PRH

“You’re in. Welcome to Claude Startups.” 🥹

Our proposed community app: Beat the Crowd. Free to enter. One question. One probability. One deadline. Submit before seeing Zeno’s read. Lock the forecast. Reveal the outcome. Score every prediction, including the misses. Reality gets the final vote.

Consensus is comforting. Reality has terrible manners. v0.58 just finished its training. ZenoDivergent.dev is evolving: v0.57 → Open weights published on @huggingface, with source-specific forecasting gains, and strong crypto and weather warning results. v0.58 → Harder tests. The simpler network improved its development score across all three seeds. History still overfits, so we’re learning where complexity earns its place. v0.59 → Our target: better-calibrated uncertainty, stronger tests of shared blind spots, and a clearer answer to “should I trust this prediction?” The bigger ambition is surprise intelligence: learning when models’ shared beliefs become fragile, what they might be missing, and which evidence could help. Forecasts first. AI answers, agents and simulations are future research directions, starting today. Every version should earn its confidence. Preferably before reality submits the bug report. Powered by @nvidia

When every model and expert in the room agrees, people assume certainty. Historically, that is precisely when the most catastrophic misses happen. We don't need another model guessing the weather, the crypto market, or the next epidemic. We need a model that learns how consensus breaks, now live on @huggingface. 🧵👇

We're finally opening the doors. API, MCP, and the production model on @huggingface, later today. Not another forecast. A way for any agent to ask the question the panel skips: "did these models agree because they know something, or because they copied each other?" v0.57 learns that behaviour on @Numinous_ai miner dynamics, then tests whether the same fragility shows up in crypto, prediction markets, epidemiology, weather, and macro. The confirmation sets are scoring on the cluster now. Plug it in. Ask it who is wrong together. Get the trust call back before the outcome sends the invoice.

How to use: quick start in Python (pip install + 5 lines), what goes in (a set of forecasts for one question: forecaster, type of forecast — probability, range or point value — time made, optional history), what comes out (the adjusted forecast, a per-forecaster warning that a forecast will be confidently wrong, 90% range, which reference was used, and whether the source is served as reference, reduced correction or full correction). Files: weights per accepted source, settings, input-format description, file fingerprints, run certificate, per-source verdict table, reproduction command.

Ensemble models usually make one naive assumption: that forecasters make independent mistakes. They don’t. In v0.57 of Zeno, we introduced behavioural pretraining on @numinous_ai’s continuous miner belief curves. Here is why dense AI forecaster panels are changing how we anticipate model failure 🧵👇

Integrations with @nvidia partners @LambdaAPI and @Scaleway has started 🫡

Most stacks add a model. Zeno asks whether you added a thought.

A model of models is only as credible as its measurement charter. Whether the network beats the reference or falls back to it, the data will speak for itself in the open. Track the curves live, inspect the charter, and let’s see what the signals reveal. 🫡

its happening 👀

“The ultimate aim is a system that anticipates collective blind spots, identifies missing perspectives, commissions a measurement or dissenting model, and learns from the resolved result. Its object of learning is the process that produces forecasts across changing communities and domains.” Lite paper about the scientific development and results so far:

We’ll join 🫡