Comment LLM results getting much worse (Score 2) 135
LLM, aka artificial summarizers, have grown horribly worse the last 12 months. Humans have slowed posting q and a in favor of ai answers, but LLM is only as good as its human material. Eg, 2 people I know each spent 2-3 hours with friends using paid Claude to fix their printer before giving up and calling me; I fixed each in 3-5 minutes by deleting the printer and re-adding; in March I asked ai if the world cup was this year and it contradicted itself answering No this year is 2026 but the world cup will be in 2026; I asked it a basic arithmetic question about modula and floor commands and it included 9-8 = -1; I just asked it when a nba team starts training camp and it replied in 2024; I asked it what the best bluetooth speaker is for iOS 26 and it explained that iOS 26 will be released in 2032 and the current version in iOS 17 but recommended a bluetooth speaker for iOS 18; I save screenshots of these horrible answers. A Carnegie Mellon study recently found ai agents are wrong much more than half the time. CEOs have the wrong mental model of what LLM does. The main creator of LLM recently said LLM is a dead end and about as smart as a house cat. You all know about this trend but I restate it because repurposing or continuing to use data center equipment for inference may not be profitable if its results are so junkie. Also data center inference chips burn out extremely fast 1-2 years from extraordinary intense initial use so the mind boggling costs never made economic sense. I have personally known several tech CEOs and their salient characteristic is fearless aggressive boldness, absolutely not intelligence. The initial impressive uncanny LLM honeymoon was like businessviagra for them and they redid their wardrobes to accommodate that but who would buy their used clothing as they shed them when the phenomenon wears off? Granted, ai finding bugs is pretty cool.