The Wikipedia article on Agentic AI says an Agentic AI system is an autonomous system that makes decisions and does things without human intervention. Robots usually have fixed rules, but AI agents analyze data and learn.
The Wikipedia one on intelligent agent emphasizes goal orientation as a defining characteristic. A thermostat, human being, organization or even geographical region is an agent if it has goals, makes decisions and acts on the decisions.
But is it useful to think of a thermostat or other simple control mechanism as an agent with goals? The goal of a thermostat is to keep a temperature constant. But is that the thermostat’s goal or the goal of its users? Isn’t saying it’s the thermostat’s goal, or that the thermostat has goals, rather than uses or a purpose, giving them capabilities that living things have?
As a general principle, it is not necessary to propose a purpose or reason to explain something.
A: Why is he running up the stairs?
B: He always runs up stairs (on other occasions and
other stairs)And some teleological explanations, eg, why the chicken crossed the road, make sense but are silly. https://mastodon.social/@drbean/116700446402745043
B’s explanation is a good explanation. Many explanations in science saying how something is consistent with or the result of a more general principle are like this, aren’t they. Avoiding teleological arguments about goals is one aim.
And this agency, is that in the first or second sense of DuplicitousAgent? I think proposing intelligent agents, they mean the first sense, doers rather than helpers.
What is agency? The Wikipedia referral page, http://en.wikipedia.org/wiki/Agency (See thread starting at htttps:hachyderm.io/@drbean/116615289209118211) has besides organizational agencies, like CIA,
Keywords appear to be autonomy, action, independence, ability, all
abstract ideas. But the words, goals, purpose, ends don’t
appear.
I don’t recognize in the language agents of https://www.tandfonline.com/doi/full/10.1080/0020174X.2024.2439995#d1e224
the sense of AI software agents above, ie, code deployed on your
computer by its developers independently pursuing its own goals. But its
definition of an agent as an entity pursuing goals with a degree of
autonomy is in line with that sense.
It focuses on goals and adds the requirement that it perform actions to achieve them.
Actions differ from behaviors by the agent being sensitive to their value in achieving their goals. Behaviors produced solely as a result of design or evolution are not actions. The designer or evolutionary lineage is the one sensitive, but/and the system itself is not, not being autonomous. Change over time does not imply sensitivity either. There must be also a set of goals to which some system function is attuned. Reinforcement learning systems are the prime example.
Mmh. Useful on the importance of focusing on actions and how goals are central to agency.
The Agentic AI Wikipedia article says that defining AI in terms of agents:
I don’t think treating AI as agents does any of these things, and giving people (real agents) and their AI tools the same status is bizarre, more appropriate for science fiction than scientific fact.
I agree purpose and goals are important. They’re fundamental to what computer processes are achieving. But ascribing agency to mechanical processes to handle their purpose and goals is unnecessary and a mistake.
People have tools. They have goals using these tools. Tools are created with a purpose in mind. People use the tools to some end.
The tools have a use, meaning they are useful. But they don’t have goals. ‘The goal of a tool’ means its use has a purpose, but the purpose is the user’s not the tool’s. Actions with the tool have a purpose. There is a reason for the action. It is the reason the tool is used.
The purpose of the hammer is to hammer nails into wood, but
When all you have is a hammer, everything looks like a nail.I think that’s relevant to purpose and agency. When you are using the hammer for another purpose, your purpose, without realizing it’s not the purpose for which tool makers intended it to be used, you’re liable to make mistakes.
Rather than using the hammer this way without thought about purpose, be very clear about your purpose, and how the way you want to use it is compatible or not with the way it is designed to be used.
Think of using a hammer to put bottle caps on beer bottles when you don’t have a bottle cap applicator. Without thought about purpose, you bring down the hammer with some force on the cap, breaking the bottle, instead of just giving it a light tap.
Our tools do our thinking for us, but we shouldn’t let them do that.
Is the AI community falling into the trap somehow, of thinking their apps with purposes have agency?
It is a truth not universally acknowledged, but as the gun lobby says,
Guns don't kill people, people kill people.That doesn’t mean everyone should have access to guns. It means it’s important to distinguish who/what has the purpose, the tool or you.
Don’t let your tools do your thinking for you. A tool is a means to an end, the goal or purpose of the tool, not an end in itself.
I don’t understand how the AI community can believe agency for their apps is the right approach.
Whose interests are served by apps having agency? Don’t obey/welcome our new overlords, Big Tech and AI.
In a curious reversal of agency for AI apps, Frankenstein’s monster, as distinct from an AI app, had agency as portrayed in the story by Wollstonecraft, but his identification with his creator by the use now of the Frankenstein name for him denies this agency in the present day.
AI community, be warned.
“Many of the truths we cling to depend greatly upon our own point of view.” –Obi Wan Kenobi
Simon Willison’s collection of definitions:
https://simonwillison.net/tags/agent-definitions
Compare and contrast with the problem of agency in OnEvolution.
Me atBack to AI