I am a programmer—a full-stack one, and for years people called me an engineer. At Rackspace they meant it as a compliment: very technical, very good. I always pushed back. I never felt like an engineer. I felt squishier than that.
My way of thinking runs toward biology. Things are less deterministic, more non-deterministic—messier in the ways living systems are messy. I think about biology more than algorithms, even though algorithms have paid the bills for however many years. The nature of brain matter, how we think, why we think, the differences between us and monkeys—that is much more interesting to me, and much more understandable.
LLMs are playing a role in all of this, but they are not the whole story. A world model is something else: the kind of thinking Yann LeCun has been arguing for—more grounded in the real world, more biological in spirit. I do not think the answer is to pick a side. The best system, in my view, is a little LLM, a little biology, and a lot of harness.
The harness is already the big thing. It is just not always said plainly: you need a harness that can use LLMs and LeCun-style world models together—a world-based methodology, not a single monolithic brain in a box.
The harness is like the spinal cord. It holds everything together. It reaches across subsystems and makes things actually happen. Without it, you have clever parts that never coordinate.
A crude heuristic I use for LeCun, JEPA, and LLMs maps onto an old split people used to talk about: left brain, right brain, and the spine between them. The folklore was never neuroscience-precise, but it was useful. One side pattern-matches language and symbols; another side grasps space, motion, and embodied prediction; the spinal cord is what keeps the organism upright and moving.
LLMs excel at the symbolic, fluent, left-hand kind of work. JEPA and related world-model ideas lean toward prediction in a more embodied, right-hand direction—what happens next in the world, not just the next token. The harness is neither. It is the infrastructure that wires them together and decides when to call which.
That framing is oversimplified. Brains are not two hemispheres plus a cable. But as a design intuition it helps: do not expect one model type to do everything, and do not underrate the boring connective tissue that makes a system behave like one mind instead of three plugins.
I still write code for a living. I still love a clean algorithm when one fits. But when I imagine where useful AI is going, I picture something squishier—biology-informed, world-grounded, held together by a harness we have barely started to build well. The interesting work is not only making models smarter. It is making the spine strong enough that different kinds of intelligence can share a body.