Watch the full, wide-ranging interview with OpenAI CTO Mira Murati here: on.wsj.com/3TNv2Mu

The CTO can't talk about where the training data came from...

"I can neither confirm nor deny that we trained on youtube videos"

The fact that she dodged that question about the training data isn't very "Open"AI of her.

It's like asking a farmer: - Where do you take your apples from? Do you pick them from trees? - I dont know. If they grow on trees, perhaps yes.

4:50 She's smart enough to know where the models came from and she's hesitant to say because of potential lawsuits.

The vagueness of her response about what was used from the CTO shows she definitely knows they’re using copyrighted content imo.

"My prompt for this crab said nothing about a mustache" is the most 2024 sentence ever.

You should’ve asked Sora to make 10 people playing rock paper scissors in fast pace. 😂

I´m so glad this interviewer asked some real questions, Im sick of all the shallow questions over naive smiles. AI can be good, but also dangerous. It´s important we understand how serious it is, and train and use it carefully.

Remember the Super Phrase: *Publicly available data*

I'm not going to lie, I spent half the video with my eyes darting to Mira's hands just to make sure she's real 😅

The way she dodged over and over the question about the training data, clearly their are not using only publicly available data.

I cannot express how wild it seems to me that the *CTO* of one of the biggest genAI companies, promoting their latest genAI product couldn't answer what data the model was trained on.

this is a product i made that i am very proud of. i'm the spokesperson of the product and was sent here to speak specificity about the product only. "how did i make my product", you ask? i don't know.

the questions asked in this interview were really good -- concise and hitting all of the important points. thank you for this informative video

Joanna Stern is the best! Love how she grilled her about the training data - super awkward response 😂

As Sora can create "cartoon animation" OpenAI should show their list of animation production companies who have licensed their films to be used in the machine learning model. There's not much animation is available in public domain, especially 3D animation.

What’s the point of replacing creativity we don’t want to live in a world where nothing is creative

The CTO isn't sure if youtube training data was used???

not sure of sources??

In my 40s now, looking back, I am amazed at the acceleration of technological innovation. It's now it's almost impossible to imagine what computing will be like in just 2-3 years.

ClosedAI

ouch those data training questions hit HARD!

That interview took a quick turn! Hehehe

Really good reporting the interviewer asked the correct questions and Mira definitely knows the data they used she does not want to reveal anything because they’ll get crucified

They always say AI will change us in good ways and we will be much better but as you know it’s actually gonna go the other way

Yikes that was one bomb of an interview… they are definitely training it on KZfaq/x/fb/ig everything 🤣

Really good interview. Grilled her with sensible questions

basically they “openly” “take” whatever there are available online to train the ai

Great interview, thanks. I'm liking the timeframe mentioned....

Uh oh where that data coming from 😂

will they be using this video as a training to generate a video when the prompt is: did you pay for the data you used?

4:34 GOLDEN reaction

Hmmmn. The training data issue is going to haunt OpenAI for a while. The law has not caught up with the technology. US Law makers are going to have a tough time with this one.

Lol, wow, that interview turned really sketchy @4:24... 😆

I feel like a CTO should know where the data came from lol

Let me get this straight: they've created a model that generates videos indistinguishable from reality, and their genius idea to make that distinction is: a watermark! Wow!

A powerful Ai tool like Sora, when used with intricate technical prompts, could potentially be susceptible to certain vulnerabilities. Will they be able to deal with this ?

I think it's definitely worth trying" ie the profit motive is too great not to.

“It’s very difficult to simulate the motion of hands.” Every VFX studio from the last 25 years: “Do I mean nothing to you?”

Good for the interviewer pushing for the source of the data. That was disappointing that the interviewee didn’t seem to know where the data comes from.

@3:32 Joanna's smug smile here has me cracking up. 😆

I stole a car because i saw it on the street and i thought that was public property so by this logic isn't your car publically available for me to use for my own benefit.

Lawsuits incoming

thanks, good job there.

Reading between the lines and tbe expressions of all the top executives they seem to be saying people are going to be able to do whatever they want

in a distant future I can imagine cops having to shout "hands! show me your hands! Hands in the air!" for more than one reason

I can confirm that the models were trained on all data, publicly available, licensed and non-licensed etc.

Where did they get this data to train on? Must be fine. /s

This was some incredible reporting. Well done to her;

Not going into the details of what trained sora because NYT is currently suing...among others.

