AI That Predicts the Future
with Ben Turtle, LightningRod AI
AI That Predicts the Future
Show Notes
Ben Turtle did not set out to build a forecasting company. He set out to answer a question that had been bothering him since his time building AI at Google: why do humans learn without labeled data, and why can't AI? The answer became LightningRod AI - a training data platform that teaches AI to predict the future using messy, real-world data, no human labelers required.
The result: a tiny model trained on public news that beats every major frontier model at forecasting real-world outcomes on Polymarket. More accurate. More profit. Better calibrated probabilities. And it was built on data every frontier model has already seen.
Every Decision Is a Prediction Problem
Ben's core framing: when you choose A over B, you are making a prediction. You predict that path A leads to a better outcome than path B. Science itself is prediction - a theory only holds water if it lets you forecast what happens next from a single observation. Most founders treat prediction as a niche capability. Ben argues it is the universal substrate of intelligence, and that pulling this principle into how we train LLMs is the most underexplored leverage point in AI.
The Chronological Grounding Method
The core insight behind LightningRod: real-world data is chronological. Documents are timestamped. News happens in sequence. Slack messages, emails, reports - all of it unfolds in time. LightningRod trains AI by playing a prediction game over this data: given everything up to timestamp A, predict what happens at timestamp B. Repeat across all the data. The AI that gets good at this game actually understands the causal factors in your domain - not just the surface patterns. That's grounded training data, and it's what separates it from synthetic slop.
The Two-Bucket Customer Framework
Every company building with AI today falls into one of two buckets, and each needs a different solution:
- Bucket 1 - Have data, can't use it: Large enterprises sitting on dense, unstructured data - company reports, clinical records, internal communications - that is valuable but not formatted for AI training. LightningRod transforms it into high-quality training sets without human labelers.
- Bucket 2 - Need data, don't have it: Startups and indie developers who want to train a domain-specific model but are starting from scratch. LightningRod pulls relevant public data - news archives, SEC filings, open-domain documents - and runs the chronological prediction game to generate grounded training data from nothing.
Where LightningRod Is Going
Until now LightningRod has operated as an enterprise-facing service, running custom training data projects for larger customers. The next phase is an SDK - a self-serve, developer-first product that puts the chronological grounding framework in the hands of any Python or ML engineer. Swipe a credit card, pay for what you use, and get the same training data quality that previously required an enterprise deal. The GTM targets Hugging Face, GitHub, and Python community forums - the places where ML engineers already live.
Glossary
Q&A
What is LightningRod AI and what problem does it solve?
LightningRod is a training data platform that creates high-quality AI training data from messy, real-world sources - Slack channels, PDFs, emails, public news - without human labelers or synthetic generation. The core method trains AI by playing a prediction game over chronological data, forcing it to learn causal patterns rather than surface correlations.
Why can't companies just use synthetic data for AI training?
Synthetic data is AI generating examples of what it already knows. It cannot add new information the model doesn't have. If you want a model that genuinely understands your domain, you need data that contains information not already in the model's weights - which means real-world, grounded data. That's the gap LightningRod fills.
How did LightningRod beat frontier models at forecasting on Polymarket?
They trained a small model on publicly available news data using their chronological prediction method - the same data all frontier models have already seen. By forcing the model to predict what happens at timestamp B given everything up to timestamp A, the model learned genuine causal relationships in real-world events. The result was more accurate forecasts, better calibrated probabilities, and more profit than models many times larger.
Who is LightningRod for right now, and who is it becoming for?
Currently it serves enterprise customers with large messy data sets who want to train or fine-tune AI models. Imminently, it's launching an SDK for Python and machine learning engineers - individual developers, startup teams, and enterprise ML practitioners who want to generate high-quality training data without building a complex data infrastructure themselves.
What is the most important thing founders should understand about prediction and AI strategy?
Every decision you make as a founder is a prediction: you're predicting that option A leads to a better outcome than option B. The companies that will compound the most in the AI era are those that build systematic prediction capabilities into their decision-making - not just using AI to execute faster, but using AI to see further. The gap between having an idea and testing it is collapsing. The new constraint is predictive clarity about which ideas are worth testing.