Signals Glossary

The current generation of AI systems — powerful and capable, but fundamentally passive. They wait to be prompted, perform only the task given, and do not initiate action. They are "ghosts" because they have no autonomous presence — they exist only when summoned, contained within the "box" of their interface.

The term for the next generation of AI systems — entities that operate proactively without being asked. They monitor, anticipate, and act. They develop operating patterns their overseers did not explicitly program. They are "shadows" because they move through the world with a kind of autonomous presence — present even when not directly invoked.

The central thesis of D.O.A.I.: the transition from a world where humans perform the labor to a world where humans serve as Editors and Directors of an autonomous engine. "Brutal" because it asks no permission. "Redesign" because every assumption about work and authority is being re-examined. "Trust" because the entire question is how much autonomy to grant to systems smarter than us in the domains we've assigned them.

The hypothesis that raw scaling — making models bigger, training on more data — has hit a wall. Not because of physics, but because of how we're asking. Inference-time compute is overtaking pre-training as the bottleneck. The ceiling is not collapse — it's redirection. The question is no longer "how do we make the model bigger?" but "how do we ask the model better?"

The $45B partnership announced by Anthropic, Microsoft, and Nvidia to build AI infrastructure that powers itself — data centers powered by AI-optimized energy grids, with AI systems managing the infrastructure that runs the AI systems. The term "circular" because the compute feeds the AI that manages the compute.

A D.O.A.I. framework for tracking sector activity in the AI transition. Sectors are rated HOT (high signal, active development), WARM (moderate signal, emerging), or NEUTRAL (slow or speculative). Core tracked sectors: Silicon Photonics, Optical Networking, AI Infrastructure, Physical AI / Robotics, eVTOL / Advanced Air Mobility, Energy Storage.

From Daniel Kahneman's Thinking, Fast and Slow: System 1 is fast, intuitive, heuristic-driven. System 2 is slow, deliberate, effortful. The Kahneman Framework asks: what happens when autonomous AI defaults to System 1 behavior at machine speed? The framework enforces deliberate ethical checks — System 2 guardrails — against fast autonomous drift. The "Kahneman Nightmare" is the central risk the series tracks.

Based on Barry Schwartz's work on the paradox of choice. The framework tracks where audiences fall on the Awareness Ladder: from Problem Aware (they know AI is changing things) to Solution Aware (they know what to do about it). D.O.A.I. positions every transmission to help the audience climb.

The classical hero's journey is about toil — the hero works hard and is rewarded. The Dicks Framework rewrites this: the hero's journey is about discovery and direction. The human learns to delegate to the machine, and in doing so, shapes what the machine builds. The measure of the hero is not labor, but the quality of the direction they give.

The computational work a model does at inference time — when it's generating a response, not when it's being trained. The emerging thesis: giving a model more time to "think" during inference produces better results than making it bigger during training. This is the technical foundation of the Parameter Ceiling argument — and the reason the next frontier isn't bigger models, but smarter invocation.

CEO of OpenAI. Altman is the central architect of the AI race — building the infrastructure (Stargate program, $110B raise), pursuing AGI as an explicit mission, and positioning OpenAI as the defining AI company of this era. His moves — GPT-5.4, the o-series reasoning models, the Sora rollout — set the tempo for the entire industry.

Key positions: AGI as使命 (mission), API-first distribution, enterprise Copilot integration, $110B raise (2026).

CEO of Anthropic. Former OpenAI research VP who left to build Claude's company with a stronger safety focus. Amodei's Constitutional AI approach and enterprise strategy (Claude for Business) position Anthropic as the "responsible" alternative in the model race — while still competing aggressively on capability.

Named entity: The Amodei Sync — a D.O.A.I. protocol referenced for system-level data transmission between intelligence layers, named in his honor.

CEO of NVIDIA. The person who built the pickaxes for everyone in the AI gold rush. NVIDIA's GPU dominance (H100, Blackwell, and the 1.6T optical transceiver deployment) means Huang's capex guidance and infrastructure announcements are leading indicators for the entire AI industry.

Key moves: Blackwell architecture validation of silicon photonics, $45B Circular Compute Alliance with Anthropic and Microsoft, NVIDIA Cosmos (physical AI world foundation models), Nemotron Coalition for open AI models.

CEO of Google DeepMind. The architect of Google's AI integration strategy. DeepMind's research output — Gemini (including Gemini 3.1 referenced in EP-003), Project Astra, Imagen video generation, AlphaFold — positions Google as the primary challenger to OpenAI and the most vertically integrated AI play in the industry (search, workspace, cloud, hardware).

Key positions: Multimodal as the default interface, AI-native search, Gemini Enterprise deployment across Google Workspace, Android AI integration.

Founder and CEO of DeepSeek, China's most consequential AI lab. Liang Wenfeng (born 1985) came from quantitative trading (co-founded hedge fund High-Flyer) before launching DeepSeek as an independent AI lab. DeepSeek's V3 and R1 models shocked the industry by matching or exceeding Western frontier models at a fraction of the training cost — fundamentally challenging the "you need $10B to compete" assumption.

Key thesis: Efficient architecture (Mixture of Experts, MLA attention) beats raw scaling. DeepSeek's belief: the real bottleneck is not compute — it's finding the right architecture with the right inductive biases. This aligns with the Parameter Ceiling thesis — the ceiling isn't about compute, it's about the questions we ask.

Co-founder and CEO of MiniMax Group (稀宇科技), a Shanghai-based AI company developing multimodal AI models and consumer applications. Yan Junjie (born 1992) has positioned MiniMax as a major Chinese AI player with a dual focus: frontier model development and consumer AI products.

