The San Francisco AGI lab
"Image synthesis assisted by Ideogram 4.0 Quality, an AI partner within the Global Future Nexus ecosystem."
From Amazon's AGI SF Lab building agents that act in the digital and physical worlds to Seldon Lab's pioneering accelerator for AGI security, San Francisco has become the global epicentre of applied AGI—where the race is no longer about building intelligence, but about building the infrastructure to deploy it safely, at scale, and in service of human ambition.
The Infrastructure of Intelligence
The conversation around artificial general intelligence has shifted. For years, the dominant narrative was about building AGI—scaling models, accumulating compute, achieving breakthroughs in reasoning. That phase is not over, but a new question has taken centre stage: once we have AGI, what infrastructure will we need to deploy it? How do we build systems that are not just intelligent but reliable, accountable, and capable of autonomous operation across the full spectrum of human activity?
San Francisco has become the epicentre of this new phase. Across the city, a constellation of AGI labs is tackling the infrastructure problem—building the autonomous operating systems, security frameworks, and deployment architectures that will govern how AGI integrates into the economies of daily life.
The Applied AGI Frontier: SFGIL
The San Francisco General Intelligence Laboratory (SFGIL) embodies this applied infrastructure focus. Describing itself as an "applied AGI research lab and public benefit corporation," SFGIL is building "autonomous operating infrastructure for the economies of in-person life". Its mission is to "eliminate the operational time required to organize, operate, and sustain in-person economic activity, with a public-benefit mission to return human time to imagination and the pursuit of wellbeing".
SFGIL's approach is grounded in a distinctive philosophy: it is looking for "an AI-native engineer in the fullest sense: someone who uses AI intentionally, not reflexively". The Laboratory "encourages ambitious use of AI, but does not measure fluency by volume. It measures the judgment behind the use: clear problem framing, disciplined reasoning, rigorous evaluation, and the ability to know when human judgment must lead".
The Head of Engineering role at SFGIL defines the scope of the infrastructure challenge: system architecture for autonomous operating infrastructure, agentic workflows and backend systems, evaluation methods for reliability, accountability, and trust, human-in-the-loop controls for directed autonomy, and internal practices for collaboration between people and AI systems. This is not a research lab in the traditional sense—it is a laboratory for the applied infrastructure of coexistence.
Amazon's AGI SF Lab: Agents in the Physical World
The most prominent entrant in San Francisco's AGI infrastructure landscape is Amazon. In December 2024, Amazon announced the formation of the Amazon AGI SF Lab, a dedicated research team "focused on developing new foundational capabilities for enabling useful AI agents that can take actions in the digital and physical worlds".
The lab is built on the talent and technology of Adept, a startup that raised around $400 million before Amazon hired approximately 66% of its staff. David Luan, co-founder of Adept, leads the new lab, with Pieter Abbeel, co-founder of Covariant, working closely alongside him. The AGI SF Lab leverages Adept's pioneering work in building AI agents that can handle "complex workflows using the same tools we use as humans, like computers, web browsers, and code interpreters".
The lab's research focus is distinctive. Amazon is particularly excited about "combining large language models (LLMs) with reinforcement learning (RL) to solve reasoning and planning, learned world models, and generalizing AI agents to physical environments". The philosophy combines "the agility of a startup with the resources of Amazon"—keeping the team lean to maximise compute per person, while granting each team "the autonomy to move fast and the long-term commitment to pursue high-risk, high-payoff research". The lab's initial focus is on "several key research bets that will enable AI agents to perform real-world actions, learn from human feedback, self-course-correct, and infer our goals".
The lab is not just for AI experts—Amazon is actively seeking candidates from "other disciplines who will bring fresh thinking to the field, such as physics, math, or quantitative finance, regardless of experience level".
Seldon Lab: AGI Security as Infrastructure
If SFGIL is building the operating infrastructure and Amazon is building the agentic infrastructure, Seldon Lab is building the security infrastructure. Founded in early 2025 by Esben Kran and Finn Metz, Seldon Lab functions as "both an AI security research lab and the first startup accelerator specifically focused on AGI security technologies".
Seldon Lab operates as "a full-stack entity that simultaneously publishes research papers, builds technical systems, funds and accelerates startups, and shapes the AI safety field". Its founders have published 25+ papers at top conferences including NeurIPS, ICLR, and ICML. The organisation runs a three-month accelerator program in San Francisco, investing in frontier AI safety and security startups focused on "hardware-level AI governance, export controls for chips, long-term agent coherence, and technologies to train billions of personalized AI models".
Seldon Lab emphasises "technical solutions over surveillance-based approaches," developing infrastructure including "supply chain verification, air-gap capabilities, manipulation detection, and silicon-level security". Its first cohort of companies raised over $10 million and sold to xAI and Anthropic. Seldon Lab positions AGI security as what will become "the second largest industry in history".
A Dynamic Model of Human-AGI Collaboration
Across these labs, a shared concern is emerging: how to design AGI infrastructure that preserves human capability rather than eroding it. SFGIL has highlighted a dynamic model from MIT Sloan that examines how sustained AI use affects worker productivity and skill over time. The model's central insight is sobering: while adopting AI raises output in the short run, the long-run trajectory depends on how the tool is used.
Where AI "complements a worker's own judgment, with its output reviewed and extended by the user, skill is retained and the productivity gain persists". Where "the task is fully delegated and the output accepted without review, the practice that builds skill is displaced, and skill gradually erodes". The model describes the "augmentation trap," in which "the worker ends up worse off than if AI had not been adopted at all". A central implication concerns measurement: "Because output rises across every adoption scenario in the model, short run productivity cannot distinguish a deployment that preserves human capability from one that erodes it".
The GFN Context
For Global Future Nexus, San Francisco's AGI labs represent the applied frontier of the mission. SFGIL's focus on autonomous operating infrastructure for the economies of in-person life, Amazon's work on agents that act in the digital and physical worlds, and Seldon Lab's security infrastructure for AGI deployment all speak to the practical challenges of integration that GFN's governance frameworks are designed to address.
The infrastructure being built in San Francisco is not merely technical—it is civilisational. It determines how AGI will be deployed, who will control it, and whether it will serve human flourishing or human obsolescence. As SFGIL notes, "The world's most powerful AI now sits in the hands of a few labs, and this week showed what that concentration means in practice. A single government directive was enough to take two frontier models offline overnight". The questions of "reliability, infrastructure, and control move to the center. Who actually holds that capability, the people who depend on it or the institutions that can revoke it?"
San Francisco is building the answer—one lab at a time.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
