Determinism, Free Will, and its relationship with AI

Picture yourself living in a cabin deep in the forest during the final week of your 3-month spiritual retreat. Your only rule has been to maintain complete silence throughout this time, yet you feel an overwhelming urge to pick up your phone and call your loved ones. Your skin crawls, your hand hovers near the phone, and your mind races: should you break your promise and call them, or wait one more week to honor your goals? After careful consideration, you decide to call. Having spent three months in meditation, you step back and wonder what truly made you pick up the phone. Did you actually make this decision, or was this outcome inevitable? Is there such a thing as careful consideration, or consideration at all? How much does your biological state influence your choices?

While I haven’t embarked on a 3-month forest retreat myself, I’ve thought about these questions for a while. In this blog post, I invite you to explore with me the possibility that free will doesn’t exist and how this idea suggests that we live in a purely deterministic world. We’ll examine these concepts through the lens of AI and neural networks, which, as a computer scientist, is how I approach these questions.

Consider a neural network (in this case, a physical one called the brain) whose initial weights are determined by genetics. You can think of this as the prior in Bayesian statistics. For the first few years of life, this neural network receives input in the form of sensory experiences—sight, smell, touch, taste, and hearing—especially sight and touch. In response to these inputs, driven initially by instinct, it acts and learns from rewards. Touch fire, get burnt. Walk into a wall, get hurt. As time passes, this neural network evolves beyond pure instinct and begins generating latent tokens, which we call inner monologue. This inner monologue can take various forms: language (thinking about what to say), visual imagery (imagining an apple), or audio (a musician conceiving a melody). These latent tokens are fed back into the neural network as input, producing either more latent tokens or actions. The outcomes of these actions are also fed back into the model as a continuous stream of information.

If you accept this description of the human brain, an interesting conclusion follows: humans have no free will. All actions are simply generated by a forward pass through a neural network, conditioned by our previous experiences, inner monologue, and genetics. Free will is an illusion…

Recently, researchers at Meta and UC San Diego implemented a very similar idea, which essentially involves training large language models to reason in a continuous latent space. It made me very happy to see that it improved upon vanilla chain-of-thought.