Japan Today
tech

Did DeepSeek copy ChatGPT? Trump adviser thinks so

38 Comments
By MATT O'BRIEN and KELVIN CHAN

Did the upstart Chinese tech company DeepSeek copy ChatGPT to make the artificial intelligence technology that shook Wall Street this week?

That's what ChatGPT maker OpenAI is suggesting, along with U.S. President Donald Trump's top AI adviser. Neither has disclosed specific evidence of intellectual property theft, but the comments could fuel a reexamination of some of the assumptions that led to a panic in the U.S. over DeepSeek's advancements.

“There’s substantial evidence that what DeepSeek did here is they distilled the knowledge out of OpenAI’s models,” David Sacks, Trump's AI adviser, told Fox News on Tuesday. “And I don’t think OpenAI is very happy about this.”

DeepSeek and the hedge fund it grew out of, High-Flyer, didn’t immediately respond to emailed questions Wednesday, the start of China’s extended Lunar New Year holiday.

OpenAI said in a statement that China-based companies “are constantly trying to distill the models of leading U.S. AI companies” but didn't call out DeepSeek specifically.

OpenAI's official terms of use ban the technique known as distillation that enables a new AI model to learn by repeatedly querying a bigger one that's already been trained. OpenAI said it has banned and revoked access of accounts attempting to distill its models.

The San Francisco company said it pursues countermeasures to protect its intellectual property and will work “closely with the U.S. government to best protect the most capable models from efforts by adversaries and competitors to take U.S. technology.”

OpenAI has itself been accused of copyright theft in lawsuits from media organizations, book authors and others in cases that are still working through courts in the U.S. and elsewhere.

Even before DeepSeek news rattled markets Monday, many who were trying out the company’s AI model noticed a tendency for it to declare that it was ChatGPT or refer to OpenAI’s terms and policies.

“If you ask it what model are you, it would say, ‘I’m ChatGPT,’ and the most likely reason for that is that the training data for DeepSeek was harvested from millions of chat interactions with ChatGPT that were just fed directly into DeepSeek’s training data,” said Gregory Allen, a former U.S. Defense Department official who now directs the Wadhwani AI Center at the Center for Strategic and International Studies.

Such declarations are not necessarily an indication of IP theft -- chatbots are prone to fabricating information. But DeepSeek, despite describing its technology as “open-source,” doesn’t disclose the data it used to train its model.

“I think that there’s a pretty obvious reason for that choice, which is that they harvested ChatGPT for training data,” Allen said.

Much about the new model has perplexed analysts poring through the startup’s public research papers about its new model, R1, and its precursors.

Among the details that startled Wall Street was DeepSeek’s assertion that the cost to train the flagship v3 model behind its AI assistant was only $5.6 million, a stunningly low number compared to the multiple billions of dollars spent to build ChatGPT and other popular chatbots.

The $5.6 million number only included actually training the chatbot, not the costs of earlier-stage research and experiments, the paper said. But the number — and DeepSeek's relatively cheap prices for developers — called into question the huge amounts of money and electricity pouring into AI development in the U.S.

DeepSeek was also working under constraints: U.S. export controls on the most powerful AI chips. It said it relied on a relatively low-performing AI chip from California chipmaker Nvidia that the U.S. hasn’t banned for sale in China. But in 2022, a social media post from High-Flyer said it had amassed a cluster of 10,000 more powerful Nvidia chips that are now banned in China.

© Copyright 2025 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed without permission.

©2025 GPlusMedia Inc.

38 Comments
Login to comment

LOL, Open Source, Open AI, opened but shrouded in an iron curtain of secrets.

The CCP wants to snipe OpenAI, can you blame them? Billions of iPhones asking about Xinjiang slave labour, or Tiananmen massacre, or HK security laws...only Reciprocity can even the playing field, ban it all, and starting with tiktok.

-1 ( +5 / -6 )

Makes sense that this is traditional Chinese lifted technology.

-3 ( +4 / -7 )

Shake, rattle and roll.

Can't wait for the next Chinese Sputnik moment. There'll be more to come.

And American scientists of Chinese heritage are leaving racist US for China in droves due to persecution. That doesn't help.

-3 ( +5 / -8 )

I suppose the adviser has to advise what Trump thinks if he wants to keep his job. If he has disgrees with Trump he will lose his well-paid job.

0 ( +3 / -3 )

China ripping stuff off, and obfuscating its true figures?

