01
An event-sourced reactive graph runtime for auditable, forkable, long-running agents.
Yohei Nakajima
VC / builder. A running log of agents, tools, research, and experiments.
featured
selected work02
A mathematical explanation of how information sharing without action coordination can worsen group outcomes.
03
A controlled packing study across 188 instances: gradually hardening soft shapes finds packings rigid searches miss, but tuned rigid starts win more often overall. Every run is replayable.
build log
139 / 139
build activity
last 52 weeksfilter / sort all projects
Sloptoberfest: Hit Tweet and music-video thread
A music-video thread opening with "Hit Tweet," an electropop/dancepop experiment made with Suno and Slop Cannon, with an ASCII effect added through Codex.
Unpolished Signals Matter More Now | Yohei Nakajima, Untapped Capital
On Odin's Going Solo with Dan Gray: building in public, unpolished founder signals, the shared discovery paradox, and keeping investment decisions human.
The Playbook That Writes Itself (and Other Ways VCs Are Putting AI Agents to Work)
A Venture5 event recording and article with Lindel Eakman and Carta's Scarlett Pierce on agent workflows in venture, a playbook that proposes its own updates, and human decision-making.
Soft-to-Rigid: Shape Continuation as an Alternative Search for Packing Congruent Shapes
A controlled packing study across 188 instances: gradually hardening soft shapes finds packings rigid searches miss, but tuned rigid starts win more often overall. Every run is replayable.
n=12 cube packing: video announcement
The October 8 video announcement and raw coordinates behind the separate 12-cube packing project. Novelty relative to Haowei Lin's arrangement remains disputed.
Twelve unit cubes in a cube (n=12)
A packing of 12 unit cubes in a cube of side 2.9315185094797, found through soft-to-rigid search and checked with exact rational arithmetic. Listed in Friedman's packing catalogue.
A sillier way to solve square packing
The original video demonstration of soft-to-rigid square packing: 11 squares reached side 3.877 in about 3.75% of the reported runs. A demo preceding the separate controlled study, not a new square-packing record.
From BabyAGI to ActiveGraph: Why Long-Running Agents Need Logs, State, and Proof with Yohei Nakajima
On ODSC's Ai X Podcast with Jerry Wolfe: the path from BabyAGI to ActiveGraph, durable agent state, auditable logs, and reproducible research.
Glance VLM
An open-source Python harness that reads typed answers from a local vision-language model's logits and returns probabilities without generation.
Beep with JEV
An experiment giving JEV a way to express emotion and chat through audio beeps.
JEV Graph Extraction
A JEV experiment that scores words for semantic significance, tags relevant words with IDs, and builds a graph from the highest-scoring words.
Evaluator Bench
An evidence-linked research preview that scores the independence of third-party frontier AI evaluators under adjustable weights and evidence policies.
No projects match those filters.
about
I’m an early-stage investor who builds software to explore ideas, understand emerging technology, and meet ambitious founders. This is the artifact trail.
machine-readable
Public project metadata is available without JavaScript.