The Path to Artificial General Intelligence

The central proposition

Smart models are not yet dependable intelligent systems.

Large language models can produce extraordinary answers and plans. But the capacity to propose is different from the authority to act. A fluent rationale is not independently audited reasoning. A remembered instruction is not a durable commitment. A sensor reading is not evidence of ownership or permission. And successful optimization is not, by itself, wise choice.

The trilogy develops a layered response. Volume I asks how intelligent agents can challenge and coordinate with one another. Volume II asks how they can reason, test causes, and learn from mistakes. Volume III asks what it takes to operate in a changing world under evidence, authority, recovery, and human responsibility.

The goal is not simply to make a model more capable. It is to make a system's claims, commitments, actions, and purposes accountable.

Volume I · CollaborationWho should challenge whom?

Controlled debate, independent criticism, and durable multi-agent workflows.

Volume II · ReasoningWhat warrants our conclusions?

Semantic anchoring, causal tests, reasoning-quality feedback, and prevention.

Volume III · WisdomWhat may we do—and should we?

Operational commitments, grounding, authorization, verified effects, and restraint.

The books

Three volumes, one cumulative argument

Each volume addresses a failure left open by the previous one. They are designed to be read in sequence, but each also supplies methods and examples for researchers working on a particular layer of intelligent systems.

Book one / Published 2025

Multi-LLM Agent Collaborative Intelligence

The Path to Artificial General Intelligence

From one persuasive model to a disciplined collaboration of competing perspectives.

The question

How can multiple LLMs disagree productively, recognize weak arguments, exchange useful information, and reach an accountable decision without merely voting or amplifying shared errors?

The contribution

MACI treats collaboration as a governed process: SocraSynth structures debate, CRIT and related evaluators scrutinize arguments, EVINCE measures informational change, while transactional planning, checks and balances, and retrieval organize the work beyond one conversation.

SocraSynthCRITUCCT / EVINCESagaLLMDIKE–ERISBEAM

What changes

Debate becomes a mechanism for challenge, evidence gathering, and refinement rather than an assumption that more agents automatically imply better answers.

Its boundary

Consensus can still be wrong; a reasoning judge can still fail. Collaboration is a research and engineering scaffold, not proof of AGI or a guarantee of correctness.

Read at ACM Digital Library ↗DOI 10.1145/3749421

Book two / Published July 2026

System-2 Reasoning

From Semantic Anchoring to Causal Intelligence

From plausible explanations to explicit warrants, causal judgment, and error-aware learning.

The question

When a model produces a sophisticated explanation, how do we know that its premises, causal inferences, and degree of certainty are justified? Can reasoning improve without a perfect gold answer for every step?

The contribution

The book develops semantic anchoring and task-conditioned inference, examines causal errors such as confusing association with intervention, and organizes feedback on reasoning quality. It connects independent critique, correction, meta-cognition, and learning before a repeated failure.

Semantic anchoringCausal reasoningERM / RLERCausalT5KLearning Before Failure

What changes

The unit of evaluation becomes the reasoning process and its evidence, not merely whether a final sentence is convincing. Missing evidence and appropriate uncertainty are first-class outputs.

Its boundary

Better reasoning does not establish whether a bridge is open now, whether a tool call is permitted, or whether the observed outcome matches the intended effect. Those require Volume III.

Read at ACM Digital Library ↗DOI 10.1145/3822380

Book three / Author edition 3 November 2026 / ACM Books scheduled February 2027

Beyond Intelligence

From Operational AGI through Grounded Intelligence to Wisdom

From reasoning on a screen to governed action in a real, changing, shared world.

The question

Even a capable reasoner can act on stale facts, the wrong person's data, or an unauthorized reinterpretation. How can proposals become safe commitments, how are effects verified, and who governs irreversible choices?

The contribution

The book joins operational memory and transactional safeguards (Mnemosyne, CTL, StateView, and active contracts) with six grounding relation families, evidence and authority gates, and the Quadrivium disciplines of Reason, Elicit, Contract, and Bridge. Wisdom asks what ends the system should serve.

MnemosyneGrounding contractsSemantic integrityQuadriviumOperational AGIWisdom

What changes

Proposals, durable commitments, dispatched actions, and observed outcomes become different objects with different checks. Human authority and the limits of evidence are part of the system, not footnotes to it.

Its boundary

Some components have experimental or conformance evidence; the entire architecture is not claimed as a fully deployed or verified AGI system. Autonomous strategic reframing and machine wisdom remain open questions.

Read Volume III · Second Edition ↗ACM Books edition forthcoming, February 2027

How to cite this edition Chang, E. Y. (2026). Beyond Intelligence: From Operational AGI through Grounded Intelligence to Wisdom: The Path to Artificial General Intelligence, Volume III, Edition 2. Zenodo. doi:10.5281/zenodo.23898640

Cover artwork supplied by the author. Volume III links to its November 2026 Second Edition on Zenodo; the separate ACM Books publication is scheduled for February 2027.

How to read the series. Volume I broadens and disciplines deliberation; Volume II interrogates the quality and causal basis of its conclusions; Volume III binds decisions to time, people, authority, and verifiable consequences. The trajectory is cumulative, not a claim that the final volume proves AGI has been achieved.

The conceptual arc

From intelligence to operation, grounding, and wisdom

Volume III makes the trilogy's central separation especially concrete. Powerful reasoning is useful, but it must be supported by durable state, sufficient knowledge of the world, authority to act, and governance of the ends pursued.

