How Macrocosmos got from a contest to liquid training
These are all the research papers the Macrocosmos team has published, each explained for people who know AI models exist but have never looked at how one is built. Macrocosmos runs IOTA, the training network on Bittensor subnet 9, and calls its approach liquid training: one model trained across scattered, mismatched machines over the public internet.inferred
Read in order, three of the papers tell one story. In 2024, miners each trained a whole model and the best one was paid. In 2025, one model was split across many miners, each paid for their part. In 2026, the handoff between those miners was made small enough for ordinary internet links. The fourth paper is a side track on how to pay miners for work no one can score. Below, each one is laid out as its first page, annotated.
The desk · four first pages, in the order they were written
Pick a page, swipe, or use ← and →Who cites whom
From each paper's reference list · hover or tap a node- Papers
- 4every one Macrocosmos listsstated
- On arXiv
- 2the others self-publishedstated
- Span
- 20 monthsAug 2024 to Apr 2026inferred
- Named authors
- 7Cruz on all three that name peopleinferred
- Compression reached
- 128xprimer and ResBM, counted differentlystated
All four as a table · newest first
| Paper | Date | Authors | In one line | What it gave IOTA and Orion | Full text | |
|---|---|---|---|---|---|---|
| ResBM: Residual Bottleneck Models for Low-Bandwidth Pipeline Parallelism | 13 Apr 2026 | Aboudib, Lopez Portillo A., Brady, Cruz | A learned encoder and decoder at each pipeline boundary shrink the handoff 128x while the residual path stays intact. | The compression block behind the Orion runs; Orion-100B used it at 64x.stated | arXiv 2604.11947 | Read → |
| Generative Adversarial Mining on Decentralized Networks | 9 Jan 2026 | Quinque, Brady, Cruz | Pay miners for quality no one can score, by making them guess whether a miner or the validator wrote each answer. | Nothing directly: it is the pay rule for Apex, subnet 1. It describes IOTA as "code attestation". | Wayback copy | Read → |
| Incentivised Orchestrated Training Architecture (IOTA): A Technical Primer for Release | 16 Jul 2025 | Quinque, Aboudib, Fonau, Lopez Portillo Alcocer, McCrindle, Cruz | One model split across many miners, paid per verified backward pass, merged by butterfly all-reduce. | The architecture itself: orchestrator, pay rule, merge, and the first bottleneck block.stated | arXiv 2507.17766 | Read → |
| LLM Pretraining: The Use-Case Blockchain Has Been Waiting For? | Aug 2024 | Macrocosmos, Taoverse, Const, Datura (no individuals named) | Subnet 9 as a winner-takes-all contest in which each miner pretrains a whole model, up to 7B. | The starting point: the primer cites it as reference [1] and names its two flaws.stated | macrocosmos.ai PDF | Read → |
Where to start
New to how models are trained: start with ResBM. Its first chapter explains pretraining, the training loop and pipeline parallelism from scratch, and the other pages link back to it.
Following IOTA's history: read the training line in order, SN9, then the primer, then ResBM, then the runs in the Orion Register.
Interested in incentive design: Generative Adversarial Mining stands on its own.
How this list was made
A paper is included if Macrocosmos is its affiliation. The four here are exactly the four in the Research menu on macrocosmos.ai, checked on 29 September 2026 against arXiv searches for each author, an arXiv full-text search for "Macrocosmos", and Semantic Scholar.stated
The first pages above are drawn, not scanned: titles, authors and affiliations are the papers' own; each abstract is paraphrased in my words, and the grey lines stand in for body text. The citation map reads each paper's reference list.
Every explainer was written from the full text. Corrections to connect@mikyo.one.