The big news is that Nvidia’s founder and CEO Jensen Huang has joined X. In his first post, he shared a letter highlighting the importance of open models in AI. It’s an insightful letter that begins by talking about the open-source software movement that started against the belief that “software would advance only if companies kept tight control over their code” and the fact that open-source “created a shared foundation of knowledge on which generations of American engineers and entrepreneurs built their institutional sovereignty.”
China v. US
Today, China has been at the forefront of the open-source and open-weights movements in artificial intelligence, and America is lagging behind. The comparison of China and America is apt here because these are the only two countries that are making meaningful progress in releasing newer AI models at an unprecedented launch cadence.
Jensen argues that a similar choice, as with open-source software, is also faced with respect to artificial intelligence today, and if open models are not promoted, we will not be able to accelerate AI diffusion across industries and enhance cybersecurity.
The letter has been signed by Nvidia, Meta, Microsoft and others urging policymakers to expand compute access and avoid premature restrictions on open models. Interestingly, the letter also talks about distillation that has been termed as a “legitimate model-development technique” and states that the practice of using a model’s outputs to help train or improve another is a widely used technique that reflects a tradition of learning from existing technologies.
In my view, the timing of the letter could not be better. We are already seeing instances where internal models of closed AI are exhibiting rogue or misaligned behaviour. All this creates heavy levels of distrust in AI in general. Also, if US companies adopt an open model approach, it will be great for the rest of the world too. Most of the innovation in AI has come from China and the US. China is already on the path of open models, and with the US joining it too, the world will indeed see a hyper-accelerated pace of development as Jensen is talking about.
Win-Win Situation
The level of AI models today has reached a certain level where they are able to provide solutions to long unsolved problems in the field of mathematics. What is happening now in the field of mathematics will happen in all the other fields in the future, be it physics, biology, chemistry etc. With open models, the countries that are lagging behind in AI development will also be able to contribute meaningfully in the overall progress. Some of the key advantages in this regard could be:
- Lower barriers to entry and cost savings – The need for massive GPUs could be sidestepped and lagging countries can start from capable open models.
- Local adaptation and relevance – Customization of models with local languages, historical data, cultural knowledge etc. can be done at a fraction of the cost. “Sovereign AI” could become a reality in practice and not just in theory.
- Shift from pure consumption to ownership of capability – The countries will own the results of running inference and fine-tuning. The knowledge and value will not be locked behind foreign APIs.
- Parallel experimentation and diffusion – Independent researchers and institutions can simultaneously adapt same base models for different sectors making innovation distributed rather than gated. Local iterations will compound the progress.
- Better defensive and evaluative capacity – Customized and localized safeguard mechanisms could be adopted while examining model behaviour. Rather than relying on external safety claims about closed systems that cannot be fully inspected, open models provide the necessary tools to understand and counter potential misuse.
Star Trek v. Terminator
We want to end up in a Star Trek reality and not a Terminator-like one. By allowing concentration of immense AI power in the hands of the few corporations, we will be inviting trouble. Today, we have LLMs, tomorrow we will have full-fledged robots, and that is the reason we must embrace an open and transparent approach.
Also, if America adopts an increasingly open model approach, it will initiate a healthy AI race in the world where all the other countries will be able to come up with their best solutions.
If a Skynet-like AI takes birth, it will be better if we know about it sooner rather than later so that we can correct its alignment within time.