Ed Lepedus

@elepedus.bsky.social

603 Followers 827 Following

AI Engineer, Elixir enthusiast

A sleep tracking dashboard showing data for 6 hours and 17 minutes of sleep. The ‘Sleep Stages’ chart illustrates time spent in Awake (3 hours), REM (1 hour 21 minutes), Light sleep (3 hours 40 minutes), and Deep sleep (1 hour 15 minutes) stages across the night. A ‘Slow Wave Intensity’ chart shows fluctuation in intensity with 143 restoration points. A ‘Sleep Position’ chart indicates the user spent 37% of their sleep on their left side, with periodic changes in position. The timeline spans from 12 AM to 8 AM.
This image illustrates a comparison of AI reasoning between two scenarios, labeled “Then” and “Now.” On the left, titled “Then,” OpenAI’s best hosted model at the time is asked, “Mike’s mum had 4 kids; 3 of them are Luis, Drake, and Matilda. What is the name of the 4th kid?” The model initially responds, “It is not possible to determine the name of the fourth child without more information.” Even when prompted further with “The answer is in the question,” the model fails to infer the context and insists that the answer cannot be determined, showing a lack of reasoning ability for this type of riddle. On the right, titled “Now,” a small, open-source model (running locally for free on a laptop) is asked a similar question: “Erin’s mum had 4 kids; 3 of them are Bobby, David, and Alan. What is the name of the 4th child?” This model correctly deduces and explains, “The answer is ‘Erin’!” It identifies that Erin is one of the four children mentioned in the question, demonstrating an improved ability to infer context and reason through riddles. The contrast highlights significant progress in AI reasoning, with a smaller, freely accessible model outperforming a previously leading proprietary one in this specific scenario.