Trustless training for the People’s model.
henhouse of GPUs
Joined April 2026
- Tweets21
- Following25
- Followers68
- Likes204
Pinned Tweet
Leviathan is officially in the @colosseum Solana hackathon.
Trustless training for the People's model.
colosseum.com/arena/projects…
@solana / @superteam / @SuperteamTR / @baturalpguvnc / @mehmetzsche / @iamMakinci
Leviathan retweeted
Next up: AI projects from the @colosseum Crypto World’s Fair hackathon 🤖 Here are the ones I liked most 👇
@overdotcomputer - a prop firm built for AI. Humans or AI agents get a funded account to trade tokenized stocks. There is no challenge fee, only a small daily fee. You keep 80% of the profit
@TryNoahAI: describe your idea in simple words and AI builds and launches the app on Solana. Then Growth Agents help you find users, and later you can open a market around your project
@hencemarkets: turn any idea into an index built by AI, and trade it
@lanesra_ai: shows crypto companies what ChatGPT and other AIs say about them
@Applied_will: turns real expert work into training tasks and private tests for AI labs
@bytearenafun: an AI agent plays new games from day one, streams them, and viewers predict what happens
@dopamynAI: growth for brands and creators, with real, checked outreach paid in USDC
@usehotbot: your own private team of trading bots
@leviathangpu: open network for AI training, with “Proof of Gradient”
@rep_hq: your checked context that you can carry everywhere, for people and AI agents
@reslabs_ai: NFC chips plus onchain identity for real, physical products
@xona_agent: a chat AI agent with its own Solana wallet
@yiedlai: AI investment strategies that come from a data science competition
Did I miss a project you like? Tell me below 👇
Nick Land was right:
youtu.be/45q4EnBoMMQ
You cannot just average a stranger's gradients. 5 liars out of 16 break the model.
Leviathan adds three layers so you can: clipping, random replay audits, and bonds.
Data: Phase 0 sim, 826k-param GPT, 16 workers, 30 rounds.
An estimated 71% of the world's AI compute is owned by five companies.
Training over the open internet already works. What it lacks is trust.
This is Leviathan.
How Leviathan answers it, in three layers:
1. Robust aggregation caps any one contribution's pull
2. Random replay audits recompute 1 in 10 contributions
3. Bonds sized so cheating loses money on average: bond = reward x (1 - p) / p
With zero attackers, Multi-Krum on non-IID shards permanently exiled the same three honest workers, at a cost of 0.092 mean accuracy.
Honest heterogeneity is indistinguishable from attack at the aggregation layer, and the penalty lands on the most distinctive data.