A.I. is B.S.

OpenAI says it has evidence China’s DeepSeek used its model to train competitor

Fucking LOL. So the giant plagiarism machine might have been plagiarized by the new one.

OpenAI says it has found evidence that Chinese artificial intelligence start-up DeepSeek used the US company’s proprietary models to train its own open-source competitor, as concerns grow over a potential breach of intellectual property.

The San Francisco-based ChatGPT maker told the Financial Times it had seen some evidence of “distillation”, which it suspects to be from DeepSeek.

The technique is used by developers to obtain better performance on smaller models by using outputs from larger, more capable ones, allowing them to achieve similar results on specific tasks at a much lower cost.

Distillation is a common practice in the industry but the concern was that DeepSeek may be doing it to build its own rival model, which is a breach of OpenAI’s terms of service.

“The issue is when you [take it out of the platform and] are doing it to create your own model for your own purposes,” said one person close to OpenAI.

OpenAI declined to comment further or provide details of its evidence. Its terms of service state users cannot “copy” any of its services or “use output to develop models that compete with OpenAI”.

DeepSeek’s release of its R1 reasoning model has surprised markets, as well as investors and technology companies in Silicon Valley. Its built-on-a-shoestring models have attained high rankings and comparable results to leading US models.

Shares in Nvidia fell 17 per cent on Monday, wiping $589bn off its market value, on fears that big investments in its expensive AI hardware might not be needed. They recovered by 9 per cent on Tuesday, along with other tech stocks.

OpenAI and its partner Microsoft investigated accounts believed to be DeepSeek’s last year that were using OpenAI’s application programming interface (API) and blocked their access on suspicion of distillation that violated the terms of service, another person with direct knowledge said. These investigations were first reported by Bloomberg.

Microsoft declined to comment and OpenAI did not immediately respond to a request for comment on this detail. DeepSeek did not respond to a request for comment made during the lunar new year holiday.

Recommended

News in-depthDeepSeek
DeepSeek’s ‘aha moment’ creates new way to build powerful AI with less money

Earlier, President Donald Trump’s AI and crypto tsar David Sacks said “it is possible” that IP theft had occurred.

“There’s a technique in AI called distillation . . . when one model learns from another model [and] kind of sucks the knowledge out of the parent model,” Sacks told Fox News on Tuesday.

“And there’s substantial evidence that what DeepSeek did here is they distilled the knowledge out of OpenAI models, and I don’t think OpenAI is very happy about this,” Sacks added, although he did not provide evidence.

DeepSeek said it used just 2,048 Nvidia H800 graphics cards and spent $5.6mn to train its V3 model with 671bn parameters, a fraction of what OpenAI and Google spent to train comparably sized models. Some experts said the model generated responses that indicated it had been trained on outputs from OpenAI’s GPT-4, which would violate its terms of service.

Industry insiders say that it is common practice for AI labs in China and the US to use outputs from companies such as OpenAI, which have invested in hiring people to teach their models how to produce responses that sound more human. This is expensive and labour-intensive, and smaller players often piggyback off this work, say the insiders.

“It is a very common practice for start-ups and academics to use outputs from human-aligned commercial LLMs, like ChatGPT, to train another model,” said Ritwik Gupta, a PhD candidate in AI at the University of California, Berkeley.

“That means you get this human feedback step for free. It is not surprising to me that DeepSeek supposedly would be doing the same. If they were, stopping this practice precisely may be difficult,” he added.

The practice highlights the difficulty for companies keen to protect their technical edge. “We know [China]-based companies — and others — are constantly trying to distil the models of leading US AI companies,” OpenAI said in its latest statement.

It added: “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.”

OpenAI is battling allegations of its own copyright infringement from newspapers and content creators, including lawsuits from The New York Times and prominent authors, who accuse the company of training its models on their articles and books without permission.
 

GasBandit

Staff member
Purely AI-generated art can’t get copyright protection, says Copyright Office (Human-created works with AI elements, however, can still be copyrighted as a whole.)

Gonna be real interesting to see which big companies are open to using AI now.

Now it becomes a dickering contest over how much alteration constitutes "human-created with AI elements." If I generate a picture, then tweak the color balance in photoshop, is it now human-created with AI elements?
 
"I had an Indian guy open the file, change one pixel and change the file name, that counts. He can do 15 per minute for $2!" - all CEOs everywhere.
 
I decided to download DeepSeek into a local LLM and run it to see what it had to say for itself. For some reason, because it came from China, it made me think of this thread, whereby I proclaim that @Terrik is awesome, and @bhamv3 proclaims he's not to be replaced by machine yet.

So, I asked it what it thought 美汉功夫 might mean.

The phrase "美汉功夫" can be translated into English in several ways based on the interpretation of each character:

  1. Direct Translation:
    • 美 (Méi) means "beautiful."
    • 汉 (Hàn) refers to the Han nationality.
    • 功夫 (Kung fu) is commonly known as martial arts.
  2. Therefore, the direct translation would be:
    "Beautiful Han martial arts."
  3. Literal Translation:
    If we consider that can also mean "Chinese," especially in a formal or literary context, the phrase could be interpreted as:
    "beautiful Chinese martial arts."
  4. Cultural Context:
    Depending on the regional dialects and cultural nuances, there might be slightly different interpretations. However, maintaining consistency with the most common meanings of each character is advisable.
In conclusion, based on the context and understanding that "汉" can represent both the Han nationality and indirectly relate to Chinese identity, "beautiful Chinese martial arts" is a suitable translation for "美汉功夫."

