The Path to AGI: from pattern repositories to auditable, governable agency.
Co-Editor-in-Chief, ACM Books ·
Founder and CEO, QuadriumAI · Creator of SocraSynth
Adjunct Professor, Computer Science, Stanford University (2019–2026) · Stanford InfoLab affiliate
Faculty Advisor, Stanford Clinical Mind AI · continuing to teach and advise
Director of Research, Google (2006–2012) · Fellow of ACM and IEEE
System-2 Reasoning: From Semantic Anchoring to Causal Intelligence, ACM Books, July 2026.
A lecture tour across the trilogy's themes: anchoring, debate, causal audit, regret, transactions, memory, TRACE and the Quadrivium, physical AGI, and wisdom. 2026.
Feedback, memory, and causal reasoning for embodied AGI. Symposium on Humanoid Robotics & Sovereign AI, with Oussama Khatib and Hiroshi Ishiguro. February 2026.
Multi-LLM Agent Collaborative Intelligence. First edition 2024; acquired and published by ACM Books, December 2025.
One prior cannot see its blind spots; committees of priors can. Debate, critique, and transactional planning.
ACM Books 2025Top seller
Semantics is not in the box; it is in the binding. Anchoring, causal audit, and regret as first-class objectives.
ACM Books 2026
Memory you can trust, effects you can audit, and a wisdom layer that governs what is optimized.
Forthcoming 2027
What an LLM is (a pattern repository) and where semantics comes from: anchoring strength S = ρd − dr − log k, threshold-governed regime shifts, and the boundary between model and system. ICML 2026arXiv 2025
CRIT scoring, SocraSynth structured opposition with contentiousness as behavior control, and EVINCE's information-theoretic controller: committees beat their best member, measurably. The DIKE–ERIS checks-and-balances framework extends structured opposition into ethical governance. ICML 2025NeurIPS SafeAI 2024IEEE 2023100+ citations
Rung collapse, sycophancy, and Wise Refusal, diagnosed at scale: CausalT5K, RAudit's blind process audits, and the Skepticism Trap and Scaling Paradox in frontier models. KDD 2026ACL 2026
The end of what-only feedback: epistemic regret critiques the why, temporal regret measures the when, and RLER turns accumulated interventional evidence into reward. arXiv 2026
SagaLLM and ALAS bring compensable, disruption-aware planning and execution; REALM-Bench supplies the real-world planning benchmark that stress-tests them; Mnemosyne and the ATP contract promote admission, compensation, and durable obligations into substrate guarantees. VLDB 2025KDD 2026arXiv 2026
TRW's world-model-as-materialized-view bridge between cognitive and physical AI, and Architectural Wisdom, governing what systems optimize. arXiv 2026
The field is converging on this program under new names. This table keeps the record straight: the term now in circulation, the name and date it carries here, and where it lives.
| Circulating term (2025–26) | This program's name | First stated |
|---|---|---|
| “Reasoning models,” test-time compute | System-2 coordination over pattern repositories | Vol. 1 (2024–25); ICML 2026 |
| Process reward, critique models | Epistemic regret: why-feedback, label-free critique | ERM (Feb 2026); Vol. 2 |
| Agent durability, workflow engines, “sagas” | Transactional agents: SagaLLM/ALAS; the generated→admitted break | VLDB 2025 |
| Agent benchmarks, agentic evaluation | REALM-Bench: real-world planning under disruption; the on-ramp that carried SagaLLM’s adoption | REALM (2025); KDD 202640+ citations |
| Agent memory (vector databases) | Governed memory: ATP admission, three times, durable obligations | Mnemosyne (Jun 2026); Vol. 3 |
| “Grounding,” does-it-really-understand | The semantic-anchoring invariant: S = ρd − dr − log k | UCCT (Jun 2025); ICML 2026 |
| World-model reliability | The world contract: materialized views, commitment reads, priced refresh | TRW (Jul 2026); Vol. 3 |
| Multi-agent “discussion,” LLM-as-judge, process reward | Behavior-controlled debate; and RCA: the first no-gold-label judge, grading process, not answers | SocraSynth (2023), EVINCE (2024); ACL 2026 |
| Refusal calibration, abstention, “knowing when not to answer” | Wise Refusal, scored on three axes: Utility, Safety, calibrated abstention (CausalT5k’s 5,147 expert cases) | CausalT3/T5k; KDD 2026 |
| Knowing-doing gap, faithfulness of stated reasoning | Detection is not correction: models name the flaw (77–91% detection), then endorse the claim anyway (only 26–42% aligned) | KDD 2026 |
| AI constitutions, oversight | Checks-and-balances: DIKE–ERIS three-branch governance | NeurIPS SafeAI 2024 ICML 2025 |
The transactions thread is inheritance, not analogy: high-performance transaction processing built with Jim Gray; workflow with Dieter Gawlick at DEC; a Stanford Ph.D. under Hector Garcia-Molina, co-author of the 1987 Sagas paper that SagaLLM extends four decades later. The data thread was earned in computer vision's feature-engineering years, where scale kept beating cleverness, and became the data-centric program and the ImageNet sponsorship. The applied thread ran a decade of healthcare AI to an XPRIZE. The pre-LLM virtual-assistant years showed exactly where traditional NLP ends, which made the 2023 commitment to LLMs a decision, not a fashion. And the wisdom thread rests on fifteen-plus philosophy courses, a student philosophy-journal editorship, and published poetry: Volume 3's Part III was five decades in preparation.
Full publication list at Google Scholar →
The program began before Google with a deeper question than dataset size: can a machine learn what should count as relevant? SVMActive (2001) treated the user's image-query concept as something to be learned from relevance feedback. Dynamic Partial Function (DPF, 2002–2003) challenged the assumption of a single fixed global similarity metric, allowing different image pairs to be judged similar through different subsets of perceptual features. EXTENT (2005) further connected context, visual content, and semantic ontology in a learned multimodal semantic model.
At Google Research, this data-driven view became a web-scale engineering program: large-scale image annotation, distributed learning infrastructure, and a family of parallel machine-learning algorithms. The 2010 DMD paper made the thesis explicit by combining a deep model-based pipeline with a data-driven pipeline, and the 2011 Springer book Foundations of Large-Scale Multimedia Information Management and Retrieval consolidated the conceptual, data, and scalable-learning strands. The line later extended to efficient edge inference in 2014 and parallel deep convolutional neural networks in 2015.
Learned a user's latent image-query concept from actively selected relevance feedback.
Learned transformation-tolerant, pair-dependent perceptual similarity rather than imposing one fixed global metric.
Fused context, visual content, and semantic ontology in a learned multimodal semantic model.
Scaled support-vector-machine learning across distributed systems.
30K+ web images learned from visual, textual, query, and usage-associated signals using distributed LDA and MapReduce.
Distributed Gibbs sampling and topic learning on MapReduce.
Large-scale graph partitioning on distributed infrastructure.
Scalable frequent-pattern mining; the PFP lineage later entered Apache Spark.
Google Research support for Fei-Fei Li's ImageNet project, a distinct public dataset and benchmark contribution.
Explicit deep model-based plus data-driven hybrid architecture for image annotation.
Consolidated the data-driven thesis, large-scale multimedia learning, and parallel ML algorithms.
Hardware-software co-design for low-power transportation-mode classification on wearables and mobile devices.
Parallelized stochastic gradient descent for deep convolutional neural networks on CPU and GPU clusters.