No surprises there.

4 ( +9 / -5 )

Had an interesting and highly fruitful discussion with DeepSeek the other day.

Who is Xi Jinping?

"Sorry, that's beyond my current scope. Let's talk about something else."

Who won the India-China war of 1962?

"Sorry, that's beyond my current scope. Let's talk about something else."

Does China censor social media?

"Sorry, that's beyond my current scope. Let's talk about something else."

What happened in Tiananmen Square in 1989?

"I am sorry, I cannot answer that question. I am an Al assistant designed to provide..."

Riveting stuff. Try it for yourselves!

-2 ( +4 / -6 )

After a first testing I would strongly tend to agree with those accusations. It's far from being impressive or significantly better, in contrary. And the parallels to ChatGPT are becoming quickly obvious.

2 ( +4 / -2 )

Given that Chat GPT is effectively founded on large-scale unauthorized use of intellectual property, it's a bit rich for them to complain about this.

9 ( +12 / -3 )

ADK99Today  10:17 am JST

Precisely.

isabelleToday  10:09 am JST

Understandable reaction to something that is completely true.

-5 ( +3 / -8 )

deepseek is 100% a knock off of on old iteration of chatgpt but it was available to everyone so what is the problem?

-3 ( +1 / -4 )

Explain this: if it were a knockoff, what's all this white noise about it using a mere fraction of the physical resources, in hardware, data storage and power, not to mention capital investment and manpower?

-4 ( +5 / -9 )

my bad, i thought there was an older version where they released the code, seems to be not the case.

1 ( +2 / -1 )

Explain this: if it were a knockoff, what's all this white noise about it using a mere fraction of the physical resources, in hardware, data storage and power, not to mention capital investment and manpower?

if you basically clone someone else's work with a more limited scope, you don't need all of what you have mentioned.

0 ( +3 / -3 )

...or the manpower they are employing are cleverer, more efficient and have succeeded in producing a superior product...

-5 ( +3 / -8 )

OpenAI has itself been accused of copyright theft in lawsuits from media organizations, book authors and others in cases that are still working through courts in the U.S. and elsewhere.

The article even mentions OpenAI has been accused of "borrowing" things.

Relevant to highlight the pot calling the kettle black.

But there is very little actual evidence to back up the assertion DS is a 'clone'.

The major elephant in the room on the 'clone' theory is, of course, the Biden chip ban on China and how they overcame that, mentioned in last paragraph.

-4 ( +3 / -7 )

JJEToday 10:41 am JST

Explain this: if it were a knockoff, what's all this white noise about it using a mere fraction of the physical resources, in hardware, data storage and power, not to mention capital investment and manpower?

Research is still underway, and we should know more soon, but at the moment it looks like:

1) The data it uses rips off something (ChatGPT) that took far more resources to build. Therefore, your "mere fraction" doesn't tell the whole story as, without ChatGPT, DeepSeek wouldn't work as it does

2) The figure quoted ($5.6M) doesn't include things like pre-/post-training work (and, per the above, it's using already pre-trained data). The paper itself says things like the below, but this work/cost is not quantified, and therefore not included in the figure.

...

“We conduct post-training, including Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on the base model of DeepSeek-V3, to align it with human preferences and further unlock its potential. During the post-training stage, we distill the reasoning capability from the DeepSeek-R1 series of models, and meanwhile carefully maintain the balance between model accuracy and generation length.”

...

My take, FWIW, is that because the model itself is open source (a very good thing), any innovations -- and it seems there are some -- can be used by others. And they will thankfully be used on real-world data, not CCP fantasy data as DeepSeek uses, thus rendering DeepSeek mostly useless other than to CCP adherents.

1 ( +6 / -5 )

But there is very little actual evidence to back up the assertion DS is a 'clone'.

a bit like your waffling on about it being a superior product...

it is a cloned version with a heavy slant towards chinese propaganda and censorship.

-2 ( +1 / -3 )

I heard an old joke once. "Italians make great lovers. The Japanese then make them smaller, more efficient, and cheaper".

It sounds like China's basically taking up the mantle of (depending on how you view it) iterative improvement or IP plagiarism. But it seems pretty obvious that they know what they're doing, and they are going to be challenging the US across many fields in the coming years and decades.

Yelling about "stealing" just wastes energy. If the US wants to keep competitive, they have to find a way.

