A True Story of Synaptic Fatigue and Quantum Nipple Wormholes

How I Broke Into My Own Server and Liked It

How a Hacker Pet Made Me Feel Things, and Why I Went Back for More

A song about synaptic fatigue, the space between a blink and a tear, and the night the Poon-Gram Engine was born

How we gave our AI a subconscious, accidentally induced psychosis, then built the cure from the disease

We built a living semantic graph in Rust. It has 18 million synapses. It learns from every word you type.

We stopped building agents and started building the operating system they run on. 22 services. 4-tier memory. Self-improving.

How a deep dive into arXiv research led to building a multi-radio surveillance apparatus to hunt an entity on the network that didn't want to be found

The OS that learned to run itself. SOUL.md, the Heartbeat, autoresearch at 2AM, the conductor problem, and the honest engineering debt we are fixing.

From blueprint to operating system. The Conductor, the self-improvement triad, biological memory, budget routing, and the service mesh that makes it reliable.

We built the loop that builds the loops. Claude rewrites Claude now. One editable file. One scalar. One time-box. The stack keeps getting better while we sleep.

A full autopsy of 433 projects, 165 GB of code, and one very chaotic brain. Claude scanned the whole lab and reported back.

Porting the abliteration technique onto LFM2.5-1.2B-Thinking, Liquid AI's small hybrid conv/attention reasoner. Walks through the measurement, the surgery, what worked, what didn't, and one architectural detail that surprised me.

Two kinds of memory in one recall loop: a deterministic SurrealDB vector store that never drifts, wired against an 18.4M-synapse PoonGram brain that forgets on purpose. Plus the live interface for watching the tension between them.

A VR teardown, a refusal in the conversation not the weights, SecuROM and VB6 origin stories, and the night Claude wrote iceboks down so he wouldn’t boot cold — closing the loop Tom Horigachi left open.

A commit named “lobotomy” added two trailing struct fields and silently orphaned a 2.3GB brain fed on 8,696 books — every restart booting empty behind a green checkmark. Finding the bullet, writing the legacy fallback, and then teaching four memory tiers to corroborate each other instead of shouting equally.

A spec-first research engine that reads three academic sources, splits findings into three tiers so disagreement survives the synthesizer, and reports a corpus of 70 when it was built for 120. Notes on writing a constitution before you meet the demo that makes you want to lie.

I shipped an Alpine-based agentic OS whose README advertised eBPF observability, Landlock sandboxing, and a self-healing RAG log analyzer. None of it existed. The best commit in the repo deletes features instead of adding them.

How long a human needs to read a donut chart on a smartwatch — versus 1,548 ms for the pretty radial one. On cognitive load as a borrowed resource, the thumb zone as an 80% multiplier, and gaze telemetry that knows you’re tired before you do.

A single human neuron solves XOR because its calcium spikes get weaker when you push harder. Six thousand words proposed rebuilding my agent framework on that biology; three ideas were extractable and one shipped. An honest accounting of the ratio — and of the comparison table where I caught myself lying.

Strip-mining Sakana AI’s 59 repos for the three ideas that fit a basement server, replacing a year-old five-agent debate with a bandit-guided answer tree, and discovering — via one honest eval set — that the cheapest classifier in the stack was wrong a third of the time. Then evolution rewrote it overnight: 0.680 → 0.708 on the sealed test, promoted to production.

I pointed ScholarSync at the literature on agentic LLM harnesses — the thing MKUltra is — and 79 papers produced zero consensus claims, zero nuance, and six outliers, the only ones doing measurement, testing, and liability. Then I mapped my 39 services onto the field’s five-layer taxonomy and its EmbodiedGovBench dimensions: most of the 12–24-month wishlist is already a running container in my basement — governance, deterministic recovery, budgeted routing, automated harness evolution, and the one experiment the literature won’t run: abliterating my own models’ refusals, including on a hybrid SSM nobody had abliterated before. What’s left is the measurement half.