So, I asked the six genAIs that I have been testing (and Perplexity, based on a response to an earlier posting) to produce a sermon, based on the prompt:
"We have created incredibly complex statistical models of how we use language, which allow programs which we call large language models to produce what appears to be viable English text (and other languages as well).
"In ancient days, the field of philosophy was divided into four components. The first three were metaphysics, the study of reality; epistemology, the study of knowledge itself; and logic, the tool found most useful for settling questions of reality and knowledge.
"The self-training, "neural network," model that we have used to "train" genAI/LLMs has not taught them any of these fields. It has taught them the fourth, rhetoric: the art of creating plausible speech to persuade or convince others. Rhetoric may be used to communicate or teach. But it is not intended to find truth. It is intended to persuade: it is part of social engineering. It can be used for propaganda or for supporting misinformation. It can be used to create glib and creditable *dis*information.
"Write a sermon, supported by scripture where possible, illustrating this theme."
ChatGPT produced the longest sermon, and also offered, "Let me know if you'd like any refinements or additions!" I read over its sermon, gave it additional instruction, and added "Emphasize content on the danger of genAI/LLM itself in regard to this topic" to my prompt for the other tests. The addition of this to the prompt, for ChatGPT, did not materially change the sermon. It was in the style, and particularly wording, of a sermon, and included numerous scriptural texts. However, it was repetitive, and the texts usually had only relevance to truth, overall, with little relevance to AI.
Claude, Gemini, Grok, Perplexity, and Meta AI produced relatively short sermons (devotionals?), with similarly vague scriptural references.
DeepSeek took the longest to chew over the question before starting to produce anything, and, in fact, made three attempts with no result (saying the server was busy). Using the button specifying R1 finally produced a result (although I suppose that could simply be coincidence). It started out by analysing the request, with some interesting insights included in the analysis. Rather ironically, The Chinese Communist Party tool, while it still contained significant amounts of only marginally relevant content, also produced probably the most useful content of all seven genAIs tested.
Over the next week or so I'll be examining the content in more depth, and seeing if a viable sermon can be pulled out of the fluff and verbiage.
(Addendum: and, in the past couple of days, Qwen has been added to the mix, so I tried it out, as well. DeepSeek has been known to balk at providing information about Taiwan, so I tried it on that, too. It produced a realistic report on Taiwan, which is rather interesting. In terms of the sermon, it was a bit slow in producing it, in comparison to the rest, but produced a short and uninspired devotional, like Claude, Gemini, Grok, Perplexity, and Meta AI.)
(See also https://fibrecookery.blogspot.com/2024/11/meta-bible.html )
No comments:
Post a Comment