Artificial system

Frontiers of Artificial Intelligence: Do Androids Dream of Electric Sheep?

In 2020, American poet Andrew Brown commissioned a student to write a poem from the perspective of a cloud looking at two warring cities. The student wrote the following poem:

“I think I’ll start to rain,

Cause I don’t think I can take the pain

To see you both,

Fight like you do.

Awesome, isn’t it?

Well, Brown’s “pupil” turned out to be a computer program, not a human.

The program, called GPT-3, is one of the most powerful AI language models ever. Created in 2020 by the research firm OpenAI, its development has cost tens of millions of dollars. Formed on 200 billion words taken from books, articles and websites, GPT-3 can generate smooth text streams on any subject imaginable.

Algorithms are everywhere these days.

Companies like Amazon, Netflix, Spotify and LinkedIn feed our personal preferences there to create targeted recommendations. But their power does not end there. Google has acquired an AI to design computer chips in just under six hours. For a human, it would take months.

It was an AI, called AlphaFold, which solved the “protein folding problem”, one of the greatest unresolved challenges in biology. More recently, Deep mind has launched its own language model called RETRO, which the company claims can beat larger models 25 times its size.

We’ve even seen the emergence of hundreds of AI artists: algorithms that paint psychedelic images, create pieces of music and compose poetry, like Robert Brown’s “student”.

So, does this mean that we are heading towards the holy grail of the field: human-level intelligence, also known as AGI (artificial general intelligence)? To answer that, let’s take a closer look at the power and limitations of artificial intelligence.

Mathematics, music and algorithms are closely related

Music and computers share a surprising common ground: the mathematical language of algorithms.

In his book The Code of Creativity: How AI Learns to Write, Paint and Think, Mathematician Marcus du Sautoy explains that classical composers often use algorithms to create musical complexity. They start from a simple melody or theme and transform it according to mathematical rules. Using math, they create variations and additional voices to build the composition.

Composers with a strong signature style prefer some mathematical models to others. Mozart, for example, often used the Alberti bass model: three notes played in a sequence of 13231323.

As a result, computer programs like Musenet Where AIVA can identify the distinct mathematical patterns of specific composers or musical genres and construct a unique composition in a similar style.

Vision and language: AI blind spots

A year after the launch of GPT-3, the OpenAI team set out to achieve a bigger goal: to create a neural network that could work in both images and text. This required two new systems, named DALL-E and CLIP.

SLAB is a neural network which, according to OpenAI chief scientist Ilya Sutskever, can create an image from any text. It can produce the desired result even if it has not encountered a particular concept before in the training.

Two anthropomorphic daikon radishes walking their pet dogs. Image generated by OpenAI’s DALL-E model. Credit: OpenAI

CLIP can take any set of visual categories and instantly create reliable and visually classifiable text descriptions.

However, like the GPT-3, the new models are far from perfect. DALL-E, in particular, depends on the exact wording of the prompt to generate a consistent image. In fact, the flexibility of language remains a blind spot for artificial intelligence.

Take this sentence: “The children will not eat the grapes because they are old.”

Unlike humans, a program cannot determine who is “old”: children or grapes. How we interpret a word or phrase depends on context, but also on a level of understanding that goes beyond simple definitions of words. A human mind can do it, but an algorithm cannot.

Besides language, algorithms struggle to see the “big picture”.

An AI that recognizes images does so by asking questions about each individual pixel that makes up the image. However, the specific combination of pixels that we intuitively recognize as a cat, for example, is different every time. The program must therefore “learn” to correlate different pixels with each other to decide whether the photo contains a cat.

This is why many websites use a reCAPTCHA system to prevent automated spamming. We can easily identify cars or bridges in random photographs, but programs cannot.

Limits of machines and Moravec’s paradox

In an article titled Why AI is harder than you think, Melanie Mitchell, professor of computer science at Portland State University, discusses the more common misconceptions around AI.

Thanks to machine learning, modern AI exceeds many of our previous expectations, but its creativity and intelligence remain narrow: its skills come from large datasets provided by humans in a given context, not from a deeper level of understanding.

This is another reason why human language cannot describe the “intelligence” of AI. As Mitchell points out, we use words like “read,” “understand,” and “think” to describe AI, but those words don’t give us a precise description of how AI actually works.

And then there is Moravec’s paradox.

Hans Moravec is an Austrian-American robotic and computer scientist and currently an adjunct faculty member at the Institute of Robotics at Carnegie Mellon University.

In the 1980s, Moravec and other pioneering scientists like Marvin Minsky and Rodney Brooks made an insightful observation: high-level reasoning is easy for machines, but simple human actions, sensorimotor skills, and menial physical tasks are incredibly difficult.

Fast forward to today, and Moravec’s observation is solid.

So, does that mean that a four-year-old is smarter than a million dollar overkill AI? The answer is yes, if we recognize the complexity of everyday tasks that we perceive as easy.

Children can easily understand cause and effect relationships based on their interactions with the world around them and their accumulated experience. Machines often fail to make basic causal inferences because they lack that context.

The so-called cognitive revolution of abstract thinking and mathematical reasoning is a relatively recent development. However, our visual perception, hearing and motor skills are “built-in codes”, the result of thousands of years of evolution.

These human abilities are just not easy to develop by narrow AI, which mainly focuses on rigid mathematical reasoning to find solutions.

Intelligence is not just in our heads

As Mitchell points out, intelligence isn’t just limited to our head – it also requires a physical body.

In the 1868 novel by Fyodor Dostoyevsky The idiot, Myshkin tells the Yepanchin ladies the story of a man who is sentenced to death but who is pardoned just before execution. But it wasn’t just a story: Dostoyevsky was putting his own personal experiences down on paper. Through a fictional character, he described what his mind and body experienced on a cold day in 1849: the unique human feeling of his own imminent death.

According to Ben Goertzel, an expert in general artificial intelligence,

Humans are as much bodies as they are spirits. Therefore, to achieve human-like AGI, it will be necessary to integrate AI systems into physical systems capable of interacting with the everyday human world in a nuanced manner.

The idea of ​​AGI alone takes us to fascinating places. But until we fully understand intelligence itself, or even our own mind, a belief in AGI is like a belief in magic: something that dazzles us, makes headlines, but that we fundamentally can not understand.

Eleni Natsi

Eleni Natsi is a technical journalist and a marketing and communications professional. She is also a member of the Advanced Media Institute of the Open University of Cyprus (area of ​​research on AI and algorithms) and a guest lecturer on the postgraduate program “Communication and New Journalism”.