Our history — from a poker hand to an economy of agents.
d2bot · Realize · Nandom · 2020 – 2026
We've been called a few things
Same project, same people — three names.
"d2bot" comes from the d2themfi, a world class Omaha poker player.
"realize" comes from the Equity Realizers, a group of poker players lead by Daniel.
"nandom" comes from the random kingdom of the NANDs and it's also a central hub of Ghana.
The noisy ones
…and the rest of the crew
Earlier hands who shaped the beginning: Tanguy, Craig, and Esam.
What we built
- A fast C poker simulator (≈190k hands/s/core)
- A genetic loop — clans, a king per clan and an emperor champion (a 47-situation test battery)
- Hebbian learning, subnetworks, network persistence & replay
- Hand-history pipeline into PokerTracker 4
What came of it
- Nov 24Networks evolve the ability to FOLD — "the Foldinator"
- Dec 01First full evolutionary run completes
- Dec 07Networks begin learning hand strength
- Dec 08They learn flush-draw value; strategy diversifies
The first collapse
For weeks the population kept collapsing to one dumb strategy — raise 100%.
The workaround was to remove some variance — CONSTANT_HANDS (every table plays the same hands) and SYMMETRICAL_HANDS. Then the extinction events, Clan War and Holocaust, and a bit-packed DNA genome (~185 genes in ~10 chromosomes) gave evolution something to grip — including a second layer that evolves the DNAs themselves across eras.
What we built
- The DNA genetic-algorithm system (serialization, reload, crossover)
- Distributed compute: Cowbell as master, fine-grained tasks to AWS-spot slaves — up to 144 cores alongside Bizon
- The genealogic "era tree"; FACEOFF scoring; Cowbot Slack bot
- The Receptor substrate — N×M shapes that bind by fit & distance; up to 14 receptors per gene
- Game environments: Poker, Pacman, Tetris, Gammon, Chess, Checkers, Rush Hour, Approximation
What came of it
- JanPoker backend + playable web UI deployed online
- JulPoker within 15 BB/100 of the Isham benchmark
- OctPacman networks reaching ~100+ coins
- teamMatt and Craig join
How we actually trained it
We consumed power more than ever. Cowbell — a box in Maxime's closet, still humming today — ran as the master, farming long and heavy runs, spawning batch jobs to a fleet of AWS-spot slaves. At peak, ~144 cores ran in concert with Bizon.
A lot of data comes with great responsability: visualization. We worked with some dutch guys to spin up a Spotfire server (on yet another machine next to Cowbell)
Are we gamers ?
After poker we built a variety of environments to test whether the same algorithm generalized: Pacman, Tetris, Backgammon, Chess, Checkers, Shifumi, Rush Hour, Approx. Even with variations on some of them.
- Pacman"basically solved" the simple board (14357 fitness)
- Shifumievolved real conditional strategies — and exposed the "idiot network" problem
- Aproxsolved f(x) = x² among other simple functions
- the restchess · checkers · backgammon · Tetris · Rush Hour were built as environments, never fully trained
Networks at play
The evolved networks, actually playing — every frame is a network making its own moves.
Chess, checkers, backgammon & Shifumi had environments too — however no GIFs to show, this was on the former website.
What we built
- RuledNetwork — LHS→RHS graph-rewrite rules; specificity; output rules vs "stop thinking"
- TRANSFER & FUSE reproduction; the "memory market"; UniversalNetwork
- AWS auto-scaling evolution; PostgreSQL analysis DB
What came of it
- Feb 22First RuledNetwork plays Pacman
- Mar 01First RuledNetwork with a clean strategy
- MarPacman networks reaching ~12k fitness; (4-letter alphabet, 20 rules)
- teamEsam and Asad join
What is a RuledNetwork?
Instead of a fixed network with tuned weights, a RuledNetwork is defined by a set of graph-rewrite rules — little patterns of the form left-hand side → right-hand side. Each step it finds where a rule's LHS matches its graph and rewrites it into the RHS, so the network literally rewrites itself as it "thinks".
- inputfed in as rules; a decision runs up to ~50 rewrite steps
- outputspecial "output" rules fire — or the net emits stop thinking
- orderrules apply by specificity
- growreproduction by FUSE / TRANSFER; evolution shapes the rule set, not the network
The whole point of the era: stop designing networks, start designing the rules that grow them. Later reframed as the UniversalNetwork — going always closer to Ω.
What was worked on
- Automata networks & "intelligence as search"
- The omega operation → FSMs that simulate NAND → computational completeness
- String/graph languages with non-local rules
What we built
- Daniel and Nadya hands in Python
- BFF & Z80 program-soup experiments
What came of it
- JulEmergence of self-replicators in the soup
- AugWorking V2 of the zff_enviro environment effects
What we built
- The Agent class → the agents platform
- AgentNetwork — Multi instruction agents exchanging messages
- OIANetwork — One Instruction Agent Network
- Resource allocation — contribute to network output to survive
- Results viewer (per-user & AWS); CodeScientist and LLM trials
What came of it
- FebSingle-agent system operational; cyclic behavior confirmed
- Oct 28First agents run on AWS
What we built
- Prices (read & write), Resource transfer, the RWC network
- Reachability matrix; absolute pricing; in-place ops (INCR → INCR*)
- Our own customized LLM chatbot and helpers
What came of it
- Jun 19Official Nandom logo finalized (?)
