The past few weeks have been a tumultuous period for Meta, with the volatility centered upon its efforts in open-source and AI. Its share price reached an all-time high of $531.49 on April 8 – up 50% since the start of the year, and nearly 6x its low in late 2022. Then, Meta had a series of major product reveals, such as Llama 3 (the next generation of its open-source large language model), the release of a set of trust and safety tools, the opening up of its Quest VR operating system, and the debut of its new AI chip.

CEO Mark Zuckerberg followed up this week on the quarterly earnings call with the not-all-too-surprising news that Meta plans to continue to invest in AI, to the tune of $35B-$40B in 2024. Still, the news – and the implication that Meta’s “year of efficiency” was over – rocked its share price, causing it to plummet 15% in a single day. This is despite Meta’s revenue being up 27% and earnings being up 117% year-over-year.

Specifically, Zuckerberg said: “[W]e should invest significantly more over the coming years to build even more advanced models and the largest scale AI services in the world… [R]ealistically, even with shifting many of our existing resources to focus on AI, we'll still grow our investment envelope meaningfully before we make much revenue from some of these new products…[W]e've historically seen a lot of volatility in our stock during this phase of our product playbook ‐‐ where we're investing in scaling a new product but aren't yet monetizing it. We saw this with Reels, Stories, as News Feed transitioned to mobile and more. And I also expect to see a multi‐year investment cycle before we've fully scaled Meta AI, business AIs, and more into the profitable services I expect as well.”

We should recall that the current momentum behind open-source large language models (LLMs) was set off by the leak of the training weights for Meta’s LLaMA (Large Language Model Meta AI) in early 2023. (Model weights store the statistical relationships across a training dataset.) With that, LLaMA became the first “really capable” state-of-the-art foundation model made available to the open-source community – vaulting Meta’s standing into becoming a leading player in AI.

Since then, Meta has released Llama 2 and its code-specialized version Code Llama as open-source models available free of charge for commercial use. It has also incorporated Llama-powered AI agents into Facebook, Instagram, WhatsApp, and Messenger.

Last Thursday, Meta announced the release of Llama 3, with industry watchers calling it “the best-performing open-source LLM available” and Meta asserting that it outperforms other models of similar size. In rankings, only proprietary frontier models from OpenAI (GPT-4), Anthropic (Claude 3 Opus), and Google (Gemini 1.5 Pro) have performed better. Llama 3 is being made widely available through Amazon AWS, Google Cloud, Microsoft Azure, Databricks, Snowflake, Hugging Face, IBM WatsonX, NVIDIA NIM, and Kaggle.

Llama 3 was trained on two 24K-GPU clusters and is available in two sizes, 8B parameters and 70B parameters. Llama’s performance has been advancing rapidly across its generations. According to Zuckerberg, the 8B-parameter smallest-sized Llama 3 is as powerful as the 70B-parameter (currently) largest-sized Llama 2. Meta hopes to release its 405B-parameter “dense model” (which is still in training) later this year. It also plans to introduce multimodality, conversations in multiple languages, and a “much longer” context window.

In conjunction with the Llama 3 announcement, Meta also released new trust and safety tools for developing LLMs: Llama Guard 2 (an LLM that provides safety guardrails for chatbot inputs and outputs), Cybersec Eval 2 (an evaluation suite to assess an LLM’s security risks), and Code Shield (inference-time filtering of insecure code generated by an LLM).

This past Monday, Meta followed with the reveal that it would be opening up the operating system that powers its Meta Quest VR headsets to 3rd-party hardware makers. Other hardware companies will now be able to create devices that can run on Meta Horizon OS, using the same mixed-reality technologies as Meta (e.g. hand, eye, face and body tracking; Passthrough; Scene Understanding; Spatial Anchors). Meta Horizon OS devices will have access to the same companion app that Quest owners use, which will be renamed the Meta Horizon app. 3 companies are already developing new devices with the Meta Horizon OS: ASUS’ Republic of Gamers is building a performance-gaming headset; Lenovo will develop mixed-reality devices for productivity, learning, and entertainment; and Xbox is working with Meta on an Xbox-inspired limited-edition Meta Quest.

Developers will be able to create mixed-reality experiences for a larger audience and ecosystem. They will have access to Meta Horizon OS’ content discovery and monetization platforms, including the Meta Horizon Store (a rebrand of the Meta Quest Store). Meta is also making it easier for developers to distribute VR apps that meet basic requirements, removing the barriers between the Horizon Store and Meta’s App Lab. Meta is furthermore creating a new spatial app framework for mixed-reality experiences, although developers can still use their familiar tools to bring their existing mobile apps onto the platform.

