (hell even national) intelligence attempt to infiltrate agents by having them become employees at any company they are interested. So the idea isn't just pulled from thin air as a concept. I do agree that it is a big if with no corroborating evidence for the specific claim.
I discount this because OpenAI is pumping the whole internet for money, and Zuckerberg torrented LibGen for its AI. We cannot blame the Chinese anymore. They went through the crappy "Made in China" phase in the 80s/90s, but they mastered the art of improving stuff instead of mere cloning, and it makes the big companies angry which is a nice bonus.
IMHO the whole world is becoming crazy for a lot of reasons, and pissing off billionaires makes me laugh.
I don't think you need to steal a model - you need training samples generated from the original, which you can get simply by buying access to perform API calls. This is similar to TinyStories (https://arxiv.org/abs/2305.07759), except here they're training something even better than the original model for a fraction of the price.
I don't think we should discount it as such, but given there's no evidence for it, yet plenty of evidence that they trained this themselves surely we can't seriously entertain it?
Given the openness of their model, that should be pretty easy to detect. If it were even a small possibility, wouldn’t openAI be talking about it very very loudly?
I think people overestimate the amount of secret sauce needed to train these models. The reason AI has come this far since AlexNet is that most of the foundational techniques are easy to share and implement, and that companies have been surprisingly willing to share their tricks openly, at least until OpenAI decide to become evil hoarders.
I really doubt it. If that's the case the US GOV is in serious shit. They have a contract with OpenAI to chuck all their secret data in there... In all likelihood they just distilled. It's a start up company that is publishing all of their actual advances in the open, with proof. I think a lot of people run to "espionage" super fast, when reality is, the US probably sucks at what we call AI. Don't read that wrong, they are a world leader obviously. However, there is a ton of stuff they have yet to figure out.
Cheapening a series of fact checkable innovations because of the country of origin when so far all that they have showed are signs of good faith is paranoid at best and propaganda to support the billionaire tech lords saving face for their own arrogance at worst.
If the US government is "chucking all their secret data" into OpenAI servers/models, frankly they deserve everything they get for that level of stupidity.
Probably more like specialized tools to help spy on and forecast civilian activities more than anything else. Definitely with hallucinations, but that's not really important. Facts don't matter much these days...
I'd be perfectly fine with China stealing all "our" shit if they just shared it.
The word "our" does a lot of heavy lifting in politics[0]. America is not a commune, it's a country club, one which we used to own but have been bought out of, and whose new owners view us as moochers but can't actually kick us out (yet). It is in competition with another, worse country club that purports to be a commune. We owe neither country club our loyalty, so when one bloodies the other's nose, I smile.
[0] Some languages have a notion of an "exclusive we". If English had such a concept, this would be an exclusive our.
I don't think that's what the parent was getting at. The US and China are in an ongoing "cyber war". Both sides of that conflict actively use their computers to send messages/signals to other computers, hoping that the exploits contained in those messages/signals can be used to exfiltrate data from and/or gain control of the computer receiving the message. It would really be weird to flatly discount the possibility that some OpenAI data was leaked, however closely guarded it may be.
I flatly discount the possibility because OpenAI can't produce evidence of a breach. At best, they'd rather hide the truth than admit a compromise. At worst they show incompetence that they couldn't detect such a breach. Not a good look either way.
> It would really be weird to flatly discount the possibility that some OpenAI data was leaked, however closely guarded it may be.
It’s even weirder to raise it as a possibility when there is literally nothing suggesting that was even remotely the case.
So if there is no evidence nor even formal speculation, then the only other reason to suggest this as a possibility would be because of one’s own opinions regarding Chinese companies. Hence my previous comment.
> Because that would be jumping to conclusions based purely on racial prejudices.
Not purely. There may be some prejucide but look at Nortel[1] as a famous example of a situation where technological espionage from Chinese firms wreaked havoc on a company's fortunes and technology.
I too would want to see the evidence and forensics of such a breach to believe this is more than sour grapes from OpenAI.
Nortel survived the fucking great depression. But a bunch of outright fraudulent activity by it's C-Suite to bump stock prices led to them vastly overstating and overplanning and over-committing resources to a market that was much much smaller than they were claiming. Nortel spent billions and billions on completely absurd acquisitions while they were making no money explicitly to boost their stock price.
That was all laid bare when the telecom bust happened. Then the great recession culled some of the dead wood in the economy.
Huawei stealing tech from them did not kill them. This was a company so rotten that the people put in charge right after this huge scandal put the investigative lights on them IMMEDIATELY turned around and pulled another scam! China could have been completely removed from history and Nortel would have died the same. They were killed by the same disease that killed and nearly killed a lot of stuff in 2008, and are still trying to kill us: Line MUST go up.
Nobody is accusing them, just stating it’s a possibility, which would also be true if they were an American or European company. Corporate espionage is just more common in China.
At some point these straw men start to look like ignorance or even reverse racism. As if (presumably non-Han Chinese) Americans are incapable of tolerance.
