Open AI copying the entire internet for their own goals, are angry that the Chinese use Open AI's model to train theirs.
Yes. That's what I call a hypocrite.
Yes. That's what I call a hypocrite.
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
Since when do two wrongs make a right?
Turning OpenAIs AI into open AI... brilliant! 🤣🤣🤣 Made me lough way too hard... the irony hits so hard...
Here is former Open AI scientist who was one among early founders/employees. Nvidia isn’t doomed but all those projections of needing hundreds of billions of infrastructure look way overhyped. We are in punch card era of AI, may be floppy. Long way to go in AI race.
Sounds like sour grapes from losers and hypocrites who’ve plundered regardless of I.P. Besides, who tells ‘the’ truth any longer?
OpenAI: We'll scrape the Internet without paying to build our model.
Also OpenAI: Deepseek scraped our model. That's crossing a line!
Also OpenAI: Deepseek scraped our model. That's crossing a line!
I didn't download it, but I checked it out on a coworkers phone. It gets right a question about IV medication administration that ChatGPT gets wrong.. 🤷🏻♂️
Last edited:
Tech Bro hypocrites who’ll no doubt be supported by liar in chief.
You can load several versions of it in Private LLM. The Qwen based model was clearly inferior to ChatGPT. The Llama 70B based model was pretty good on my tests. It gave some interesting results on par with o1, though o1 was still a hair better in phrasing. It gave some interestingly bad results too. Whether the results are good or bad, the text it generates while pondering its answer is entertaining.
Here's the worst result DeepSeek-R1-Distill-Llama-70B gave me.
Q: define innuendo
A: “Innuendo” appears to be a misspelling or typo of the Spanish word “yendo,” which is the present participle of the verb “ir” (to go). In context, it would be correctly used as “yendo,” as in “Vámonos yendo” (We’re going).
Last edited:
Were you asking it about Tiananmen Square by chance? It definitely doesn’t want to answer any questions about that.
Whilst we’re quoting cliches, "What’s good for the goose, is good for the gander"
I've also seen cases in which DeepSeek pretended to be Claude, so I guess they didn't limit themselves to models from OpenAI. Anyway, anybody surprised by that? Chinese do what Chinese do.
Well, the great american AI peddlers have trained on copyrighted data without permission. Now they complain when someone pirates from them?
Tired cliches won’t wash any longer, especially on this one.
How can they copy Claude or open AI. Heck they didn’t even use labeled training data. It’s sour grapes from Open AI and Microsoft. What they are alleging is deep seek queried Open AI models for some training data. This opens a big legal problem for open AI, they can’t accuse some one of ripping the very thing others did.
I don't see that DeekSeek did anything wrong.
I haven't used any AIs aside from what's built into the iPhone. But if it can't handle basic word definitions, how can I trust it for more complex things? It seems to me that any answer would need to be verified by other means.
This is my biggest take away, DeepSeek figured out how to do it cheaper and more efficient.
I expect other AI companies to revise their billions of dollars worth of infrastructure plans for the future.
Microsoft is one of the biggest losers. They were getting stake in AI companies with cloud GPU credits. Microsoft didn’t give money to Open AI, they gave the cloud credits. You don’t need to spend billions on hardware.
There’s non so blind as those who don’t want to see.
It has done much better at math and coding, very helpful to use it locally with visual studio code. Much better than meta Llama models which are open source and were trained with lot more compute and cost.
This is just the beginning, huge barrier to entry in costs for training is starting to crumble.
We don't know if their cost claim is accurate. It could be, but has to be independently verified before believing the claim. We know the model works well because it is independently verified but the cost has not been.
Also, it's possible it costs less because they might have taken a shortcut and built from OpenAI's work. Meaning, if they started their model from scratch more like what OpenAI did, it likely would have cost much more. This gives them the benefit of OpenAI's expenses and work without having to recreate it. I'm personally okay with that, just as I'm okay with OpenAI using all sources they did to train their models. I'm just stating that the direct costs are only part of the story.
Edit: Some discussion and estimates here that suggest the stated cost is likely accurate but only part of the total cost.
Last edited:
