And why LLMs fundamentally can’t. There’s a capability we keep pretending AI already has: The ability to reset. Not retry. Not re-prompt. Not “let’s think step by step.” A real reset. LLMs Don’t Reset — They Commit LLMs operate like this: context → next token → next token → next
Everyone is excited about world models. The idea is compelling: learn a latent representation of reality, predict forward, and make better decisions than systems trained on surface-level data. But in healthcare, most world model efforts are going to fail. Not because the idea is wrong — but because the implementation will
Two very different ways to build software with AI AI is rapidly changing how software is written. Over the past couple of years, a new style of development has emerged where developers describe what they want and the AI generates the code. Many people call this vibe coding. It is
Every week we see claims that large language models are approaching “superhuman intelligence.” They pass bar exams, ace medical questions, summarize research, and explain quantum physics better than most professors. In terms of raw cognitive output, LLMs already exceed the capabilities of: * medical specialists * legal scholars * research scientists * mathematicians * historians
We all talk about emergent properties in evolution, intelligence, and now AI. But most explanations focus on surface-level mechanics—complexity, mutation, selection pressure, etc. Those matter, but they’re not the root cause of emergence. Emergence appears when two deeper conditions exist: **1. A Subconscious + Conscious Divide (Selective Focus Without
We’ve reached a strange moment in AI where the world is both mesmerized and terrified by large language models. They write code, draft emails, pass exams, and even feel conversationally “smart.” But beneath the surface, every LLM—no matter how powerful—shares one fundamental truth: 👉 LLMs are next-token predictors.
If you follow modern AI for robotics and agents, you’ve probably noticed the phrase “world model” popping up everywhere. In 2025, Meta’s V‑JEPA 2 made that trend impossible to ignore: a self-supervised video world model trained on over a million hours of internet video, then adapted to
I’ve spent a lot of time being annoyed at missing data in healthcare. Patient monitoring streams that drop out for hours. Vitals that come in bursts. Labs that appear once every few days. EHR time series that look like someone took scissors to them. For years my mental model
I’ve been thinking a lot about how little we genuinely understand about the forces that shape our lives, especially when it comes to health and human connection. The U.S. Surgeon General recently published an eye-opening advisory on what he calls the “epidemic of loneliness and isolation,” and the