If you want to dive into the details of generative AI technology (and you do, you do) there are some academic papers to look at: Murray Shanahan explains large language models (11 page PDF) as "generative mathematical models of the statistical distribution of tokens in the vast public corpus of human-generated text." Simon Willison offers an illustrative example. The generative AI we're using today is based on foundational models that can be applied on a wide range of tasks; this looong report (214 page PDF) describes the advantages and risks. One risk is described in detail in another paper: "in the most severe case, known as 'detached hallucinations', the output is completely detached from the source, which not only reveals fundamental limitations of current models, but also risks misleading users and undermining trust." We can detect when we're hallucinating; the AI cannot.
Today: 0 Total: 16 [Direct link] [Share]
