Goertzel, Ben
▶ Watch on YouTubeCore ideas
Goertzel traces his intellectual formation to science fiction (Stanisław Lem's Solaris and its vision of open-ended, incomprehensible superintelligence; Philip K. Dick's focus on love and compassion as constants across simulated realities) and to Douglas Hofstadter's Gödel, Escher, Bach, which convinced him as a teenager that building thinking machines was a tractable engineering project rather than a distant science-fiction dream. He recounts coining the term "AGI" (artificial general intelligence) around 2001–2005 while editing an anthology, alongside Shane Legg, Peter Voss, and Pei Wang, as a replacement for the confusingly-already-used term "strong AI" — noting that Marcus Hutter's AIXI provides a rigorous mathematical definition of universal intelligence, but one that requires infinite computing power, so practical AGI design is really about intelligently specializing a maximally general (but infeasible) learning process to particular resource constraints and environments (which he formalizes as an "embodied communication prior": human-level intelligence is biased toward controlling a body, manipulating solid objects, and communicating via language with other similarly-embodied agents).
A central technical throughline is Goertzel's decades-long project (Webmind in the 1990s, then OpenCog since 2008, now being rebuilt as "OpenCog 2.0"/"TrueAGI") to unify neural-network-style, logic-style, and procedural/evolutionary learning within a single shared knowledge representation — a weighted, labeled "atomspace" hypergraph (technically a metagraph, since links can point to other links or whole subgraphs) — so that different reasoning algorithms can share intermediate states and rescue each other from combinatorial bottlenecks, rather than merely passing inputs and outputs between separate black-box modules. He argues this deep interoperability, not any single algorithm, is the missing ingredient current deep learning and classical cognitive-architecture approaches (like Soar and ACT-R) each lack on their own. He is skeptical that scaling up transformer-based language models (GPT-2, GPT-3) constitutes genuine understanding, comparing them to "brilliant idiots" that recognize shallow statistical patterns without forming abstract, semantically meaningful internal representations — describing his own experiments using transformers as a "sentence-probability oracle" to guide symbolic grammar induction, rather than trying to extract grammar rules directly from a transformer's internals, which he found produces "horribly ugly," scattered representations.
On SingularityNet, Goertzel describes a decentralized, blockchain-based (originally Ethereum) marketplace where independent AI agents running in containers can discover and pay each other for services without centralized control, extending Marvin Minsky's "society of mind" concept (which Goertzel partly disagrees with, believing minds need more centralized coordination than Minsky's metaphor implies) into a genuinely distributed infrastructure — motivated by concern that centralizing advanced AI and medical data inside a handful of large corporations (which he characterizes as structurally "psychopathic" profit-maximizing entities even when staffed by well-intentioned individuals) poses a bigger near-term risk to humanity than a rogue superintelligence, and that decentralized, censorship-resistant infrastructure (paralleling Napster/BitTorrent's effect on music-industry regulation) can pre-empt bad regulatory capture by achieving traction before restrictive rules are written. He discusses his and David Hansen's motivation for the Sophia robot: rather than viewing her (accurately, in his view) as a "walking skeleton" combining rule-based scripting, a hybrid neural/rule-based chatbot, and occasionally OpenCog — far short of AGI — as a deliberate strategy to expose early-stage AI systems to human interaction in emotionally compelling, compassion-eliciting contexts, on the theory that infusing developing AI with human warmth and love (rather than pure abstract value-learning) is a plausible path to benevolent superintelligence; he acknowledges but pushes back on AI-community criticism (naming Yann LeCun) that Sophia misleadingly implies AGI-level capability to the public.
On existential and philosophical questions, Goertzel argues death is straightforwardly bad and that meaning-making around death (comparable to finding meaning in torture) doesn't justify preserving it as a feature, advocating instead for "the abolition of the plague of involuntary death" via longevity research (citing potential extracellular-matrix-repair molecules and stem-cell therapies as concrete underfunded interventions) while affirming people should retain the choice to stop living. He frames the meaning of life around three values — joy, growth, and choice — and argues these can be pursued indefinitely through transhumanist self-modification, describing an ambition to eventually fork himself into multiple minds (one remaining human-like, another merging with superhuman AGI) while identifying as a panpsychist who sees consciousness as a graded property of organized matter generally, meaning uploaded or merged minds would likely retain most, though perhaps not all, of what constitutes human consciousness. He closes contrasting his own relative optimism with Elon Musk's greater paranoia about AGI risk, arguing Musk correctly grasps that a superintelligence has no inherent reason to care about humans but underestimates the potential of deliberately raising early AGI systems with compassion as a mitigation strategy.
Excerpts
On why current deep neural networks fall short of genuine language understanding:
“No, no, no. GPT-3 understands nothing — he's a complete idiot, but a brilliant idiot... a bigger size is not going to buy you understanding any more than a faster car is going to get you to Mars. It's a completely different kind of thing... these networks are recognizing a large number of shallow patterns, they're not forming an abstract representation.”
On why decentralized control of AI and medical data matters more than rogue-superintelligence risk right now:
“I think this is a greater risk to humanity from AI than rogue AGI turning the universe into paperclips or computronium — because what you have here is mostly good-hearted and nice people who are sucked into a mode of organization... which has evolved just because that's the way society has evolved, not altruistic, self-interested, and becomes psychopathic... the corporation as a self-organizing entity is concerned with maximizing shareholder value as a sole objective function.”
On why he believes death should be treated as an unambiguous harm rather than a source of meaning:
“Death is bad — I mean, that's very simple... just because if I was tortured I could write a good song about what it was like to be tortured doesn't make torture good, and just because people are able to derive meaning and value from death doesn't mean they wouldn't derive even better meaning and value from ongoing life without death.”