From: allin
Meta has been described as the first major technology company to fully embrace open-source AI [00:05:44]. This approach is characterized as “scorched Earth” [00:05:39], aiming to make new markets economically unviable for competitors, thereby allowing Meta to have more influence within a robust ecosystem [00:07:45].
Key Developments
- Llama Models: Meta’s Llama model was reportedly leaked last year, with some speculating it was a covert release [00:05:47]. More recently, Llama 3 was released and quickly ranked among the top two models on Hugging Face’s leaderboard, a trusted platform for benchmarking [00:06:05]. Developers have noted Llama 3 is faster and “less preachy” than ChatGPT 4, despite being slightly lower in quality [00:06:17]. Its context window, initially smaller than OpenAI’s models, was quickly expanded to 96,000 tokens [00:16:33]. The open-source nature of Llama 3 allows for rapid deployment; for example, Groq deployed it in the Groq Cloud within 14 hours of its announcement, enabling over 100,000 developers to build upon it [00:10:33].
- Meta Horizons OS: Meta announced the open-sourcing of the Quest operating system, now called Meta Horizons OS, which powers its mixed reality and VR headsets [00:06:28].
- AI Assistant Integration: Meta is integrating its AI assistant chatbox into Instagram, Facebook, and WhatsApp, which will effectively put a “modern search engine” in front of three billion users [00:06:40], potentially capturing 10% of the search market and creating unrivaled advertising data [00:18:05].
Strategic Reasoning
Meta’s open-source strategy aims to disincentivize venture capital firms from investing hundreds of millions in foundational model development companies, thereby limiting competition in the core foundational model market [00:12:47]. This “scorched earth” tactic is intended to make the economic value of foundational models negligible [00:10:00], reinforcing Meta’s existing economic moat in monetizing its family of apps [00:11:03].
This strategy echoes Google’s Android model, where Google open-sourced its operating system to ensure its services were not disadvantaged by closed, proprietary systems [00:12:12]. For AI, Meta’s strategy benefits from its vast amounts of human feedback data (clicks, comments, likes, shares) for reinforcement learning [00:09:07].
Market Impact and Stock Performance
Meta’s stock dropped by as much as 16% after its Q1 earnings report, despite beating estimates [00:06:54]. This downturn is attributed by some to the market’s negative reaction to Meta’s substantial spending on infrastructure, particularly its allocation to Nvidia for inference, an area where Nvidia’s GPUs are considered miscast compared to specialized solutions like Groq [00:20:39]. The market expects Meta to understand the distinction between AI training and inference costs and allocate capital more efficiently [00:21:49].
The rapid iteration and open-source nature of models like Llama 3 put closed models on their heels, as they struggle to compete with the speed and cost-effectiveness of open-source alternatives [00:10:46]. This dynamic suggests that the economic value of foundational models is disintegrating, shifting the focus to who can build on top of them the fastest [00:10:23].