On Stochastic Libraries
Large Language Models as library-like Knowledge-Synthesizers.
Last year i wrote a short piece (in german) on Wishful Mnemonics, how AI-Systems are less “intelligent“, and more like a stochastic tool to retrieve and mash up knowledge of all kinds. I keep refering to AI-stuff as “stochastic liberaries“, and I think this term describes the nature of AI-Systems pretty closely.
Now Jaron Lanier wrote a piece in The New Yorker on the mythologization of AI-systems as “intelligent“ and the hidden psychological risks of the anthropomorphization of this tech. In the piece, he writes that “there is no A.I.“, describes the tech as an “innovative form of social collaboration“, “a technology that is non-repeating“ and as a mashup-machine for knowledge based on human labor. Writing about AI this way is describing them as “stochastic liberaries“ by other terms.
I keep refering to that piece from last summer again and again, and as i gained a ton of international readers since then, switching to the english language in the process, i want to have an english language version of that post, so i reworked the piece and translated it.
Wishful Mnemonics
Large Language Models as library-like Knowledge-Synthesizers — a new framework for the coming AI-revolution
People who work directly with AI-technology instinctively feel that these machines are fundamentally different from what we understand as "human intelligence" in a neuropsychological context and we currently see many different terms for these tools, that will soon fundamentally change the working world.
These algorithmic machines that are currently emerging are not intelligent, but they provide access to synthetic recombinations of human knowledge, reminiscent of a new type of library whose combined content is navigated by text input. This new form of library doesn’t provide access to individual human works, but rather continuations, amalgams, remixes, modulations, and interpolations between nodes in a network of existing knowledge, "in the style of Banksy, trending on artstation."
One of the most beautiful terms for this new technology, i think, is "wishful mnemonics," coined by Beth Carey: an AI system not as an intelligent and thus capable agent, but as a "wish-fulfilling memory aid," which poetically describes exactly what these machines do.
Another term thrown around on AI-Twitter describes this technology as a "synthesis engine," which illustrates the tech-evolutionary step from search engines like Google to a machine that is able to present existing knowledge in new, algorithmically generated variations, such as by analyzing our knowledge of protein folding and practically calculating all 200 million possibilities in which proteins can fold or by extrapolating our knowledge of chemical weapons into 40,000 new variations of the same.
Psychologist Alison Gropnik recently wrote about these technologies describing AI systems as a step in the cultural evolution of knowledge-tools. Modern Large Language Models have been trained on gigantic collections of existing human knowledge, and they store billions of diverse associations in their databases. This puts these AI models much closer to cultural technologies like libraries, writing, or language, than intelligence itself — all of these technologies serve to provide, mediate, and recombine existing knowledge.
This framing allows for new thinking about upcoming regulations of the technology, such as in the context of copyrighted works in the datasets of image generators. For example, there’s an internationally applied obligation of legal deposit for national libraries, according to which copyright owners must provide copies of their publications — raising the question of whether and in what form AI systems that make collected human knowledge accessible can or should be managed akin to a national library.
In my opinion, at least the large, all-encompassing "statistically stochastic knowledge synthesizer libraries," the so called "foundational models", could be operated by the public sector to ensure safety, ethical production and prevention of abuse. Running foundational models by the public would also ensure data transparency and work against the "black boxing" of this tech. I’m not sure or convinced that this approach would be practical or feasible to do, but i think it would provide the most stability and transparency.
This wouldn’t mean that private knowledge synthesizers are not possible; all variations are conceivable, from highly commercial and very expensive niche models for synthetic image generation in specifically trained illustration styles for a modular AI Photoshop of the next generations to small AI-Office-assistants working on top of public AI-systems to, as Yann LeCun put it, "save typing".
The classification of LLMs as a stochastic library — a cultural tool providing access to existing knowledge, analogous to the internet, printing, and writing — eliminates esoteric associations of anthropomorphic properties such as "intelligence" or "consciousness" from these novel database technologies, and enables a language that allows for regulation without falling into the trap of anthropomorphization, truly highlighting the technological possibilities of these new knowledge synthesizers.
Alison Gopnik also gave a short 15minute presentation on LLMs as cultural technologies at the Simons Institute in Berkeley:
Recent work on large language models has focused on whether or not they are analogous to individual intelligent agents. I argue that instead we should think of them as cultural transmission technologies, by which accumulated information from other humans is passed on in a compact form. This makes them aanogous to other human technologies such as writing, print, libraries and internet search, and arguably language itself rather than as intelligent systems.
Currently reading a book by Krajewski about the history of index cards. LLMs seem to be related to the boxes and the way texts were produced by positioning ideas within the constellation of crossreferenced associative memories. As a game of references, hermeneutic labour of prompting and reframing results is not as passive as it seems. What is advertised as intelligence is the result of user-collective archival coupling.