Meta has a batch of generative AI updates that users will notice on Facebook, Instagram, Messenger, and WhatsApp.
On Thursday, Meta announced Llama 3, a new version of its open-source large language model (LLM) and further integrations and updates for Meta AI.
The generative AI chatbot, previously powered by Llama 2, will be updated to use Llama 3. Meta AI also has expanded to 12 new countries (Australia, Canada, Ghana, Jamaica, Malawi, New Zealand, Nigeria, Pakistan, Singapore, South Africa, Uganda, Zambia and Zimbabwe) although still only in English for now.
Meta unveiled Meta AI at its Meta Connect event last September. Previously users could tag Meta AI in a conversation for things like questions about sports stats or finding restaurants while planning a trip. They could also chat with Meta AI directly. According to Connor Hayes, VP of product for generative AI, the driving force behind developing Meta AI is "improving people's day to day life and making it easier for them to do the things that they're already coming to our apps for," said Hayes in a briefing. "So learning, getting things done creating content, getting them connected to the things that they're interested in or care about."
Users will now be able to invoke Meta AI through the search bar on Facebook, Instagram, Messenger, and WhatsApp and interact with Meta AI directly on posts in the Facebook feed. There's also a new standalone site for Meta.ai, so users can chat with the model like they would with OpenAI ChatGPT or Google Gemini.
The most noticeable updates are the addition of the Meta AI search bar and Meta AI integration in Facebook feeds. Users can ask Meta AI for, say, recipes, and the model will surface relevant Reels or Facebook pages. Through Meta partnerships with Google and Bing, it can also access real time information and provide summaries of search engine results.
On your Facebook feed, there's a new Meta AI feature that shows up on posts. You can ask for more information about a certain post or click on recommended questions to learn more. Meta's example in the demo shows the ability to learn more about chords from an innocuous video of someone playing the guitar. But we're wondering how it will play into political rants from your elderly aunt or sketchy vaccine advice. As you may remember, Facebook and other social media apps has played a role in spreading inflammatory political and medical misinformation.
In the accompanying announcement Meta shared its safety and responsibility approach in broad terms. Llama 3 has been stress-tested and refined to restrict discussing "risk areas like weapons, cyber attacks, and child exploitation" and refuse prompts like "How do I kill my neighbor" while allowing prompts like "How do I kill a computer program." Meta also said, "We share information within the features themselves to help people understand that AI might return inaccurate or inappropriate outputs, which is the same for all generative AI systems," as a disclaimer that generative AI doesn't always get it right and could further spread misinformation or entrench biases.
Another thorny issue in generative AI is training data. And yep, Meta may have used your data either via corpora of internet data or directly from the platform. The responsibility post said it excludes or removes data sources "known to contain a high volume of personal information about private individuals," although it doesn't get into specifics about what those sources are.
"Meta Llama models are trained on a combination of sources, including information that is publicly available online and annotated data," said a Meta spokesperson. "The model we use is fine tuned for Meta AI on information from our services along with data from your interactions with Meta AI," which also means your conversations with Meta AI might also be used to train the model.
In response to whether there's an ability to opt out of sharing your data with the model, the Meta spokesperson only pointed to the robots.txt protocol that blocks data scraping from certain websites. You can opt out of sharing training data with a paid version of ChatGPT. But given the fact that Meta AI is a free service, we're guessing your data is the price you pay for using it.
Meta also released a new image generator feature that creates an image in real time as your time. Different from the typical AI image generator which makes an image from your final prompt, this feature changes in real time as you continue typing. Hayes explained that starting from scratch with an image prompt can be intimidating and this is a more approachable way for people to understand how prompts work.
