1 week ago
106
It is likely that you’re in the midst of the ultimate juggling act, with a decision yet to be made. That is, which artificial intelligence (AI) chatbot or assistant to begin using, more regularly. A conundrum many are struggling with. If the present choices of OpenAI’s ChatGPT, Microsoft’s Copilot, Google Gemini and Perplexity AI’s curation of multiple AI models with the Pro subscription weren’t enough, add Meta AI to the list. To that effect, Mark Zuckerberg’s Meta may be making adoption simpler, by seamlessly appearing within WhatsApp, Instagram, Facebook and Messenger, alongside meta.ai on the web.
Meta AI integration across WhatsApp, Instagram and Facebook. (Official images) Meta’s latest AI model, the Llama 3, is what Meta AI is built on. That, according to the company, is the reason they’re confident this foundational open-source model will outperform competition. In fact, Meta is rolling out two different Llama 3 models. One will be the underlier for consumer facing AI chatbots in Meta’s apps.
HT launches Crick-it, a one stop destination to catch Cricket, anytime, anywhere. Explore now!
The other, a larger and multimodal version, will soon be available on platforms including Amazon Web Services or AWS, Databricks, Google Cloud, Hugging Face, Kaggle, IBM WatsonX, Microsoft Azure, NVIDIA NIM, and Snowflake. Extensive hardware support too, including chips by by AMD, AWS, Dell, Intel, NVIDIA, and Qualcomm.
While functionality may vary slightly depending on which app or service you access Meta AI from, but the capabilities portfolio is extensive. For example, the search functionality which is expected to be a standard spread across Meta’s apps, can point you to web search results from Google Search as well as Microsoft Bing – as far as my knowledge helps, Meta AI is the only AI assistant thus far, to do so. ChatGPT and Copilot rely on Microsoft Bing, while Gemini expectedly points to Google Search.
There is a text to image generator tool which goes beyond just that – Imagine can generate AI images as you may be typing a message in WhatsApp, and this is now rolling out as part of a beta test for now, but only in the US. There’s integration with the Facebook Feed for more context or information about something you just saw in a post.
There is a certain level of confidence that Meta is clearly projecting, with some benchmark tests that they point to suggesting Meta AI’s Llama 3 has a performance advantage in the 8B and 70B parameters in instruct model and pre-trained model tests – the former compared with Gemma and Mistral, while the latter compared with Google’s Gemini Pro 1.5 and Claude 3 Sonnet. OpenAI never disclosed parameter specifics for GPT-4, and the next generation GPT-5 is expected to roll out in the coming months.
For Meta, foundational strength and accuracy will be important, as it would want to avoid the mistakes Google made in succession with Bard and then Gemini, including earlier this year when the latter got the diversity balance all wrong in image generations – some generations included pictures of Nazis and the Founding Fathers of America, as people of colour. Google’s Prabhakar Raghavan later clarified a series of “tuning” issues with context and caution. OpenAI’s GPT, which also is the basis for Microsoft’s Copilot assistant, has also often been criticised for incorrect information or limited context.
As AI model parameter naming suggests, 8B is categorised as a very large model, with 8 billion parameters used for training, while 70B is an even larger model with 70 billion parameters. The thumb rule is – the more parameters an AI model is adept with, the higher the scope of learning complex patterns.
“Improvements in our post-training procedures substantially reduced false refusal rates, improved alignment, and increased diversity in model responses. We also saw greatly improved capabilities like reasoning, code generation, and instruction following making Llama 3 more steerable,” the company explains the technical details.
Meta confirms that the training data used for Llama 3 consisted of data sets seven times larger than those used for Llama 2, and that includes four times more code too. All data was collected from what Meta says are “publicly available sources”, though they don’t get into the specifics.
There is the sense that Llama 3’s 8B and 70B models are just the beginning, with models as large as 400B parameters, in training. Expect those to be rolled out for specific use cases in the coming months.
Meta seems to be taking it slow with availability. Additional specifics, beyond the beta availability of Imagine in the US, point us to English language support that’s rolling out today in the US, Australia, Canada, Ghana, Jamaica, Malawi, New Zealand, Nigeria, Pakistan, Singapore, South Africa, Uganda, Zambia and Zimbabwe.
HT reached out to the Meta team for guidance about Meta AI’s availability in India, and we are yet to hear back.
