AI News v8 2024 - Gemma, Sora, Redit, Gemini AI, Nvidia

Welcome to  AIAW News, a special segment of the Artificial Intelligence After Work (AIAW) Podcast, where we bring you the latest and most impactful developments in the world of AI. This week, we explore innovations from industry giants, revealing how AI is reshaping our digital experiences.

1. Google has recently launched Gemma, a new suite of open models aimed at promoting responsible AI development. These models, inspired by Google's Gemini project, are designed to be lightweight and accessible, coming in two sizes: Gemma 2B and Gemma 7B. They are optimized for various AI hardware platforms, including NVIDIA GPUs and Google Cloud TPUs, and are compatible with major frameworks like JAX, PyTorch, and TensorFlow. Google has emphasized responsible AI use by providing tools for safety and model debugging, alongside extensive safety measures during the model's development. Gemma is available globally for a wide range of AI applications.

2. Google has paused its Gemini AI image generator project due to inaccuracies. Gemini, a multimodal large language model developed by Google’s AI research labs DeepMind and Google Research, comes in three variants: Ultra, Pro, and Nano. However, there have been concerns regarding its performance, especially in the wake of its underwhelming initial launch and a promotional video that turned out to be heavily doctored. Despite claims of Gemini outperforming OpenAI's GPT-4 on various benchmarks, there have been instances where Gemini struggled with basic facts, translations, and coding suggestions. This has led to skepticism about its capabilities and effectiveness.

3. Reddit has reportedly entered into a content licensing deal with Google, involving the use of AI technology. Specific details about the deal, such as the nature of the content and the terms, are currently not disclosed. This partnership could potentially involve the use of Google's AI tools and technologies for content creation or moderation on Reddit, leveraging the capabilities of Google's AI systems like the recently introduced Gemini. Such collaborations typically aim to enhance user experience and content management through advanced AI solutions.

4. Andrey Karpathy, a prominent AI developer, has left OpenAI for the second time. Karpathy's departure marks a significant event in the AI community, given his contributions and stature in the field. The reasons behind his departure and his future plans are not specified in the available information. OpenAI, known for its groundbreaking work in AI and machine learning, has been at the forefront of AI research and development, with Karpathy playing a pivotal role in some of its major projects.

5. NVIDIA has announced the enhancement of Large Language Model (LLM) applications on Windows PCs through the integration of NVIDIA RTX systems. This development aims to supercharge AI applications, providing more power and efficiency for complex AI tasks. NVIDIA's involvement in AI has been growing, with the company offering hardware and software solutions that cater to various AI demands. This particular move indicates NVIDIA's commitment to making AI more accessible and powerful for a broader range of users and developers.

6. OpenAI has introduced Sora, a new generative video model. Sora represents a significant step forward in AI technology, showcasing the ability to generate video content. The model is currently being tested by a select group of safety testers, with wider availability expected in the future. OpenAI's continuous innovation in the field of AI demonstrates its commitment to advancing the capabilities of AI technology and exploring new frontiers in content creation and generative models.

Follow us on youtube: https://www.youtube.com/@aiawpodcast

Transcript

Goran Cvetanovski: 0:04

It's time for AI News brought to you by AI AW podcast.

Henrik Göthberg: 0:11

Last, you know, last season, everything was spinning, you know, a new announcement and new large language model every week. So we said, okay, let's see what is the go-slip with the latest news for this week. Excellent, we all have like one or two picks each, and then we try to do it like speed day, you know speed like the news right, so do you want to start?

Anders Arpteg: 0:38

Anders, on the same day as we had the last podcast, you know last Thursday. You know a number of news stories happened and I guess we it feels like so old news now, so it's almost embarrassing to talk about. Sora is one, gemini 1.5 is another, the Jepa is the third. All of them on.

Henrik Göthberg: 0:59

I think they're all worthy of a couple of minute snips from you. Anders, do you want to take one of them? We want to add one. Yeah, which one?

Erich Hugo: 1:08

That company in Gothenburg that has just let all their creatives go and going over to AI. So a web design company or no video game company in Gothenburg last week let everybody go or not most of their creatives and said we're going to do this with AI because it's better for share all the value.

