H. Sureja | October 5, 2026
A February 2026 whitepaper from Craig A. Kaplan, CEO of iQ Company, argues that artificial general intelligence can be reached safely by connecting millions of people to cloneable AI agents that each carry one person’s expertise and ethics. The 117-pagedocument, “Advanced Autonomous Artificial Intelligence,” describes an architecture called SCAN-II and is hosted on SuperIntelligence.com. It is a design description. Its preface says it is not formatted to journal standards, and across its 96 text pages it reports no experimental results, pilot data or user counts. Its central claim is also its most definitional: the network counts as AGI from the day it launches.
“AAAI” in the title means Advanced Autonomous AI. It has no connection to the AAAI research conference.
What is the SCAN-II architecture?
SCAN-II stands for Safe, Customizable, Architecture and Network, Integrated, and Improving. It maps to five subsystems, and the paper says partial implementations compromise both capability and safety.
Customization turns a base language model into a personal agent, called an AAAI, using a person’s social media, email, documents, purchase history and direct conversation. The paper expects existing LLMs to serve as the base.
Architecture represents all problem solving as search through a problem space, following Allen Newell and Herbert Simon’s Human Problem Solving theory.
Network is a marketplace where agents and humans take on clients’ problems. It covers cloning, three supervision levels, solver matching, reputation and payment.
Integration aggregates training data and ethics from many agents, converts successful solution paths into reusable procedures, and supports one-human-one-vote ballots on ethical questions.
Improvement applies learning at every level, with a stated commitment that humanity’s survival probability must increase monotonically as the system improves.
How does the WorldThink protocol move a problem through the network?
The mechanism is a problem tree. A problem is defined by states, operators that change states, a goal test and evaluation functions that rank promising branches. Every attempt is recorded, including dead ends, with contributor, timestamp and outcome. In the paper’s
example, bringing clean water to an African village splits into sources, infrastructure, community buy-in, labor and testing, which hydrogeology, engineering and anthropology agents pursue in parallel while humans handle local politics.
The WorldThink protocol would run the marketplace: a client posts a problem with success criteria and payment, solvers are matched by reputation, the system validates a solution, and payment is released. The paper’s steps for the serial case run from problem submission to compensation, and the collaborative case carves sub-problem rewards out of the main reward. Solvers earn royalties whenever a later solution reuses their work.
Three implementations are offered. An Ethereum version stores the tree in logs and pays through smart contracts, though the paper acknowledges gas costs, limited throughput and storage cost. A centralized version is faster and cheaper but requires trusting the operator. A hybrid anchors critical events on-chain. In the on-chain version, solvers vote on competing solutions through token curated registries and stake tokens on their choice, forfeiting them if the solution fails the client’s criteria.
What does the paper claim, and what supports it?
Here is the table, ready to copy into your editor:
| Claim | Where | What the paper provides | Claim state | What would verify it |
|---|---|---|---|---|
| AGI-level performance exists at network launch | Secs. 3.3, 5.4 | Argument from definition: humans on the network can always fill gaps | Asserted by definition | A defined task suite with cost, latency and quality against expert baselines |
| Data, not compute, is the binding constraint | Sec. 1 | Assertion; no cited dataset | Asserted | Measured gains from private expert data versus other sources |
| A customized agent carries its owner’s expertise and values | Secs. 4, 10.2 | Five-stage process, illustrated with a fictional Paris cafe expert | Designed | Blind comparison of agent output with its human owner |
| Harmful customization is screened out | Sec. 4.6 | Block customization that “99% of humans would consider harmful” | Designed | Independent red-team results |
| Democratic ethics aggregation resists capture | Secs. 7.5, 7.6 | One-human-one-vote; three aggregation methods | Designed | Independent adversarial test, including fake-identity attacks |
| Safety and speed point the same way | Sec. 11.4 | Argument that human participation provides both | Asserted | Comparative timelines and incident data, which do not exist yet |
| PredictWallStreet powered a top-ten market-neutral fund by 2018 | Sec. 5.2 | Author statement, no source | Author-reported, unverified | Fund or regulatory documentation |
Claim states are TLM’s classification of the paper’s text.
