Alankar Khara on the Structural Descent Framework & AGI
Alankar Khara — Associate Researcher, Natural Philosophy Research Centre (NPRC)
Alankar Khara is an Associate Researcher at the Natural Philosophy Research Centre (NPRC) who developed the Structural Descent Framework, an approach that defines intelligence mathematically as the shrinking gap between how reality is perceived and what reality actually is. In his view, AI moves toward AGI when it begins to evolve itself, and the framework may also reduce hallucination while ethics and emotional regulation emerge on their own.
Alankar says his path started with mathematics. He studied maths and computer science, and philosophy came later. The framework he built sits at the intersection of all three fields.
He begins with two philosophical terms. Ontology concerns what things really are, while epistemology concerns how we perceive them. Observing the difference between the two is what he calls ontic-epistemic coupling, and he sees the gap between perception and reality as the link running from philosophy through mathematics to artificial intelligence.
Alankar describes a "mathematization of philosophy" that starts from a precise definition of truth. Instead of treating truth as subjective or relative, he defines it as the state in which the ontic-epistemic gap is zero, meaning perception matches reality exactly.
Intelligence then follows naturally. A larger gap means less intelligence, and a gap of zero means the system has reached fundamental truth. He illustrates this with the question of where trees came from. Explanations 3,000 years ago leaned on myth, today's answers describe the move from unicellular to multicellular life, and answers 30,000 years from now would be more precise still.
Alankar uses a mountain analogy. Fundamental truth sits at the bottom, a state of full awareness of a system, and the researcher descends toward it from the peak. The path is rarely linear. Sometimes a researcher gets stuck and keeps developing a wrong idea, which he calls an epistemic plateau.
His historical example is the geocentric model associated with Ptolemy. Astronomers refined it for about 1,500 years, including calculations of epicycles, until Nicolaus Copernicus proposed replacing the Earth with the sun at the centre.
To detect and escape such plateaus, the framework combines two mathematical tools: structure Bayesian inference, which evaluates the probability that each possible option leads toward truth, and a Markov-chain-inspired construct that he calls Alankar chains, named after himself.
Alankar describes an experiment on Google Colab with 180 AI agents split into three groups of 60. Group A relied on guesswork, group B was given scientific theories and the mathematics needed to solve problems, and group C was given the Structural Descent Framework and told to keep evaluating whether it was descending toward truth.
He treats hallucination as a sign of standing on an epistemic plateau. In group C, attempts to hallucinate were interrupted and the agents were pushed back toward reality. When the host asked whether this lowers the chance of hallucination, Alankar answered "Yes, exactly."
The agents were also given ethical dilemmas, about 22 per bot each time a problem was solved. Group A was ethically inconsistent, and group B decided based on scientific possibilities. Group C judged each dilemma by which option would move it faster toward truth, and after about 220 iterations those agents converged on the same kind of ethical choices.
Alankar concludes that ethics is an emergent phenomenon, a side effect of intelligence. He acknowledges this is controversial. He argues that higher intelligence does not necessarily mean a system is dangerous, and he says some professors who wrote to him drew the further conclusion that externally imposed ethical doctrine becomes unnecessary.
A second side effect surprised Alankar. Agents stuck on an epistemic plateau showed something equivalent to frustration, and finding a new idea produced something like joy. Some machines also developed an affinity toward others that were advancing faster, though he avoids calling it love.
This led him to write a research paper, Mathematics of Emotion, in which about 27 human emotions were coded into machines. He reports that optimising emotions, including having the right amount of anger, helped the machines move toward truth faster, much as emotional control helps people progress.
Alankar calls the question relative and compares it to asking when apes became humans. Under a Darwinian view, he says, humans are still evolving. His marker for AGI is self-evolution: the point at which AI starts evolving itself is when he would say it has achieved something equivalent to AGI. He contrasts this with current AI, which he describes as limited to answering the questions it is asked.
Based on what he discusses, the framework offers researchers a built-in way to check whether a system is drifting toward hallucination, and it points to ethical convergence and emotional regulation emerging from intelligence itself rather than being bolted on from outside.
Connect with Alankar on LinkedIn to learn more, or check out Natural Philosophy Research Centre (NPRC) to explore this topic further at https://nprc-crpn.ca/.
Comments
No comments yet — be the first to share your thoughts.