Free Lesson
Replace Jev with Your Own LLM, Then Beat It at Its Own Game
30 min
Oct 7, 2026 2:00 PM
By continuing, you agree to Maven's Terms and Privacy Policy.
What you'll learn
How you describe a situation can matter more than the model
Jev, plain chat models, and hand-written rules all played Pac-Man from the same text. All scored in the same range.
You may not need a specialized model to replace Jev
A small open model with no extra training outscored Jev at Pac-Man. On harder questions, Jev still leads.
Owning the model gives you choices Jev cannot offer
A tiny model trained for Pac-Man scored more than twice Jev and runs on a phone, at the cost of general skill.
Why this topic matters
Jev made fast, cheap multiple-choice decisions practical inside products, such as routing tickets or checking policies. We tested it in a real-time Pac-Man game, because playing unseen games is Jev's own showcase. In our test, how you describe each situation in text shaped the score more than the model did. An ordinary open LLM matched Jev (232 to 178 pellets), and owning the model gives you speed and size choices a hosted API does not.
You'll learn from
Yan Wang, PhD
Co-founder of Superlinear Academy · AI at Samsara, Microsoft, Pinterest
