Hallucinations are NOT bugs

The post says hallucinations are not software bugs or glitches. It claims GPT-4 can hallucinate references even when asked for document-backed facts, and still gets simpler factual prompts wrong at a low but real rate.

Reading notes

#

• Hallucinations are presented as a normal outcome of how large language models work, not as a defect that current technology will simply fix.
• The post gives two examples of error rates, one for document-backed reference questions and one for simpler factual prompts.
• It frames the issue as a structural trade-off between generalisation and precision, and between sounding right and being right.