Skip to content
Research
Notes

It knew: predicting a number an LLM can only guess

Where will the ball land? An LLM guesses, the textbook formula ignores air resistance, and PredictLM predicts it from 600 past throws — zero-shot, no equation. The capability your AI agents are missing, in 25 seconds.

Zero One Research2 min read·#predictlm#demo#agents#mcp

Large language models are brilliant with words and brittle with numbers. Ask an agent "where will this ball land?" and it does one of two things: it guesses, or it writes Python that uses the textbook projectile formula — which ignores air resistance and is wrong for any real ball.

PredictLM does neither. It learns the real relationship from data and predicts.

What you're watching

A real air-drag physics simulation. We give PredictLM 600 past throws — each one (angle, speed, mass, ball diameter) → landing distance — and then ask it to predict where a new, unseen throw lands.

  • 🧠 LLM, no tools — guesses. No grounding in the data, so it's off.
  • 📐 Textbook formula — ignores air resistance, so it overshoots every time.
  • ✦ PredictLM — predicts from the 600 examples, zero-shot, no equation. Across 60 unseen throws it hits R² = 0.994.

It never saw the physics. It learned the real flight from data.

Why this matters for agents

This is the capability missing from most AI systems. When an agent needs a number it doesn't have — a price, a demand forecast, a risk score, a sensor reading — it shouldn't hallucinate it or bolt on brittle calculation code. It should call a model that learns the pattern from your data and returns a calibrated prediction.

PredictLM is a tabular foundation model: give it a handful of example rows and it predicts the missing value in a single forward pass — no training, no tuning. It runs locally (the Mini model fits on a laptop), it's open-weight (Apache-2.0), and it's exposed over MCP, so your agents can call it as a tool.

Try it

The ball flight is a controlled air-drag physics simulation chosen to make the point visible — an LLM and the vacuum formula both fail where data-driven prediction succeeds. For benchmark numbers on real OpenML tasks, see the PredictLM v1 launch.