Goran Cvetanovski: 1:28

So that is the local news.

Erich Hugo: 1:32

It's going to have global impact.

Anders Arpteg: 1:35

That's a good one. That could be a follow up perhaps to the impact of having these kind of developments happening in the world.

Erich Hugo: 1:41

It's really happening right now.

Henrik Göthberg: 1:43

Let's park that back to deeper topic.

Erich Hugo: 1:46

But one of Eric's news was did you hear about the Well, for me it's basically I got involved more in the new menstrual release. I was very impressed with it and I'm so frustrated with them because they cannot. I mean Europeans suck at making consumer products and if you think about how Windows 95 conquered Africa, they didn't go after, they went after making consumers use them. If you want your AI to succeed, get consumers to use it. They will take it into the businesses for you, and I wish that menstrual or some European AI company will bloody make a consumer product that people will be so excited about as open.

Henrik Göthberg: 2:33

So the core topic here is like it's not as easy as chart GPD to use as a new.

Erich Hugo: 2:38

It's not at all. Not at all and it's not as fun.

Henrik Göthberg: 2:44

But back to your stories, then, andes.

Anders Arpteg: 2:47

Which do you?

Henrik Göthberg: 2:48

prefer. I think you need to make a snippet on all of them and then you can choose one a bit more, but I think they're all worthy of snippets actually.

Anders Arpteg: 2:57

Yeah, even though, speaking about the week old news, is feeling really bad. Okay, but let's do, sora then yeah, let's talk.

Henrik Göthberg: 3:04

We cannot not talk about Sora.

Erich Hugo: 3:05

Have you seen it? We have played around with it.

Anders Arpteg: 3:08

It's not available yet.

Erich Hugo: 3:09

No, no no, but the guy sent me videos from it and stuff like that. It was just amazing.

Anders Arpteg: 3:14

So, in short, it's a new text to video model that was released by OpenAI and on the same day as Google made their big release, but it seems like they are competing here, trying to make a new story happening on the same day just to out-compete the other. Anyway, it's amazingly good. It can generate videos up to like minutes with coherence to the prompt, and the quality of the video is much better than anything else, and it's really hard even to see that it is generated. It really looks like more or less photography or a proper shooting that has occurred?

Henrik Göthberg: 3:50

Yeah, did you hear the gossip that they think it has a lot to do with the Unreal engine?

Anders Arpteg: 3:58

They have released a technical paper and they literally state in it it is not using a physical engine and I do believe them, so I don't believe it, but I do believe they use a physical engine like Unreal to produce training data for it.

Erich Hugo: 4:13

It's amazing. Can I just add something? Do you guys follow the TV program now on SVTcom called Svensk historia? So did you see the last episode?

Henrik Göthberg: 4:23

No, I didn't see the last episode.

Erich Hugo: 4:24

So in the last episode they actually started because they're telling the history of Sweden and because there's been such a massive evolution of tools that create moving videos. They actually created AI videos in Swenskistoria telling the story of Sweden. Ai videos and I mean that quality sucked compared to Sura. But do you guys realize the impact this is going to have on the creative industry? I mean, if they can make an engine right now that does three minutes of high quality video, disney is fucked. I mean, anybody will be able to tell a story. You can take a Robert Heinlein science fiction novel, feed it to an AI and it will make you a movie. How epic is that.

Anders Arpteg: 5:01

It's crazy, and it's just the start of it. Yeah that is the story, right, exactly.

Henrik Göthberg: 5:05

So you know but you can also argue they're way ahead of anyone else.

Anders Arpteg: 5:11

Or yes, but there are competitors like Runway ML, et cetera. They are behind and of course now OpenAI completely shuts down their whole business model. So they will have a hard time to compete with OpenAI going forward. But just to get a bit technical, perhaps because Jan Likkun you know he was a bit, I think, upset because they released the BJEPPA and they did it afterwards and he said the Sura approach was a complete dead end and I think this is actually the mistake of Jan Likkun. So I disagree with Jan Likkun here. She's not often.

