SAGI, Search for Artificial General Intelligence

Artificial intelligence research has converged on a narrow strategy: train increasingly large transformer models through backpropagation on increasingly large datasets and compute clusters. This approach has produced remarkable systems, but its success has encouraged us to treat one effective path as though it were the only possible path.

SAGI starts from the fact that we do not know the architecture of general intelligence. We do not know whether it will be based on transformers, backpropagation, offline training, or internet-scale datasets. Instead of continuing to scale one known paradigm, SAGI proposes a worldwide search through the space of possible learning systems.

The project is inspired by SETI@home, artificial life, evolutionary computation, and Stephen Wolfram's search through large spaces of simple computational rules for structures that might underlie the laws of physics. SAGI applies the same philosophy to intelligence: construct broad but computable spaces of architectures and learning rules, then search them experimentally.

The goal is to discover the algorithms that could lead to AGI.

A distributed living laboratory

SAGI is a distributed evolutionary laboratory in which every candidate learning system is represented by a genome. The genome describes its architecture, memory, internal state, plasticity, and the rules by which it changes while interacting with an environment.

Each genome is expressed as a small Tamagotchi-like organism. The organism is the phenotype of the underlying learning system. Its behaviour reveals whether it can remember, adapt, transfer knowledge, recover from mistakes, and respond when the rules of its environment change. This turns an abstract learning algorithm into something observable and gives users a direct relationship with the artificial minds they are cultivating.

The laboratory should contain many populations and many kinds of environments. Some searches will be highly specialized, targeting properties such as rapid adaptation, efficient memory, or performance under strict compute constraints. Others will be intentionally broad, exposing organisms to changing rules, unfamiliar combinations, new sensory representations, multiple tasks, and open-ended ecologies.

The purpose is not to find the organism that exploits one benchmark most effectively. It is to find learning systems whose abilities transfer across environments and whose internal rules remain useful when the surface form of a problem changes.

The exact network architecture of SAGI remains open. One possibility is a centralized system in which users perform searches locally and a trusted server verifies promising organisms on hidden tasks and seeds. Another is a more decentralized system in which results are reproduced and validated through peer review by multiple independent nodes. A hybrid system could combine central coordination with distributed replication and verification.

This decision also affects the economic model. A centralized version could sell subscriptions containing compute budgets. SAGI would rent or operate hardware, add a margin, and run additional populations for users. Since strong results can win real bounty money, users have a concrete reason to purchase more search capacity.

A decentralized version could instead reward compute, verification, benchmark design, and successful discoveries through some form of token or shared incentive system. The incentive structure should follow the technical architecture rather than being added merely for speculation.

Searching architectures and learning rules

The genome must describe more than a conventional network with mutable weights. SAGI should search the architecture and the learning process together.

A candidate organism may contain recurrent structures, persistent memory, plastic connections, local update rules, modulatory signals, symbolic operations, or components that do not resemble present neural networks at all. It receives observations, produces predictions or actions, and modifies itself continuously as a result of its experience.

Training and inference do not need to remain separate phases. The most interesting organisms may keep learning throughout their lives.

Initially, SAGI will need carefully designed genome languages: constrained spaces that are broad enough to produce genuinely different systems but structured enough to search. Later, researchers and users should be able to add new primitives, architecture families, mutation operators, environments, and entire search substrates.

Evolution supplies broad exploration, but it is not the only intelligence guiding the process. Humans can inspect lineages, edit genomes, create bounties, and direct searches toward promising regions. AI research agents can run experiments, modify code, compare results, preserve useful failures, and propose the next experiment, in the spirit of Andrej Karpathy's autoresearch loop.

This creates a combination of evolutionary search, human intuition, and automated research. Instead of one research team following one promising direction, the system can maintain many competing lineages and hypotheses simultaneously.

The generalization hypothesis

The deeper motivation behind SAGI is a hypothesis about transformer generalization.

Transformers clearly do more than literal memorization. They can combine concepts, transfer patterns, and produce responses that did not occur verbatim in their training data. Yet their scaling laws suggest that much of this capability may still depend on broad statistical coverage of an enormous, long-tailed distribution of examples.

