From: allin
The development of Artificial Intelligence (AI), particularly Large Language Models (LLMs), has sparked a debate about the merits of open source versus closed source models. This discussion centers on how these models are developed, distributed, and accessed, with implications for competition, innovation, and accessibility [00:17:20].
OpenAI’s Approach
OpenAI, a leading AI development company, has a unique position in this debate. From its founding, OpenAI never claimed to be open source in the traditional sense [00:15:07]. Instead, its initial claim was to be “open access,” emphasizing safety and non-differential or non-controlling access to its technology [00:15:10]. This principle is considered the “Genesis of the word open” in its name [00:15:20].
Reed Hoffman, a co-founder of Inflection AI and an early founding investor in OpenAI, notes that OpenAI’s structure evolved [00:07:41]. It began as a 501(c)(3) non-profit with philanthropic support aimed at ensuring open access to AI as an instrumental technology for humanity [00:20:29]. However, to scale and raise significant capital (e.g., a $600 million commercial round), OpenAI created a commercial entity [00:21:01]. This entity functions as a subsidiary of the non-profit, allowing for commercial benefits while maintaining the non-profit’s mission control [00:22:42].
Meta’s Open Source Strategy
In contrast, Meta has adopted an open source AI strategy. This decision is seen as a way to catch up, given that Meta was “far behind OpenAI and Google Microsoft” in the AI race [00:14:38]. Meta’s approach involves training models like Llama and making them available, leveraging a closed internal system that doesn’t rely on selling tokens [00:15:32]. Other entities like Mistral AI are also developing competent open source models [00:15:44].
The Future of AI Model Competition
Reed Hoffman believes there will be “winners all over the place” in the AI model competition, encompassing both open source and proprietary models [00:15:25]. He anticipates a significant demand for inference chips as the market scales [00:09:50]. This demand could force companies like Nvidia to decide between maintaining high prices and margins or adapting to a more competitive market [00:09:54].
The key question in “pure model competition” is when AI development will hit an asymptote to scale, potentially around GPT-6 or even later [00:16:15]. This long-term bet on scaling is what OpenAI, Anthropic, and hyperscalers are making [00:16:35]. Even if smaller, more specialized models gain traction, larger models will remain instrumental in training them [00:16:50].
Hoffman dismisses the idea of a single “God model” that dominates the entire market, akin to “Sauron’s ring” [00:17:51]. Instead, he envisions a future with “networks of models” and “traffic control and escalation” [00:18:09]. Larger models, while powerful, will always be more expensive [00:18:34]. This cost differential creates opportunities for smaller, specialized models to handle specific tasks like language translation more efficiently [00:18:51]. The use of “blends of models” and agents is likely to become “quickly universal” [00:19:07]. This multi-model approach creates ample room for startups to innovate in specific applications and verticals [00:17:11].
IP and Business Models
The intellectual property (IP) question regarding LLMs training on copyrighted content is complex [00:26:01]. While content creators should benefit economically from their work, AI training, which is akin to reading, could be considered fair use [00:27:08]. Hoffman suggests updating terms of service or copyright law to address this evolving landscape without blocking innovation [00:27:36]. He advises news organizations not to hold out for money on the training side, as synthetic data will eventually reduce the importance of any particular data source [00:28:22]. Instead, they should focus on freshness, brand, and establishing ongoing economic arrangements [00:28:30]. Both OpenAI and Microsoft have agreed on the importance of fairly apportioning economics in this new phase [00:28:44].
The massive AI infrastructure buildout is driven by the understanding that AI is a new computing platform [00:14:04]. Every device with a CPU or GPU will become more intelligent, making it crucial for hyperscalers to be fully engaged [00:13:51]. While some discussions about this buildout disregard traditional ROI analysis (e.g., “digital God” yielding “trillions” in returns), companies like Microsoft, under Satya Nadella, blend strategic insight with sensible return on capital [00:10:55]. They rationalize capital expenditure by focusing on expected revenue and business outcomes in areas like Office and Cloud, rather than just abstract notions like AGI [00:12:47]. The challenge is to invest intelligently without “spending like drunken sailors” [00:13:01].