Open Source Generative Artificial Intelligence: Challenges and Opportunities

Generative AI, a field dominated by proprietary models from large tech companies, is being disrupted by a new wave of open-source models.

Supporters argue that open-source has significant advantages, such as enabling broader access, promoting innovation, and increasing transparency. Many believe that open-source models will dominate the market.

However, this conclusion is not obvious.

Open-source generative AI is fundamentally different from the open-source movement that provides tools like TensorFlow, MySQL, or Kubernetes. Open-source dominates in these areas because the required investment can be achieved through crowdsourcing. But generative AI requires data and compute power, both of which are becoming increasingly expensive and difficult to obtain for most open-source players.

A few large companies are investing heavily in generative AI models and almost monopolizing the GPU market. With competition from these proprietary companies, they may offer their models at cost or low profit to establish market share. In the short term, the price of remaining GPU capacity may rise to a level that is uncompetitive for open-source users.

In addition, the value created by proprietary models and open-source models may be asymmetric. Efforts are being made to reduce the scale of generative AI models so that they can be computed on edge devices like smartphones or autonomous vehicles. However, the greatest value will lie in models with agency, or multiple models working together, that can perform inference, control other software, and take action in the real world.

This capability is likely to remain the exclusive domain of resource-rich private companies. Some companies are already building agents that can perform inference using models like GPT-4 and operate tools. This is something that open-source generative AI models currently cannot fully achieve.

Meta is an exception as they have decided to open-source their LLaMA series of models. However, there may be strong opposition to open-sourcing such powerful models due to the potential for abuse.

Although Meta argues that generative AI should be open-source to prevent power from being concentrated in the hands of a few, there is a viewpoint that locking down generative AI may be a better choice, especially once large models have agency.

Recently, at a closed-door Senate forum convened by Chuck Schumer, a Democratic senator from New York, Meta's founder Mark Zuckerberg mentioned the capabilities of LLaMA 2 in detail when questioned about the security of open-sourcing such powerful technology, including its ability to produce anthrax (a deadly toxin).

While the debate continues, no consensus has been reached. A group of researchers from Collaborations Pharmaceuticals recently asked a proprietary generative AI system called MegaSyn to generate toxic molecules, resulting in the generation of numerous toxic molecules, some of which are similar to known neurotoxins.

Open-source models allow access to the parameters within the model (i.e., the weights between artificial neurons that influence predictions), enabling anyone to experiment freely with the model.

Although creators of open-source models set up some safeguards, researchers claim that these safeguards can be bypassed in a matter of days, allowing the model to do whatever a user wants. Given this, open-source may struggle to beat proprietary models without some scientific breakthroughs (e.g., models with fewer parameters that have competitive performance) or commitments from powerful private companies like Meta to outsource future generations of models.

While accepting control of generative AI by large tech companies is challenging, it may be the only way forward without such breakthroughs or a transfer of funds.