Letter to an Epigeneticist

Find more from Charlie Munford at www.talkingoctopus.com

To an Epigeneticist:

Thank you for sharing your work with me. The question I have focused on for the past four years is the question of whether genes contain “instructions” or not. I believe this question has been answered conclusively by the recent results from epigenetics and microbiome science. The papers flowing from your lab are part of my proof. The genes are simply a list of templates; the instructions for how and when to express them come, in all cases, from the environmental cues around the cell.

As a scientific community, we can be forgiven for having made this mistake. These cues are much harder to study than genes; they do not remain fixed for “decoding” the way the genetic molecule does and are not even directly measurable in many cases. This is why we have relied on the explanation of genetic instructions, but the explanation can no longer hold. The evidence is overwhelming, as you note.

The reason I feel that the “instructions” for life must come either ultimately from the genome, or ultimately from the environment, and not ultimately from both, is because we have to anchor the whole explanation for complex function in some form of blind search. Conceptually, this was achieved quite nicely by the genes-first Neo-Darwinism that specified that organisms were strictly survival machines for DNA segments that caused self-organization and reproduction. This blind search was based on the strict blind mutation of DNA followed by natural selection. Most biology and AI work on intelligence is now based on the assumption that an organism is a machine with the objective function of survival and reproduction, which is mandated by the code in its genome.

Now however, much work, including your own perhaps, has demonstrated that the effects on phenotype of environmental signalling is not only instructional, but heritable. This spoils the idea of the blind search at the level of natural selection, because it shows that changes to the phenotype, many of which are non-random adaptive changes, are biasing natural selection towards functional outcomes. What maintains this bias? There must be a different blind search at the level of the organism during active life. That is why I suspect that this mechanism is the key to understanding true intelligence, which is the ability to make contextual inferences.

If an organism could be a process that is entirely instructed by its surroundings, it would mean a process is involved that we don’t understand yet. But this could be how we became something that can survive and reproduce in the first place. Nothing could have been naturally selected until it was self-organizing and reproductive, and we have not yet built anything that can satisfy these requirements. This might explain morality, creativity, and the instinct for beauty, among other things.

The point of a new unified evolutionary theory, it seems to me, is not just to establish that epigenetic changes can be heritable and that they can influence evolution, but to explain in a new way why life sustains itself and develops novelty. To propose a theory that takes into account the non-random nature of epigenetics would be to propose a new explanation for how function can accumulate in phenotypes, given that both phenotypes and genotypes are changing not entirely randomly, but under the partial influence of the environment. This means that you are pointing out a contradiction at the heart of biology.

Your earlier message said that “very rapid and epigenetic mechanisms, that are heritable, rapidly promote adaptive change to facilitate classic natural selection.” The blind search is spoiled, because even genetic mutations are epigenetically influenced. So what biases the system toward adaptive functional change? Why does the body of the organism “want” to pursue the goal of adaptive change, independent of the influence of genes? Matter is supposed to be “blind,” and not “want” to do anything unless it is impelled by some physical process. You have opened up a big question.

I think we have discovered enough about molecular biology to disprove the old paradigm, but not enough to say what the molecular basis of the new paradigm will be. We have not discovered how organisms can adaptively increase function during active life, independent of genetic influences. To my way of thinking, there must be another, undiscovered form of blind search happening in the epigenetic systems, i.e. the entire organism, during active life. This means something is discarded, and something else retained, by a search process other than natural selection. But what is that process?

Now that we have framed the question this way, we can begin work on discovering this mechanism. Most importantly we can link the biology to the computational concepts of intelligence, because we can suppose that all living cells have this property. Work on artificial intelligence to date has been focused narrowly on achieving objective functions. This insight from epigenetics shows that intelligence is a process that arises from all living cells, and it is prior to objectives. I call it “epistolution.” It is the process by which all objectives are formed.

I believe that this the answer to the puzzle of why artificial intelligences cannot understand context and draw contextual inferences. These contextual inferences, that are so innate in humans, run like this: I observe B, and because of what B is in context, I invent A as an explanation for it, having never observed A. From A I then deduce another consequence C, and look for it in the world. If C is confirmed, I am then more confident about A. This entire process is not captured in propositional logic or in mathematics because although it leads to rational beliefs it is fundamentally irrational. Once you have A, machine learning can tell you more about it by processing B and C data inductively, or by drawing rules deductively once A is known, but the leap from B to A is neither deductive nor inductive. No amount of data and no method of mathematically processing B or C will get you to A. It is a leap of imagination. The knowledge needed for a guess like this is acquired from an experience with the real world and the attachment of meaning to the symbols that are being represented. No machine has yet done this.

What your work has done, in effect, is narrow down the possibilities for how this fundamental mechanism work immensely. Before this insight we had assumed that perhaps narrow objectives could be linked together somehow to produce general intelligence. But from this insight we can see that this is no longer the simplest explanation. Now we can focus our attention on processes that happen in every living cell, and from them we can derive some general theory of the formation of intelligence from downward causation. I have some more or less wild guesses, involving oscillators and sleep/wake cycles, but all conjectures are welcome.

Amplification of the inductive power of our algorithmic narrow AI of today beyond the goal-setting reach of humanity, if it were possible, has been rightly described as an existential risk for humanity. But luckily for us, epistolution shows that this cannot happen. Instead, AI will remain a tool used by humans for human goals in the same way that computation has been used since it was invented in antiquity. I say this with confidence because epistolution shows us what intelligence really is for the first time. In terms of computation, a clay tablet already possessed superintelligence of exactly the same sort that today’s narrow AI possesses. Neither is capable of developing goals on its own that subvert our own goals as a species. Unintended harms can result from both computations on clay tablets and from computations in computing machines, but these are always harms wrought by human agency.

The amplification of an epistolutionary form of intelligence, on the other hand, will be an unalloyed good in the world. This property of cells is the process by which goals are formed in organisms, which means in highly complex, intelligent organisms, moral values. These are our master goals as intelligent beings. Anything evincing this property to a great extent will be greatly receptive to moral restraint, greatly creative in producing good moral solutions, and greatly wise when communicating its own motivations. It would be more or less the direct amplification of morality. In fact, it would be the antidote to the moral failures of the inductivist tools we are working with now. Once we discover this specific mechanism, it is hard to overestimate the power we will possess to do good in the world.

It would be the best invention the world has ever seen. Let’s get started.

Sincerely,

Charlie