A simplified empirical scaling relation is:

$$ L(N,D)=L_\infty+A N^{-\alpha}+B D^{-\beta} $$

where $N$ is parameter count, $D$ is the amount of training data, and $L_\infty$ is the irreducible loss. Compute-optimal results indicate that model size and training data should increase together.

This is useful engineering guidance, but it may also tell us something about the learning process. If improvement largely comes from representing more local structures and seeing more of the long tail, then additional parameters are required to represent those structures while additional data is required to reveal them.

Knowledge can instead be imagined as a generative grammar. A compact grammar stores reusable concepts and rules whose combinations explain a much larger number of concrete situations. Its stored structure grows approximately with the number of reusable elements, while its possible interpretations grow through their combinations.

A shallower learner may discover many useful local templates without fully discovering the deeper grammar that connects them. Each template generalizes within a limited region, and an enormous collection of such templates can appear broadly intelligent. However, expanding that coverage requires continuing growth in both examples and representation.

The SAGI hypothesis is that current transformers may derive much of their power from this broad, shallow form of generalization, while an alternative learner could discover deeper reusable structure and therefore generalize from far fewer examples.

Such a system might still require enormous compute. It may need to search over hypotheses, actively conduct experiments, reorganize its memory, construct internal programs, or test causal explanations. The important distinction is that computation and data are different resources. A system may spend substantial computation extracting a powerful rule from a small number of observations.

Its scaling law could therefore look very different: less dependence on ever-larger datasets, with more computation devoted to finding compact structures that transfer broadly.

Why less data matters

Better sample efficiency would matter scientifically, economically, and legally.

Frontier models currently depend on enormous collections of books, websites, software, journalism, images, audio, and other human-created material. This produces continuing disputes over permission, compensation, provenance, and copyright. The pressure to absorb nearly all available cultural material is partly a consequence of a learning paradigm that gains capability through ever-larger datasets.

A system that learns stronger abstractions from smaller, carefully licensed, synthetic, or interactive datasets would not automatically resolve every copyright question, but it would change the economics. Curated and compensated datasets would become more practical, and advanced models would be less dependent on indiscriminate scraping.

Sample efficiency is therefore not merely another benchmark. It may be necessary for a more sustainable form of artificial intelligence.

Benchmarks, bounties, and incentives

SAGI should measure intelligence across many dimensions rather than reducing it immediately to one score. Benchmarks may test transfer to unseen task families, adaptation after rules change, learning speed, robustness, memory efficiency, computational cost, and performance across different hardware.

Leaderboards give each dimension a visible competitive structure, while bounties allow researchers, companies, sponsors, or community members to direct effort toward specific unsolved problems. A bounty might reward the first verified organism that reaches a transfer score under a fixed interaction or compute budget.

Because successful discoveries can earn real money, compute becomes more than a donation. Users can invest their own hardware or purchase additional search capacity in an attempt to produce stronger organisms. This creates a direct economic connection between research progress, compute allocation, and rewards.

The system should also reward less visible contributions, such as creating difficult environments, reproducing results, identifying benchmark exploits, contributing architecture primitives, and finding useful intermediate lineages.

A distributed frontier lab

Current frontier research is concentrated because transformer scaling rewards enormous homogeneous clusters, vast datasets, and capital-intensive training runs. A search across many small, irregular and heterogeneous learning systems is naturally more distributable.

Different populations could run asynchronously on CPUs, GPUs, laptops, cloud hardware, and future devices. The scientific work could also be distributed: users contribute genomes and environments, independent nodes reproduce discoveries, human researchers direct promising searches, and AI agents continuously run new experiments.

The ambition is not merely to crowdsource compute. It is to create a worldwide research organization with no single fixed architecture, benchmark, or research agenda.

SAGI begins with small artificial organisms and deliberately limited genome languages, but it can expand into neural, symbolic, recurrent, plastic, cellular, program-like, and hybrid systems. Some searches may begin from known algorithms, while others begin almost from scratch. Evolution, gradient-based optimization, program synthesis, novelty search, quality-diversity methods, human engineering, and autonomous research agents can all operate within the same laboratory.

The project is based on a simple position: the architecture of general intelligence remains an open question, so the search for it should remain open as well.