Anders Arpteg: 5:47

It's funny, you're a fanboy, I know, but in this case he said you know they are operating in pixel space and if you actually read the technical report, they are not. So they do have an encoder that moves from pixel space to an internal latent space and they do the generation of the video inside that laven space of a space time frame. So the patches that they take from the images in a video is both on the spatial grid, so to speak, x and Y, but also on the time dimension, and that is the patches they remove and the model has to train to create and in that way, by moving and doing the inference and prediction and training inside the latent space, they are actually doing exactly what BJEPPA is doing. And I think this is really interesting and I think this is one of the true reasons for them being able to do this at a level that no one else has done before.

Henrik Göthberg: 6:43

And is that the segue to BJEPPA?

Anders Arpteg: 6:47

Yes, if we want to do it quickly, yes.

Goran Cvetanovski: 6:48

But they can do a lot of stuff.

Anders Arpteg: 6:50

You know they can't only do text to video. They can also take another video and continue it. They can merge or morph between two videos. The social implications of this is massive I want to see in order to see some people doing proper videos, sorry movies using Sora.

Erich Hugo: 7:08

I mean you're gonna have a kid up in North Luleå making his first movie. I mean it's great, I love it.

Anders Arpteg: 7:15

I think it's. I also like that they do take it serious from an ethical point of view. So now it's not released in the total public, it's released to red teams, red teaming people that are looking into the potential ethical impacts.

Erich Hugo: 7:26

It's gonna be abused. It's going to be abused. That's just the way it is.

Anders Arpteg: 7:30

But they try to at least minimize it. Yeah, and that's at least good. They try to right.

Erich Hugo: 7:34

They'll grok. It's a hand on it.

Anders Arpteg: 7:36

Maybe, maybe. Yeah, that would be fun. But anyway, they try to minimize the ethical consequences Because I think we all realize the potential abuse that this could have.

Erich Hugo: 7:46

Well, we know it, because in chat GPT right now, when you ask the question, I'm not allowed to or I'm not. I'm a language model and I can't answer that question. No, you're an American server boy. You don't want to answer that question?

Anders Arpteg: 8:00

Yeah, well, they do try to the best at their abilities to add God rails to it and they are you know, at least if you believe them trying to do that as good as they can.

Anders Arpteg: 8:11

They won't, you know, be 100 percent, of course, but at least they're trying to be ethical here. Ok, that's. The Jeff is basically the same story. It's a continuation of the Jeff paper. That generally could put it out like a couple of years ago in 2022, I believe. And then they had the IJEPA for images and our video, the JEPA, and once again they move from the pixel space into a Latin space. They do the inference there, very similar to Sora, and then they can move back.

Henrik Göthberg: 8:42

So what is the VJEPA? Is it an article where they detail quite deeply how to do it? What is the bottom line here?

Anders Arpteg: 8:50

So this is META, this is Jan Likun, and they put out everything. They have a proper research paper actually explaining how it works and they even have, I believe, the source code animals. I think they are released, they will at least. So they are being much more open with the method, the science, open with the model, open with the code Brilliant.

Erich Hugo: 9:13

And that is really good. Even though I don't like META, I love that.

Henrik Göthberg: 9:16

Yeah, and we will come. This is sort of opening up.

Anders Arpteg: 9:19

Yeah, and they will release it under the common creative license, so free to use for whatever reason. Wonderful.

Henrik Göthberg: 9:25

Can I now take the next news as a segue? Because at the same time this week, or even I think it was yesterday, google released GEMMA, gemma, have you heard about GEMMA?

Goran Cvetanovski: 9:37

No.

Henrik Göthberg: 9:38

So basically, this is GEMMA, this is what they then Google called Introducing New State of the Art Open Models. And now it gets tricky, because then this is sort of Google's play to be more open, right, but now we come into the next and we can think about is this good or how will it work? I mean, like for me the news is not so much only about what GEMMA can do and all that, but it's actually that actually right now there is a they're starting to become a spectrum on what is open source versus what is called open models, which is actually not, and this is not open. It's like open weights Marketing. So basically, in one way, you can, you can, you can fiddle with the weights in the model, but you but you don't really get to the real code or the real stuff. So I think the interesting stuff here is actually At least they're going a bit more open compared to GEMMA.