Meta has planted its flag on being a champion of open-source. Zuckerberg has reiterated that in the grand battle between open vs. closed models, “In this next generation, Meta is going to be the open model and I really want to make sure that the open model wins out again.” Meta’s official stance is that “openness leads to better, safer products, faster innovation, and a healthier overall market.”

Meta is looking to position itself as the anti-OpenAI. OpenAI, despite its name, does not open-source its family of GPT models. Its closed nature has been a hot-button issue – just last month, OpenAI’s co-founder Elon Musk sued it for breaching a “founding agreement” that OpenAI would not keep information private for commercial purposes.

In the world of LLMs, being “open” doesn’t necessarily mean the LLM can be inspected and built upon, or that it is made available under an established open-source license. Being open often just means releasing the training weights, which both Meta’s Llama models and Grok from Elon Musk’s xAI company have done. At the same time, Meta has been criticized for using bespoke licenses that are not very permissive and engaging in “openwashing.”

One of the biggest questions for Meta is how they will make money from their massive investments in open-source and AI. Zuckerberg has said, “I’m basically very inclined to think that open sourcing is going to be good for the community…[O]ne of the reasons why I'm philosophically so pro open-source is that I do think that a concentration of AI in the future has the potential to be as dangerous as it being widespread… [I]t's probably also pretty bad for one institution to have an AI that is way more powerful than everyone else's AI.”

Zuckerberg and Meta – after having gone through the metaverse hype crash and pain of layoffs in 2022 and the year of efficiency in 2023 – are likely to be pragmatic and ROI-driven with respect to AI. Meta’s fleet of GPUs – which will reach 350K GPUs by the end of 2024 – was and continues to be a massive investment. Much of the fleet is being dedicated to inference (rather than training), which can be enormously expensive given the scale of Meta’s consumer-facing social platforms. (Meta’s new AI chip is part of its push to diversify away from Nvidia.)

If we unpack the different areas where Meta will benefit, they can be bucketed in 3 categories: (1) Enhancing its consumer products and ad platform, (2) reinforcing its leadership in growth arenas, and (3) making its internal operations more efficient.

With respect to the first category, Meta is already working to incorporate AI into its social platforms, where even small improvements in personalization, recommendations, and user engagement can move the needle in big ways. More personalized ads and better targeted recommendations have obvious benefits for Meta’s core business, as well as automation for advertisers. Meta also has its eye on business messaging and paid content/ads in AI interactions. An AI search toolbar (which Meta is piloting now) could make its platforms more of a destination for people looking for product information or shopping, who might try their questions there first instead of heading to Google or Amazon. Zuckerberg believes Meta AI is already “the most intelligent AI assistant that you can freely use.”

On the second category, open-sourcing Horizon OS could help Meta become the Android/iOS of the mixed-reality space, propelling the metaverse forward, opening up new revenue streams (e.g. Horizon Store), and finally allowing Meta to move out from under Apple’s thumb. In the realm of AI, Meta is aiming to be “the leading AI company in the world.” Providing the frameworks, LLMs, and benchmarks that everyone is using means that state-of-the-art advances developed elsewhere have more “natural interoperability” with Meta’s systems and are more likely to “snap into place.” Meta is also considering allowing people to pay to use bigger AI models and more compute, or it could take a cut when large tech firms resell Meta’s models.

On the third, being a leader in open-source AI can help Meta attract scarce top-tier talent. It also means Meta can more quickly incorporate learnings from the research community, capitalize on emerging use cases, and leverage new technologies in its stack. For instance, Zuckerberg is expecting efficiency gains in running its Llama models from leveraging improvements from the open-source community (similar to what happened with the first Llama model). Given how expensive the models are to run, these savings could run into the billions of dollars. Meta is already using AI across its operations, in areas like data center optimization and developer coding.

Meta is wary about giving away the store. Zuckerberg has said, “We're obviously very pro open source, but I haven't committed to releasing every single thing that we do.” 3rd-party developers, for instance, cannot use Llama 3 or its output to improve any other LLM, per the Llama 3 license.

There is a growing belief in the industry that open-source AI could erode the competitive moats of proprietary AI. In that case, companies will have to establish a “competitive moat not based on the capabilities of their model” – and it is here that Meta is positioned particularly well to benefit.

May 12 2023 (3 Shifts): Will open-source LLMs wipe out AI’s competitive moats?

Apr 14 2023 (3 Shifts): Proprietary enterprise LLMs vs. open-source LLMs