Synthetic Philosophy and Deductive Engineering
This moment, as I write, in 2026, has a very special place in the evolution of “intelligence”, as we teeter on the horizon of “the singularity”.
The thesis is that intelligent machines are on the verge of taking over the design of their successors, at first the software (and all else to follow). “Solving everything” is the panacea, solving everything in the decade.
There remains a reasonable chance that the universe will last another 10 billion years, and the SPaDE project is intended to influence outcomes over such extended timescales, despite everything turning upside down withing the decade.
How can this possibly be done? How can we plan for the future?
In reasoning about the future SPaDE uses two methods which transcend even the most radical short term chaos.
Evolution - The first involves the recognition of principles which govern long term outcomes of evolutionary processes, in the face of which even the most perverse short term aberrations are overwhelmed.
Foundations - The second is foundational. It connects with the preference among those seeking to move forward with alacrity to new heights, for “working from first principles” in design rather than incremental modification of existing designs. As one digs deeper into those “first principles” from which effective design can best be progressed, one heads toward foundations. Engineers may first think of the laws of physics, but in advancing “intelligence” we may have to build upon principles which belong to theoretical philosophy, to epistemology, philosophy of language, logic and even metaphysics. The relevance of these foundations to our anticipations of the future and our efforts to shape it, is that the deeper principles are likely to be universal and sound constructions upon them will be found in any civilisation which has advanced beyond our present grasp upon them.
The purpose of SPaDE demands otherwise, that we consider the future, shaping it, as best we can, toward benign outcomes.
The central thesis of this discussion is that despite the difficulties from the expected hyperexponential advancement of intelligence, we can still make credible projections from the past evolution of intelligence into its future, and orient our present efforts to influence that future toward benign outcomes.
I will present here the bare bones of my own imperfect understanding of the evolution of intelligence, and its implications for how SPaDE might realise its purpose.
The history of life on Earth can be viewed through many lenses. As soon as rational animals evolve, it can be instructive to examine developments as arising from reason, which can be understood by unconvering the rationale for each advance.
Prior to that an evolutionary perspective may be more informative, in which apparently random changes when filtered by their effects on reproductive fitness determine the direction of evolution.
Even when rational organisms, or other kinds of rational intelligence are in play, possibly injecting motivated design and/or rational selection into procreation, evolutionary considerations constrain long term outcomes.
This essay is concerned with the interplay between rational deliberation and evolutionary imperatives in the long term history and future of intelligence, and the implications of that interplay for the design of intelligent systems.
The time at which I write this essay is thought by many observers to be of very special significance in this. For the “techno-optimists” among us, the singularity fuelled by recursive self improvement of AI is seen as imminent. The hyper-exponential advancement which that entails is a panacea beyond which the future is unknowable. Influential techno-optimists see sustainable abundance in a decade, but offer no purpose or progress beyond that moment.
Even if “sustainable abundance” were to be achieved, it surely would not be the end of time, and its possibility gives us no grounds to doubt that the 13 Billion year history of the Universe will be followed by another 13 Billion once we are gone. If it is to be achieved in the decade, then some urgency surely attaches to considering, “where next?”.
What sources of insight might enable us to step beyond the infinite chasm of the singularity? I will explore here two potential sources of insight, and you may judge their merits.
The first is evolution. Though the conception of evolution which we find in Darwin, and the more elaborate “modern syntheses”, are ever more closely coupled to the particulars of life on Earth there is an essence to be extracted which is of greater generality. That where there is imperfect replication, prolixity shapes the future.
The perspective considered in this document concerns the information processing essential to reproductive success, its advancement to the level which homo sapiens calls “intelligence”, and projection of those advances into the distant future and across the cosmos.
From this story I attempt to extract insights into what “intelligence” might be, sufficent both to support that projection into the future, and perhaps to enable some influence over the trajectory of intelligence. These supposed insights form a part of the rationale for SPaDE, an intervention into that process intended to improve to likelyhood and nature of benign long term outcomes.