Anders Arpteg: 10:36

This is more open than GEMMA. Yes, at least it's a good direction.

Henrik Göthberg: 10:39

It's a good direction.

Erich Hugo: 10:40

But I but, and I appreciate your optimism and and and your empathy in terms of that. The fact of the matter is is that we are at an inflection point now. Now, in the next two to three years, if we don't decide the future of AI with an, if the European Union can stop being so reactive and rather be more proactive, ai is going to go to the power of a few shareholders.

Anders Arpteg: 11:05

No question, we've been speaking about the AI divide for so long and I think this acceleration between the divide of the few selected companies, the few superscalers that we have, is just continuing to accelerate, and I'm the first one to say this is not a good direction.

Henrik Göthberg: 11:19

We were the ones. We actually I don't know, but they've been talking about a discussion about the digital divide, right.

Erich Hugo: 11:28

Yeah, but before you go there, I've got one new snippet that I want to add. Yeah, we finished with that.

Anders Arpteg: 11:33

Oh, maybe I think GEMMA you know, or GEMMA you know. When I say GEMMA, I think about the son of Anarchy you know the mother of that quote GEMMA. So anyway, yeah, okay, I think my reflection was interesting.

Henrik Göthberg: 11:47

They go in this direction. Beware that is a distinction between open source and open models and open weights. So it's interesting. Yeah, I didn't want to linger on that more than that.

Erich Hugo: 12:00

But if you think about it, if you do a search quickly on, did you guys see what Nvidia released yesterday? Nvidia just released a desktop LLM.

Henrik Göthberg: 12:10

Yeah, the desktop LLM.

Anders Arpteg: 12:11

Yes, the Ortex thing, yeah, but that was like a week ago as well.

Erich Hugo: 12:15

Okay, sorry it's going so fast. Well, I got up and I thought that this is just epic. You know, this actually really gets so many more people playing with it. You know this is bringing LLMs to the masses, you know, in a way that none of the other programs are doing it.

Anders Arpteg: 12:33

Still logging it into Nvidia shapes, but still yeah.

Erich Hugo: 12:36

True, but again I'm going to use your argument. They're trying.

Anders Arpteg: 12:41

You know they're moving in the right direction. Potentially.

Erich Hugo: 12:43

But this is what people what I don't think that people understand this is a commercial gamer move, not gamer as in games, but this is a commercial decision that's going to determine the future of Nvidia, because they want people to get into this space in the year and they want to develop year. They don't want people to go to other things because they've got plans. Nvidia has got plans, every company's got plans and they're fighting for the users now and they should right.

Anders Arpteg: 13:12

Yes, they are forced to.

Henrik Göthberg: 13:15

So how do you see this game? What are they getting at? They want you to basically build your own LLM PC or use to your they want people to budge on their computers at home.

Erich Hugo: 13:25

Okay, and exactly the way Windows 95 was released in the world in 95. Windows basically made an operating system ubiquitous so that all of a sudden, everybody started using that and they started building other things on it. I mean, they built, they made. Windows became available and all of a sudden Excel became the tool of choice for people to actually manage their companies. You know, and still, after 40 years, every single CFO in the company uses Excel. This is Nvidia's play to get people to budge on the desktop with their tools, so that in 40 years, people use Nvidia tools to.

Henrik Göthberg: 14:03

So this is Now. This is at our eyes. Right in front of us, we have a Windows 95 move made by Nvidia in 2021.

Erich Hugo: 14:12

Business moves don't change.

Anders Arpteg: 14:14

But the Nvidia has been doing this for a very long time. The whole CUDA environment that is powering the Nvidia shape has been always in this way. This is how they lock people into the Not only hardware, but the software.

Henrik Göthberg: 14:26

This is another software play to make the hardware, but isn't it the consumer play that is interesting here is in exactly the same way.

