By Javier Surasky
Understanding,
at least at a basic level, what an AI does is a window of opportunity to deepen
our grasp of its advantages and dangers, and also to identify possible entry
points where each person and profession can contribute.
Let us then
carry out a first "dissection" of how any AI operates, to identify its main
elements.
The first
step is to understand that any AI interacts with the non-virtual reality in a
reciprocal exchange that, in principle, modifies both: the ""real reality"" (let's' call it RR) provides technology and data to the AI's' ""virtual reality"" (VR), which in turn returns information (or performs actions) that can
transform the RR.
This gives
us a first, somewhat obvious step worth mentioning: AI operates within and acts
upon a larger ecosystem that is RR, becoming what we call an "agent" to reflect
that it has the capacity for action. Both ChatGPT and a self-driving car are
agents. Google defines "agent" as "software systems that use AI to achieve
goals and complete tasks on behalf of users. They display reasoning, planning,
and memory, and have a level of autonomy to make decisions, learn, and adapt." This definition is crucial as it provides several of the elements we will
examine below.
Before
that, we must include something the definition leaves out: perception, meaning
the agent's ability to receive context, that is, to receive inputs at any
moment, based on which it builds its own experience history, which in turn will
inform its decisions. Just as with a person, immediate perceptions, combined
with its historical experience, will guide the AI system's actions.
The bridge
between what the agent perceives and what it does is its function, what it has
been trained to do (play chess, drive a car, translate texts, etc.). The
function, of course, takes shape from the program that creates the agent.
It is to be
expected that any AI will orient its actions toward what it considers good and
move away from what it considers bad. But what is good and what is bad? For an
AI, good is what brings it closer to its assigned goals, and bad is what moves
it away from them. If an AI is given the mission to scam people, then "good" for it will be to scam as many people as possible, and "bad" will be failing to
do so. We have already discussed in a previous post that AI is a means, not an end
in itself. The AI's intention always comes from the humans who create or use
it. Achieving its goals is the measure of success for any AI, and thus requires
that the goals be measurable as a precondition. To measure success, the most
common approach is to design metrics based on the desired results in RR, rather
than on how the agent should behave.
Combining
what we have just seen, any AI should be understood as a rational agent with a
rationality oriented toward achieving the goals for which it was created, based
on its perceptions and perception history, and seeking to maximize its success
measure.
Given the
importance of perceptions for successful results, AI may take actions aimed at
"refining" its future perceptions. To do this, it gathers and seeks
information. Consequently, the usefulness of the information it holds will
depend on the rationality with which the agent was endowed. Here lies one of
the most important elements in AI development: the agent's ability to learn
from new perceptions combined with its history, transforming raw data into helpful
information for rational decision-making. This learning ability gives AI its
autonomous nature: it does not depend exclusively on the information provided
by the programmer. It partly explains the existence of "black boxes" that
prevent us from knowing exactly how an AI reaches a decision.
As a
result, any rational AI agent includes four fundamental components:
- The learning element: to
improve itself.
- The performance element: to set
the actions the AI will take.
- The critical element: to review
the agent's performance and, if necessary, modify the performance element
to improve its success measure.
- The problem generator: suggests
actions that may lead the AI to explore innovative experiences, preventing
it from closing in on those it has already mastered. It encourages the AI
to remain unsatisfied with its current expertise and constantly explore
new paths.
And this is
where the idea of an "algorithm" plays a central role. Algorithms are simply
process organizers, sequential or iterative. A cooking recipe is an algorithm
that guides the agent (the cook) through a step-by-step process to
achieve a desired result, with the success measure being the diners' satisfaction with the
prepared food. It's worth noting that success is best measured by changes in the
environment rather than by the activities performed. We have all dealt with
algorithms in our lives. Did you study the calculation of the greatest common
divisor in primary school? That is one of the earliest "coded" algorithms in
history.
The
algorithm is the foundation (the soul?) of AI, the manager of the processes
that link its perceptions, history, goals, success measure, and rationality to
achieve the best rational result it can obtain. That is, the result may not be
perfect or the absolute best, but it is the best the AI can deliver given the
described elements. Search and planning are AI subfields focused on finding
sequences of actions that allow agents to achieve their assigned objectives.
But agents
are not only made of data and programming. They also have a physical support
(chips, cables, boards), which we call AI's architecture. Thus, AI development
work can be described as the path toward programs that, based on current architecture limitations, exhibit rational
behavior while minimizing code and expanding the ability to perceive, create
histories, and, from there, maximize success based on their rationality.
Of course,
we are leaving out aspects such as the types of environments in which AI
operates, its ability to perceive them fully or partially, the continuity or
change in their conditions, among many others. Nor have we analyzed how these
elements change, or may change, when we speak of weak AI, strong AI,
generative AI, or agentive AI. But we now have a better idea of "what "AI does, and that is no small thing.