AI: Getting Specific About Origins

Agüera y Arcas wants to get more specific about origins. How does pattern emerge from randomness? How does code emerge from an unorganised soup of molecules?

In approaching these questions, he takes his cue from Turing and von Neumann, whose experiments anticipated the discovery of the molecular structure of DNA in 1953. The 1936 Turing machine established a minimalist prototype for computational function with the simple components of a coded tape and a read/write head. Von Neumann brought in a focus on embodied computation, where the components of the machine or body are part of what is written.

This is where Agüera y Arcas situates his work. His breakthrough came from adopting a programming language, devised in 1993, called “Brainfuck”. With just eight command symbols, Brainfuck set the parameters for a controlled experiment, in which Agüera y Arcas and his team used 64 byte tapes coded with “junk” drawn from a soup of code and data.

In the experiment, two tapes are selected at random, joined end to end, and run to test for interaction patterns. Then it’s rinse and repeat. The tapes are returned to the soup, and two more are run.

At first, nothing much shows up amidst the randomness. But after a million or so repeats (not massive in computing terms) the magic starts to happen. Loops appear. Patterns emerge. At around the five million mark, the non-functional code or “Turing gas” transforms itself into a “computorium” of replicating code.

In lectures, Agüera y Arcas shows a screenshot of this on his laptop: a vertical line down the centre of the field of data marks the “phase transition”. The image is reproduced on the cover of his book, as an emblem of the paradigm shift he is tracking.

If the transition to replicating code is indeed an expression of what is happening in the development of life forms, the theory of natural selection may lose its claim to primacy as the explanatory model for evolution. Richard Dawkins enthusiasts, hang on to your hats.

Agüera y Arcas does not engage in a polemical critique of Dawkins, but his book brings Margulis, an early adversary of Dawkins, into the centre of the arena. The pair faced off in a public debate in Oxford in 2009, where Dawkins’ popularised concept of the “selfish gene” came under pressure from Margulis’ theory of symbiogenesis, literally genesis through combination or fusion.

The Dawkins account is based on a Darwinian view of natural selection through competitive advantage; Margulis was drawing on research into the formation of microorganisms through combinations of mitochondria and chloroplasts, once independent life forms.

It was survival of the fittest versus a vision of biological complexity generated through endosymbiosis, a relationship in which one organism lives inside another, potentially resulting in a new life form – or, as Agüera y Arcas sees it, an impetus towards “fit” understood as pattern completion, rather than “fitness” understood as advantage.

Prediction and function

Agüera y Arcas’ central concepts are prediction and function, which work together to explain intelligence as the development of functional complexity through predictive pattern completion.

He is erasing a familiar conceptual boundary here: intelligence does not prompt function, it is function.

Intelligence, he argues, is a property of systems rather than beings, and function is its primary indicator. A rock does not function, but a kidney does. This is demonstrated simply by cutting them in half. The rock becomes two rocks, but the kidney is no longer a kidney.

So does a kidney have intelligence? Or an amoeba? Or a leaf? These questions are opened up, along with the question of whether Large Language Models have intelligence, which may a better way to frame it than asking whether they are intelligent.

Agüera y Arcas is not alone in taking an affirmative position. Influential biologist Michael Levin runs a research laboratory at Tufts University, where he and his team study the functional correlations between natural organisms and synthetic or chimeric life forms in search of “intelligence behaviour in unfamiliar guises”.

Their declared goal is to develop modes of communication with truly diverse intelligences, including cells, tissues, organs, synthetic living constructs, robots and software-based AIs.

Such an approach steers a course between the stochastic parrots view and biologist Rupert Sheldrake’s theory of “morphic resonance,” which proposes that organic form is a manifestation of memory, resonating through generations as genetic heritage. Agüera y Arcas avoids both Sheldrake’s intuitive and telepathic orientations, and the hard-headed constraints of mechanistic determinism.

The thesis presented in What is Intelligence? is unfamiliar rather than intrinsically difficult. Much of the explanation is easy enough for the general reader to follow, though Agüera y Arcas has a tendency to veer into more the technical and abstract terrain of programming, as if addressing an insider audience. The extensive glossary does not include standard programming terms, such as logic gates, gradients, weights and backpropagation.

At over 600 pages, What is Intelligence? is a marathon read and it is encumbered by tangential excursions. I’m not sure why Agüera y Arcas needs to go into the history of industrialisation, or anthropological studies of the Pirahā people of the Amazon. This is a book for dipping into rather than swallowing whole.

But its ideas are important. They may well be part of a major transformation in our thinking about where human intelligence sits in the rapidly evolving environment of AI.

- Author: Jane Goodall, The Conversation