Anders Arpteg: 14:32

I don't think it's changing at all. It's just making sure that media can use their hardware in an even broader way, do you remember?

Erich Hugo: 14:38

Okay, when did you get involved in the IT world? When did you start?

Henrik Göthberg: 14:44

For me, yeah, a little bit more in around the 99. Okay, so in 1980.

Erich Hugo: 14:51

Well, you look so young, thank you. So I remember in the 90s when people asked you so what do you work with? And you said I work with computers. They win. We got you. It was such an inclusive term of everything. It didn't matter if you were a programmer or if you were techy fixing anything like that, it was the biggest term of saying what you did I work with computers. And that included the internet, included mainframe computing, which is now called cloud, and so on and so on. And why was it easy for me? It was easy for me. I mean. Sure I did add a bit of, let's say, windows 3.1 experience before, but Windows 95 just made it so easy, you know. And all of a sudden I could build a website. And I built the e-commerce website in 1995 in a small town, in Bloomfontein, you know, because it was easy, because I had this platform to make it happen for me. And this is Nvidia's play. Now. It's easy. All of a sudden you can have a budget.

Anders Arpteg: 15:51

And another thing is Are you happy for that or do you think it's a bad move?

Erich Hugo: 15:55

No, I want more people to get involved because I believe educating people is the key to saving humanity. Educate people as much as you can and even if you have to use bad tools sometimes, like Windows 95, it got so many people. And then Sweden made a great. When I came to Sweden they had attacks and rebate for people getting a PC at home. I don't know if you guys remember that MPS, yeah, and most companies that deals with Dell but what they managed to do is that they managed to get a PC into basically every single home in Sweden. The difference that that made in Swedish society is Canning it, and this cannot be underrated.

Anders Arpteg: 16:35

I'm still very annoyed when I have to buy a laptop and I pay the Microsoft tax for it. Basically, I'm forced to buy the Windows operating system.

Erich Hugo: 16:42

But you never use it. But don't you buy a Linux computer?

Anders Arpteg: 16:46

I do. This is a Linux computer, but I have to pay Windows license for it anyway, unless you.

Erich Hugo: 16:51

Well, every iPhone that you pay for, you get Give some money to Nokia and to Ericsson. So, yeah, yeah, yeah, I mean there's some.

Henrik Göthberg: 16:59

Let's wrap up the news section. Did we have? That was the last second news Go around. Do you want to add any news?

Goran Cvetanovski: 17:07

Yes, just two of them that you missed actually.

Anders Arpteg: 17:11

It was a crazy week last week. It was, but this is super new.

Goran Cvetanovski: 17:17

I don't get why it is like this, but it doesn't matter. So Google had to pause the Gemini AI image generator because, basically it started, it refused to make pictures of white people, so all the historical what is called images of white people were replaced with native Indians and a colored one. It went woke on them, so now they're pausing this this is actually very fresh. It's coming out from today, so it's very good. The second thing is that but that's what Hollywood does.

Erich Hugo: 17:49

So what's the big deal? Now? Ai is so. Organic intelligence is allowed to do it, but artificial intelligence is not allowed.

Anders Arpteg: 17:55

Exactly.

Goran Cvetanovski: 17:57

We should not treat AI the same. And then this was very interesting for me, so really actually says has come into agreement with Google to sell their data to Google so they can Google constrain their models on Reddit, this game, this has been going on for some time now, this is sick. Yeah, this is sick and yeah let's see how Reddit is gonna do after this.

Henrik Göthberg: 18:20

It's interesting.

Erich Hugo: 18:21

They're selling their data, which is actually our data of course, and like I say that they don't care, that let somebody.

Goran Cvetanovski: 18:30

This is the most important line for me, right when people are talking about like well, you cannot put value on data. Well, actually, you have seen the contract with all of it owner, owner. Google my money. It worth 60 million 60 million dollars per year. Okay, according to one of the sources, that is how much the data is good or bad.

Erich Hugo: 18:51

I think you can see some positives with this as well. I am Super optimist.

