As we saw in the previous post, the book Free Agents by Kevin Mitchell deals with the origins of human consciousness and free will. In a brief epilogue, the book addresses the topic of strong artificial intelligence—the real kind, which doesn't yet exist—and formulates some hypotheses about the possibility of its becoming feasible.
It emphasizes that one of the most active branches of research in AI, especially in recent years, is the field of artificial neural networks, which has led to advances such as Large Language Models (LLMs). It compares these neural networks in our programs with those that exist in our brains and in the brains of many animals more or less similar to us. It says that we are witnessing impressive advances in fields such as image recognition, text prediction, speech recognition, and language translation, based on the use of deep learning, remotely inspired on the architecture of the cerebral cortex.
These systems respond spectacularly to certain
types of requests, but fail (also spectacularly) with other types of questions.
Since ChatGPT appeared a little over three years ago, many of us have witnessed
some of these failures, several of which I've mentioned in this blog. The
questions that break down these systems are those that present novel
scenarios, which
are not represented in the training data. In contrast, for humans, this kind of
questions is easy to answer.
Why is this? According to Mitchell, because our
programs have been designed in a completely different way than living beings,
and are subject to limitations inherent in that design. To approach general
artificial intelligence, we shouldn't focus on current applications, but rather
on the intelligence of animals, which are capable of facing novel and
uncertain environments and
applying their past experience to predict the future—a future that includes the
outcomes of their own possible actions. This is very different from what LLMs do, which just
predict the next word.
Another characteristic of natural intelligence is
that it requires few resources. A living being is small; even a blue whale is
small compared to our gigantic data centers. A living being uses very little
energy, far less than the megawatts of our data centers. An animal cannot train
itself with millions of data points, nor does it have the time to exhaustively
computing its
behavior. If it does, it risks being captured by a predator or losing its prey.
One of the most useful characteristics of natural
intelligence is the ability to relate cause and effect. What does it take to do
this? Could machines do it?
To understand causality, a living being notices
that an event X is always followed by an event Y. But this can be due to two
reasons: either X is the cause of Y, or they are statistically correlated (see
my blog post titled Correlation
or causality). How can we distinguish between them? By acting on the world,
modifying the conditions, blocking event X, and checking if event Y disappears.
If this happens, it is very likely that there is causality. Otherwise, what
existed was statistical correlation. Mitchell expresses it this way: The
hypothesis has to be tested. Causal knowledge thus comes from causal
intervention in the world.
It is clear that current AI programs are not
capable of causally interacting with the world. Therefore, they are not
prepared to behave even like intelligent animals, much less like human beings.
If we want to move towards strong artificial intelligence, a radical paradigm
shift will be necessary.
[A]rtificial
general intelligence will not arise in systems that only passively receive
data. They need to be able to act back on the world and see how those
data change in response. Such systems may thus have to be embodied in some way:
either in physical robotics or in software entities that can act in simulated
environments.
The two systems he proposes are precisely those I
have used in my science fiction novels related to strong artificial
intelligence. Intelligent robots appear in two of the novels in my Solar System
series (Operation
Quatuor and Operation
Viginti). And in Jacob's
Ladder there are intelligent software entities that operate in simulated
environments. Not bad, for an author that doubts that this goal can be achieved,
at least in the short term.
This is the conclusion of the epilogue:
In summary,
evolution has given us a roadmap of how to get to intelligence: by building
systems that can learn but in a way that is grounded in real experience and
causal autonomy from the get-go. You can’t build a bunch of algorithms and
expect an entity to pop into existence. Instead, you may have to build
something with the architecture of an entity and let the algorithms emerge. To
get intelligence you may have to build an intelligence. The next question
will be whether we should.
The epilogue of Mitchell’s book is only five pages
long, but it is worth reading.
Thematic Thread about Natural and Artificial Intelligence: Previous Next
No comments:
Post a Comment