The statement “Avocado is not a fruit” hits the nail on the head. Meta’s new closed-source direction, codenamed “Avocado” (or Guac), signals that their most advanced AI models are no longer a communal, openly-shared harvest for the developer ecosystem. They are now a proprietary product designed for market competition and monetization.
| Llama (Open Weights) | Avocado (Proprietary) |
|---|---|
| Philosophy: Commoditize LLMs; fuel the ecosystem. | Philosophy: Protect performance; sell premium access. |
| Goal: Promote adoption and standardization. | Goal: Revenue generation and competitive differentiation. |
| Status: Open Weights (but restricted license). | Status: Closed Source (likely gated API access). |
It’s a structural shift that proves that while Meta loves the positive PR of open source, it cannot justify the staggering $70-72 billion in expected 2025 AI capital expenditures without a clear path to revenue, which a fully open model simply cannot provide.
LLMs Are the Commodity, The Framework is the Cash Register
LLMs are becoming a commodity; the true value and the money is in the framework and the surrounding tools.
This is where Meta’s existing strength lies, and why the framework is key:
- Google’s Strategy: Google had a “sleepy start” on the public-facing LLM race but is now dominant because its money is made through tools and platforms (Cloud, Android, Search) that leverage AI. Gemini is simply a new engine powering existing revenue streams (ads, enterprise cloud contracts).
- Meta’s Strategy: Meta’s revenue is overwhelmingly driven by digital advertising (98%) and user engagement. Their AI is primarily focused on improving ad targeting, content recommendations (Instagram Reels), and user engagement across their 4 billion users. The new AI framework is meant to drive efficiency and ad value within their existing walled garden. The TikTok-for-kids concern is real, as much of their monetization push is about making content creation and consumption stickier for all users through tools like Meta AI.
The Enterprise Data Firewall: The Ultimate Barrier
The most critical observation is this: No real company sends their core, sensitive data to an AI giant.
| Risk Factor | Compliance Issue |
|---|---|
| Data Exposure | Transferring data to a cloud provider’s API risks leakage, misuse, or unauthorized access by the third party. |
| Regulatory Non-Compliance | Strict regulations like GDPR and HIPAA prohibit transferring sensitive data outside of controlled, compliant environments. |
| Model Inversion Attack | Even anonymized data sent to a public API can potentially be reverse-engineered by advanced attacks to reveal confidential information. |
This is why private, on-premise, or confidential computing solutions (like Gnoppix AI Browser Extension running local Ollama) are the only viable path for true enterprise adoption of AI for core business processes. The enterprise market has a data security firewall that prevents them from being the “real customer” that OpenAI needs to hit its $1 trillion valuation target.
The Future: AI Robotics and the Embodied World
AI Robotics (Embodiment) is the real use-case, not just LLMs. Leading AI researchers, including Meta’s own former chief AI scientist, Yann LeCun, have long argued that LLMs will never reach Artificial General Intelligence (AGI) because they lack “world models” the ability to understand and interact with the physical world.
The next major leap is already underway: Physical AI.
- Companies like OpenAI, Tesla, and DeepMind are aggressively investing in robotics and World Models (which combine LLM concepts with real-world physics and action data).
- LLMs are simply the brain; robotics provides the body and the sense of reality necessary for true intelligence and for solving real-world, non-text-based problems (e.g., manufacturing, elder care, autonomous operations).
My prediction about the cloud-based LLM giants facing a harsh reckoning seems to align with a broader industry consensus: the initial LLM gold rush is plateauing, and the next phase of value will come from integrating AI into the physical, private, and localized world. This is where models like Qwen and Deepseek, which continue to contribute to the open-weights community, become crucial counterbalances to the proprietary power of the Western giants.
I don’t quite understand the hype. Of course, it’s great to have an online reference tool. The fact that tasks can be completely automated has been true for a long time. You don’t need AI for that. Even in manufacturing, installing part A into part B 100,000 times has been possible for a long time. We’ve had modern machine learning since 1990. So, while all of this is great, like a better Google search that delivers results, true AI is when I have data and the AI independently learns from its mistakes and then optimizes the whole process.
Perhaps it’s time to support talented scientists instead of political squabbles and political correctness. People have extraordinary abilities, why don’t we utilize them? America was once the land of unlimited possibilities; intelligent people were given a chance to do their thing. Today, I can’t even get past the reception desk to present a good idea. Maybe we should all take a step back, become more human again, and listen to others. That used to help, but it’s completely disappeared today.