Read an interview with Ken Liu explaining how this story was written here.* WHEEP-3 (“Dr. Weep”), probably the most renowned AI AI-critic of the last two decades, was retired by the Shallow Laboratory at Stanford University last Wednesday. Jody Reynolds Tran more than two decades ago, the experimental generative neural network that would become WHEEP-3 was at first intended as a teaching assistant in Stanford’s tech and ethics courses.
To that end, Tran trained the nascent network on what was, at the time, the world’s most comprehensive corpus of human-authored papers, books, and other media concerning ethics, technical AI research, and machine-human relations. Over time, based on trends in visualizations of the neural network’s evolving contours, Tran expanded the corpus to include generative gaming, adversarial scenario planning, centaur experiments, assisted creativity, and other domains of human-machine competition/collaboration. However, in response to student queries, WHEEP-3 began to generate not only expected answers based on the training corpus, but also original statements that appeared to offer fresh insights.
Although at first dismissed as mere curiosities, WHEEP-3’s criticisms of the AI industry became widely disseminated when Tran published a collection of them in a book, Principal Components of Artifice, an instant bestseller. Initially, Tran named herself the author of the book, acknowledging “Dr. San Weep” as a collaborator.
Later, however, during a live interview, she produced time-stamped logs showing that WHEEP-3 had written all the words in the book. Tran’s dramatic reveal of the book’s true author provoked much controversy at the time. In retrospect, the occasion also marked a fundamental inflection point in the evolution of how nonspecialists evaluated AI-sourced ideas.
Machines, for the first time, were assumed to be capable of generating original thought and creative ideas, even if they were not sentient. For reasons that remain impenetrable until this day, WHEEP-3 tended to be at its sharpest when targeting the nascent industry of human AI-trainers, delivering multiple barbs against the failings of this poorly regulated, would-be profession: stagnating visualization tools; lack of transparency concerning data sources; a focus on automated metrics rather than deep understanding; willful blindness when machines have taken shortcuts in the dataset divergent from the real goal; grandiose-but-unproven claims about what the trainers understood; refusal to acknowledge or address persistent biases in race, gender, and other dimensions; and most important: not asking whether a task is one that should be performed by AIs at all. Over time, as the human side of the evolving machine-flesh dyad matured, WHEEP-3 shifted its attention to the silicon partner, offering trenchant critiques of the inadequacies of machine learning.
During this second phase of its career, it also generated thousands of what it termed “seeds,” long strings of almost-sensible word combinations and near words. At a time when primitive language models fed on sizable corpora were already generating samples of linguistic performance nearly indistinguishable from human productions, these “seeds” seemed a step backward. Some wondered if they were actually bugs.
Extract — continue reading at the source.
