Revolutionizing Intelligence: The Unveiling of Meta's CM3leon AI

0:00What can 160 years of experience teach you about the future? When it comes to protecting what matters, Pacific Life provides life insurance, retirement income, and employee benefits for people and businesses building a more confident tomorrow. Strategies rooted in strength and backed by experience. Ask a financial professional how Pacific Life can help you today. Pacific Life Insurance Company, Omaha, Nebraska, and in New York. Pacific Life and Annuity, Phoenix, Arizona. If you ask anyone on the street what the most popular, what the most well-known AI models are, you're going to hear things like ChatGPT, maybe Bard, you hear things like MidJourney.

0:38And I think that one area, one company in particular that is not getting a lot of attention is Meta. Meta is a absolute pioneer in the AI space. They've been doing this for a long time. Today on the podcast, we're going to be talking about a major new announcement from Meta, a major new AI model they've just announced that does text and image generation. and we're going to talk about what its implications are, what its innovations are, and how we think this is going to change things up. So without further ado, let's jump into it. I think traditionally there's been a division between AI image generators and AI text generators.

1:10You know, we have OpenAI's ChatGPT, which does text, and merging these two capabilities has been a little bit challenging. So despite OpenAI's release of the multi-modular GPT-4 earlier this year um it still is something that the industry has struggled with at large so meta's new it's called cm3 eleon like like c melon um and that's their new model essentially aims to bridge this gap by enabling the output and generation of text and images the tools allows for the creation of captions also known as image to text generation as well as high resolution images um i mean this It's kind of funny because they're obviously met as the owner of Facebook and Instagram.

1:54So technically on this model, you could get it to generate an Instagram image and the caption in one thing, which is kind of funny. But I think a lot of existing AI image generators on the market, like Stable Diffusion and Opening Eyes Dolly, employ diffusion models to generate images. So this process involves adding and removing Gaussian noise from training data. but in contrast meta's approach is completely different it actually involves supervised fine-tuning which is a technique used to train text-based transform of models using a data set and they've actually done this using a data set of license images and captions from shutterstock so the reason they're able to do the captions for the images is because when all the metadata that they got from shutterstock who they you know purchase the the license to use their data set every image has a caption, so that's kind of interesting.

2:43And this allows for the interpretation of really complex text and objects, which results in better alignment with user inputs. So Meta's researchers emphasized the importance of supervised fine-tuning in training large language models like Chai Chibuti. However, its application in multi-modular settings has been largely unexplored till now. So the result is a text-to-image generation process that produces quote-unquote more coherent imagery that better follows the input prompts that's according to meta so i think to kind of illustrate this capability the company shared examples of highly compositional images generated by cmelon based on prompts like a small cactus wearing a straw hat and neon sunglasses in the hara desert which to be honest i saw the image that was generated by that and it was like it was accurate right it did it did the thing but it was it felt a little bit like dolly too.

3:37It was like it wasn't like it wasn't something like you get out of mid journey. And so I think a problem we might see right now is that with some of these image generators specifically, they're trained on data sets that have, you know, they've bought the rights to the data set. They trained on the data set. Something like mid journey was trained on everything. And so because it has been trained on literally everything, probably data it wasn't supposed to. It has just such a, the results are so much better. And so sometimes I feel like it might feel like we're taking a little step backwards when we're getting some of these new models, but essentially these new models are going to be essential for, you know, enterprise customers because enterprises are not going to be able to use something that comes out of mid journey because of copyright issues and other things like that.

4:24So I think, you know, we might feel like we're taking a step backwards as these models get tuned and built better and kind of go through the process. But eventually, I believe that they'll still be able to achieve the same results as MidJourney. And eventually, we'll have completely copyright free images where you own the copyright when you generate it. And it's just as good as MidJourney and better. And it can do anything you need. I think what's really interesting is that this new model by Facebook or by Meta was actually able to generate a really realistic, detailed human hand, which is, of course a problem that these AIs always have is generating hands.

5:01Cmelon also excels in text generated image editing so it's actually surpassing previous models like InstructPix2Pix. The tool's ability to really recognize both contextual and visual content allows for what I would say is definitely more accurate and precise editing based on text prompts. So something else that's I think really interesting meta recently showcased a feature by demonstrating kind of the text guided image editing using vimear's remount or like vimear has a painting that's called girl with pearl earrings and they use that as the initial input and then they uh started kind of generating images in response to text prompts like you know put a pair of sunglasses and you know she should look like 100 years old and so they were able to actually manipulate the picture itself which was really cool.

5:50In addition to its fairly impressive capabilities, Cmelon stands out for its significantly reduced compute power requirements. This is something I've been saying this since the beginning. This is something that Meta absolutely excels at. You can say what you want about the fact that they haven't released any consumer products yet or that they're open sourcing a lot of stuff, but Meta consistently puts out these AI models that are using way less compute power. And I think this is a big step. Like when you do an open source project, you don't want to open source something as powerful as, you know, GPT-4 that took three months to train the thing on, you know, crazy supercomputers and uses a lot of compute.

6:32You want this thing to be able to run on the most efficient resources possible. And so I think that's something that Meta does really, really good. I think that this achievement holds the potential to promote greater equity in the AI space, an issue that critics have long highlighted. C-Mellon operates on five times less computing power than comparable models. So Meta has actually received a lot of praise for this and also for fully licensing its data set from Shutterstock, which demonstrates that exceptional performance can still be achieved using a distinct data distribution from what previous models have relied on.

7:09previous models like mid-journey are just scraping every image they can get. So I believe by promoting transparency and kind of sharing their work, Meta's aiming to kind of get this whole collaboration and innovation thing going. They're really, I think over the last few months, they've been introducing a lot of different generative AI features across their platform. And so these include AI-generated stickers on Messenger, an AI sandbox for Facebook's advertisers. They have an AI powered video generation system. So with Cmelon kind of getting unveiled, I think Meta has showcased its commitment to really just pushing the boundaries of AI and tech, unlocking a lot of new possibilities for creation of text and images, while also addressing critical challenges in the field.

7:54Love them or hate them. I really do believe that Meta has done a good job with the way they are open sourcing things, the way they're going about that, coming up with a lot of really innovative tech. And I really just want them to kind of jump in the field and come out with some consumer products. People have been hinting that that is coming very soon. So that's going to be a really interesting area to follow because if meta jumps in the race, we have a meta, Google, OpenAI slash Microsoft, we have some very, very big players. All we'd be waiting for at that point is kind of an Amazon or Apple Play to kind of come in.

8:26And now we have a lot of really powerful tech competing against each other. And I think that would be incredible for the ecosystem, moving beyond essentially just fragmenting as much as possible, commoditizing AI as much as possible. I think amazing things will happen for consumers and we'll get a lot of really great options.