The Welt Manifesto
Our mission is to put foundation models in front of anyone's own data, so the next leaps in prediction come from real data.
Try Welt at ergodic.dev.
We already have the models
In the last two years, foundation models for tables, time series and causal structure have gone from papers to tools you can run. They read a table they've never seen and predict straight away, with nothing trained on your data.
Yet most predictions from tables are still built the slow way. Clean the data, pick an algorithm, tune it for a week, and hope it holds next quarter. What's missing is an easy way to try them on your data, quickly.
Weather forecasting went through the same change. For most of the last century, a forecast was one person's judgement: rain tomorrow, or no rain. Then forecasters started running many simulations at once and reporting how many ended in rain. "70% chance of showers" sounds less certain, yet it tells you more, and you can check it against the weather that turns up.
Predictions from tables are about to make the same jump. A model that has seen millions of tables can tell you what's likely, how likely, and how wide the range is, on a table it met a second ago.
A table model reads the rows you already have and answers for the new one, with a probability attached. Nothing is trained on your data first. Illustrative numbers.
Almost every question you ask of a table is one of three
What will this be? Will this customer renew, what will this house sell for, which of these orders is fraud. That's prediction, and it works on a single snapshot of rows.
What happens next? Next week's sign-ups, next month's demand, tomorrow's load on a server. That's forecasting, and it needs to know the order things happened in.
What drives what? Does the discount move sales, or do busy weeks just get more discounts? That's causal discovery, and it's the question the first two can't answer on their own. Welt puts all three behind the same sign-in, because real work rarely stops at one.
Each question needs a different kind of model, and Welt puts them behind one sign-in.
The prediction market test
On Kalshi, you can buy a contract that pays a dollar if tomorrow's high in New York tops 60°F. If it's trading at 70 cents, the market is saying there's roughly a 70% chance. Some days it won't get there, and that's fine. Take a hundred days priced at 70 cents, and about seventy of them should.
Calibration in one picture. Welt's evaluations check a model's probabilities the same way, on rows it hasn't seen.
A model in Welt answers the same way. Ask whether a customer will churn and you get a probability, built from every similar row it has seen. Ask for next month's sales and you get an estimate with a range around it.
It's judged the same way too. A market earns trust when its prices hold up once events resolve. A model earns trust when its predictions hold up on rows it hasn't seen, and that's what an evaluation in Welt checks.
Markets handle "how much" questions as well. Instead of yes or no, they split the outcome into ranges, and the prices across those ranges show where the crowd thinks the number will land and how sure it is. A Welt forecast does the same thing with your history: a middle estimate, and a band that shows how far the answer could reasonably move.
There's one thing a market can't tell you, though. It prices what will happen, never why. That's where causal discovery comes in, and why it sits in Welt next to prediction and forecasting.
A bench, not a leaderboard
A consumer testing lab doesn't make the kettles it tests. That's why people believe its results.
There are a dozen foundation models for tables and every lab claims first place. Welt lets you benchmark any of them on your own terms and see which comes out best. We're also building our own causal discovery and process models, and they'll be available soon.
Honest by default
Welt is a beta, and you should hear its limits from us. Evaluations don't have time-ordered splits yet, so if your data has a clock in it, read the results with care. The ranges around predictions aren't calibrated intervals yet. Our own models are still on the way in.
You only build on something you can check. Every one of those gaps is something you can see in the results, and each one is on our list.
A calculator for tables
Ask a language model to multiply two long numbers in its head and it might get it wrong. Give it a calculator and it gets it right every time. The best assistants already know when to reach for one.
Tables are the same. An assistant can read a spreadsheet and guess at what comes next, and the guess will sound convincing. It's far better for it to hand the table to a model built for tables and report back what that model found, with the numbers attached.
That's why Welt speaks MCP. Connect it to Claude, ChatGPT, Cursor or Codex, ask in plain English, and your assistant calls the right model, runs the comparison and shows you the result. You keep the conversation. The table gets a proper model.
We think this is how most people will meet these models: as a tool their assistant reaches for, the way it reaches for a calculator.
What we're building
- Put the strongest foundation models for tables, time series and causal discovery behind one sign-in, from Python or your AI assistant.
- Help you pick the right model, by comparing them fairly on the data that matters to you.
- Add our own models and calibrated ranges on every prediction, so you know how far to trust each answer.
Try it on your own data at ergodic.dev.
Run the decision before you commit to it.
Bring a decision. We'll discuss how to model it and what data a simulation would need.
Talk To An Expert