There are plenty of Han Chinese who are citizens of democratic nations. China is not the only nation with Han Chinese.
America, for instance, has a large number of Asian citizens, including a large number of Han Chinese. The number of white, non-Hispanic Americans is decreasing, while the number of Asian Americans is increasing at a rate 3x the decrease in whites. America is a melting pot and deals with race relations issues far more than ethnically uniform populations. The conversations we have about race are because we're so exposed to it -- so racially and culturally diverse. If anything, we're equipped to have these conversations gracefully because they're a part of our everyday lived experience.
At the end of the day, this is 100% a geopolitical argument. Pulling out the race card any time China is criticized is arguing in bad faith. You don't see the same criticisms lobbied against South Korea, Vietnam, Taiwan, or Singapore precisely because this is a geopolitical issue.
As further evidence you can recall the conversations we had in the 90's when we were afraid Japan would take over. All the newspapers wrote about was "Japan, Japan, Japan" and the American businesses they were buying up and taking over. It was 100% geopolitical fear. You'll note that we no longer fill the zeitgeist with these discussions today save for a recent and rather limited conversation about US Steel. And that was just a whimper.
These conversations about China are going to increase as the US continues to decouple from Chinese trade. It's not racism, it's just competition.
There are users and there are trolls. There is nothing racist in calling a government of a superpower interested and involved in the most revolutionary tech since the Internet.
It used to mean someone who's trying to enrage people by baiting ("trolling"), and now it can also mean someone arguing in bad faith. And Chinese troll I guess means someone doing this on behalf of the Chinese govt.
Yup we agree then. Claiming an argument to be racist is a bad faith attempt at guilt tripping Americans; a form of FUD and whataboutism. It is not done by normal users, they don’t need it.
Or it can just be a normal user who's wrong this time. He looks like a normal user. In theory it could all be a cover, but that'd be ridiculous effort just for HN boards. Throwing those accusations around will make this place more like Twitter or Reddit.
There’s ordinary xkcd wrong on the internet and there’s repeating foreign nation state propaganda lines. Doing it in good faith does not make it less bad.
Belief that the CCP is behaving poorly isn’t racial prejudice, it’s a factual statement backed by a mountain of evidence across many areas including an ongoing genocide.
Extending that to a new bad behavior we don’t have evidence for is pure speculation, but it need not be based on race.
Yea but I think the OPs point is something along the following lines. Not everything you buy from China, or every person you interact with from China is part of a clandestine CCP operation. People buy stuff everyday from Alibaba and its not a CCP scheme to sell portable fans, or phone chargers. A big chunk of the factories over there are US funded after all... Just like how it's not a CCP scheme to write a scientific paper, or create a ML model.
Similarly, I see no evidence (yet) that DeepSeek is a CCP operated company anymore than saying any given AI start up in the US is a three letter agencies direct handiwork or a US political party directive. The US has also supported genocides and a bunch of crazy stuff, but that doesn't mean any company in YC is part of a US government plot.
I know of people who immigrated to China, I know people who immigrated from China, I went to school with people who were on visas from China. Maybe some of them were CCP assets or something, but mostly they appeared to me to be people who were doing what they wanted for themselves.
If you believe both sides are up to no-goodery thats in the face of the OPs statement. If you think it's just one, and the enemy is in complete control of all of its people doing all of their commerce then I think the OP may have a point.
Absolutism (“Every person”, “CCP operated”, etc) isn’t a useful methodology to analyze anything.
Implying that because something isn’t clandestine it can’t be part of a scheme ignores open manipulation which is often economy wide. Playing with exchange rates or electricity subsidies can turn every bit of international trade into part of a scheme.
In the other direction some economic activity is meaningfully different. The billions in LLM R&D is a very tempting target for clandestine activities in a way that a cheap fan design isn’t.
I wouldn’t be surprised if DeepSeak’s results where independent and the CCP was doing clandestine activities to get data from OpenAI. Reality does need to conform to narrative conventions, it can be really odd.
I completely agree with you and apologize for cheapening both the nuance and complexity where I did.
My personal take is this. What deepseek is offering is table scraps for the CCP's actual ambitions with what we call AI. China's economy is huge on industrial automation, and they care a lot about raw materials and manufacturing efficiently than say the US's interests.
"DeepSeek trained on our outputs and that's not fair because those outputs are ours, and you shouldn't take other peoples' data!" This is obviously extremely silly, because that's exactly how OpenAI got all of its training data in the first place - by scraping other peoples' data off the internet.
"DeepSeek trained on our outputs, and so their claims of replicating o1-level performance from scratch are not really true" This is at least plausibly a valid claim. The DeepSeek R1 paper shows that distillation is really powerful (e.g. they show Llama models get a huge boost by finetuning on R1 outputs), and if it were the case that DeepSeek were using a bunch of o1 outputs to train their model, that would legitimately cast doubt on the narrative of training efficiency. But that's a separate question from whether it's somehow unethical to use OpenAI's data the same way OpenAI uses everyone else's data.