On Open Source LLMs and Capitalism
TLDR:
Open source promotes diversity, reduces bias, and democratizes innovation
A few companies have a disproportionate amount of control with closed source models
There’s a spectrum of steps these AI companies can take to promote model transparency, but ultimately open source requires a reimagined business model
Yann Lecun, the Turing Award winner and Chief AI Scientist at Meta, said something on a recent podcast that I’ve been thinking about a lot: open sourcing enables diversity.
It’s something that people need to hear more of in the AI craze. At the moment, there’s a small group of AI companies training models on conceivably all of human knowledge. They fine tune based on what they think makes sense and prescribe guiding principles and guardrails for the model to follow (i.e., like this). The models are based on ideas of what you probably want, but they likely didn’t ask you. These underlying large language models (LLMs) by companies like OpenAI, Anthropic, and Google are closed source, meaning that the source code of the model is not made available to the public. Individuals do not have the ability to access, modify, and use these models for their specific use cases.
On the one hand, I can appreciate why companies who develop closed source LLMs want to hold their code close to their chest. GPU chips are expensive and AI talent is in high demand. Businesses want to make a return on their investment and may have capital structures where investors frankly expect returns. There’s also the threat that open source models, where individuals can control how to tune the model, can be used for nefarious reasons. Some individuals have even claimed that it’s not very difficult to remove safety measures from models.
But on the other hand, closed source models prevent all of us from utilizing our collective knowledge in a myriad of ways that are incomprehensible to a few AI companies. Yann called closed source models a danger to democracy. Open sourcing gives individuals the freedom to innovate, enables much wider distribution, and provides fine tuning flexibility. It ultimately allows individuals to make things that people want.
To be clear, I’m not advocating for all closed source models to be released and made available for public consumption tomorrow. What I am saying is that there is a spectrum of steps companies can take to facilitate LLM transparency. These can be as simple as releasing model datasets to releasing prior versions of the models themselves. Aside from the tech stack, open sourcing to me means setting up the processes and frameworks to make it successful for all parties. This includes commercial licensing, support infrastructure, and security measures.
AI systems are inherently biased, and sometimes companies get it really wrong. But open sourcing is one path I see forward that will require business model innovation and new processes in order to celebrate our diversity and solve our unique problems.
