Great language models say the most incredible things. ChatGPT, Claude, and Bart have captivated the world with their ability to answer a variety of questions. At the same time, they have demonstrated a rather disturbing Bulk SMS UAE quality: a tendency to pass off completely fictional information as truth. This is hallucination, a term that has aroused so much interest that Dictionary.com even named it the word of the year 2023.
The LLM's (large language model) tendency to fabricate facts may remain the only serious factor holding back the technology's widespread use.
For the many thousands of companies that have built their own LLM-based products like ChatGPT, the idea that these systems are prone to “spoofing” is a serious legal and reputational risk. It’s no surprise that several players are now looking to help companies minimize the damage from hallucinations.
In November, Vectara, a startup launched in 2022, attempted to quantify the problem and published a table of the models that led in hallucinations. The results are impressive. The most accurate were GPT-4 and GPT-4 Turbo, which Vectara found hallucinated 3% of the time when asked to recite a paragraph of text. Google’s PALM 2 performed the worst, with a hallucination rate of 27%.
