#29: Tokens and Coins

Can ChatGPT speak the Truth? The answer might surprise you...

This post is an updated and edited version of a post originally made on Twitter. All images generated by FLUX.1 schnell fp8 txt2img model.

I. Bullshit and Truth

On a fundamental philosophical level, present generative AI technology—LLMs, diffusion-based image generation, etc.—is automated bullshit generation. I mean “bullshit” in the philosophical sense, put forward by Harry Frankfurt in his text On Bullshit. I quote:

The fact about himself that the bullshitter hides… is that the truth-values of his statements are of no central interest to him; what we are not to understand is that his intention is neither to report the truth nor to conceal it.¹

To put the claim simply, generative AI is indifferent to truth. This is not a novel observation, but it is one that bears repeating, as an entry-point into questions about the nature of these tools. The frame begets an obvious follow-up question: what is it that distinguishes these language models from humans, who do have some capacity to “report the truth”? What does it mean to speak “truth” anyway?

Nietzsche’s essay “On Truth and Lies in a Nonmoral Sense” tackles these questions of truth directly (boldface mine):

What then is truth? A movable host of metaphors, metonymies, and; anthropomorphisms: in short, a sum of human relations which have been poetically and rhetorically intensified, transferred, and embellished, and which, after long usage, seem to a people to be fixed, canonical, and binding. Truths are illusions which we have forgotten are illusions- they are metaphors that have become worn out and have been drained of sensuous force, coins which have lost their embossing and are now considered as metal and no longer as coins.

Is it a coincidence that for Nietzsche, metaphors, which are for Nietzsche the basis of knowledge, are “coins”, akin to how LLMs consume and generate “tokens”? Are we ultimately dealing with markets, mediums of exchange at the level of knowing? are LLM tokens convertible to “human” “coins” and vice versa? And when does “truth” enter the picture?

II. On Tokens

How can we distinguish between tokens and coins? How does conversion work? At first blush, it strikes me that tokens exist in relation to a particular, closed economy. I go to the arcade, I exchange my USD coins for tokens, and then I use those tokens to exchange again for a specific set of outputs, such as getting to play an arcade game.

This arcade game is, in theory, of immediate value for me, and the tokens are, to use crypto language, burned once I have received the arcade game as output. In most cases, I cannot convert back from tokens to coins². The arcade itself received payment in coins when I exchanged them for tokens, so it has already made its profits, and it would represent a pure loss to allow token purchasers to ask for coins back: it is a one-way, closed economy.

Coins, on the other hand, have a universal quality. I can purchase with my coins “anything”, including other kinds of coins, i.e. FOREX (FOReign EXchange). The epistemological equivalent here, to extend Nietzsche's metaphor, of FOREX, is that a text in one language can be translated into another language and retain its value (perhaps even increasing its value, such as in the case of Strachey's translations of Freud's ouevre). Thus “truth” for Nietzsche has an inherent history of universal value in relation to human communication and action, and then the face wears away to eliminate that value, coin becoming a “truth” as it becomes something parroted and un-thought, no longer useful in the market.

So we can look at the LLM infrastructure's use of “tokens” as being closer to the arcade than to the global market. Once my coins (the inspiration of my prompt) have become tokens (the prompt itself), what I receive in exchange is more tokens (of output), that in theory are of immediate use value to me, like the arcade game (e.g. I ask an LLM how to do X, it tells me how to do X, then I go do X following its steps).

If we inspect the structure more carefully, LLMs are somewhat akin to the “pusher” game, where you feed it your tokens and hope to get a large payout in kind (if your token “pushes” more tokens into the collection tray than you put in), but where the game itself can also be consumed and enjoyed for its own sake. In this enjoyment of the flowing of token inputs and variable token outputs, we witness an unmistakable gambling element: will THIS prompt be the one to get me the output I like or want? “Maybe a little tweak will do the job?” I think as I feed another token into the slot. But as we know, the house always wins (unless you can count cards, in which case the casino either eliminates you or hires you).

III. On Coins and Counterfeits

What is missing is that the field of universal human exchange, the realm of action affecting the great web of human relations, takes place with coins and not tokens. Most people can tell if I'm attempting to pay in tokens when they're expecting coins (i.e. they know if I'm sending them LLM-slop), and they're likely to reject it, unless they are not wise to these things and overlook the “counterfeit” nature of my offer.

It appears some realms of human exchange already relied on tokens prior to the invention of LLMs. Many “formal” domains of interaction, such as law, academic journals, etc., require a certain allocations of tokens upfront in order to accept your offer of coin beneath, your attempt to change the world in your own small way (assuming such a coin exists; if not, you have the classic “nonsense paper”).

This textual demand for a certain degree of formality has the same abstract structure as a bribe: I pay you gatekeepers something (convert some coin to tokens, for your internal, perhaps institutional economy) to listen to what I really have to say (accept my coin in a properly human-to-human exchange).

As mentioned in the coin pusher example, the way LLMs work is unique in that they accept tokens to generate more tokens, rather than necessarily generating something of direct use value. The economic result is that token-based bribes become cheaper: if I need to write in scientific form to get my idea taken seriously, then it costs a lot less when I can ask ChatGPT to write in the form I want.

It's currently unclear whether these LLM tokens are worth-less to the mafia! They seem to have approximately the same inherent structure as the tokens that came before it, in the sense that ChatGPT used well to approximate the academic paper form can generate results indistinguishable from something written by hand, but if the intent of the bribe is extraction from the submitter (e.g. to demonstrate some degree of commitment) rather than wealth for the recipient, it makes sense that an LLM would eliminate much of the extractive element, and thus threaten the entire token economy.

The eventual outcome may well be a world where tokens have become useless, because they are so cheap, and only coin will matter. In this case the LLM ironically renders itself worthless in the domain of action, becoming a search engine for information of no exchange value, no capacity to move the world, a speaker of Truth and only Truth, the great genie conjuring forgotten illusions, coins now considered as metal and no longer as coins, generative AI relegated to the domain of pure work³. The active economy of human knowledge will route itself around them. The academy and other arcade-like institutions will erect new barriers to assess commitment. The meta will shift. It's already happening.

1. Excerpt from https://thevimblog.com/2018/11/10/harry-frankfurts-on-bullshit/

2. Ironically, inconvertibility from tokens back to coins is often not the case in crypto; perhaps the term "token" should be reserved for soulbound assets only, and we should call the rest something else.

3. I am using “action” and “work” here in the sense Hannah Arendt defines in her text The Human Condition. Briefly, “work” brings a pre-imagined output into existence, whereas “action” is a pure event, of speech or deed, that produces unpredictable ripples across the present condition of Man, the “web of human relations” being her term to describe the plane of action proper.