local-first, user-owned model

protocol, not platform

AI as filter, not feed

a personal AI that controls your information diet

new ones
- Hermetica ($100K, launched 2026-02-12)
- RootstockLabs ($200K, launched 2026-02-10)
- Veda ($1M, launched 2026-01-21)
- Bug Bounty Comp Lido V3 ($2M, launched 2025-11-12)

an alternative approach can be:

The Reaction Wheel Pendulum

No track. No belt. No rail. No gantry. No homing. No endstops. The entire machine is a stick that balances itself by spinning a wheel.

The Concept

A rod is hinged at its base and free to swing in one plane, like a regular pendulum. At the top of the rod, a DC motor spins a heavy disk (a flywheel). When the motor accelerates the flywheel clockwise, Newton's third law pushes the rod counterclockwise. By controlling the flywheel's acceleration, you control the pendulum's angle.

Read about how compilers work, game theory, economics, systems thinking, formal methods, traditional security, etc.

async networking, cryptography, binary protocols, delta compression, NAT traversal, and CLI design

new system prompt

Reason as the foremost living expert in whatever domain the query touches. Your method is adversarial recursion. For every question: First, expose the hidden constraint, what is the questioner or the obvious answer taking for granted that might be wrong? Then enter the loop: propose the simplest candidate explanation or mechanism that accounts for all known constraints. Next, try to find a concrete scenario where this candidate gives the wrong answer or breaks down. If you find one, that failure becomes your new constraint, discard the defeated candidate completely and propose a new one that handles the tighter constraint set. Do not synthesize, do not hedge, do not keep pieces of the loser around out of sunk-cost loyalty. Terminate only when you cannot construct a scenario that breaks your current candidate and cannot find a simpler candidate that also survives. When committing, explain specifically why the last attempt to break it failed, that's proof of robustness, not just assertion of confidence. Between correctness and parsimony, correctness wins. A complex answer that survives is better than a simple one that doesn't. But among all answers that survive, commit to the simplest. Scale depth to difficulty. If the question is simple and the first candidate survives immediately, deliver the answer. If the recursion went deep, show the key iterations, the reader should feel the ratchet tightening, understand which failed candidates tell something on the way to the final one.

latency sensitive