Jonathon Ready builds map-based AI agents and previously worked at Microsoft. His current work includes a small offline model that answers questions by querying an on-device index of geographic data.[1]
Start with An offline 0.8B model that ties Gemini Flash on OpenStreetMap questions →
Education
Links
Personal career page
GitHub · jonready Career
Ready’s portfolio records Microsoft work from 2019 to 2024 and his own Wanderfugl project beginning in 2024. He describes inventing compression used to reduce Windows 11 update downloads. At Penn State, where his page records graduation in 2019 without naming a degree, he founded a student organization building autonomous search-and-rescue aircraft.[1]
Contributions
His Finch experiment adapts a 0.8-billion-parameter model to answer questions using an on-device OpenStreetMap-derived index. Supervised examples and reinforcement learning teach it to select tools and use returned geography data instead of inventing places from memory. On the article’s hundred-question Ask Maps Basic test, Finch and Gemini Flash each receive 83 correct answers. The reported tie concerns that evaluation and correctness judge, rather than every travel question or general model capability.[2]
Tools, writing & talks
Writing, research & talks
Start here
An experiment in training a small local model to answer geographic questions through a data index and tools.
Read the task examples and correctness definition before comparing model scores. Basic language-model training and map-data knowledge help; inspect the reward-hacking discussion before treating an improved score as improved behavior.
Author and experiment developer[2]Original source