Ben Goertzel: Algorithmic Chemistry for AGI
Overview
Ben Goertzel presented a two-part talk at the Russian AGI community seminar. Part one covered origin-of-life simulations. Part two connected these dynamics to AGI via algorithmic chemistry.
Core Ideas
Origin of Life Research
Life criteria: metabolism (system keeps itself going, absorbs energy, uses it to move, build, fix, maintain) and reproduction (system spawns new systems).
Autocatalysis: a network of chemicals where each catalyzes reactions between others. Chemicals lift themselves up by their own bootstraps. The Eigen-Schuster hypercycle is a known example.
Getting to metabolism via autocatalysis is the hard part. Simulations have not been run at sufficient scale. The primordial sea was big. Without that scale, autocatalysis is statistically improbable to congeal.
Anthropic-Principle Simulation Approach
Standard approach: start from disordered chemical brew, let it run, hope autocatalysis pops out.
New approach: take initial condition (disordered brew, no autocatalysis) and terminal condition (autocatalysis emerged). Find the shortest path between them. Simulations show this path is not hard to get.
Conclusion: spontaneous self-organization of autocatalysis is not inherently difficult. The issue is scale. You need the scale of a whole primordial ocean.
Morphic Resonance and Anti-Resonance
Morphic resonance: patterns that have occurred are more likely to occur again. Plain vanilla chemistry explains this. No exotic physics needed.
Morphic anti-resonance: once a pattern occurs too much it gets inhibited. Pattern occurring a bit gets stronger. Pattern occurring too much gets weaker.
With properly tuned morphic resonance and anti-resonance, the distribution of near-autocatalytic sets becomes much healthier. New autocatalytic sets pop up easier and become more prevalent.
Algorithmic Chemistry
Algorithmic chemistry: programs rewrite other programs. Catalytic programs make rewrites easier. A big inter-rewriting program soup.
History:
- Walter Fontana (1980s): Lisp codelets rewriting each other in simulated primordial soup.
- Kristinn Thórisson: RepliCode rewrite rules and ERA cognitive architecture controlling humanoid robots.
This paradigm has never been rolled out at GPT-2 scale or above. Until now.
Hyperon as Platform
Hyperon infrastructure now supports:
- Atom spaces as algorithmic chemistry soups
- Distributed Atom space across multiple machines
- SingularityNET/ASI chain coordination without central owner
The infrastructure is ready for large-scale algorithmic chemistry experiments.
Autocatalysis in Cognitive Context
An autocatalytic set of thought patterns is a self-reinforcing belief system or set of procedures that help each other do things better.
Mental morphic resonance: grab what worked, keep doing it. Mental morphic anti-resonance: if a pattern becomes too prevalent, de-reinforce it to let other things get a chance.
Mental reproduction: a knowledge subnetwork that works in one context spawns copies that try similar things in other contexts.
Comparison with Other AI Paradigms
- Abductive reasoning: creativity via unexpected analogies and combinations
- Evolutionary program learning: mutate and combine patterns that worked before
- Algorithmic chemistry: may have even more creative power than the above
Role of LLMs
LLMs enable running these simulations in spare time. Ideas come from humans. LLMs structure code and analyze results.
LLMs are not yet smart enough to generate core ideas (like the shortest path approach) but are essential amplifiers.
Unique Historical Moment
Getting AGI ideas right, ethics right, creativity right, reasoning right, control and guidance of AGI systems: this is perhaps existentially important to our species.
After singularity, humans will not be the best problem solvers. The work matters now while humans can still contribute meaningfully.
Key References
- Eigen-Schuster hypercycle
- Walter Fontana algorithmic chemistry papers
- Kristinn Thórisson RepliCode and ERA
- Oparin model (water droplets as scaffolding for metabolic systems)
- Anthropic principle
- Rupert Sheldrake morphic resonance
- Hyperon system with Meta language and MM2
- SingularityNET platform