The Metaverse is Dead

Business Summary

• Meta’s metaverse strategy appears to have been shelved, according to the presenter. The project had consumed substantial resources without demonstrating a path to sufficient returns. The presenter suggests investor pressure and the cost of Reality Labs’ losses were likely factors, but this is an interpretation rather than a confirmed explanation.
• Meta is shifting its attention toward generative AI, a fast-growing area following the success of OpenAI’s ChatGPT. Mark Zuckerberg was quoted announcing a new top-level product group to bring together Meta’s generative-AI teams.
• Planned AI applications span Meta’s existing products, including creative tools and experiences for Facebook, Instagram, WhatsApp, and Messenger. These could include text-based interactions, image filters, ad formats, video, and eventually AI personas.
• Meta announced LLaMA language models in several sizes: 7B, 13B, 33B, and up to 65B parameters. The presenter suggests that smaller models could be cheaper to develop and easier to run locally.
• Distribution and infrastructure may be strategic advantages. The presenter argues that Meta’s scale and existing infrastructure could help it compete with OpenAI, while smaller models could make local deployment more practical.
• Openness and user control could be differentiators. The presenter favors free, open-source, self-hosted AI, which could let users fine-tune models on their own data without sharing it with a platform provider. Whether Meta’s models will be meaningfully open or restrictive remains uncertain.

Frameworks, Processes, and Playbooks

• Reallocate investment from a costly initiative to a more promising market. The video portrays Meta’s shift from metaverse spending toward generative AI as a response to large losses and investor pressure.
• Build product capability through a centralized group. Meta announced plans to consolidate its generative-AI teams and apply the technology across multiple products.
• Use a tiered model portfolio. Models with different parameter sizes could serve varying cost, hardware, and capability needs.
• Compete through deployment options. The presenter sees local and self-hosted use as a potential advantage over models that depend on costly infrastructure or restrictive APIs.

Metrics and Figures Cited

• Reality Labs losses: The presenter cites cumulative losses of about $26 billion, then later refers to $24 billion. The figures are inconsistent in the subtitles.
• LLaMA model sizes: 7B, 13B, 33B, and up to 65B parameters.
• GPT-3 comparison: Described as having more than 175 billion parameters.
• Hardware requirement for running a GPT-3-scale model: About 700 GB of VRAM, according to a Stack Exchange post referenced by the presenter.
• GPT-3 development-cost estimate: About $10 million, as stated in the video.
• Revenue, adoption, or profitability targets and timelines: None are provided for Meta’s AI products.

Examples and Actionable Takeaways

The presenter contrasts Meta’s situation with SpaceX’s early failed Falcon 1 launches. Expensive setbacks may be easier to sustain when a company or founder is willing to absorb them, highlighting how capital providers’ tolerance for losses can shape strategy.

For companies moving into AI, the video’s implied considerations are to:

• Control development costs.
• Match model size to available infrastructure.
• Integrate tools into existing products.
• Decide how much user access and data control to offer.

The presenter argues that self-hosted models can support customization and privacy, but this is an opinion rather than a demonstrated business outcome.

Presenter and Sources

• Presenter: Mental Outlaw
• Sources referenced: Mark Zuckerberg’s comments on Meta’s generative-AI product group; Meta’s LLaMA announcement; OpenAI and GPT-3/ChatGPT; and a Stack Exchange post about VRAM requirements.