Is the AGI claim anything more than a definition?
The paper says AGI-level performance “is available immediately upon network deployment.” The argument is that if agents contribute little, humans on the network can still address any problem a human can, and humans “by definition, possess human-level general intelligence.” As agents learn, the paper says, they take over more of the work.
That describes a human expert marketplace with AI assistance, and it counts as AGI only under the paper’s own definition. What a reader can test is narrower: how well, how fast and at what cost the network solves problems against expert baselines. The paper reports none of those numbers.
What does Kaplan’s cited track record cover?
The paper cites Kaplan’s “issued U.S. patent” on Online Distributed Problem-Solving as having reduced the underlying theory to practice. Public records show US7155157B2, filed September 20, 2001, granted December 26, 2006 and assigned to IQ Consulting Inc. It expired on October 23, 2021. Its described system coordinates large groups of human experts through threaded discussion and a reward matrix over a tree-shaped problem space, and it does not involve AI agents. A related application, US20070160970A1, is listed as abandoned.
The paper also says Kaplan built PredictWallStreet by 2006, and that by 2018 it powered atop-ten market-neutral hedge fund. iQ Company’s research page lists PredictWallStreet publications from 2001, 2008 and 2013, and SuperIntelligence.com says Kaplan’s designs powered trading systems that executed billions of dollars in trades. TLM could not find independent documentation of the fund ranking. The record supports experience coordinating human contributors. It is not evidence about networks of AI agents.
Where does the safety argument depend on untested assumptions?
The safety case is structural: screening at customization, a Three-Organ Test at each decision, reputation and staking on the network, democratic ethics aggregation, and continuous improvement. Several dependencies deserve testing.
The Heart test checks the owner’s values. The Three-Organ Test (Brain: logic, Heart: ethics, Gut: similarity to past bad outcomes) checks each action against “the ethical parameters established during customization,” meaning the individual owner’s. Collective values enter only later, through aggregation.
One human, one vote needs a way to count humans. The paper says an owner of many clones gets “an economic advantage, but not an ethical advantage.” Its authentication options (passwords, biometrics, multi-factor, hardware tokens)establish who holds an account. The paper does not describe how it would confirm that each account belongs to a distinct person. Its Integration takeaways also describe voting as weighted by “reputation and stake,” while the body text specifies one-human-one-vote. The two statements should be reconciled.
Dilution assumes honest, independent contributions. The paper says a harmful contribution is “diluted by millions of ordinary contributors.” Coordinated manipulation would test that.
Platform concentration sits uneasily with distributed authority. The paper says network effects “suggest that the AAAI network may tend toward a single dominant platform,” while arguing that spreading ethical authority avoids a single point of failure. A dominant platform would also set the platform standards and choose among the aggregation methods the paper describes.
Autonomy and oversight pull apart. The design has agents take on “progressively more” work while safety relies on human involvement. The paper also says problem formulation and decomposition depend on humans, and that the Heart test, screening and voting all draw on human input.
Guarantees. The introduction contrasts scaling approaches that offer “no architectural guarantees” with this design, while the conclusion says the architecture “does not guarantee a positive outcome; no design can.”
The design also has real strengths: an auditable record of every attempt, attributed credit and royalties, three supervision levels, human review of high-stakes decisions, and open acknowledgment of blockchain trade-offs and of democracy’s costs in noise and inconsistency.
What would settle it?
A public pilot with numbers. Evidence that moves this from design to result includes a network of stated size solving a published task set at disclosed cost and latency, a blind comparison of customized agents with their owners, publication of the aggregation and evaluation methods, a stated way to verify one-human-one-vote, and an independent adversarial test of the ethics design. As of October 5, 2026, aaai.com redirects to SuperIntelligence.com.
