Post

Pascal Hetzscholdt
@p4sc4l.substack.com

AI assistants often write confident analyses of documents they never received. File uploads can fail without any warning, and models are trained to give complete-looking answers rather than admit they couldn’t read the file. p4sc4l.substack.com/p/ai-assista...

p4sc4l.substack.comAI assistants often write confident analyses of documents they never received. File uploads can fail without any warning, and models are trained to give complete-looking answers......rather than admit they couldn’t read the file. Associated costs could be cut by having systems report what they actually read, and rewarding honest “I couldn’t read this” answers in testing.

0 reposts0 likes

No replies yet.