I nearly bet a client project on “world models”
At 10 p.m. last Wednesday, I was revising a freelance proposal in a Hangzhou guesthouse: use AI to predict whether a client would change requirements next week, cancel the budget, or even reveal a competitor’s next move. I ran more than 600 small experiments around the idea. In the end, I put it in a drawer. Not because the concept wasn’t elegant, but because it was too far from real business.
What JEPA actually does, and who is using it
JEPA is a class of methods that teaches AI how environments change over time. It goes beyond identifying what is in front of it and tries to infer what might happen next. Haddad Al-Chaer documented 600 experiments on his blog; his use case looks more like research than quoting prices for clients. What most founders encounter today is “predictive AI” packaging: forecasting demand, email responses, or project delays. That is not a full JEPA system.
What it costs to replicate this today
A research-grade version has a difficult-to-eliminate price: model training and cloud compute can run from hundreds to thousands of dollars per month, plus several weeks of continuous work. The technical barrier is finding someone who can handle model training, experiment evaluation, and compute scheduling. The first step isn’t buying tools; it’s opening the paper or project documentation and understanding what it predicts, how it is evaluated, and whether it fits the business. You may also discover that reading the material already takes more effort than the potential return. This tool isn’t for everyone, and skipping it for now is fine.
Here’s how I’d choose by stage
Just starting: I wouldn’t spend the budget. I’d forecast demand through customer interviews, past orders, and a simple spreadsheet. With 1–2 clients: I’d put a limited set of past emails and project records into an ordinary AI tool, ask it to list “possible changes” and “questions that need confirmation,” but not let it make commitments for me. Scaling up: Only if forecasting had become a repeated bottleneck would I bring in a technical cofounder for a small validation, comparing predictions with actual outcomes using one month of data. I only realized this after getting stuck myself: for most of us, shortening the feedback loop is a better deal than predicting the future.