DeepSeek is tremendously good news for Apple, as it implies significantly more powerful AI tools will be able to run locally on device, particularly given Apple Silicon's unified memory.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
OpenAI Alleges DeepSeek Used Its Models for AI Training
- Thread starter MacRumors
- Start date
- Sort by reaction score
That is because it was trained on some synthetic data generated by ChatGPT. They didn't steal openAI proprietary tech, they just used chatgpt to generate synthetic data they used to train their model which is against OpenAI's terms of service. Same way, It is against the terms of service of most websites to not use their data to train AI models but openAI does it anyway. Their video models are trained on data they stole from youtube creators too. Without data, these models cannot be trained.
LOL. Half of this thread is people whining about the fact that AI companies use their data that they put on the web without permission. The other half is happy that an AI companies data that they put on the web gets stolen too.
I have thousands of pages of data on the web and I'm sure they got used for AI training too. It's unfair but I don't know what the big deal is as these data comprise a miniscule and likely irrelevant and likely redundant part of the value of an AI company. I put the data out there for everyone to see and use. So why would I be mad that somebody is using it? If the only value of the data is to get people to visit my website than my website and service seem to not have much inherent value. It's not that AI companies take the data as is and sell it. They use the data to train a model. The data was out there for people to use it.
Whatever the reasoning is, the ship has sailed 30 years ago when everybody thought it was ok to steal music and jpeg and movies from the web. The stealing of data from the web is now not only done by individuals, but by small and large businesses and its accepted by governments and societies. Will be almost impossible to change this worldwide.
I have thousands of pages of data on the web and I'm sure they got used for AI training too. It's unfair but I don't know what the big deal is as these data comprise a miniscule and likely irrelevant and likely redundant part of the value of an AI company. I put the data out there for everyone to see and use. So why would I be mad that somebody is using it? If the only value of the data is to get people to visit my website than my website and service seem to not have much inherent value. It's not that AI companies take the data as is and sell it. They use the data to train a model. The data was out there for people to use it.
Whatever the reasoning is, the ship has sailed 30 years ago when everybody thought it was ok to steal music and jpeg and movies from the web. The stealing of data from the web is now not only done by individuals, but by small and large businesses and its accepted by governments and societies. Will be almost impossible to change this worldwide.
This most likely has to do with fact they trained on GPT2 model which was open sourced anyways. Most of the LLm models use GPT2 or modified GPT2 architecture. The data from open AI was used for distillation.
Thank you for that video. It was a good jumping off point for my (layman’s) further research.
No, we should not be naive, but profit and greed are also not a working long term approach for a society. We should want to change that approach.
Exactly. The Tech. Bros. steal whatever they want and act amazed that it's going on with their 'work'.
This reminds me of how, in the '00s, Facebook hypocritically threatened legal action against foreign copycat social networking websites when Facebook itself was a copycat of the social networking website HarvardConnection, which was invented by Divya Narendra, Cameron Winklevoss, and Tyler Winklevoss.
Probably true, but... I mean... it's kinda funny hearing OpenAI complain about having its content stolen. Pot meet kettle.
Also for Meta, and other open source LLM models. Bad for Open AI, Microsoft, and new AI czar. A gut check for Nvidia hype.
I think the statement isn't accurate.
If they used "distillation"; which is using the output of ChatGPT or Llama or any other AI to train DeepSeek, then the resources they claim for training and the cost is a complete lie.
The aggregated cot is whatever was used to train the model they got output from + whatever development they had to pay for.
You can't claim it only took x amount of time and resources to train a model when you used the other model to feed yours.
I'm not saying right or wrong to use the other model to train, but don't claim you had this ground breaking efficiency in resources when you didn't.
Exactly
It's not the stealing, it's claiming it's your own work and then claiming you were able to do it more efficiently than others.
Did OpenAI receive permission to unilaterally scrape content from content creators ?
After benefitting from the billions and billions of work hours that went into creating the data they trained their models on for the cost of downloading it, they have the audacity to actually take their own patently absurd ToS seriously. I hope a court correctly invalidates those ToS, because they are literally impossible to meet. The entire internet is overflowing with chatGPT output, which can't be watermarked and is practically impossible to filter out. Nobody can train a model any longer without stumbling over chatGPT-generated text eventually, so yes, literally every model has some sense of chatGPT-style responses as the standard AI assistant tone. It's actually pretty hard to shake it off at this point.
They used chatGPT (or whatever other model) to generate data for fine-tuning, which is basically standard practice. Everybody does it, yet nobody manages to be as efficient as Deepseek. Also, the fine-tuning really doesn't make a big difference in performance, the magic sauce is usually in the pre-training, that's where the model gets smart. There is no incorrect accounting here.
Great World War 4 is going to be the war of Large Language Models, because anyone calling this AI is clearly misinterpreting what this is.
When will people realise, modern age Chinese aren't innovators, they blatantly copy (poorly) and steal everything
OpenAI says it has uncovered evidence that Chinese AI startup DeepSeek used its proprietary models to train a competing open-source model, potentially violating the company's terms of service.
The discovery centers around a technique called "distillation," where developers use outputs from larger AI models to train smaller ones. The practice is common in AI development, but OpenAI claims DeepSeek crossed a line by using it to build a rival model.
"The issue is when you take it out of the platform and are doing it to create your own model for your own purposes," a source close to OpenAI told the Financial Times.
DeepSeek's R1 reasoning model has attracted widespread attention in the tech industry for achieving comparable results to leading US models at a remarkably low cost. The company claims it spent just $5.6 million on development, which is a fraction of what companies like OpenAI and Google typically invest. The app this week reached the number one position on Apple's App Store free charts in multiple countries, including the US.
Asked about OpenAI's allegations in an interview with Fox News, White House AI czar David Sacks didn't mince his words.
"There's substantial evidence that what DeepSeek did here is they distilled knowledge out of OpenAI models, and I don't think OpenAI is very happy about this," he said.
The controversy has already had market implications. Nvidia saw its shares drop 17% on Monday, wiping a one-day record $589 billion off its market value, as investors questioned whether expensive AI hardware investments might be unnecessary if companies can achieve similar results with fewer resources.
According to Bloomberg, OpenAI and Microsoft reportedly investigated and blocked accounts in August for suspected terms of service violations, and they now believe these accounts were associated with DeepSeek. Both companies have declined to provide specific details about their evidence.
Article Link: OpenAI Alleges DeepSeek Used Its Models for AI Training
