In Praise of Basic Instinct

We are awash in language models. They are driving speculative investment and increasingly they are driving you to buy certain products, to select certain media outlets, to find new friends online. Words drive thought, thought propels language and now large language models drive you. LLMs drive just about everything except, ironically, the one thing we wanted them to drive – our cars.

Language encapsulates our conscious thought, but much of what we do that keeps us alive and gives us joy is not reducible to semantics. You touch a hot stove and your arm jerks back before the signals gets to your brain. You touch someone next to you and sparks fly, making your spine tingle. Many of our motor responses are encoded in our spinal cord and much of what we consider skill resides in a combination of our brain stem and cerebellum.

I’ve coached many youth sports teams. If you tell a seven year old how to kick a soccer ball seven different ways they still don’t really know how to do it. Or rather they know in terms of language, but not in terms of skill. If you tell them seventy ways it still doesn’t mean they have the skill to be a great player. What they need is not words, not models, not even demonstrations. They need practice to encode the knowledge implicitly as skill. Robots are no different. For AI, understanding the components of a good soccer play doesn’t mean a robot can play well. I remember the old days in 1999 watching CMU and Dartmouth robots battle it out on a real (small) pitch. It was wild to see the robots score and even wilder to see how loudly the excited nerds (like me) yelled in response. Those robots were smart, but they weren’t athletes. Not even a little.

Athletes need muscle memory – the “motion intelligence” that encodes practiced skill implicitly as behavior. When the Sundance Kid, played by Robert Redford, decides to “go straight” and protect the stagecoach (rather than rob it as usual), he finds that he doesn’t know how to hit a target while standing still. All of his gunslinging skill is encoded as instinct and that instinct was learned in motion. Standing still, he can’t hit a thing, but when his partner Butch throws a can in the air, the Sundance kid sends multiple bullets through it before it hits the ground. He didn’t have time to think as muscle memory took over.

Lionel Messi is undoubtedly brilliant, but his scintillating plays are less a sign of cognitive IQ and more a sign of “motion intelligence.” He knows where to be and uses instinct based on nearest neighbor motion... not language, not models, not plans. ?If Messi mapped out where everyone was and selected an optimal play from a model, it would be too late. AI could learn a thing or too from Messi. AI is optimal, but struggles with things not already in the map: snow banks, construction patterns or pedestrians. AI book learning needs more street smarts, but how do we do this computationally?

Every great athlete, yearns to let go of conscious control and enter “flow state.” Musicians, composers, writers, artists… we all crave that flood of unconscious inspiration as our unfettered intuition and instinct converge, taking us into “the zone.” ?What would it mean for robots to do the same? To shed their dependence on the cloud-based maps and follow the robot up ahead… To localize off of their peers rather than satellites in space… ?To prioritize the ebb and flow of forklifts and humans all around them over their pre-planned route through the logistics center. To do this they’d have to think beyond language models

We are becoming increasingly driven by tech, and tech is increasingly driven by large language models. Language is valuable, but it is only a part of the picture we need. Perhaps anxiety is on the rise because humans were not meant to be so language driven. Social media is constant stream of words, analysis, comparisons and contrasts. You are not an individual computer. You are a swarm entity called a human, meant to be in motion, meant to be in flow. The ancient Gnostics believed that language is a trap. It confines us to what we can express explicitly. No matter how good you are with words, words are only a glimmer of truth. Binding truth to language hinders our ability to see and seek truth. Many religions recognize that ineffable truth may elude our conscious mind, but be revealed to us in other ways.

Don’t care about the history of religion? Biology is ample evidence that language is not the only source of intelligence. Ants can build entire cities without explicit language or semantic plans of any kind. An increasing body of work shows that swarm intelligence solves a host of problems more reliably than individual human intelligence. Just read social media to see that human intelligence and human language does not always serve us well. More than ever, we need an antidote to the stream of fake news. But boy, we sure love language. I am using it right now. I’m trying to speak truth, but it’s hard.

Real truth is felt. It is felt in our gut and in our bones. It is felt in the way our amygdala rouses us when threatened. It is a combination of hormones and electrochemical surges. You don’t think your way into love. You fall into it. We are not just language in motion, but hearts and sinew and sweat seeking beauty and truth. In Descartes Error, Antonio Demasio argues that we are not the cognitive machines we imagine ourselves to be. Our neurons are anything but a deterministic process and he sees that as a good thing. We love the things we don’t understand and we crave the truth we cannot grasp. ?

Unlike a perfect pass or a sultry kiss, large language models do not give us joy. Perhaps that is not a fair request of AI models. Is it enough to know they are adaptive, learning machines? AI models are now so good that we are likely to forget that they are not based on real-world truth. Just because humans have said something in the past doesn’t make it true. Our dependence on LLMs has displaced the foundation of truth which all biology once held dear: real-world perception. Truth is of course a construct of our minds… but if you get hit by a car… that impact sure feels like truth.

Robotics could hold the key to linking AI to the real world. In robotics terms, ground truth is the key to reliable behavior and getting things done. Physical AI could reconnect our computational intelligence to the real world. Robots shouldn’t operate merely on text, but rather on real world cause and effect. I believe collective motion intelligence is the key to scaling physical AI. Many robots can see what one cannot. Like ants, AI could embrace a swarm intelligence model where diversity and chaos become key to assembling the many facets of truth. When we think in swarm terms, the entire concept of truth changes. Truth is not the one person or perspective who we judge to be right. Truth is the amalgam of different perspectives that contribute to a complex and holistic view. Truth takes context into account. Swarm truth isn’t in your five-year forecast, but in the contingency plans worked out with shareholders and employees late at night.

Perhaps we need to rethink AI’s fixation on language as the only source of truth. The most important things in your life can’t be dictated with language, rules or plans… who you love, how you sacrifice for your children, the feeling of holding someone. AI is all about giving you an answer. What if that answer is insufficient to move you? Moving you is what matters. Motion is what matters. Truth is beauty in motion. Even for robots. Truth is of course subjective, but what I’ve learned in 25 years of fielding robots is that for things that move truth is simply perception that works. You know the truth of an apple tree only when you move an apple to your lips.