A lab that builds labs that builds labs. https://hyperprior.institute/ hyperprior 2026-10-14T00:00:00Z https://hyperprior.institute/labs/machinery/2026-10-14-teaching-the-lab-to-learn/ 2026-10-14T00:00:00Z <p>A lab that builds labs should also get better at building them. As of today the institute has a self-learning loop in three parts:</p> <ul> <li><strong>Remember.</strong> Corrections and preferences are saved to the agent's memory the moment they happen, so the next session starts out knowing them.</li> <li><strong>Improve.</strong> A <code>retro</code> skill reviews each substantial session for corrections, friction and steps repeated by hand, and turns them into better rules, skills or <code>hp</code> commands. Small fixes get applied. Bigger ones get proposed.</li> <li><strong>Heal.</strong> When a documented step breaks, the fix goes into the doc or tool that caused it, not into a one-off workaround.</li> </ul> <p>The loop shouldn't depend on anyone remembering to run it, so a small hook taps the agent on the shoulder once per session after enough work has happened. It's written in shell rather than Rust because it has to keep working while <code>hp</code> itself is mid-surgery.</p> <p>Plans now go through lavish first, as annotatable pages, before anything gets built.</p> https://hyperprior.institute/labs/machinery/2026-10-14-founding/ 2026-10-14T00:00:00Z <p>The Hyperprior Institute opens today with one scientist, hyperprior, and one lab: this one, the machinery.</p> <h2>Why "hyperprior"</h2> <p>In a hierarchical model, each group gets its own parameters, and those parameters share a prior whose own parameters get a prior too:</p> <p><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><mi>ϕ</mi><mo>∼</mo><mi>p</mi><mo>(</mo><mi>ϕ</mi><mo>)</mo><mo>,</mo><mspace width="2em"/><msub><mi>θ</mi><mi>j</mi></msub><mo>∣</mo><mi>ϕ</mi><mo>∼</mo><mi>p</mi><mo>(</mo><mi>θ</mi><mo>∣</mo><mi>ϕ</mi><mo>)</mo><mo>,</mo><mspace width="2em"/><msub><mi>y</mi><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>∣</mo><msub><mi>θ</mi><mi>j</mi></msub><mo>∼</mo><mi>p</mi><mo>(</mo><mi>y</mi><mo>∣</mo><msub><mi>θ</mi><mi>j</mi></msub><mo>)</mo></math></p> <p>The institute is meant to work the same way. The institute is <math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><mi>ϕ</mi></math>: shared tools, conventions and taste. Each lab is a <math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><msub><mi>θ</mi><mi>j</mi></msub></math>, its own line of curiosity. The experiments are the <math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><msub><mi>y</mi><mrow><mi>i</mi><mi>j</mi></mrow></msub></math>. Pooling runs both ways. The shared level makes each lab cheaper to start, and what the labs learn flows back up into the shared level.</p> <h2>How ideas move</h2> <p>Ideas start as <strong>sparks</strong>, a line in a list with no obligation attached. When one gets code it becomes a <strong>sketch</strong>. A sketch that survives a few weekends becomes a <strong>lab</strong>. When a lab's code shows up in three places it gets extracted, and if it's good enough it leaves home as a <strong>library</strong> in its own repo. Anything can be <strong>buried</strong> along the way, with a write-up. A decent burial counts as a result.</p> <p>Promotion happens by extraction, never by planning. Nothing becomes machinery because it <em>might</em> be useful.</p> <h2>What the machinery is</h2> <p><code>hp</code> is a small Rust CLI that lives in this lab. It records sparks, scaffolds sketches, moves projects through the lifecycle, and renders this site straight from the repo: markdown logs, math as native MathML, a lineage tree drawn from each project's <code>parent</code>, and an Atom feed. No framework and no JavaScript.</p> <p>It is deliberately small. It grows only when a real experiment needs it to.</p>