Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

OpenAI Launches o3-mini, a Cost-Efficient Reasoning Model Rivaling DeepSeek

“For STEM”… can it do math? It was my understanding these are predictors not calculators.
 
Nothing is for free.
If it's free you need to ask yourself what the real cost is? Am I the product?
You'd think after decades of Google, Facebook, etc profiling you until they know you better than you know yourself and then selling you to their highest bidder the public would get wise to "free".
So a Chinese company comes out with a LLM that's on par with the best in the West and offers it for free - because they are generous people by nature?
No. They want to get to know you.
I installed DeepSeek and it wants to know my name and email address. It's got my IP address so knows my location.
For me, and I re-sell ChatGPT integrated in to my software, it's exactly like the current ChatGPT 3 in how it answers and even how it formats those answers.
The API isn't much different from OpenAIs either.
It's all too similar. I think there's more to DeepSeek that we've still to find out.
 
  • Like
  • Love
Reactions: gusmula, cyb3rdud3, KeithBN and 1 other person
Don't understand why the user is supposed to care about this? They (we) type in a prompt and expect an answer. I don't really care how it does it or how efficient it is. It's like getting an explanation how a website works and why it loads faster now every time there is a TYPO3 or WordPress update.

Just release it ...
 
For people with very tight available time like me, I missed one crucial bit in the article: which one is more powerful for coding, o1 or o3-mini? It says the later is faster, but faster tends to mean less powerful. If it’s faster and more powerful, it’s not clear where o1 sits.
 
turbineseaplane said:
DeepSeek is OSS and can be run locally and it will gladly answer all the things its creators don't want it talking about
Click to expand...


BS. Deepseek censors when running locally too.

Tbf Llama can also censor, but its censorship is due to the model’s incompetency and hallucinations and not by design.

Wiping historical events from history makes a model an extremely bad Trojan horse.
 
  • Like
Reactions: D_J and joshmcx
Zwhaler said:
That's just the web version. The full open source model doesn't censor.
Click to expand...

Two errors.

The locally run model also censors. I just did a run again now to confirm.

Second error, the model weights are open. The code is not open source.

Neither Meta nor Deepseek open source their code.
 
  • Like
Reactions: D_J and joshmcx
ThomasJL said:
When I say that ChatGPT plagarizes authors, I mean that they take other people's work that has been made available on the Internet, and then plagarize it by rewording it just enough to circumvent copyright laws.
Click to expand...

That’s not the way machine learning or the transformer model works. The models do not contain any works, only the training data does.

You can ask the best models to output text in the style of Bukowski or Hemingway and it still won’t read like Bukowski or Hemingway. A bad impersonation at best, and a bad impersonation is not plagiarism.

The models are trained on labelled text to learn how to predict sequences of words and patterns.

All they need is copious amounts of text. You could have a dataset without any books in them. The dataset could just be Reddit posts or Macrumors posts. The models will still learn the same linguistic patterns and still be able to output fictional content such a novels.

That doesn’t mean they will be good of course. Only tech nerds and dimwits look at the novels produced with ChatGPT or Llama and think it is good writing.
 
  • Like
Reactions: joshmcx
Arsenikdote said:
That also means that you don't use it correct? Because if you use it, even without paying, you are still allowing them to earn revenue from you, and while you are not writing a check to them, you still provide them revenue and thereby support them plagiarizing authors. I am not making an accusation, I just want to make sure you understand that by using the platform you ARE contributing to the issue...if you use it, which I am not saying you do.

Also, when you say plagiarizing authors, are you meaning everyone who "contributes" or specifically authors with materials under a real copyright?
Click to expand...
Copyright is automatic for prose, there is no ”real copyright”. And shifting the blame isn’t the way to go here.
 
  • Like
Reactions: freedomlinux and KeithBN
ArPe said:
That’s not the way machine learning or the transformer model works. The models do not contain any works, only the training data does.

You can ask the best models to output text in the style of Bukowski or Hemingway and it still won’t read like Bukowski or Hemingway. A bad impersonation at best, and a bad impersonation is not plagiarism.

The models are trained on labelled text to learn how to predict sequences of words and patterns.

All they need is copious amounts of text. You could have a dataset without any books in them. The dataset could just be Reddit posts or Macrumors posts. The models will still learn the same linguistic patterns and still be able to output fictional content such a novels.

That doesn’t mean they will be good of course. Only tech nerds and dimwits look at the novels produced with ChatGPT or Llama and think it is good writing.
Click to expand...
Transformer models are trained to predict next tokens in a text. Optimally, they'd always guess right and if they always guessed right, they'd have de-facto memorized the training data. The larger the models get, the more they will be enabled to do that. You can right now, even with very humbly-sized models, ask to quote the first sentences in prominent books and they'll succeed with a high likelihood. So yes, in part, they absolutely do contain works, some more than others. They don't just passively consume books with little retention (like maybe we'd imagine a human would), they actively attempt to memorize everything and compress it as much as they can. The intelligence is in the act of compression.
Your point about "copious amounts of text" is misleading. The quality of the training data matters. A model trained purely on reddit post will likely be incredibly stupid. There's a reason why openAI and everyone else includes a massive corpus of copyright-protected contents in their training. It's the most valuable part.

Point being, I don't think that using copyright-protected material as training data under fair use is an easy thing to argue. The only reason why nobody wants to tackle the issue is because everybody is panicking that if you don't allow it, others will, and then you're screwed.
 
  • Like
Reactions: Grey Area, haruhiko and KeithBN
On the bright side, the font size in the chat bar in the latest update to the ChatGPT app on macOS is finally large enough to be readable.
 
ThomasJL said:
I would never pay OpenAI because that would be allowing them to profit by plagarizing authors who OpenAI neither credits nor financially compensates.

In other words, I don't pay thieves for stolen property. Thieves like Sam Altman should be in jail for theft, not financially rewarded for theft.
Click to expand...

I suspect DeepSeek did the same thing, directly or indirectly; if they had been paying authors that would have been news.

I tried DeepSeek for some technical questions and it mixed nonsense with correct information.
 
CapitalIdea said:
DeepSeek does the same thing for 97% less, so unless o1 is 34x better, it’s a worse value.
Click to expand...
No. It doesn’t. DeepSeek is not multimodal. DeepSeek cannot search the web. Its answers are nowhere near as complete as o3. ChatGPT does not censor its answers to comply with the political mandates of the Chinese Communist Party.
 
  • Like
Reactions: LeeW
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