Hacker News Reader

Comment by qwertox | Hacker News Reader

"We had difficulties accessing OpenAI, our data provider." /s

Rumour was that DeepSeek used the outputs of the thinking steps in OpenAI's reasoning model (o1 at the time) to traing DeepSeek's Large Reasoning Model R1.

More like a direct (and extremely dubious) accusation without proof from Altman. In reality those two models have as little in common as possible, and o1 reasoning chain wasn't available anyway.

Maybe they also do that, but I work with a class of problems* that no other model has managed to crack, except for R1 and that is still the case today.

Remember that DeepSeek is the offshoot of a hedge fund that was already using machine learning extensively, so they probably have troves of high quality datasets and source code repos to throw at it. Plus, they might have higher quality data for the Chinese side of the internet.

* Of course I won't detail my class of problems else my benchmark would quickly stop being useful. I'll just say that it is a task at the undergraduate level of CS, that requires quite a bit of deductive reasoning.

This is completely irrelevant without knowing if you are effectively prompting each model. Your workflow may just be suitable for a particular model and not others. And tuning a workflow for each model is tedious. I seriously doubt there is ANY class of problem DSR1 can solve that OAI's third tier model can't at this point (o4-mini).

Deepseek published thinking trace before OpenAI did, not after.

I don't think so. They came up with a new RL algorithm that's just better.

Better how? DeepSeek has never held the top spot in any aggregated benchmark. The Chinese bot armies are certainly better at convincing the internet they are trailing western models, despite the fact that in practical use this is not the case and at this rate, likely will never be. If AI progress is exponential, so is falling behind. What doesnt change however, is holding your ace card and dropping it when the time is right. China is competing with the public western models. Western SOTA labs are competing with their previous unreleased SOTA model. Publicly, China is ~3 months behind. But in reality, they are much further behind, and will never catch up.

Be mindful of what this means. A kid in his garage fine tuning a model can "catch up" to SOTA models for most use cases. For actual "frontier" work that requires SOTA levels of intelligence, there are only 3 companies in the race. None of them are from China or Europe.

OpenAI used literally all available text owned by the entire human race to train o1/o3.

so what?

This, my guess is OpenAI wised up after r1 and put safeguards in place for o3 that it didn't have for o1, hence the delay.

I think that's unlikely.

DeepSeek-R1 0528 performs almost as well as o3 in AI quality benchmarks. So, either OpenAI didn't restrict access, DeepSeek wasn't using OpenAI's output, or using OpenAI's output doesn't have a material impact in DeepSeek's performance.

https://artificialanalysis.ai/?models=gpt-4-1%2Co4-mini%2Co3...

almost as well as o3? kind of like gemini 2.5? I dug deeper and surprise surprise: https://techcrunch.com/2025/06/03/deepseek-may-have-used-goo...

I am not at all surprised, the CCP views AI race as absolutely critical for their own survival...

Not everything that's written is worth reading, let alone drawing conclusions from. That benchmark shows different trees each time the author runs it, which should tell you something about it. It also stacks grok-3-beta together with gpt-4.5-preview in the GPT family, making the former appear to be trained on the latter. This doesn't make sense if you check the release dates. And previously it classified gpt-4.5-preview to be in a completely different branch than 4o (which does make some sense but now it's different).

EQBench, another "slop benchmark" from the same author, is equally dubious, as is most of his work, e.g. antislop sampler which is trying to solve an NLP task in a programmatic manner.

The benchmarks are not reflective of real world use case. This is why OpenAI dominates B2B. As a business, its in your best interest to save money without sacrificing quality.

"Follow the money."

Businesses are pouring money into the OpenAI API. This is your biggest clue.

It would be hypocritical to criticize DeepSeek if this is true, since OpenAI and all major players in this space train their models on everything they can get their hands on, with zero legal or moral concerns. Pot, meet kettle.

Not too sure why you are downvoted but OpenAI did announce that they are investigating on the Deepseek (mis)use of their outputs, and that they were tightening up the validation of those who use the API access, presumably to prevent the misuse.

To me that does seem like a reasonable speculation, though unproven.

I still find it amusing to call it "misuse". No AI company has ever asked for permission to train.

Exactly because it's phrased like the poster knows this is the reason. I wouldn't downvote it if it was a clear speculation with the link to the OAI announcement you mentioned for bonus points.

Reply on news.ycombinator.com