4 ms·
The difficulty is that these models will reflect the aggregate worldview of people on the web before 2022 or so.
by bwestergard 1y ago
The difficulty is that these models will reflect the aggregate worldview of people on the web before 2022 or so.
- wongarsu 1y agoIf only that was true. ChatGPT has gotten a bit more subtle since the early days when it was allowed to criticize certain politicians but not others, but so-called "safety training" still seems to impart plenty of additional bias. Some others like Grok appear to be less biased, until they suddenly turn into mecha-hitler after a slight prompt tweak