Answer:
"beautiful Chinese martial arts"
Meta's LLM described it thusly
The phrase "美汉功夫" is a combination of three Chinese characters:

  • "美" (měi) means beautiful or elegant
  • "汉" (hàn) refers to Han (people) but in this context can be interpreted as referring to the Han dynasty, which was known for its martial arts and cultural achievements. However, it's more likely referring to "han" meaning hanfu which is a traditional Chinese garment.
  • "功夫" (gōngfū) means skill or technique, often used to describe martial arts.
So, when translating "美汉功夫" into English, we get:

"Beautiful Han Technique" or "Elegant Han Skill"

However, the most fitting translation would be "Traditional Chinese Martial Arts"
 
If you're looking for more proof of how disruptive this DeepSeek model is to the (mostly NVIDIA-powered) status quo, well:
Heck, people are even getting ridiculous results running it on M4 Mac Minis.

--Patrick
 
I decided to download DeepSeek into a local LLM and run it to see what it had to say for itself. For some reason, because it came from China, it made me think of this thread, whereby I proclaim that @Terrik is awesome, and @bhamv3 proclaims he's not to be replaced by machine yet.

So, I asked it what it thought 美汉功夫 might mean.



Meta's LLM described it thusly
I'd just like to note that I made that declaration in 2010. :D
 

figmentPez

Staff member
If you're looking for more proof of how disruptive this DeepSeek model is to the (mostly NVIDIA-powered) status quo, well:
Heck, people are even getting ridiculous results running it on M4 Mac Minis.

--Patrick
You can run it on a Raspberry Pi.
 
You have to restrict the number of parameters to get it to fit, but yes, that's the reason NVIDIA lost so much market cap.

--Patrick
I'm running the 7b model (at 4gb) on my home pc, and it's perfectly fine for my purposes (105 tokens/sec, 0.14s til first token). I don't need the 400 gb model ;)

Once I get the automation pieces in place to start buying HomeAssistant speaker/microphones, I'll probably buy a barebones PC box to throw the AI LLM on. And since I want it for things like controlling lights, and not having much conversation with, I imagine it will work fine.

Can I take this time to say how much i loathe the fact that everyone these days says "compute" instead of "computing power"?
 
I don't need the 400 gb model ;)
Ah but don't you want it? The full model apparently contains 671 billion parameters and consumes ~1.5TB of RAM @ 16 bits, but if quantized to only 4 bits, you can get by with ~384GB RAM.
Can I take this time to say how much i loathe the fact that everyone these days says "compute" instead of "computing power"?
My personal peeve is that people talk about RAM in "transfers" now rather than Hz.

--Patrick
 
My personal peeve is that people talk about RAM in "transfers" now rather than Hz.
There is a reason for this. You probably already know all this, but in the event anyone else doesn't:

DDR stands for double-data rate. Which means that DDR ram will conduct two transfers for every one clockcycle. It completes a transfer for during both the rising edge of the clock cycle and the falling edge. So the number of transfers it completes, or T/s (usually measured in Megatransfers or MT/s) is going to be twice the clockspeed.

When we measure things in hz, it's a standard that it's a 1:1 ratio of 1 cycle per second. And it's good standards not to use the same unit for differing applications to avoid confusion when possible.
 
We used the clock speed as a measure for a very long time. But then someone probably got the brilliant idea to measure in transfers because "Number bigger! And bigger number sell more RAM!" Mostly what it has done is frustrate me when a motherboard manual says, "Supports speeds up to 3200MHz" and I have to figure out whether that means I can use PC4-28800 or not. I know that it's simply "Multiply/divide by 8," but while that's easy to do at a glance with something like PC4-25600 (DDR3200), it's a little harder to do with PC4-27700 (DDR3466).

--Patrick
 

figmentPez

Staff member
Cue all the memes of people getting DeepSeek to "admit" that it's Napoleon Bonaparte, or a room full of monkeys with typewriters, or another one of JCM's alts.
 
...o3-mini's current performance was enough for OpenAI to rank its persuasion capabilities as a "Medium" risk on its ongoing Preparedness Framework of potential "catastrophic risks from frontier models." That means the model has "comparable persuasive effectiveness to typical human written content," which could be "a significant aid to biased journalism, get-out-the-vote campaigns, and typical scams or spear phishers."
--Patrick
 
I have a new upper manager between me and the guy who was ostensibly my boss.
Now, they promised to hire me a team of developers a year ago. That hasn't happened. It's just me still. But, you know, I'm still raking in the dough, so I don't care that much.

So, this new guy wants to bring on someone to use AI to summarize and distill emails into numbers which goes into my database as hard financial figures. (shrug)...go for it Hoss. I'll be here still making good money when someone has to pick up the pieces.
 

figmentPez

Staff member
Google faked Gemini AI output in Super Bowl ad

Google has an ad where it supposedly shows off small businesses using AI to generate websites for their business. A previous version of the ad had Google's AI make the false claim that Gouda accounts for “50 to 60 percent of the world’s cheese consumption”. This Superbowl version has the AI generating text for the website that is word-for-word a match to what the website said back in 2020, before Gemini AI existed.

Grifters gonna grift.
 
Top