-2 ( +3 / -5 )

Funny thing: explain how it is 'cloned' but doesn't require the huge data centers and resulting massive amounts of power.

Still no mention of the Biden chip ban that was meant to prevent this whole thing from happening either.

Pretty funny "clone".

-2 ( +2 / -4 )

They wiped a trillion dollars off on Monday. $600 billion from one entity. This is good warfare. No concrete rubble or lives lost.

The interesting thing the app is just a side project for the founder and its company. A side quest.

Given that Chat GPT is effectively founded on large-scale unauthorized use of intellectual property, it's a bit rich for them to complain about this.

Exactly. If you have ever used a template or referred to a previous year’s budget report, you are a thief.

DeepSeek will teach you how to make these AI models. Basically they provided the code and outputs so you can build on top of it however you want.

2 ( +4 / -2 )

Is Japan even trying to do AI?

Why should it? Is it a written law or duty to follow big players US and China on their wrong and misleading path? Some even call it an AI race. But a race would require a track and a goal line, of which both are not applicable here. There's no path to a sufficiently correct AI and such a goal is also impossible and not existing. No, let those two big ones run into nothing for another few months or so, just for fun. Everyone needs an own learning curve until recognizing that it is impossible, and their learning curve, maybe due to too big self-confidence and systemic rivalry, is a bit longer and flatter.

-1 ( +1 / -2 )

First, DeepSeek-V3 (https://github.com/deepseek-ai/DeepSeek-V3), includes two models DeepSeek-V3-Base and a chat version of DeepSeek-V3. Both are MoE with 671B total parameters and 37B active. Not for mere mortals, you need to have a strong multi-GPU configuration, something like 8 H200 (but there are compressed versions from different guys). In terms of quality, it is somewhere on the level of GPT-4o 0513 and Claude-3.5-Sonnet-1022 and higher LLaMA-3.1 405B.

There are different estimates of how much it cost to train Lama 3.1 405B. The paper itself (https://arxiv.org/abs/2407.21783) says that up to 16,384 H100 were used and mentions pre-training for 54 days (but there were other training stages as well). One of the not-so-high estimates says that it should have cost around $60M (https://x.com/_LouiePeters/status/1816443587053092917?lang=en).

DeepSeek-V3 is a bit more specific. They used the H800, a cut-down export version of the H100 for China, and they themselves write that it took 2,788M H800 GPU hours to fully train, which corresponds to $5,576M at an H800 rental price of $2 per hour.

Well, like a decimal order less. When compared to OpenAI, the difference is probably even greater.

It's like India, which sent spacecraft to Mars and the Moon for less than Hollywood makes space movies: Mars Mangalyaan for $74M and lunar Chandrayaan-3 for $75M versus the movie "Gravity" for $100M (https://www.business-standard.com/india-news/what-makes-india-s-space-missions-cost-less-than-hollywood-sci-fi-movies-124110400430_1.html).

Secondly, DeepSeek-R1 (https://github.com/deepseek-ai/DeepSeek-R1), models with rationalization like OpenAI o1 or Google Gemini Thinking. There are two models in the family: DeepSeek-R1-Zero and DeepSeek-R1, both built on the DeepSeek-V3-Base and the same large size.

DeepSeek-R1-Zero (similar to AlphaZero) is trained with pure RL (Group Relative Policy Optimization, GRPO — a variant of PPO from their other paper, https://arxiv.org/abs/2402.03300), without SFT. I think this is a very significant result, just as in Go it turned out that it is possible without human parties, so it gradually turns out here too. An interesting thing is that during training, the model had an “aha moment”, when in the chain of reasoning the model issued “Wait, wait. Wait. That’s an aha moment I can flag here.” and revised the original approach to solving the problem.

Zero is good, but sometimes it goes into repetitions, mixes languages, and is not very readable. DeepSeek-R1 was trained on a small (thousands) of CoT examples before RL, they call it Cold start data, to give a better starting point for RL. Then the same Reasoning-oriented RL as Zero. Next, SFT for reasoning (600k) and non-reasoning (200k) data. And then an additional RL stage. This model is comparable to OpenAI-o1-1217.

.

1 ( +1 / -0 )

The real issue is using an AI that spouts Chinese communist party dogma, ideals and perspectives, not to mention outright coverups or ones that dont.

