I ran my local LLM for hours and watched it get dumber in real time
…This then hits a quirk in how transformer models work. They are measurably worse at recalling information in the middle of a long context window, preferring to read the start and the…