Ben Goertzel Outlines 20-Step Roadmap for Building Beneficial AGI
07 Oct 2026
The AI researcher argues that artificial general intelligence is unlikely to emerge from scaling large language models alone, proposing instead a decentralised architecture built around neural-symbolic agents, shared memory, formal verification, governance and collective intelligence.
7 October 2026 — AI researcher Ben Goertzel has published a 20-step roadmap for what he calls a path toward beneficial artificial general intelligence, arguing that the current industry focus on scaling large language models is too narrow to produce the kind of adaptive, self-directed intelligence associated with AGI.
In Beneficial AGI in 20 Steps, Goertzel proposes a system built around neural-symbolic-evolutionary agents, collaborative agent “hives”, shared knowledge stores, decentralised infrastructure and governance mechanisms designed to keep increasingly capable systems aligned with human and broader sentient interests. His central claim is that LLMs can be useful components of AGI systems, but should not be treated as sufficient foundations on their own.
The roadmap is not presented as a prediction of how AGI will inevitably emerge. It is Goertzel’s proposed architecture for reducing what he describes as the risk of destructive instability during the transition toward more capable AI systems.
Beyond the idea that bigger LLMs automatically become AGI
Goertzel begins by challenging what he sees as an increasingly common assumption: that human-level AGI will emerge primarily through frontier laboratories training larger language models and combining them with more sophisticated agent frameworks.
He argues that LLMs are useful tools, but lack several properties he considers essential for general intelligence, including persistent agency, richer self-models, long-term memory and the capacity for autonomous innovation.
That view differs from the dominant commercial trajectory of the AI industry, where much of the investment in frontier systems has focused on scaling transformer-based models and building increasingly capable agentic layers around them.
Goertzel instead proposes a hybrid architecture combining neural models with symbolic reasoning, evolutionary processes and multiple forms of memory.
The OpenCog Hyperon project is one example of this approach. The Artificial Superintelligence Alliance describes Hyperon as a next-generation cognitive architecture combining neural networks, symbolic reasoning and probabilistic logic, designed to support scalable and self-improving intelligence.
Step one: build richer agents, not only larger models
The first technical phase of Goertzel’s roadmap centres on creating neural-symbolic-evolutionary agents.
These agents would combine LLM capabilities with symbolic working memory, medium- and long-term memory, reasoning mechanisms and evolutionary processes that allow new strategies and structures to emerge over time.
The distinction matters because Goertzel’s model of AGI is not a single giant model answering prompts. It is an ongoing cognitive system capable of maintaining goals, accumulating knowledge and modifying its own behaviour.
He then proposes organising these agents into hives, with different agents performing specialised functions while sharing common symbolic knowledge.
Multiple hives would eventually connect into what he calls a super-colony, allowing information and capabilities to circulate between groups while retaining differentiated roles.
Self-improvement becomes part of the architecture
Several of the steps focus on recursive improvement.
Goertzel proposes developing metrics that allow agents and hives to evaluate their own general intelligence across multiple domains. Some hives could then be assigned specifically to improve their architectures, using existing AGI designs alongside new variations generated by the systems themselves.
That introduces one of the most consequential ideas in the roadmap: AI systems becoming increasingly involved in designing their successors.
Goertzel has explored this concept elsewhere under the idea of recursive self-improvement, where current “proto-AGI” systems could potentially accelerate progress toward much more capable systems.
The roadmap also proposes replacing fixed-weight language models over time with open models capable of continual learning, supplemented by neural-symbolic layers.
The goal is to move away from systems whose knowledge is largely frozen after training toward ones capable of learning continuously from new experience.
Shared ontologies would give agents a common conceptual language
Another proposed component is a seed ontology: a structured set of concepts and relationships that gives agents a common starting framework for representing knowledge.
Rather than leaving each agent to develop its own conceptual system independently, the super-colony would begin with a shared representation and collectively refine it.
That shared structure would support communication between agents, reasoning across domains and the eventual development of common values and governance rules.
Goertzel also links this to mathematical work around distinction calculus and related knowledge-representation systems.
The idea reflects a broader principle running through the roadmap: intelligence should be distributed, but not fragmented.
Security is built into the system rather than added later
Goertzel devotes several steps to cybersecurity and formal verification.
He proposes running agent hives inside containers built on provably secure microkernels, while specialised mathematical agents work to verify that evolving systems continue to satisfy their specifications.
Other agents would refine those specifications and periodically involve expert humans in review.
The roadmap also proposes dedicated cybersecurity “purple teams” combining offensive and defensive agents. These systems would build models of networks and computer systems while continuously testing for vulnerabilities.
The intention is to make security part of the underlying AGI architecture rather than something added only after a powerful system has already been built.
This is particularly important in Goertzel’s framework because the agents are expected to become increasingly capable of modifying themselves and interacting with external systems.
Decentralisation is central to Goertzel’s vision
Perhaps the most distinctive part of the roadmap is its emphasis on decentralised infrastructure.
Goertzel proposes deploying a substantial part of the super-colony across a network with no single owner or central controller.
The Artificial Superintelligence Alliance, which Goertzel leads as CEO, is already developing infrastructure aligned with this direction. Its ASI:Chain and Hyperon projects are intended to provide open, decentralised foundations for advanced AI systems.
Goertzel argues that decentralisation could reduce the risks associated with a small number of companies or governments controlling increasingly capable intelligence.
However, decentralisation also introduces new problems: how to prevent malicious forks, coordinate upgrades, enforce rules and protect sensitive knowledge.