Anders Arpteg: 18:55

No, but otherwise in a reddit would be dead. I think in this way it's actually way, that's the invisible hand.

Erich Hugo: 19:01

I mean, then reddit has to die and the new service has to come up.

Anders Arpteg: 19:05

Yeah, but that's the new service would be Google and one of the tech giants.

Erich Hugo: 19:08

They can feel that so.

Anders Arpteg: 19:09

That's why it's. It's kind of dangerous unless you can actually have the kind of old content producers out there to be alive, which old content produces reddit in this case.

Erich Hugo: 19:22

Okay. So all this newish content producers, but otherwise you know they would be dead.

Anders Arpteg: 19:27

I mean because otherwise people would, just you know, be asking chat to BT or Gemini questions.

Erich Hugo: 19:32

Why? Why Can't read it? Just say okay, guys, we're making money out of you and here's a credit credit at Reddit, you know, and here's a, here's a store, because you've served credit doesn't make money.

Anders Arpteg: 19:45

They're losing money. Now you know they're losing users, but now they're gonna make money out of this.

Goran Cvetanovski: 19:50

Sorry, now they're gonna make money out of this.

Anders Arpteg: 19:51

Yes, that's a good thing, right.

Goran Cvetanovski: 19:53

Yes.

Henrik Göthberg: 19:53

I'm saying good, but they're ultimately making money on me.

Anders Arpteg: 19:57

Of course, and the alternative to just have Googles and the superscalers of the world to make the money would be even worse.

Erich Hugo: 20:03

No, but I think that that's how the invisible and has to work. That's why Google reacted so fast when open AI launched is because, all of a sudden, here's somebody that's eating their pie, and the nice thing about the world that we live in right now is new companies. Is not that difficult to start with a great idea. It's not that difficult to actually start eating somebody else's pie.

Anders Arpteg: 20:21

I think it is. The divide is actually increasing. I think it will be increasingly hard. So unless we can actually have some value on the data and the superscalers have to pay for it, we will see a very quick death of a lot of companies, and I think this is the quick birth of companies.

Anders Arpteg: 20:37

I do believe that, but unless you know who can really build a shat model of the future, it's very few companies that can. This is a big problem of the AI divide going forward, so I think that's why we will see an increased concentration to power, of power to the superscalers, unless we can find a way to spread around. I think this is actually a move that at least rights to. Why is Google?

Erich Hugo: 21:01

doing this.

Anders Arpteg: 21:02

Why are they doing this? Because they want to have the data. Why?

Erich Hugo: 21:05

do they want the data to train a model? Why do they want to train the model?

Anders Arpteg: 21:09

No, come on Follow me, of course, to make money from their shat models.

Erich Hugo: 21:14

Exactly, and they have got a corporate mission that says we have to dominate the global scale.

Anders Arpteg: 21:20

But you can't really think that the alternative of not paying Reddit is better.

Erich Hugo: 21:26

No, I'm saying that. What about paying Hendrik for being on Reddit?

Anders Arpteg: 21:30

Good luck doing that, but Reddit would die even quicker. Do you remember? Do you see? Reddit would die even quicker than that.

Erich Hugo: 21:36

But then that's the invisible hand, I mean, then somebody has to come up with new ideas, somebody has to budge for a new idea to make money.

Anders Arpteg: 21:47

But unless you have the proper conditions to do so, we will see an increased concentration of power going very quickly, I think, in the Western world.

Erich Hugo: 21:56

yes, but I think that not in the developing world, not in the global south, I don't think so. I think already there are tools that are. Like I told you about the mobile money A bunch of Zanzibarian fishermen, basically were the first users who can build a foundational model in the future. But they didn't care about building a foundational model, because they had a function that they wanted. They wanted to basically sell their fish.

Henrik Göthberg: 22:19

But let's step out of this conversation because it's literally. What we are seeing is then, in one way, we are seeing a concentration of power which is dangerous, and maybe we need to fix some of this in the middle, and maybe if. Reddit is surviving a little bit longer. It's good, but there is also a bigger picture here that we are really seeing two worlds.

Erich Hugo: 22:42

It's a part of it.