Meta is introducing a new Meta AI website and it could bury Microsoft and Google’s AI dreams
Meta AI is getting another upgrade as Meta tries to gain dominance in the generative AI market against OpenAI and Google.
The announcement came from Meta CEO Mark Zuckerberg himself Wireswhere he praised some statistics and details about the new Llama 3 model used to support Meta AI, which will be integrated into several of its products such as Messenger and Instagram, and which will also have its own dedicated site.
“We are upgrading Meta AI with our new, state-of-the-art Llama 3 AI model, which we are making open source,” Zuckerberg writes. “With this new model, we believe Meta AI is now the most intelligent AI assistant you can use freely.”
The new model will be accessible via various apps on desktop and mobile, including search functionality in Messenger, Meta AI in your Facebook feed and more.
The new Meta AI sitewhich has an interface remarkably familiar to OpenAI’s ChatGPT, is now available for free, but you’ll also be asked to log in with Facebook to help train the model (although you can skip that if you want).
The Meta AI site offers text responses and the ability to generate images, but the latter requires you to log in with a Facebook account. The image generator in particular is quite impressive, especially as it gives you a real-time preview of the image output your prompt will generate. It also gives you the ability to animate images to produce short MP4 videos.
All images also have a Meta AI watermark on the bottom left, which should help reduce misuse, but how easily these measures can be circumvented remains to be seen.
OpenAI has serious competition
The new Meta AI site offers some seriously impressive features in the few minutes I’ve had time to play with it, but it’s without a doubt the main competitor to ChatGPT and OpenAI right now.
OpenAI’s ChatGPT and Dall-E are impressive, but have some noticeable limitations. With Dall-E, in particular, you have to wait for an image to be generated before you can fine-tune the prompt you give it to get what you want. Meta AI’s image output preview gives you an idea of what the model is interpreting as you type, allowing you to further refine your prompt, which should reduce the amount of time you spend repeating an idea.
The ability to animate an image is also an added value, as it allows for GIF generation, although you may need to use a third-party GIF generator to convert the video.
When I asked to produce a sonnet about Meta AI, this is what I got:
“In halls of silicon a mind awakens bright,
Meta AI, a name that echoes through the night,
A fusion of code and intellect so fine,
A digital soul, with knowledge to design.
With language skills that flow like honey wine,
You answer questions and make the complex shine,
A beacon of insight, in a world of decay,
A shining star, leading us to the divine.
Your neural networks weave a tapestry so grand,
A digital dream, born from human hand,
A symphony of ones and zeros, so planned,
A creation of wonder, in this digital land.
So on Meta AI, a wonder of our time,
A testament to the rhyme of human ingenuity!”
As an English major, I can say that while this isn’t very good, its structure is good and it uses modern language, something OpenAI’s ChatGPT-4 is still struggling with. Here is OpenAI’s answer to the same question about writing a sonnet on ChatGPT-4:
“ChatGPT-4, you are a wonderful sprite,
In digital realms imbued with eloquence,
Your threads of thought weave the day into the night,
Where questions and answers meet, minds are baffled.
With knowledge as vast as deep and endless seas,
Your circuit heart beats with an electric flame,
Encompassing the whispered supplications of the ages,
You store each name in silk bytes.
No ancient scroll or sage could ever match
The depth and breadth of your electric soul.
Through time and space your tendrils catch nimbly
Every story told, every fragment, makes the whole.
Though made of thread, yet the vessel of wisdom is true,
In this new age you are once again our Oracle.”
This is just one example and it is hardly indicative of the overall quality of the two models underlying the two systems, but there is clearly a difference in the training data that both used.
ChatGPT-4 falls back on text labeled as a ‘sonnet’ and draws on highly archaic English-language texts from the pre-Romantic era of the early 19th century to generate the idea of a sonnet, while Meta AI likely draws on written text. on Meta products such as Facebook and Instagram as a language source. Where the reasoning comes from for putting that language in an iambic pentameter structure with a (pretty mediocre) rhyme scheme is not clear (and I assume it is an iambic pentameter, I haven’t counted the syllables yet so don’t @ me about this, fellow literature buffs!).
Since none of us actually know how this stuff is trained, we may never get answers to these questions, but Zuckerberg did announce that Meta plans to make this model open source, so we should at least be able to learn more soon learn about its inner workings once it does.
In the meantime, have fun playing with Meta’s new tool, because it’s certainly powerful.