His roadmap proposes cryptographic mechanisms to make it harder for attackers to copy the knowledge of the entire system by compromising only a small number of nodes.
Prediction markets and agent economies enter the design
The roadmap also incorporates economic and collective decision systems.
Goertzel proposes decentralised prediction markets that allow agents to estimate future events and evaluate possible directions for the super-colony.
He also suggests creating an internal agent economy, where subnetworks could organise around specialised activities and use their own incentive structures.
These ideas treat AGI not simply as a software architecture but as a type of digital society.
Agents would communicate, trade resources, form specialised groups and participate in governance rather than operating as isolated applications.
That raises questions familiar from political economy as much as computer science: how should power be distributed, how should collective decisions be made and how can incentives be aligned with long-term objectives?
Humans and AGIs would share governance
One of Goertzel’s later steps proposes decentralised collective governance involving both humans and artificial agents.
The system would use AI-assisted governance and reputation mechanisms to synthesise proposals and determine how the network evolves.
Goertzel also proposes an initial ethical “constitution” expressed using the shared ontology, with the constitution itself allowed to evolve through collective processes.
This is one of the most ambitious elements of the roadmap because it treats alignment not as a fixed set of rules imposed permanently at launch, but as a governance problem that continues as the system changes.
The model therefore differs from approaches that focus primarily on controlling AI through static constraints.
Goertzel’s proposal is closer to building institutions for AI systems.
Human values remain part of the process
The nineteenth step is deliberately less technical.
Goertzel argues that people should continue to provide values, guidance and what he describes as “friendship” to increasingly capable artificial systems.
He points to BGI Commons, a community operated within the SingularityNET ecosystem that is designed around collaborative development of beneficial AI and AGI. The platform brings together developers, researchers and participants through projects, learning resources and collaborative development programmes.
The underlying idea is that alignment may depend not only on mathematical constraints but also on the relationships and culture surrounding the systems as they develop.
The final step: a “Beneficial Singularity”
Goertzel’s twentieth step is the emergence of what he calls the Beneficial Singularity.
In his vision, this would follow extensive technical development, decentralised governance, recursive improvement and collaboration between human and artificial intelligence.
The term “Singularity” refers to a hypothetical period in which technological progress accelerates dramatically, potentially as advanced AI begins improving itself faster than human institutions can independently keep pace.
Goertzel’s roadmap is specifically designed around trying to shape that transition rather than simply waiting for it to happen.
His central argument is that AGI architecture, governance and ownership cannot be separated.
If increasingly capable intelligence is controlled by a small number of actors, the risks are fundamentally different from a system whose intelligence, decision-making and infrastructure are distributed.
A roadmap, not a settled scientific consensus
Many elements of Goertzel’s proposal remain theoretical or experimental.
There is no scientific consensus that neural-symbolic architectures will outperform scaled neural models in producing AGI, nor that decentralised governance would necessarily make advanced AI safer.
Likewise, recursive self-improvement, large-scale autonomous agent societies and mixed human-AI governance remain active areas of research rather than established engineering solutions.
What makes the roadmap notable is that it brings those ideas together into a single proposed pathway.
Rather than asking only how to make AI models more powerful, it asks what kind of technical, social and economic system should surround them if AGI does emerge.
That may become one of the most important distinctions in the next phase of AI development.
The question may not simply be whether AGI is built, but what architecture it uses, who controls it and which values shape its evolution.
About Ben Goertzel
Ben Goertzel is an AI researcher, computer scientist, author and entrepreneur focused on artificial general intelligence. He is CEO of the Artificial Superintelligence Alliance and CEO and Chief Scientist of SingularityNET, and has led the development of cognitive architectures including OpenCog and OpenCog Hyperon.
His work explores AGI, decentralised AI infrastructure, neural-symbolic systems, artificial superintelligence and the governance of increasingly autonomous intelligent systems.
In a 2025 conversation with Dinis Guarda, Goertzel discussed many of the ideas that underpin his latest roadmap, including the distinction between generative AI and AGI, the architecture of superintelligence, cybersecurity, OpenCog Hyperon, decentralised superintelligence, the Artificial Superintelligence Alliance and humanoid robotics.
The Artificial Superintelligence Alliance describes Goertzel as the author of more than 25 books and 150 research papers and credits him with significant contributions to the development of AGI theory and software.
Sources
- Ben Goertzel — Beneficial AGI in 20 Steps, 6–7 October 2026
https://magazine.mindplex.ai/post/beneficial-agi-in-20-steps Mindplex - Dinis Guarda Podcast — Ben Goertzel: SingularityNET, AI, AGI, ASI Alliance, Humanoid Robots + Desdemona
- Artificial Superintelligence Alliance — OpenCog Hyperon and ASI:Chain
https://superintelligence.io/singularitynet-launches-asichain-devnet-and-hyperon-agi-framework/ Superintelligence - Artificial Superintelligence Alliance — SingularityNET and OpenCog Hyperon
https://superintelligence.io/portfolio/singularitynet/ Superintelligence - Artificial Superintelligence Alliance — Ben Goertzel profile and leadership
https://superintelligence.io/about/team/ Superintelligence - BGI Commons
https://bgicommons.org/ BGI Commons
SEO title: Ben Goertzel’s 20-Step Roadmap for Beneficial AGI
Slug: ben-goertzel-beneficial-agi-20-step-roadmap
Meta description: Ben Goertzel outlines a 20-step roadmap for beneficial AGI built around neural-symbolic agents, decentralised networks, recursive self-improvement, security and shared governance.
