Tilelli, a new AI model developed by Tilelli Lab in Marrakech, Morocco, addresses the common problem of AI hallucination by admitting when it does not know an answer. Unlike many large language models that generate confident but false responses, Tilelli stops generating text when its internal certainty drops below a threshold, according to the lab's published research.
In independent testing, Tilelli achieved a 0.93 accuracy rate (out of 1.0) in identifying out-of-distribution or nonsensical questions. In separate tests designed to catch the model bluffing, it correctly chose to remain silent 9 out of 10 times. The model's own response when asked about its identity is: 'i am small but try to be honest.'
Tilelli runs entirely offline on a standard laptop or desktop computer, requiring no internet connection, graphics card, or subscription. It installs in minutes and occupies about 39 megabytes of storage. The cost to fully retrain Tilelli from scratch, including compute time, is less than twenty dollars.
The lab publishes all test results transparently, including cases where Tilelli performed worse than simpler alternatives. Tilelli is not designed to compete with large AI systems like GPT or Gemini; it focuses on specific, focused tasks where honesty about limitations is critical.