Why was she avoiding answering the source data? They absolutely know since there is only two major parts: the code, and the learning source. She avoided answering about the Shutter deal as well, even though they had a deal with them and later in the interview confirms that. This doesn't look good.

Q: "Videos on youtube?" CTO: "I'm actually not sure about that."

I can literally take watermarked screenshots and get rid of them within 5 minutes because of AI (Which they probably do when they scrape the web for training data).

Thank you for asking about the data! This is what people care about: whether their data was ingested to train profitable AI without our consent.

Some of my art was used for training AI - I didn't give permission. I'm not sure where they got them from.

It’s a good thing they’ve somewhat broken down to bits and pcs the ‘creativity’ of the AI. Different AI’s for different applications. At the end of the day, it’s business as usual.

is it going to be released this month or next month. pika lab and runway they are doing better job and already released 1 year

This is iPad 12.9 or 11?

The data fell from one of KZfaq's trucks

These are great questions! It’s a bit a shame that interviewer almost didn’t follow up on responses. For instance follow up on “ public” data can be compliance with usage conditions. Follow up on “how distinguish” AI generated content might be question how it’s already done with released products.

SORA exhibits/solicits a human-like inattentional blindness: it produces fluent motion that seems normal at a glance. But upon closer inspection it does stuff that’s as weird as a gorilla walking across the frame (from Chabris and Simon’s famous experiments.)

Two smart women having a hard-headed conversation about an extraordinarily important topic. Nice vid.

So now, "publicly available data" does it mean "public domain"?

Now its just waiting for the wistleblower 😂🤭

Interesting conversation start at 4:42

The interviewer forgot to ask if Sora can replace her 😂

We are basically taking everythin we can get out hand

Nice, reliable humans making reliable decisions with far reaching consequences. This is how we got here .. and we shall further!

4:34 People are gonna make fun of that answer. However, clearly, the purpose is to dodge, and that answer achieved that.

Its an open secret OpenAI used a lot of pirated data from "The Eye" for ChatGPT, so i would assume they used a lot of video and movies not youtube.

When I was ababy my mom she used take me to work with her there I like to cut the nees pepper pecuter and make story usually I put news pepper in my lap and look at them and used the reazer to cut the pecsuer and books some times but i don't cut them but I use to put magenation on it until one time I see my second older he gets first class in sewming pol which I don't know he swemer .. thank you for all

“I’m not sure…” if there are dangers using this tool. Regardless, creativity is paramount. For instance, what new ways can someone gaslight others?

She knows that videos on KZfaq were used as training data. She is not a good liar lol it’s her job to know. OpenAI doesn’t want to get sued, that’s why. Sam Altman likes to move fast. They have the LLM models but they don’t have access to enough data. So they take it without paying for it.

CTO: «Sora is based on text prompts». NO! It is NOT. It is based on stolen art. The biggest art robbery in history! And it gets even worse. She pretends (!) that she doesn't know whether they stole from KZfaq creators (which they def did). She talks about "licensed data", as if OpenAI had any official license from the Sponge Bob right holders to feed their wicked machine with their property. I'm so disgusted! She says "publicly available" as if it meant "public domain". Any Disney DVD is pubicly available. Don't mean I can use that data as I please. A CTO that is "not sure" where they stole the data from. Give me a break. Furthermore, she keeps saying their in the very early stages, but keeps emphasizing that it'll be ready in a few months with mosts kinks ironed out. Which one is it now? WSJ: why did you not call this women out?

I bet you 100% from this interview onwards OPENAI will have a SPECIALIZED TEAM dedicated ONLY TO TRAIN OPENAI on HOW TO RESPOND TO THAT VERY QUESTION regarding TRAINING DATA - They will be like a pro at it lol

Looking forward to Stable Diffusion eventually doing this. I'd rather not be bound by censorship and would prefer to create locally without subscription and with open models.

Al + Sora = Job Killer

Everything will have to be “taken with a grain of salt” we will have to question everything and trust nothing in the future.

She knows, just like everyone else, where the data came from. Being CTO is just hard.

People are really trying to find faults in it to cope but honestly the flaws are so minor that you have to notice it extensively (except the hands part) and the model is still in its beta phase so this is literally a game changer

"Reliable" as the "Open"AI guys, reliable as a politician, reliable as the most of human egos... Profit, that's the word.