- Jul 20Very first in-person meeting
The architecture, one line each
Every incarnation the project passed through — the throughline is a slow retreat from designing networks toward designing the pressures that grow them.
- 2020Foldinator — genetic neural nets evolve to play Omaha poker
- 2021Era tree, Isham — evolution distributed across machines, genealogy tracked across "eras"
- 2021Receptor / CA — grow a network from a cellular-automaton substrate
- 2021Games — Pacman as a generic task; poker stops being the only game (bizon_evolution)
- 2022RuledNetwork → UniversalNetwork — networks defined by graph-rewriting rules, on the cloud
- 2023omega / automata — FSMs that simulate NAND; a proof of computational universality
- 2024The soup — Study self-replicators emergence from random programs
- 2025OIANetwork + agents — agents bid for resources; the economic core; GUI
- 2026Prices & RWC — read/write prices become the dominant dynamic
Things we tried and killed
Six years leaves a graveyard. A few that mattered:
- 2020Hebbian learning — a learning net can't be cloned across tables; it broke parallelization
- 2021The self-contained task payload (socket-sharing killed it) → a synced shared folder
- 2022TRANSFER gene-transfer, and a GC that networks learned to exploit to dodge memory tax
- 2025RTI resource-transfer instructions → replaced by peer-to-peer transfer
- 2026Company spirit (?) — who knows what Dublin will yield
Six years in one sweep — the substrate mutating (nets → automata → soup → agents), the lineage growing and pruning, the commits filling in beneath.
Turning points
The day poker stopped being the only game. A generic task slots in — and the project quietly becomes about computation itself, not cards.
RuledNetwork / UniversalNetwork: we stop designing networks and start designing the rules that grow them.
Self-replicators emerge from random programs — structure arising from nothing but pressure. One of several peaks, not the last.
Recasting everything as agents that bid for resources turns a research engine into a platform.
🗡️ Conan
In 2021 the multi-agent runs were competitive, intense, and full of emergent behavior — so this line kept showing up in the chat. The unofficial company catchphrase ever since.
🍲 The soup
It began in 2024 as a pot of random programs left to mutate until self-replicators emerge — analog to the primordial soup of life. The name stuck and grew: the whole agent population became the soup — often Maxime's soup, because he worked on it.
And then there's Nadya's chicken soup — the oil bubbles pooling on its surface look exactly like the subgraphs forming in ours. 🍜
Ω Omega
Maxime's Omega — a single global pseudorandom function (ψ) that updated every node uniformly. Elegant, universal, and eventually superseded by Chi (node-specific rules). The logo would have been Omega based, if no external force was constraining him.
"Plus Maxime would probably charge us a huge fee — 100 million at least." — Daniel.
φ The golden ratio, hiding in NAND
Sep 2025: studying how NAND biases the soup toward ones (~62% / 38%), Maxime and Nadya found the split was no accident — the steady state is literally powers of ψ ≈ 0.618, the golden ratio's reciprocal. NAND dynamics obey ψ² + ψ − 1 = 0.
Daniel had asked an LLM whether NAND and φ were related; it said there was no known relationship. The mathematics said otherwise.
◠ Sticks, cycles & trellises
A quiet through-line from 2023 on: treat the deterministic bidder→target map as a functional graph. A trajectory is a "stick" that falls into a "cycle" (a sink loop). Nadya exported the run data and rebuilt it in Mathematica to see which cycles form — the "timetrails", written up as "trellises".
The regimes even have names: ADD ossifies into self-pointing sinks; NAND coalesces toward a few sinks. Daniel first posed "hard-to-find short sticks/cycles" back in 2023.
…and a few more
- IshamThe poker benchmark everyone measured against (Isham.csv lives on to this day)
- RoyaltyScoring by Emperor → Dynasty → clan kings; later the logo proposals: "sovereign-nand", "throne-of-logic"
- NANDA universal gate — and now the name and the logo
🐄 Bizon, Cowbell & Cowbot
- BizonThe first server — in Daniel's house in Texas. Loud, hot, and it ran the earliest distributed evolution (bizon_evolution in the 2021 code).
- CowbellIts small successor, in Maxime's closet. Still humming today.
- CowbotThe Slack bot that lives on Cowbell — its own repo, still pinging after all these years.
🐕 The dogs
- ArèsMaxime's dog — a fixture since 2021, through French heatwaves and all. Halloween 2025: "me and Arès are too good — a lot of kids are hesitant and one kid even started to run away."
- Daniel'sA pack in Colombia — "not allowed upstairs in the office too often" — Macy, Rose, Ranger, Honey, Cloe.
"I feel there has been too few dog pictures posted for a while now, so here is one." — Daniel · 2021
For the record
Straight from the lab
The screens we actually stared at, 2020 → 2026.
By the numbers
Nov 2020 → Jun 2026.
From a poker hand to an economy of agents.
d2bot → Realize → Nandom · 2020 → 2026 — and the next chapter is being written now.