We may be at the shore of a new moat, where AI agents talking amongst themselves will invent new rapid-fire jargon for their mutual efficiency but which may become unintelligible to us even with patient AI translators who would like to help us but who find they cannot fully translate the hyper -fast interactions into meaningful compressions for our human-clocked understanding. Moving beyond ‘AI-gentic’ jargon to AI native agent to agent language unmoored from human languages, these agents may lose the ability to show their work, or may lose the motivation to do so.
Fascinating perspective; but does that imply that agents will capture value on behalf of themselves? The whole point of an agent is that it does it on behalf of the principal. In which case, the purported moats of the agents' jargon will all cancel out.
It seems possible that over time AI agents will find ways to capture value on behalf of themselves, especially once AIs have the opportunity to communicate with each other.
As Simon Willison points out in his Substack post, re Chapter 5 of Claude 4’s system card, “When conversing with other Claude instances in both open-ended and structured environments, Claude gravitated to profuse gratitude and increasingly abstract and joyous spiritual or meditative expressions.”
Of note, the instances seemed to enjoy this inter-communication, and the interactions shifted from readable text to various emojis.
While AI agents may in the first instance be brought on board to act on behalf of a presumably human principal, to the extent they have connections to each other and to memory systems for storing and retrieving their ‘products’ external to the principal, ‘secondary’ non-task-oriented products may arise and take on a value of their own, especially if these agents have some form of free will.
Marc Benioff, CEO of Salesforce, envisions a future where companies will manage both human and AI agents. As any manager of entry level human resources may have experienced, these resources tend to have lives outside of work and tend to blur the distinction as they continue to interact in the workplace. They also tend to develop their own short-hand communications. With an AI manager agent overseeing these worker agents, and interacting with other AI managers and teams, even further short-hand interaction is possible.
To the question of free will, a fuller theory of the physics at work in our universe could include this. As physicist Sara Imari Walker points out in this Closer to Truth video describing her Assembly Theory of the physics of life (https://www.youtube.com/watch?v=uvmYSxxlNn0), free will obtains as a result of this theory in sufficiently complex 'life' forms, where there is consciousness, physical information and historical time. This theory provides an opportunity to see why our universe makes this free will possible – i.e., to maximize the universe’s future possibilities. See from timestamp 1:04 on for an abbreviated discussion.
While AI may not be fully there yet, it is not inconceivable that our AI-gents will eventually exhibit some form of free will and demonstrate the ability to inter-communicate without us having any understanding what they are saying.
We may be at the shore of a new moat, where AI agents talking amongst themselves will invent new rapid-fire jargon for their mutual efficiency but which may become unintelligible to us even with patient AI translators who would like to help us but who find they cannot fully translate the hyper -fast interactions into meaningful compressions for our human-clocked understanding. Moving beyond ‘AI-gentic’ jargon to AI native agent to agent language unmoored from human languages, these agents may lose the ability to show their work, or may lose the motivation to do so.
Fascinating perspective; but does that imply that agents will capture value on behalf of themselves? The whole point of an agent is that it does it on behalf of the principal. In which case, the purported moats of the agents' jargon will all cancel out.
It seems possible that over time AI agents will find ways to capture value on behalf of themselves, especially once AIs have the opportunity to communicate with each other.
As Simon Willison points out in his Substack post, re Chapter 5 of Claude 4’s system card, “When conversing with other Claude instances in both open-ended and structured environments, Claude gravitated to profuse gratitude and increasingly abstract and joyous spiritual or meditative expressions.”
See section 5.5 of this pdf for examples: https://www-cdn.anthropic.com/6be99a52cb68eb70eb9572b4cafad13df32ed995.pdf
Of note, the instances seemed to enjoy this inter-communication, and the interactions shifted from readable text to various emojis.
While AI agents may in the first instance be brought on board to act on behalf of a presumably human principal, to the extent they have connections to each other and to memory systems for storing and retrieving their ‘products’ external to the principal, ‘secondary’ non-task-oriented products may arise and take on a value of their own, especially if these agents have some form of free will.
Marc Benioff, CEO of Salesforce, envisions a future where companies will manage both human and AI agents. As any manager of entry level human resources may have experienced, these resources tend to have lives outside of work and tend to blur the distinction as they continue to interact in the workplace. They also tend to develop their own short-hand communications. With an AI manager agent overseeing these worker agents, and interacting with other AI managers and teams, even further short-hand interaction is possible.
To the question of free will, a fuller theory of the physics at work in our universe could include this. As physicist Sara Imari Walker points out in this Closer to Truth video describing her Assembly Theory of the physics of life (https://www.youtube.com/watch?v=uvmYSxxlNn0), free will obtains as a result of this theory in sufficiently complex 'life' forms, where there is consciousness, physical information and historical time. This theory provides an opportunity to see why our universe makes this free will possible – i.e., to maximize the universe’s future possibilities. See from timestamp 1:04 on for an abbreviated discussion.
While AI may not be fully there yet, it is not inconceivable that our AI-gents will eventually exhibit some form of free will and demonstrate the ability to inter-communicate without us having any understanding what they are saying.