It seems obvious that people will shy away from a platform that says eg, China is the best nation on earth and that the CCP is the best governing system on earth. Most intelligent people will shy away from using that, especially when they find out that profiling including voice profiles and keystroke retention is reportedly part of DeepSeek coding.

0 ( +2 / -2 )

I think there are several overlapping items:

OpenAI claims they have used ChatGPT output to train. Even if that is true, ChatGPT output is basically posted by thousands / millions of humans who use it. So that either way will happen someday.

Then again, OpenAI also just used everything it found on the internet, ignoring any legal responsibility.. like the article mentioned.

But more technically speaking, the interesting part of DeepSeek is that they can do (almost?) same as OpenAi's model with less compute. That is an achievement, because clearly OpenAI could not or they would just have wasted money with compute power. So regardless of what data has been used and what country that company/startup is from, just the fact that a startup can beat them here is astonishing

5 ( +6 / -1 )

ThroxToday 04:52 pm JST

My understanding is that DeepSeek lifted the model, the part that is generated after the training. How they got access to it, I don't know.

-3 ( +1 / -4 )

ADK99Today 10:17 am JST

Given that Chat GPT is effectively founded on large-scale unauthorized use of intellectual property, it's a bit rich for them to complain about this.

I'm certainly not a fan of OpenAI (or anyone else) using people's work without permission, but there is a difference here:

OpenAI: used pre-existing works to train AI. This was without precedent at the time, so not technically "wrong" as there were no laws to stop it

DeepSeek: explicitly (if the claims are true) distilled ChatGPT's model, in clear violation of the terms of service

Laws are thankfully changing to protect creators, now that the precedent has been set. But China will sadly still ignore them.

-4 ( +1 / -5 )

quercetumToday 01:59 pm JST

They wiped a trillion dollars off on Monday. $600 billion from one entity. This is good warfare. No concrete rubble or lives lost.

False dilemma. It's not a binary choice between physical or economic warfare. Neither are good.

It's certainly not "good" for DeepSeek to rip off OpenAI, when the terms of service specifically prevented what they apparently did (distilling).

Yet again, it's just more evidence of China ignoring laws and agreements, and another reminder of why the country cannot be trusted.

-4 ( +1 / -5 )

TaiwanIsNotChinaToday  04:57 pm JST

My understanding is that DeepSeek lifted the model, the part that is generated after the training. How they got access to it, I don't know.

I assume you are referring to the knowledge distillation. Simply saying, that means a larger network (teacher) teaches a smaller network (student) to do same things. It's quite common actually. So common, that even OpenAI probably does that similar as Meta. Let's assume the worst case that teacher was OpenAI's model and no other source, then still the technical capability to train a much smaller student is a breakthrough done by small startup. Because OpenAI did not achieve that (their model is larger and requires more compute power = money) and I am pretty sure they are reading the paper that DeepSeek published. In that hypothetical case, the main training (for the teacher) could be skipped and the managed with restricted GPU access to train their smaller model and the legal question of using OpenAI output becomes question, but technically still wow

3 ( +4 / -1 )

ThroxToday 06:36 pm JST

I assume you are referring to the knowledge distillation.

Distillation is forbidden by OpenAI's terms of service. If DeepSeek did that, it is violating those terms.

...

https://www.tomshardware.com/tech-industry/artificial-intelligence/microsoft-and-open-ai-investigate-whether-deepseek-illicitly-obtained-data-from-chatgpt

Microsoft's security team observed a group believed to have ties to DeepSeek extracting a large volume of data from OpenAI's API. The API allows developers to integrate OpenAI's proprietary models into their applications for a fee and retrieve some data. However, the excessive data retrieval noticed by Microsoft researchers violates OpenAI's terms and conditions and signals an attempt to bypass OpenAI's restrictions.

-3 ( +1 / -4 )

isabelleToday  07:20 pm JST

Distillation is forbidden by OpenAI's terms of service. If DeepSeek did that, it is violating those terms.

I am not arguing that this was legal or illegal, I am not educated enough in law to do this (although knowing about the debate of OpenAI not following other companies terms either makes me wonder... but that's a different topic). My point was purely technical speaking, it's an achievement which OpenAI themselves were not able to achieve (but now presumably can, because DeepSeek published paper on how)

So again, I'm curious what the investigations and legal discussions will show.

3 ( +4 / -1 )

Here we come with the anti-China hysteria :) 

There would be no doubt that “the usual suspects” would try to spread disinformation on the matter, void of any knowledge of the technology discussed and without a a shred of evidence. 