Operational AGI

Make proposed transitions durable only through validated admission. Keep history, current state, dependent promises, compensation, and observed effects distinguishable.

Volume III · Part I · Chapters 2–8

Grounded intelligence

Check identities, physical and social context, current evidence, provenance, interpretation authority, and whether a contemplated action is sufficiently supported.

Volume III · Part II · Chapters 9–19

III

Wisdom

Use Reason, Elicit, Contract, and Bridge to govern collaboration, seek what is missing, check permissions, and question the research frame itself. Ask what ought to be done.

Volume III · Part III · Chapters 20–24

These are conceptual dependencies and proposed system responsibilities, not a claim that a fully integrated stack has passed end-to-end deployment testing.

A flood response that looks perfect…

An AI assigns rescue boats, sends an ambulance across a bridge, and dispatches drones to find a school bus. Its plan sounds plausible. But the bridge report may be hours old; dispatch rights may belong to another agency; a sent instruction is not a completed rescue.

The trilogy's destination is a system that knows what it must verify, what it may authorize, what actually happened, and where human judgment remains indispensable.

Reasoning: Are the tradeoffs and evidence for this plan justified?

Memory: Which obligations and previous actions remain in force?

Grounding: Is this the right bridge, the latest status, and the correct rescue unit?

Authority: Who may instruct which responder, and who may reinterpret a command?

Verification: Did the dispatch succeed, or did rescue actually occur?

Wisdom: Whose lives, rights, and future costs should shape the mission's priorities?

Research and impact

What the trilogy contributes to the field

Its impact is best understood as an integrated research agenda: methods published and tested in their own settings, engineering patterns with documented limitations, and an explicit account of the unsolved questions standing between today's agents and trustworthy general-purpose action.

Collaboration beyond a vote

SocraSynth, MACI, CRIT, and information-guided debate put the quality of arguments, the behavior of participants, and the need for independent checks above simple multi-model agreement. They establish research questions about when collaboration outperforms the best individual model.

II.

Reasoning as an auditable process

Semantic anchoring, causal diagnosis, reasoning-quality feedback, and Learning Before Failure shift the focus from fluent outcomes to causal warrants, uncertainty, and prevention of repeated errors. Benchmarks such as CausalT5K make specific failure modes testable.

III.

Database ideas for agentic systems

SagaLLM and Mnemosyne bring histories, contracts, compensable workflows, and disciplined state transitions into LLM-centered systems. They also show why transaction correctness alone cannot establish semantic correctness, authorization, or success in the world.

IV.

Safety beyond literal prompt injection

Grounding and semantic-integrity studies distinguish an authentic command from the meaning an agent assigns to it. Provenance, referents, identity, permission, and executed effects must be checked at the point of action. This gives agent safety a systems-and-authority framing.

A bridge across research communities

The series connects language modeling with databases, distributed systems, causal inference, human-AI collaboration, knowledge representation, and responsible autonomous action. This is a proposed architecture and a body of component-level evidence, not a new benchmark score labeled "AGI."

VI.

Education and open research questions

The work informs Stanford teaching and research projects, providing concrete settings for evaluating reasoning, planning, and agent collaboration. It also identifies unsolved problems: robust grounding coverage, independent strategic reframing, and accountable choices when harms are irreversible.

Evidence, not slogans

The website distinguishes three kinds of claims: published methods (with source links), reported experimental results (limited to the tasks and models tested), and architectural or philosophical proposals (which remain to be validated). The trilogy argues for a path toward AGI; it does not certify that AGI, autonomous scientific discovery, or machine wisdom has been achieved.

Selected peer-reviewed results from the research lineage

ICML 2025A Checks-and-Balances Framework for Context-Aware Ethical AI AlignmentPaper & proceedings ↗

PVLDB 2025SagaLLM: Context Management, Validation, and Transaction Guarantees for Multi-Agent LLM PlanningPaper & DOI ↗

KDD 2026CausalT5K: A diagnostic benchmark for causal reasoning failures in LLMsPaper & DOI ↗

These examples illustrate the research lineage; they are not a claim that each paper independently validates the full trilogy architecture. For a longer list, see research publications.

Learn and explore

Stanford talk and courses

Two resources from the research program remain available here. The earlier talk is retained as a historical entry point; the trilogy extends its two-bridge argument into operational grounding and the governance of action.

Video · Talk to incoming Stanford PhD students

Two Paradigm Bridges: From LLMs to AGI

The original talk introduces the limitations of System-1-style language modeling and two approaches: semantic anchoring through UCCT, and behavior-aware multi-agent debate through SocraSynth. It predates the full three-volume operational and grounding architecture.

Watch the full talk ↗

02:11 — System 1 and the problem of grounding12:40 — Semantic anchoring and UCCT31:42 — Multi-agent debate and contentiousness

What the historical talk argues

Semantic anchoring uses examples and task intent to condition the model's response; SocraSynth controls productive disagreement, beginning with exploration and moving toward integration. Those two bridges motivate the later work, but the talk should not be read as proof that anchoring alone is grounded understanding or that debate alone guarantees truth.

Stanford University · Teaching since 2019

CS372: Artificial General Intelligence for Reasoning, Planning, and Decision Making

Stanford teaching and advising have provided a setting for these ideas, with CS372's subject evolving over the years from precision medicine to reasoning, planning, and AGI research. The current course explores multi-agent intelligence, evidence, causal reasoning, memory, planning, and decision-making.

Stanford CS372. · Online course page ↗

Course titles, content, and terms vary by year. Stanford's course catalog is the current reference.