The future of foundation models is closed-source

🌈 Abstract

The article discusses the future of foundation models, exploring two contradictory narratives - one where AI centralizes and another where it decentralizes. It examines the current state of open-source and closed-source AI models, the motivations behind Meta's open-source strategy, and the potential implications for developers, model builders, and national security.

🙋 Q&A

[01] Two Narratives About the Future of Foundation Models

1. What are the two contradictory narratives about the future of foundation models?

• One narrative suggests that AI will centralize, with scaling laws holding and value accruing primarily to scaled, closed-source players.
• The other narrative suggests that AI will decentralize, with foundation models having no moat, open-source catching up to closed-source, and many competing models.

2. How do both narratives seem true today?

• There are powerful closed models as well as a thriving ecosystem of open-source models, with Llama-3 recently putting open-source on the map of GPT-4 class models.
• An unusual open-source alliance has formed among various stakeholders who want to avoid centralized control and regulatory capture.

[02] Open-Source vs. Closed-Source AI

1. What are the potential downsides of open-source AI according to the article?

• Open-source AI will become a financial drain for model builders, an inferior option for developers and consumers, and a risk to national security.
• Closed-source models will create far more economic and consumer value over the next decade.

2. How does the article compare the cost, model quality, and data security of open-source and closed-source AI models?

• Cost: Open-source models have hidden costs, and the closed-source cost curve is coming down rapidly.
• Model quality: The paid, closed-source versions are generally better than the free, open-source versions.
• Data security: Using open-source models on-prem or via third-party hosting may not be safer than using third-party LLMs in the cloud.

3. What are the potential national security risks of open-sourcing AI models?

• Open-sourcing AI model weights arms military adversaries and economic competitors, as it allows them to access and potentially misuse powerful AI capabilities.
• Rogue actors could use open-source models to generate cyberattacks, bioweapons research, and other dangerous content at scale.

[03] Meta's Open-Source Strategy

1. What are the key reasons behind Meta's open-source strategy according to the article?

• Meta wants to avoid being beholden to another tech platform, as it was with Apple's closed ecosystem.
• Open-source AI enhances Facebook and Instagram by improving social feed algorithms.
• Meta can use the extra H100 GPUs it has purchased to train and open-source AI models.

2. What are the potential limits to Meta's open-source strategy according to the article?

• Meta will likely stop open-sourcing their AI models at some point, either due to the cost or safety concerns.
• As Meta's models become more differentiated based on proprietary data and usage, they will want to capture that value in their closed products rather than open-sourcing it.

[04] The Future of Open-Source and Closed-Source AI

1. What is the author's overall perspective on the future of open-source and closed-source AI?

• The bulk of the value creation and capture in AI will happen using frontier capabilities, which will be in closed-source models.
• Open-source models will have a place for smaller, less capable, and configurable workloads, but they will lose the capital expenditure war as their ROI continues to decline.
• Closed-source model providers will be better positioned to drive sustained, long-term progress in AI due to the capital-intensive nature of foundation models.