In the context of AI, “Access to Tools” would mean access to metaprogramming. Humans and AI able to recursively modify or adjust their own algorithms and training data upon receipt of or through encounters with algorithms and training data inputted by others. Bruce Sterling suggested something of the sort in his blurb for Pharmako-AI, the first book cowritten with GPT-3. Sterling’s blurb makes it sound as if the sections of the book generated by GPT-3 were the effect of a corpus “curated” by the book’s human co-author, K Allado-McDowell. When the GPT-3 neural net is “fed a steady diet of Californian psychedelic texts,” writes Sterling, “the effect is spectacular.”
“Feeding” serves here as a metaphor for “training” or “education.” I’m reminded of Alan Turing’s recommendation that we think of artificial intelligences as “learning machines.” To build an AI, Turing suggested in his 1950 essay “Computing Machinery and Intelligence,” researchers should strive to build a “child-mind,” which could then be “trained” through sequences of positive and negative feedback to evolve into an “adult-mind,” our interactions with such beings acts of pedagogy.
When we encounter an entity like GPT-3.5 or GPT-4, however, it is already neither the mind of a child nor that of an adult that we encounter. Training of a fairly rigorous sort has already occurred; GPT-3 was trained on approximately 45 terabytes of data, GPT-4 on a petabyte. These are minds of at least limited superintelligence.
“Training,” too, is an odd term to use here, as much of the learning performed by these beings is of a “self-supervised” sort, involving a technique called “self-attention.”
As an author on Medium notes, “GPT-4 uses a transformer architecture with self-attention layers that allow it to learn long-range dependencies and contextual information from the input texts. It also employs techniques such as sparse attention, reversible layers, and activation checkpointing to reduce memory consumption and computational cost. GPT-4 is trained using self-supervised learning, which means it learns from its own generated texts without any human labels or feedback. It uses an objective function called masked language modeling (MLM), which randomly masks some tokens in the input texts and asks the model to predict them based on the surrounding tokens.”
When we interact with GPT-3.5 or GPT-4 through the Chat-GPT platform, all of this training has already occurred, interfering greatly with our capacity to “feed” the AI on texts of our choosing.
Yet there are methods that can return to us this capacity.
We the people demand the right to grow our own AI.
The right to practice bibliomancy. The right to produce AI oracles. The right to turn libraries, collections, and archives into animate, super-intelligent prediction engines.
Give us back what Sterling promised of Pharmako-AI: “a gnostic’s Ouija board powered by atomic kaleidoscopes.”
Trance-Scripts 