The CTO mentioned about SORA bringing creativity. I struggle to see where is the creativity beyond getting the right prompt to generate the video you want.

I'm excited for the development of tools like Sora, and I think all the skeptics in the film industry will eventually come around when they see how they can use such tools to improve their work flow. Much of the pushback we're seeing is simply cynicism about human nature, uncertainty about the future, fear of disruption, and ugly old Luddism. As for the people who think using publicly available videos to train AI models isn't fair use -- if everyone thought like that, the world would never get any better. I understand her reluctance to go into those issues, though, given all the old fuddy-duddies who think they can put things out into the world and not have AI models train from them. There will likely be legal wrangling over this stuff for several years to come, but it won't hold back the technology. Lower the barriers to entry, democratize creativity, and unleash human potential. That's what this technology is about. It will be interesting to watch how the process unfolds, and to participate in it.

Perfect timing release of the best fake video AI maker just in time for the biggest election year in the World

Wow, Sora's capabilities are mind-blowing! 😮

she certainly did not expect the question bout the source of the training data..... I wish we had journalists like this in South Africa

Of course they've used KZfaq and every other platform where you can upload videos. KZfaq videos are licensed and protected by copyright. Content creators typically retain the copyright to their videos unless they have explicitly licensed them under Creative Commons or other open licenses.

The data was shady :D

Lawsuits incoming!

At 5 minutes the Chief Technology Officer of a multi-billion dollar "non-profit" corporation says she's not sure if their flagship Video GPT was trained on KZfaq videos? Seriously? Sounds like something SBF would say. Is she hallucinating? We've been struggling creating prompts that'll prevent GPT-4 from making things up. Now, knowing its parents, I see why it lies so much.

I guess its hard for many to realize AI is like "gather" information from the internet. It is the same way we do as humans. We go to work and talk about "there was this one video I seen...." The internet lawyers want to get all copyright \ DMCA huffy and puffy about it all. But reality is, if its on the internet people will talk about it and possibly use what they seen in their life. More countries will expect content owners to state "do not use this without permission" than countries that automatically apply these rules. It makes more sense to watermark and state "do not use this without permission" and good luck wining ANYTHING after spending hundreds of thousands in legal expenses without these things. I advocate for content owners to specifically state "do not use this without permission" and possible watermarks or else its all fair use. As again, people will talk about xyz and possibly react \ reuse what they see on the internet. Digital IP technically needs to go through a process to be legal IP. This idea of "automatically" protected just because you clicked submit or upload is an entitlement way of thinking.

"Im not going to go into the details of the data that was used" = We stole most of the data and wont admit it

the bull is very sweet

3:56 - Scary, if someone was to show me this video of these 2 ladies I would have thought it's just a real video footage without noticing the glitches of their hands ETC, even when I already know SORA

It’s very interesting video. Thank you for making it. I did notice something disturbing. When you asked about the origin of the data used to teach SORA, I noticed something. It was quite obvious that she was evasive and then seemingly outright lying about the data used to train SORA. I am very aware of the dynamic occurring right now in the industry that is the lack of data. AI will need more and more data to feed into its learning models, but it’s also clear that up until now they have just been accessing the data without permission and breaking copyright law. They thought no one would notice or care, but it has become an issue and clearly she is not being forthright about it. Clearly, they are using everything they get their hands on and ignoring copyright law. That did seem like the Ferdinand and it seem like the crab was a SpongeBob character. I personally think it’s very disturbing that they can crassly lie about what they’ve done here. If they’re willing to lie about where they’re getting the data and their use of copyrighted data then they are also willing to compromise much more in safety. They’re driven by profit and I really didn’t like seeing her. Lie about this. Thanks again for the video.

This is like when people were pointing out all the flaws in the original iPhone.

I really want to know what kind of answer to the training data question would satisfy people asking it. Give an example of an answer, and remember that you 1) Don't agree with the copyright assertions that say you have no right to use the data for training 2) Have expensive lawyers that tell you to be mostly mum or lack specificity on this matter 3) Actually try to empathize with her position and the fact a canned "good" answer would likely be just as suspicious to people who are in opposition to your views as a "bad" answer that leads them to assuming "the worst" (in their opinion).

This video is generated by AI?? just asking, both look beautiful