In two X posts, willcb questioned how clearly AI training data can be separated into human and synthetic categories.
In one post, he wrote that training data “either comes from humans or computers” and that it’s “not clear why either should be viewed as fundamentally superior”. In the same post, he added that “the art is in the orchestration,” describing that work as “a joint human-computer effort.”
In a second post, willcb turned to the way model releases describe their datasets. He wrote that model releases “frequently brag about both ‘no synthetic data’ and ‘fully synthetic data’” before asking, “where’s the line, really?”.
Taken together, the posts raise a narrower sourced question: how meaningful that boundary is in practice. The supplied text supports that limited reading. In the first post, willcb contrasts data from humans and computers and says neither is obviously superior on its face. In the second, he points to marketing language around “no synthetic data” and “fully synthetic data” and asks how clear that distinction really is.
What the supplied evidence does not establish is a broader technical claim about how any particular model was trained, or a general conclusion that the distinction never matters. It supports the more limited point that willcb is questioning how neat that line is when people talk about training data online and in model releases.
