Ben Goertzel Outlines 20-Step Roadmap Toward Beneficial AGI

On October 6, SingularityNET founder Ben Goertzel published a thread series laying out a systematic, actionable 20-step roadmap for "how beneficial AGI can actually emerge" (the OpenBGI plan), and disclosed the OmegaHive recursive self-improvement loop his team is actually building. The overall scheme pushes back against the industry's implicit consensus, advocating a path to AGI and even ASI via decentralized agent swarms and a neural-symbolic-evolutionary architecture. It's worth attention because it offers a complete engineering route distinct from the scaling race among frontier LLM labs.

Confirmed

• Goertzel explicitly rejects the near-consensus view that "AGI will inevitably come from frontier labs training ever-larger models with better harnesses," calling it shortsighted and wrong; he argues LLMs lack creativity and true autonomy, making them suitable only as components and building tools for AGI systems, not a complete seed (step 2 of the plan)
• On architecture: step 3 builds neural-symbolic-evolutionary agents with symbolic working memory, medium- and long-term memory, running reasoning and creative evolution (his Hyperon is described as near-complete and already producing real-world impact); step 4 builds a swarm of functionally specialized agents sharing a symbolic knowledge base; step 5 gives the swarm a default motivational system, paired with the GOLEM architecture and Iter Omega runtime to handle goal drift during recursive self-improvement
• Step 8 has the swarm self-upgrade using AGI design corpora and its improvisations as material, with progress measured by the general intelligence metrics established in step 7; step 9 swaps the fixed-weight LLMs in the agent loop for open-weight LLMs wrapped in a neural-symbolic shell with continual learning
• Step 10 provides the swarm with a "seed ontology" to collaboratively improve in practice; an accompanying blog post covers "calculus of distinctions," a new mathematics Goertzel developed with LLM collaboration, aimed at self-improving agent minds that can learn, fork, and merge
• Step 11 addresses "cybersecurity doomsday": Goertzel judges that new-generation LLMs have made vulnerability discovery nearly universal (mentioning related moves by OpenAI and Anthropic), and his roadmap bets on formal verification and "correct by construction"; step 12 further assembles a "purple team swarm" of red-team and blue-team agents that build and share a system world model with neural-symbolic AI, reinforcing security through continuous adversarial dynamics
• The underlying substrate is a decentralized "superswarm" network (arrangements at the hardware and compute level), with the overall goal of recursive self-improvement toward ASI; at the AGI-26 conference Goertzel observed many attendees already using agent swarms to assemble papers (including his Hyperon team's papers) into an AGI codebase, and his team's OmegaHive/OmegaClaw loop is putting LLMs and related technologies together in practice

Why it matters

• This is a first-hand, systematic engineering roadmap from a veteran AGI researcher, in sharp contrast to the mainstream "scale the LLM" route, providing a concrete reference for watching the decentralized, neural-symbolic, and recursive self-improvement directions
• Its safety claims—formal verification, correct-by-construction, red/blue adversarial swarms, and the GOLEM architecture to stabilize goal drift—directly address the two major risks of LLMs democratizing vulnerability discovery and RSI goal changes, offering valuable input for AGI safety discussions

2026-10-06 ~ 2026-10-06 · 12 related posts

• [source] Ben Goertzel lays out a 20-step path to beneficial decentralized AGI — bengoertzel · 2026-10-06
• [source] Ben Goertzel: LLMs are a tool for AGI, not the complete seed — the frontier-lab consensus is short-sighted — bengoertzel · 2026-10-06
• Goertzel's step 3: neural-symbolic-evolutionary agents with symbolic memory beyond plain LLM loops — bengoertzel · 2026-10-06
• Goertzel on safe recursive self-improvement: GOLEM meta-architecture meets Iter Omega runtime — bengoertzel · 2026-10-06
• [source] Ben Goertzel's OmegaHive: seeding recursive self-improvement from agentic coding toward ASI — bengoertzel · 2026-10-06
• Goertzel's step 8: agent hives tasked with upgrading their own intelligence against AGI metrics — bengoertzel · 2026-10-06
• Goertzel's step 9: swap fixed-weight LLMs for open-weight models with continual-learning neural-symbolic caps — bengoertzel · 2026-10-06
• Ben Goertzel unveils 'Distinction-Calculus', new math for forking, merging, self-modifying agent minds — bengoertzel · 2026-10-06
• Ben Goertzel's plan to avert the cybersecurity apocalypse: formally verified software co-authored by agent hives — bengoertzel · 2026-10-06
• Ben Goertzel proposes purple-team hives of cybersecurity agents to complement correct-by-construction — bengoertzel · 2026-10-06
• Ben Goertzel Launches BGI Commons With an AI 'Constitution' and Weekend AGI Hackathons — bengoertzel · 2026-10-06

1 near-duplicate retellings: bengoertzel