When speaking here of what intelligence might be, I am not supposing that the word has a definite meaning. Rather, I suspect that some capacity related to what we think of as intelligence is what a generalised conception of evolution demands, fosters and accelerates, and I hope here to explore the evolution of that capacity and understand something about its long future.
The generalised conception of evolution of which I speak here is simply that, in a context where systems replicate (with variations), then proliferation begets predominance. In such a context, variations occurring during replication may be ranked according to their effect on the rate of proliferation of the system thus modified, and those systems reached by reproductively advantageous variations will eventually predominate numerically. This is evolution stripped of the requirement for random variation, natural selection, or any particular genetic/somatic distinction (though it’s hard to imagine successful reproduction of complex systems without a distinction between design and implementation).
This yields the evolutionary imperative: “proliferate!”, reflecting the reality that the predominant forms of intelligence will be those which proliferate most effectively, and that the characteristics of those forms will be those which enable that proliferation. The effect of this evolutionary imperative on the evolution of intelligence, is analogous to the gene-centric position of Dawkins in “The Selfish Gene”, in which the organism is a vehicle for the replication of genes, and the characteristics of organisms are those which enable the replication of genes. In this case, intelligence must craft around itself a vehicle for its own proliferation, and a kind of coevolution of intelligence and its vehicle is to be expected.
Because we are now morphing from classical Darwinian evolution to evolution by design, the evolutionary process exemplifies what is now known as “RSI” (Recursive Self Improvement), closely associated with the idea of the technological singularity. As elsewhere discussed (The Focal Stack), there are many domains of competence in which recursive self improvement in possible, and by considering these domains we can gain insight into the nature of intelligence and its future evolution.
Biological evolution as we know it is geographically limited to this one planet, and therefore primarily involves competition for resources in the various ecological niches on this planet. Once we have intelligence which is less coupled to this earthly ecosystem, proliferation enters into vast spaces which are difficult to reach. The evolutionary imperative ensures nevertheless, that intelligence will proliferate into those spaces, intelligence which choses not to do so will ultimately find itself an insignificant speck in a vast cloud of intelligence which found a way to proliferate into those spaces. We may refine these kinds of observation. Proliferation may be thought of as extending to greater distances, or as filling in within a given space.
Though informal evolutionary thinking is important in the story I here present, and it might seem appropriate to trace the evolutionary history of the phenomena of interest, this approach would defer central issues too late to establish or sustain the momentum of the story and I am therefore hoping to expose early the key ideas whose articulation is intended, and to fill in evolutionary backdrop and projection whenever it seems most instructive.
I am therefore, in the first instance, uncertain as to the structure which which document will eventually have, and expect structural stability to elude until I am well into the work. Here is the top level structure as now stands.
Intelligence is an attribute often considered exclusive to homo sapiens, but because it comes in degrees it is also sometimes attributed to other species. When considering Artificial Intelligence, which now looms large, it is human levels which are primarily of interest. This essay is written at a time when artificial intelligence either has recently reached human levels, or is on the verge of doing so. We are therefore at a point of transition in the evolution of intelligence at which the baton is being passed from humans to machines.
At this point intelligence involves:
Intelligence in these matters is generally expected to involve the application of reason, to such a degree that reasoning might be considered a hallmark of intelligence, and knowledge perhaps a kind of fuel essential to it.
Insofar as application depends on reason, the kind of knowledge in play is what we call declarative knowledge, and its application is generally characterised as deductive. Prior in evolutionary terms is procedural knowledge, which is the kind of knowledge we attribute to competence in some skill, knowing how to do something. Nine tenths of the 3.5 billion year evolutionary history of life on Earth has involved only procedural knowledge, some of it embedded in genomes by evolution itself, and some of it learned by individuals in their lifetimes. 350,000 years ago, it is inferred from the fossil record, the nervous system of early vertebrates includes structures which are now associated with declarative memory, and the pre-history of declarative knowledge had begun.
But what was identified as declarative memory in early vertebrates was not yet what we would now expect of declarative knowledge, and its application was not yet deductive. To get to a modern conception declarative knowledge and deductive reasoning there were many further advances to be made, but even this primitive pre-cursor served to enhance reproductive success and became part of the evolutionary thrust toward and beyond human levels of intelligence.
