Toggle sidebar Toggle sidebar
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

TylerL said:
OpenAI is accusing someone of taking their proprietary AI, and turning it into an…open AI.

They’re going to twist themselves into knots explaining why it’s ok for them to scrape everyone’s copyrighted data, but NOT ok for someone to scrape theirs.
Click to expand...
Turning OpenAIs AI into open AI... brilliant! 🤣🤣🤣 Made me lough way too hard... the irony hits so hard...
 
  • Like
  • Angry
Reactions: navaira, nt5672, jhfenton and 5 others
IIGS User said:
I downloaded the app and I was using it a bit last night. I went to fiddle with it this morning and it got brain freeze. Couldn’t even answer basic questions.

I smell a lot of hype. The short sellers probably made giant bags of money the past few days. They’ll throw those winnings back into the market and ride the rebound.

I think it’s a flash in the pan, much the same as AI in general. Until it’s smarter than the cat, it’s just a bunch of hype.
Click to expand...

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:
  • Like
  • Wow
Reactions: Beautyspin and Chuckeee
IIGS User said:
I downloaded the app and I was using it a bit last night. I went to fiddle with it this morning and it got brain freeze. Couldn’t even answer basic questions.

I smell a lot of hype. The short sellers probably made giant bags of money the past few days. They’ll throw those winnings back into the market and ride the rebound.

I think it’s a flash in the pan, much the same as AI in general. Until it’s smarter than the cat, it’s just a bunch of hype.
Click to expand...

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:
  • Haha
  • Like
Reactions: navaira, Luke MacWalker, z4co and 2 others
IIGS User said:
I downloaded the app and I was using it a bit last night. I went to fiddle with it this morning and it got brain freeze. Couldn’t even answer basic questions.

I smell a lot of hype. The short sellers probably made giant bags of money the past few days. They’ll throw those winnings back into the market and ride the rebound.

I think it’s a flash in the pan, much the same as AI in general. Until it’s smarter than the cat, it’s just a bunch of hype.
Click to expand...
Were you asking it about Tiananmen Square by chance? It definitely doesn’t want to answer any questions about that.
 
  • Like
Reactions: Luke MacWalker, Astrohunter, Razorpit and 4 others
MLVC said:
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.
Click to expand...
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.
 
  • Like
Reactions: Chuckeee and Banan911
vantelimus said:
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).
Click to expand...
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.
 
  • Like
Reactions: kamyk35, Razorpit, grover5 and 1 other person
TechnoMonk said:
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.

https://twitter.com/x/status/1872362712958906460
Click to expand...
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.
 
  • Like
  • Disagree
Reactions: canadianreader, Chuckeee, big_papa_johnny and 3 others
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.
 
Crowbot said:
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.
Click to expand...
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.
 
  • Like
Reactions: revfife and neuropsychguy
TechnoMonk said:
Actually what deep seek did was far from copy paste. They leveraged RL to cut through labeling and intensive training. In fact, one of they key founding scientist of Open AI called it beginning of a new direction.
Open AI was touting how others can’t train without huge compute advantage and resources. Deep seek showed, you need few million to train instead of billions.
Click to expand...
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.
https://www.reddit.com/r/OpenAI/comments/1ibw1za/comment/m9lnp6e
 
Last edited:
  • Like
Reactions: mode11, MacDarren, Mescagnus and 1 other person
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back