Facebook and Instagram’s parent company Meta has entered the AI race with a bang.
The social media giant recently revealed its own natural language-processing system, known as LLaMA, or Large Language Model Meta AI. This is the tech that underpins chatbots like ChatGPT and Microsoft’s Bing AI, which can converse with users to provide answers to complicated questions in a human-sounding way.
So, will Meta’s new language model pave the way for an Instagram and Facebook chatbot? Well, according to the company’s boss Mark Zuckerberg, Meta has even bigger plans for AI than that. Think automated messages, TikTok-style video filters, and possibly even photo-realistic virtual reality avatars.
There’s just one problem: Meta’s AI was leaked to the public last week, meaning it could be used for nefarious purposes. The emergence of AI-powered phishing scams on LinkedIn indicates how the tech can go awry in the wrong hands.
Good or bad, the implications of a generative AI built by a company that boasts 2.96 billion monthly users is huge.
Here’s what you need to know about Meta’s AI ambitions.
What is LLaMA?
Large-language models are the latest buzzword in tech. They’ve already been developed and harnessed by the likes of Microsoft, Google, and OpenAI.
Meta, on the other hand, doesn’t currently boast a system that you can talk to - at least not in the UK. Instead, it has released a quartet of language-processing systems in varying sizes, known collectively as LLaMA, for researchers to use and improve.
There are plenty of issues with AI that academics could address, from bias to misinformation.
Like other AI models, such as OpenAI’s GPT-3, LLaMa is trained on pieces of words, or “tokens”, scraped from the internet. In a nutshell, Meta says its systems are more lightweight than the competition, making them more cost-efficient and easier to run for researchers.
The biggest versions of LLaMA were trained on 1.4 trillion tokens, including data from Wikipedia, books, academic papers from ArXiv, GitHub, Stack Exchange, and other sites.
Who can use LLaMA?
Last week, Meta’s language model was shared on notorious forum site 4Chan, where a member posted a torrent file for the tool.
The incident marks the first time that such a vast AI model has been leaked to the public. OpenAI, for instance, typically charges for training runs of its GPT-3 model — which Microsoft and Snapchat are using to power their chatbots.
Of course, that allows it to keep a tight grip over the system, theoretically preventing bad actors from accessing it. Whereas Meta’s AI model is now out in the wild, and available to all.
Currently, it’s unclear what impact the leak will have. After all, only the most technically savvy users with specced-out computers, that cost an arm and a leg to run, will be able to use it.
How will AI be used on Instagram and Facebook?
In the wake of ChatGPT’s success, every big tech firm is chasing a slice of the AI pie. Not to be left out, Meta CEO Mark Zuckerberg announced in late February that his company was creating a group dedicated to generative AI.
Meta has been working on AI for years, but now it would pull together all the teams focused on the tech from around its cavernous organisation and put them in one unit, Zuck explained in a Facebook post.
The team will be tasked with delivering “AI personas” and experiences with text, like chat in WhatsApp and Messenger; images, such as Instagram filters and new types of ads; and video, along with other updates that combine all three formats.
Meta is behind the curve when it comes to these features. Putting aside Microsoft’s Bing chatbot, even in the sphere of social media, rival TikTok is already using AI to create viral filters for short videos – the tech may have been used for its controversial de-aging filter that recently became a talking point online. While Snapchat just released a chatbot powered by OpenAI’s GPT-3 that users could interact with.
Longer-term, Zuck promised some “really futuristic experiences”. We can hazard a guess as to what that means. As of right now, Meta’s big, money-draining project is the metaverse - a 3D version of the internet brought to life through technologies such as VR and augmented reality. The company has a road map for the metaverse that includes more VR headsets, and smart glasses, which Zuckerberg claims will one day replace smartphones.
So how will generative AI fit into the metaverse? AI is already ingrained in gaming, but it could also help to make non-playable characters that you interact with and compete against even smarter. In addition, AI can be used to upscale video game graphics to produce more realistic environments and depictions – a use case that could work wonders for more immersive tech such as VR.
We also know that Meta is developing lifelike “codec” avatars that are a far cry from its current, cartoonish representations of VR users.
These avatars will boast better control over their eyes, facial expressions, hairstyles, and appearance, the company said in 2021. Meta has also demoed the ability to simulate how the avatar’s hair, skin, and clothing could react to different lighting conditions and environments.
Does Meta have a chatbot?
Though Google’s Bard and Microsoft’s Bing AI are getting all the attention, Meta has dabbled with chatbots in the past.
The company previously made a talkative AI known as BlenderBot available to the public to test out in the US. Meta has already collected 70,000 public conversations from the research project to help developers to smooth out any wrinkles — of which there appear to be many.
Like any AI, it’s prone to mistakes as it gathers most of its knowledge from information harvested from the web. Among its numerous reported falsehoods, BlenderBot insisted that Donald Trump was still the US president.
But, Meta’s AI ambitions aren’t limited to chatbots. Back when voice assistants like Alexa were taking the world by storm, Meta released its own take on the tech known as “M”. The AI helper was housed within Messenger, and could handle things like transferring payments between friends, restaurant reservations, and fetching quotes from local contractors such as builders.
Alas, the bot was overshadowed by Alexa and Google Assistant, forcing Meta (then known as Facebook) to shut it down in 2018.