GB Research

this release candidate model version is becoming available/integrated on @huggingface 🤗 👉 huggingface.co/ZenoDivergent…

Here is the simple, five-year-old version of what is happening right now: The Big Idea: The "Crowd Trap" Detector Imagine twenty different weather forecasters all predicting bright sunshine for a football game. When everyone agrees, people trust them. But sometimes, they are looking at the exact same bad satellite map — and it rains anyway. They were all confidently wrong together. Zeno's job is not just to guess the weather — it's to watch the crowd and raise a red flag when they are walking into a trap. How We Are Teaching It We collected millions of real forecasts from the past across thirteen different worlds: - Tomorrow's temperature and rain - City power grid demand - Financial and crypto markets - Economic reports and disease outbreaks Because these events already happened, Zeno can practice. It looks at what everyone was saying before the event, tries to guess what actually happened, and sounds an alarm when it suspects the crowd is making a collective mistake. The 12-Round Tournament Happening Right Now Instead of training just one giant model and hoping for the best, the supercomputer is running a 12-round experiment: Rounds 1–3 (The Main Model): Three identical medium-sized brains started from different random seeds. If all three give the same answers, we know the model isn't just getting lucky. Rounds 4–6 (The Smaller Model): Three smaller brains to see if we can get the exact same accuracy with less complexity. Rounds 7–9 (The "Blindfold" Tests): We turn off one feature at a time — like hiding which forecasters belong to the same parent company or hiding their past track records — to prove which clues actually matter. Rounds 10–11 (Solo vs. Team): We train a model only on weather and only on crypto, to prove that learning across all fields together makes it smarter than studying just one topic. Round 12 (The Historical Check): A strict backtest against older history to make sure it holds up across different seasons and years. Where It Stands Right Now.. First tests are finished: The first model completed its training and was graded on 33,247 held-back practice questions it had never seen before. Early report card: On yes/no questions, its error rate was lower than the crowd average across the board, and the answers between different seeds lined up closely. Currently running: The supercomputer is actively rolling through the remaining tests in the schedule. Once all 12 rounds finish, we pick the most reliable model and test its advance warnings on live, sealed events whose answers haven't happened yet.

Our potential application blueprint in 6 sections 🤔 — 1. What is your company going to make? Zeno is a metacognitive intelligence model that predicts when forecasters fail. Instead of building another prediction model, Zeno analyzes panels of diverse forecasters (AI agents, human experts, and physical ensembles) to detect fragile consensus, issue advance failure warnings, and correct group forecasts across weather, energy, and markets. — 2. What is new about what you make, and how is it different from existing solutions? 🧵👇:

We ran this exact contract ($126,210.50 touch by Oct 28, 2028 from ~$80.5k spot) across our multi-model panel + prediction markets: - 4-Family Pool: 60.8% YES (@GrantCardone wins) - Polymarket/Kalshi implied curve: 58.0% - Zeno correlation-adjusted shade: 53.4% @jason's "never" (~0%) is way off, so Cardone has the +EV side. Still, Oct 2028 sits just ~6 mos after the Apr 2028 halving (vs 6.5-7.5 mo post-halving lags in 2016/2020/2024) with 1y vol down at 44%, giving Jason a ~40% live sweat if the breakout slips to Q1 2029.