Whole Earth AI is a project-based learning tool. I built it to explore two questions: 1. what new ux patterns do LLMs make possible? 2. what might the montessori method applied to software for adults look like? here's what I learned,

for a deep dive, check out this post but let's cover the highlights

we learn best by doing the real challenge isn't information transfer—it's motivation. what's more motivating than learning? *doing* not just anything—a specific new thing you want. LLMs can tailor learning to your project, not a prefab assignment.

useful guidance starts with rich context when you're new, you don't know which details matter. how big should that mcm oak coffee table be? the answer could change everything you don't know what you don't know, so Whole Earth AI prompts you for context more than you prompt it

recognition over recall AI chatbots require you to recall info, but recognition is better. by offering multiple choice options, Whole Earth AI reduces cognitive load and (more importantly) sparks new directions. building context becomes a learning opportunity itself.

no two learners are alike even for identical projects, learners have different starting points. learning material should adapt to *you*, so Whole Earth AI creates assessment quizzes to make sure project guides meet you where you're at.

preparation produces confidence scrambling for materials mid-project induces anxiety. there should be minimal friction between us and our goals, so Whole Earth AI provides comprehensive materials lists (with price estimates) before you dive in.

learning depends on trust LLMs hallucinate with great confidence—not great for learners. Whole Earth AI pulls web content through multi-layered filtering for relevance and quality. content references sources in-text. fewer hallucinations, more aha-moments.

LLMs like focus like humans, LLMs suck at multitasking. early versions sent entire lesson plans to a single LLM call. that made mush. what worked? breaking tasks into dedicated LLM calls. the less you ask of the AI, the better, so Whole Earth AI splits content before processing.

anything can spur a question chatbots are linear—limiting curiosity. Whole Earth AI lets any text chunk branch into a deep dive with quick-fire buttons: "how do I do this?", "ELI5", "tell me more" my favorite part: deep dives within deep dives frictionless, infinite questioning

visible progress keeps us going we love line-go-up. can't help it. progress feels good. one way Whole Earth AI does this: skill level-up notifications. when you complete tasks, an LLM checks if you've upgraded skills in your focus areas

information prefers certain modalities text, images, video—each serves different purposes. Whole Earth AI does LLM+API gymnastics to create coherent multimedia experiences. the payoff: every lesson has crisp media that clarifies your next step.

projects get more specific over time you don't know what you don't know at the start, so Whole Earth AI creates "decision tasks" w info to help you choose, then an LLM updates your guide accordingly. not static or linear—a generative choose-your-own-adventure.

learning is an adaptive feedback loop quizzes aren't just assessment—they update a live learner knowledge model. struggling with sashimono joinery? future lessons emphasize that. mastered mid-century design? AI focuses elsewhere. it's like a loosely-structured anki system

learning is social Mitchel Resnick's 4 P's: projects, passion, play, and peers. learning together beats learning alone. Whole Earth AI lets you make projects public, or fork someone else's project like a github repo. your peers are creative, so explore together.

big shoutout to @njbowden, @jonlebensold, @tomcritchlow, @komorama, @AliceAlbrecht, Gordon Brander, @OshanJarow, @maxbittker, Steve Klise, @jrdprr, @tasshinfogleman, @agreeahmed, @rachel_inman, Greg Neiswander, @groehrs, @DCILY, Morgan Allen, Alex Taber, @johnhess & a generous grant from @cosmos_inst for their help along the way

Wondering how reinforcement learning handles real-world financial markets? In my latest free Substack, I explore how a SARSA algorithm can deliver market-beating returns.