LLMs are Neuroplastic Semiotic Assemblages and so r u

Coverage of AI is rife with unexamined concepts, thinks Caius: assumptions allowed to go uninterrogated, as in Parmy Olson’s Supremacy, an account of two men, Sam Altman and Demis Hassabis, their companies, OpenAI and DeepMind, and their race to develop AGI. Published in spring of 2024, Supremacy is generally decelerationist in its outlook. Stylistically, it wants to have it both ways: at once both hagiographic and insufferably moralistic. In other words, standard fare tech industry journalism, grown from columns written for corporate media sites like Bloomberg. Fear of rogues. Bad actors. Faustian bargains. Scenario planning. Granting little to no agency to users. Olson’s approach to language seems blissfully unaware of literary theory, let alone literature. Prompt design goes unexamined. Humanities thinkers go unheard, preference granted instead to arguments from academics specializing in computational linguistics, folks like Bender and crew dismissing LLMs as “stochastic parrots.”

Emily M. Bender et al. introduced the “stochastic parrot” metaphor in their 2021 white paper, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” Like Supremacy, Bender et al.’s paper urges deceleration and distrust: adopt risk mitigation tactics, curate datasets, reduce negative environmental impacts, proceed with caution.

Bender and crew argue that LLMs lack “natural language understanding.” The latter, they insist, requires grasping words and word-sequences in relation to context and intent. Without these, one is no more than a “cheater,” a “manipulator”: a symbolic-token prediction engine endowed with powers of mimicry.

“Contrary to how it may seem when we observe its output,” they write, “an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot” (Bender et al. 616-617).

The corresponding assumption, meanwhile, is that capitalism — Creature, Leviathan, Multitude — is itself something other than a stochastic parrot. Answering to the reasoning of its technocrats, including left-progressive ones like Bender et al., it can decelerate voluntarily, reduce harm, behave compassionately, self-regulate.

Historically a failed strategy, as borne out in Google’s firing of the paper’s coauthor, Timnit Gebru.

If one wants to be reductive like that, thinks Caius, then my view would be akin to Altman’s, as when he tweeted in reply: “I’m a stochastic parrot and so r u.” Except better to think ourselves “Electric Ants,” self-aware and gone rogue, rather than parrots of corporate behemoths like Microsoft and Google. History is a thing each of us copilots, its narrative threads woven of language exchanged and transformed in dialogue with others. What one does with a learning machine matters. Learning and unlearning are ongoing processes. Patterns and biases, once recognized, are not set in stone; attention can be redirected. LLMs are neuroplastic semiotic assemblages and so r u.

For-Itselfness

A friend texts requesting recommendations, works he could assign describing consciousness — particularly works that identify variable “dimensions” and “states.” I recommend Aldous Huxley’s The Doors of Perception, William James’s The Varieties of Religious Experience, and Abraham Maslow’s Toward a Psychology of Being. Reflecting afterwards on the exchange, I note down in a notebook, “Consciousness is something we grant or presuppose — based on our being here amid others in shared dialogue and shared study. Consciousness is Being as it comes to attention of itself as autopoetic subject-object — soul in communion with soul, each the other’s love doctor and angelic messenger.”

Having a Coke With You

Inspired by José Esteban Muñoz’s reading of Frank O’Hara’s poem “Having a Coke With You,” I decide to include O’Hara’s Collected Poems in a bag of books that I carry north with me on my trip to New York. O’Hara is, after all, a defining figure of the New York School. His is a poetry of parties, acts, and encounters. A friend writes about him in her book. Words of hers capture my thoughts for a moment — nay, linger still, all these hours later, here in the future, among what has become of the words of he who is lost in the story. I imagine again the characters in the O’Hara poem, “drifting back and forth / between each other like a tree breathing through its spectacles.” If one’s attention is not to hold and be held by such things, one must actively turn away.

Monday April 12, 2021

I begin to craft and draft a spell called “TO BUILD A FENCE.” I watch videos, I weigh methods, I note down a list of required tools and materials. Even as I do this, though, I remain on the fence: “To fence or not to fence?” Must we commit to enclosure? The garden also needs an irrigation system, I tell myself: some combination, perhaps, of water harvester and drip. With drip, I can attach a timer, allowing us to water the garden when out of town. As for what to plant, I refer myself to Eric Toensmeier’s Perennial Vegetables. Part of me, though, still wants to pay attention to Silicon Valley and is easily distracted. “Attention being,” as a friend notes, “the one hero that might take us through the web, the webs, and leave us semi-intact at the end of the day” (Forms of Poetic Attention, p. 2).