The Art of Knowing
The Art of Knowing
When LLMs started to feel useful, the excitement around the tech was palpable. They were surprisingly good at answering personal questions. They felt natural to interact with and, for a while, it was easy to imagine them capable of handling much harder problems, like research and automation.
Spending time inside these systems, though, made me think more carefully about what kinds of intelligence they represent, and what kinds they don’t.
Most of the tasks we face each day are not purely language problems. Some require spatial or dimensional reasoning. Others depend on logic, planning, or emotional awareness. Much of what we do cannot be fully verbalized at all. We act, observe the result, and adjust. We change our behavior depending on the situation, the goal, and the constraints around us. Intelligence shows up differently depending on the task.
Our brains are good at this because they constantly reshape themselves. Useful connections get reinforced. Irrelevant ones fade. Knowledge learned in one setting gets carried into another, often in unexpected ways. That ability to extrapolate, to be creative rather than repetitive, is what makes intelligence feel alive rather than mechanical.
That difference kept bothering me.
While working on my Ph.D. at U.C. Santa Barbara, I moved between fields studying intelligence from different angles. The details varied, but one thing stayed consistent. Intelligence is not static. It is adaptive. It plans across different levels of abstraction. It learns from interacting with the world, not just from consuming information. It evaluates its own behavior and changes based on experience.
That perspective made it feel natural to explore energy-based approaches to intelligence.
Energy-based models offer a way to represent goals, constraints, and behavior without tying intelligence to a single modality like language. An energy function can be attached to many different latent spaces, each shaped by real-world data. Those spaces can represent physical systems, logical structures, workflows, or environments that evolve over time. What matters is not predicting the next output, but understanding whether a system’s state makes sense given everything else.
For a long time, I thought I would stay in academia to explore my ideas about energy-based AI models. Then I met someone who convinced me that these systems needed to be built, not just studied. Business makes you accountable to reality in a way whitepapers never do.
AI is still in its infancy. In the years ahead, it will become part of almost every human endeavor. The question is what kinds of systems we decide to trust, and where. Language models are not a mistake. They are very good at what they do. But they are just one kind of intelligence, and many problems require another approach.
We are not trying to replace existing tools, and we are not trying to replace people. Language models remain extremely useful, and they will continue to be. Intelligence, though, is broader than any single approach. AI should exist in many forms, each suited to different kinds of work.
It is still strange, and still exciting, to be working on questions this basic and this practical at the same time. That combination is what keeps me here.
When LLMs started to feel useful, the excitement around the tech was palpable. They were surprisingly good at answering personal questions. They felt natural to interact with and, for a while, it was easy to imagine them capable of handling much harder problems, like research and automation.
Spending time inside these systems, though, made me think more carefully about what kinds of intelligence they represent, and what kinds they don’t.
Most of the tasks we face each day are not purely language problems. Some require spatial or dimensional reasoning. Others depend on logic, planning, or emotional awareness. Much of what we do cannot be fully verbalized at all. We act, observe the result, and adjust. We change our behavior depending on the situation, the goal, and the constraints around us. Intelligence shows up differently depending on the task.
Our brains are good at this because they constantly reshape themselves. Useful connections get reinforced. Irrelevant ones fade. Knowledge learned in one setting gets carried into another, often in unexpected ways. That ability to extrapolate, to be creative rather than repetitive, is what makes intelligence feel alive rather than mechanical.
That difference kept bothering me.
While working on my Ph.D. at U.C. Santa Barbara, I moved between fields studying intelligence from different angles. The details varied, but one thing stayed consistent. Intelligence is not static. It is adaptive. It plans across different levels of abstraction. It learns from interacting with the world, not just from consuming information. It evaluates its own behavior and changes based on experience.
That perspective made it feel natural to explore energy-based approaches to intelligence.
Energy-based models offer a way to represent goals, constraints, and behavior without tying intelligence to a single modality like language. An energy function can be attached to many different latent spaces, each shaped by real-world data. Those spaces can represent physical systems, logical structures, workflows, or environments that evolve over time. What matters is not predicting the next output, but understanding whether a system’s state makes sense given everything else.
For a long time, I thought I would stay in academia to explore my ideas about energy-based AI models. Then I met someone who convinced me that these systems needed to be built, not just studied. Business makes you accountable to reality in a way whitepapers never do.
AI is still in its infancy. In the years ahead, it will become part of almost every human endeavor. The question is what kinds of systems we decide to trust, and where. Language models are not a mistake. They are very good at what they do. But they are just one kind of intelligence, and many problems require another approach.
We are not trying to replace existing tools, and we are not trying to replace people. Language models remain extremely useful, and they will continue to be. Intelligence, though, is broader than any single approach. AI should exist in many forms, each suited to different kinds of work.
It is still strange, and still exciting, to be working on questions this basic and this practical at the same time. That combination is what keeps me here.