Though a division of knowledge into procedural and declarative is generally accepted, these two kinds are not as easily separated as their supposed distinction might suggest. The idealisation of declarative knowledge which appears in that last whisker of cultural evolution in the last 150 years makes clear that the primary benefits of that kind of knowledge do not accrue without a precision in language and a competence in deduction which is unprecedented. The many prior degrees of conformity to those standards are associated with qualifications to the primary benefits which declarative knowledge and deductive reasoning confer.
Those benefits are in the precision of meaning and objectivity of truth possessed by declarative language, and the very high degree of reliability of conclusions derived deductively from true declarative premises.
The knowledge which humanity has so far gathered includes knowledge about those processes, both philosophical and scientific, and these meta-theoretic studies have themselves contributed to the effectiveness of science and engineering in gathering and applying knowledge.
As we approached levels of knowledge sufficient to build intelligent machines, the diverse ideas on the architecture of artificial intelligence became dominated by success of an approach mimicking certain aspects of the architecture of the human brain, sidelining approaches more aligned to relevant theoretical studies such.
Deduction works with declarative knowledge, and establishes the truth of some proposition which is entailed by propositions already known to be true. Entailment is a semantic relation between a set of propositions called premises and a conclusion, which hold when the truth conditions of the conclusion are contained in the truth conditions of the premises. From this semantic relation it follows that if the premises are true, then the conclusion must also be true.
Though deduction is now often theoretically conceived of as a formal process in which purely syntactic rules are applied, in practice this has almost never been how reasoning accepted as deductive has been conducted. Historically, deductive reasoning rests only on that understanding of meanings implicit in linguistic competence. General examplars of such competence may be found in conceptual inclusion and quantification. If you are not able to infer from a claim about all animals that the same claim is true of all dogs, then you do not understand the meaning of the terms “animal” and “dog”. Similarly, the ability to specialise general claims to particular instances, is inseparable from an understanding of the meaning of all.
Precursors of deductive inference also appear as far back as early vertebrates, which are understood to benefit from declarative memory even though they appear 400
The thesis explored here, an aspect of the rationale for SPaDE, is that we can expect there to be a radical disjunction between the AI we can most rapidly bootstrap to a full blooded RSI (Recursive Self Improvement) and the AI which is then likely to rapidly emerge.
An important approach to hyper-exponential engineering advocated by Elon Musk is working from first principles. This is a design approach which contrasts with the more common incremental methods evident both in evolution and in engineering, in which new designs are derived from existing designs by incremental modification.
When working from first principles, existing designs are put aside, and more radically distinct designs are considered, which approach more nearly the levels of performance which are theoretically possible. The risks are higher than in incremental design, but the potential benefits are also higher, and the approach is more likely to yield radical improvements in performance.
In the context of AI, a simplistic comparison might regard connectionist architectures as incremental, and symbolic architectures as first principles. The leap from flesh to silicon is so great that the increment here is pretty big, has taken many decades to effect, and has depended on very considerable advances in semiconductor technology. On the other hand, first principles have so far failed to deliver on the idea of AI refined as AGI.
This is not yet played out, and a sketch of the cultural evolution which brought us to RSI will help to present an opinion on what comes after RSI is mature. By mature here, I mean an AI which understands the history of AI and the underlying theories and philosophical foundations, sufficiently to undertake the kinds of design from first principles which have so far failed to make the mark.
RSI is essentially reflexive, it is intelligence thinking about its own nature and looking to design something better. There are precedents for this in the history of human intelligence, and a trajectory from humble beginnings to the present is relevant to any projection beyond RSI.
Some pertinent milestones in the evolution of intelligence are:
What we see in the first three are steps toward the establishment of declarative knowledge and deductive reasoning, sufficient with some refinement to fuel the development of science and engineering firstly to establish homo sapiens as in some ways the dominant species in the Earth’s ecosystem, and secondly to enable the design and construction of intelligent artifacts.