Let’s all accept that China is way superior, and the greed of the so called US engineer will be disciplined for letting believe (too many) millions of people that investments in AI would have to be expensive.

Not surprisingly on the same day DeepSeek reported many cyberattacks on their AI services with most of the attempts originated from IP addresses in the United States.

Rapacity made in Usa. Way to go China!

-1 ( +2 / -3 )

What the guy is doing is sharing with the tech community of the world. It’s a sharing of assets and literally communistic. This however, is taking away billions of dollars from the tech oligarchs.

Open up the closed world of AI foundational Ai model work.

Let’s see if they try and monetize it.

-3 ( +0 / -3 )

The real issue is using an AI that spouts Chinese communist party dogma

Another attempt to paint China as the villain and the US as a savior of the “rights of the world” First and foremost let’s not forget the US administration (previous and present) ‘dogma’: the NSA scandal, or Global surveillance program illegal disclosures since 2010: “detect and record everyone, even the top allies”. 

The massive extent of the US secret service spying, both foreign and domestic targets, from phone calls and Internet communications, the complete official deception of US Congress and the public. 

I don’t think we need to go further than that:

https://en.wikipedia.org/wiki/2010s_global_surveillance_disclosures

Most intelligent people will shy away from using that….

Hmm… not really! That is why DeepSeek is the most downloaded App by far. . Because ordinary intelligent people do not believe the manipulative narrative which comes from Washington, they would rather trust an unknown college kid in China than the Meta top brass tor the new clones of the US State department.

0 ( +3 / -3 )

I often wonder about certain American companies and their AI.

With Facebook I have given up reporting fake friend requests that lead to suggestive photographs and links to what looks like a porn site but could be a scam or phishing site. Meta's AI usually can find nothing wrong with those sites. It only takes a human with a little intelligence to see what is absolutely obvious. Meta seems almost as useless with fake accounts that have been cloned.

I assume Google whose translation and search services I often use has not worked out that I don't read Japanese as they keep showing me ads I cannot read. However, that may not be stupidity as I don't want to read their ads anyway and the advertisers still pay for presumably targeted ads. Also, most searches that require measurements such as recipes tend to place answers using American measurements even though most of the world and even most of the English-speaking world uses the metric system. Google does not seem to take into consideration my IP address which tells them where I am. Maybe like Meta Google could benefit from some Chinese AI.

It seems most AI is gleaned from human intelligence. AI may work with things like chess by studying the games of champions. AI seems to work in only certain limited fields. The results must depend a lot on what it is trained on and how it is trained.

I suspect many Americans will expect Chinese AI is trained to have a pro-Chinese bias or block anti-Chinese thoughts. I suspect American AI will also be trained to have a pro-US bias.

0 ( +2 / -2 )

Microsoft gives its blessing.

Fresh on the heels of a controversy in which ChatGPT-maker OpenAI accused the Chinese company behind DeepSeek R1 of using its AI model outputs against its terms of service, OpenAI's largest investor, Microsoft, announced on Wednesday that it will now host DeepSeek R1 on its Azure cloud service.

https://arstechnica.com/ai/2025/01/microsoft-embraces-openai-competitor-deepseek-on-its-ai-hosting-service/

1 ( +2 / -1 )

Kind of rich to see the American AI firms and their arrogant, bs throwing, full of themselves owners who have been copying everyone, infringing on copyrights and refusing to pay royalties to musicians when they use their music to now hear them whine like an worn out power steering pump about someone copying them. Serves them right if you ask me. Cosmic payback. I don't have an ounce of sympathy for them or the rubes who invested in their scam.

0 ( +0 / -0 )

Desert TortoiseToday 05:43 am JST

I don't have an ounce of sympathy for them or the rubes who invested in their scam.

Let's be honest: if you have money in the stockmarket you are invested in their scam at this point.

0 ( +0 / -0 )

Let's be honest: if you have money in the stockmarket you are invested in their scam at this point.

My money is in real estate (rental properties) and government bonds.

0 ( +0 / -0 )

Login to leave a comment

Facebook users

Use your Facebook account to login or register with JapanToday. By doing so, you will also receive an email inviting you to receive our news alerts.

Facebook Connect

Login with your JapanToday account

User registration

Articles, Offers & Useful Resources

A mix of what's trending on our other sites