This presentation explores the evolving landscape of generative AI, moving beyond the simple "open vs. closed" binary to analyze the critical nuances of the Open-Weight Era. While flagship models like GPT and Gemini remain "Black Box" API-only services, a new spectrum of openness has emerged that is redefining AI research and commercialization.

We provide a rigorous framework for categorizing modern LLMs based on three distinct pillars: Open Weights, Open Data, and Open Source. Through this lens, we examine four key tiers of the current ecosystem:

- Closed API-Only: The proprietary "Black Box" flagships.
- Restricted Open-Weight: Models like LLaMA and Qwen, where weights are accessible but governed by MAU caps or RAIL licenses.
- Permissive Open-Weight: The "Shovel Sellers" like Mistral and Falcon that offer commercial freedom but keep training data private.
- Truly Open: Projects like OLMo and LLM360 that prioritize full reproducibility, scientific auditability, and trust.

By evaluating the (de)centralization of computing power alongside these licensing dynamics, we illustrate how "True Reproducibility" remains a rare but vital goal. Attendees will gain a deeper understanding of how intellectual property, commercial interests, and the ethics of open collaboration are shaping the next generation of AI development.

- Talk presented at State of Open 2024 in London, UK: https://www.youtube.com/watch?v=tVTtyLuiYQk
- As a keynote at the "I LOVE Tech" event in Timisoara, Romania: https://www.slideshare.net/slideshows/unlocking-ainavigating-open-source-vs-commercial-frontiers/266821723
- At AI_Dev Europe Summit in Paris, France: https://sched.co/1c1mU