From: redpointai
Peter Welinder, VP of Product and Partnerships at OpenAI, offers insights into the evolving landscape of open source AI models versus proprietary models. OpenAI, known for developing leading proprietary models like GPT-4, ChatGPT, and GitHub Copilot [00:00:09], aims to enable as many builders as possible to create products on top of its technology [00:02:44].
Value Accrual in the AI Ecosystem
Welinder believes that the majority of value in the AI ecosystem will ultimately accrue at the application layer, across various applications [00:02:21]. OpenAI’s strategy aligns with this by minimizing “value extraction” at the model infrastructure layer, keeping prices very low to foster widespread development [00:02:50]. This approach has included significant price cuts, such as a 70% reduction for some models and the release of GPT-3.5 Turbo at ten times cheaper than the original 3.5 model [00:03:07].
Instead of building all the necessary tools and applications, OpenAI focuses on refining its core models and trusts the broader developer ecosystem to invent and build the best tools on top of their technology [00:04:07]. Value capture, in this view, will primarily happen at the application layer through standard business competitive moats like network effects and branding [00:05:12].
OpenAI’s Product and Development Philosophy
OpenAI intends to remain at the platform level, providing general applications like ChatGPT that are not designed to be specialized AI teachers, doctors, or lawyers [00:06:18]. Their role is to provide the intelligence and reasoning capabilities, enabling integrations with other products and proprietary content, and allowing developers to customize models with domain expertise [00:06:55]. Features like plugins in ChatGPT exemplify this direction, allowing connection to external services [00:07:26].
OpenAI’s prioritization is heavily focused on the core models themselves, making a strategic “big bet” on large language models (LLMs) and their generality [00:08:39]. This includes dedicating most personnel, compute resources, and GPUs to training these models [00:09:00]. Past ventures, such as robotics or AI for Dota 2, were exploratory and eventually scaled back to concentrate efforts on LLMs as confidence grew in that direction [00:09:27].
Open Source AI Models
Welinder acknowledges the recent explosion of activity in open source AI models, particularly following Meta’s release of Llama [00:14:35]. While he believes open source models will eventually catch up in the very long run, he suspects there will always be a category of AI systems (often proprietary) that are significantly better [00:16:10].
Comparing it to desktop Linux versus Mac OS X and Windows, Welinder suggests that open source AI models may not receive the same level of investment in intricate details to match the performance of proprietary systems [00:16:30]. The substantial capital and engineering required for training and running these models at scale are difficult to replicate in an open source environment [00:17:02]. Therefore, companies investing heavily in these models are unlikely to open source them due to investment and safety concerns [00:17:24].
Despite this, Welinder is “very excited” about open source development [00:17:59]. He sees its utility in:
- Pushing research forward: Open source models allow researchers to experiment with new approaches and training methods [00:18:10].
- Specific product areas: Smaller open source models are suitable for devices, edge deployments, or on-premise solutions [00:18:31].
Proprietary AI Models
Welinder argues that for applications requiring the best performance and reliability, proprietary models will remain the preferred choice. He predicts these models will continue to be several years ahead of their open source counterparts for the foreseeable future [00:19:09].
One key distinction is the “generality” of proprietary models. While a specialized open source model might suffice for narrow tasks like summarizing specific information, products tend towards generality over time [00:20:14]. For broader applications, such as a customer service bot handling edge cases, a smarter, more general model is necessary [00:20:41].
Competitive Risk for OpenAI
Welinder does not view the open source AI movement as a significant competitive risk to OpenAI’s business model [00:21:49]. He believes most value will derive from the “smartest models” because these enable the creation of the best products and can tackle the most economically valuable problems [00:21:54].
Companies that choose less capable models will face a competitive disadvantage [00:22:28]. OpenAI’s primary focus remains on advancing the smartest models to enable applications like co-pilots for various professions (e.g., lawyers, medical providers) and eventually for scientific discovery [00:22:56].
Strategic Open Sourcing
OpenAI selectively open sources auxiliary models that enable broader use of their smarter models. An example is the Whisper model, which provides highly accurate audio transcription. OpenAI open sourced it to allow developers to feed accurate transcriptions into large language models, thereby building new applications, without expecting to generate significant revenue from speech recognition itself [00:24:39]. The decision to open source is based on enabling more use cases rather than direct monetization [00:25:14].