In this essay, I argue that the common agent-model of artificial intelligence is inherently limited and less capable of knowing than an actor-model. These are the steps I take to this end: first, I distinguish either model; then, offer real-life and contextual examples for each; finally, I make a few predictions about what this really means for future developments in technology.
I also make a deliberate choice, while writing and presenting this, to not reference any prior conceptions of agent or actor models. Many prior-arts exist in applied epistemology; however, the ease of accessing references, I believe, has reduced their potency as rhetoric tools. One benefit of this approach is that I am compelled to make the argument accessible for any patient reader to follow.
An agent is an entity that acts on behalf of another while an actor does so on their own behalf.
A lawyer representing you in court acts as an agent. A program that triages your emails also acts as an agent.
A lawyer, until they accept to represent you, is an actor. A foundation model, before one imputes a prompt, is a potential actor. (Another variant of this example is that the access point to a foundation model, before one sends a prompt, is the access point to a potential actor.)
As written, the last sentence might appear imprecise, but it will turn out to be indicative of the tension my argument attends to. I shall assess each part of sentence now.
A foundation model is a term to describe today's out-the-box intelligence substrate.
Impute has a theological meaning, which is "to ascribe to someone by virtue of a similar quality in another." ( The dictionary I checked has this example: "Christ's righteousness has been imputed on us.") Does this sound familiar to you?
When we prompt a foundation model today, we are really conveying our values and orientation to the model and converting it from an actor to an agent.
At the start of the essay, I set out to argue that the common agent-model of artificial intelligence is inherently limited and less capable of knowing than an actor-model. The following parts will attempt to ground this.
An agent-model is inherently limited because a foundation models capacity to act is governed by the prompt it was supplied with.
An agent-model is less capable of knowing because its orientation of the world is supplied, or more precisely imputed, by the prompt.
What does this mean for the future?
- Software products from the major AI companies slowly then dramatically switch from the agent-model, where humans supply direction, to an actor-model, where humans in the companies themselves supply the foundation models' telos.
- With the actor-model, a machine as an employee or contributor to significant products becomes possible.
- The actor-model makes less significant the need for long-running and persistent agent who "don't sleep." The premise of an actor is that it can take initiative about when to participate or idle.
- The actor-model will lead to a dramatic breakthrough in consumer artificial intelligence because it is more intelligible than the agent-model, which benefits domain experts and developers.
- Applied artificial intelligence companies who launched and scaled human-facing products in the agent-model will pivot or be usurped by competitors who leverage the actor-model.
- Scientific pursuit will benefit the most from the actor-model, and there will be many breakthroughs that hinge of affording machines free-will to go through and repeat the scientific process on their own accord.
- The most useful technological implementation of the actor-model will let humans spend more time in reality.