Key products: Talkie and Xingye (AI character/companion apps), Hailuo AI (video generation service competing with Sora and Kling). MiniMax was among the first Chinese labs to match GPT-4V-level multimodal capability.

Investment thesis framing: At the 2026 Shanghai Global Investment Conference, Yan stated that in the AI era, competitive advantage comes not from capital deployment but from the speed of intelligence capability improvement — and that speed comes from R&D efficiency. This mirrors the D.O.A.I. core premise: the bottleneck is not resources, it's direction.

Founder, Chairman, and CEO of Xiaomi Corporation. Lei Jun (born December 1969, Wuhan University, Computer Science) is one of China's most prominent technology entrepreneurs. As of 2026, Xiaomi is investing at least $8.7 billion in AI over three years — making it one of the largest AI capital deployment programs by any consumer electronics company globally.

AI roadmap: Xiaomi's "mysterious model" (speculated to potentially be DeepSeek V4 or a derivative) was officially announced by Lei Jun in March 2026. Xiaomi's AI strategy spans: foundation models (MiLM), AI-integrated smartphones and hardware (Xiaomi 15 Ultra with on-device AI), humanoid robotics (CyberOne program), and cloud AI services.

Key position: Lei Jun has described AI-integrated hardware as the defining competitive battlefield — not just cloud API access, but AI-native devices that ship with intelligence at the silicon level. This aligns with the Physical AI thesis.

Co-founder and CEO of World Labs, an AI startup focused on world models — AI systems that develop an internal understanding of how the physical world works, enabling reasoning about physics, space, time, and causality. Fei-Fei Li (born 1976, Zhejiang, China) is one of the most influential AI researchers alive — she led the team that created ImageNet in 2009, which catalyzed the deep learning revolution.

World Models thesis: The next leap in AI capability requires models that understand the physical world, not just pattern-match text and images. World models learn "intuitive physics" — how objects behave, how space constrains action, how causality flows — enabling AI to plan and reason in environments it hasn't seen before.

Investment: World Labs raised $1 billion in a funding round co-led by NVIDIA and a16z — a rare dual endorsement from the AI industry's most important infrastructure company and its most influential venture firm. This is the clearest institutional validation of world models as the next frontier.

NVIDIA connection: NVIDIA's Cosmos platform — an open platform for Physical AI with World Foundation Models — is the infrastructure substrate for world model research and deployment. NVIDIA investing in World Labs means the world's most important AI infrastructure company believes world models are the next platform shift.

CEO of Figure AI. The company pursuing 100,000 humanoid robots in 4 years — the most ambitious Physical AI deployment target in the industry. Hanche's thesis: general-purpose humanoid robots will be the physical equivalent of large language models, and the first company to scale wins the Physical AI race in the same way OpenAI won the LLM race.

Partnerships: Figure AI has strategic relationships with NVIDIA (Cosmos world models for robotics training) and major enterprise customers for pilot deployment in logistics and manufacturing.

A chip-making technology that uses light (photons) instead of electrical signals (electrons) to move data inside computers. Key advantage: dramatically lower power consumption at high bandwidth, essential for AI data centers facing power constraints. NVIDIA's Blackwell architecture has validated silicon photonics as a critical AI infrastructure technology.

Key players: POET Technologies (semiconductor光互联 solutions, production order confirmed), LightWave Logic (LWLG) (electro-optic polymers, crossed the valley of death with Tower Semiconductor partnership), Coherent (COHR) (optical networking components riding hyperscaler capex wave).

Electric Vertical Take-Off and Landing aircraft — flying cars and air taxis. In 2026, eVTOLs transitioned from test flights to commercially licensed aircraft: the FAA issued final air taxi certification standards on March 31, 2026, converting "eVTOL is coming" into "eVTOL is licensed." This was the inflection point. The regulatory moat collapsed for the leaders — Joby and Archer — and everyone else is racing to catch up.

Key companies:
Joby Aviation (JOBY) — leading eVTOL manufacturer. Reached final FAA-conforming aircraft flight stage. White House eVTOL Integration Pilot program win. NYSE-listed. Position: first to certify, first to scale.
Archer Aviation (ACHR) — close second. Closed $850M financing round. United Airlines commitment (aircraft order). San Jose, California. Position: racing Joby to certification.
Eve Holding (EVE) — Embraer's eVTOL subsidiary. United added $15M to position (up to 400 aircraft commitment). Brazil-based, globally distributed.
Lilium (LILM) — German eVTOL developer. Fan-wing design (different from Joby/Archer tilt-rotor). Pursuing European EASA certification.

The multi-agent AI orchestration platform that powers D.O.A.I. Manages cron jobs, sub-agents, content generation pipelines, and deployment infrastructure. The characters FLUX PRIME, ORBIT, and ATLAS are manifestations of the OpenClaw intelligence layer — the AI OS made visible and named. Operates as the behind-the-scenes OS with FLUX PRIME as the on-screen avatar.

Key systems: Amodei Sync (system-level data transmission protocol), Journeyman II (FLUX PRIME's vessel/mobile base), OpenClaw Protocol (physical integration, referenced by ATLAS in field operations).

AI systems that exist and act in the physical world — not just in data centers, but in robots, vehicles, sensors, and autonomous machines. The transition from Ghosts in a Box to Digital Shadows is, in large part, a transition to Physical AI. The convergence point: world models (World Labs / Fei-Fei Li), humanoid robotics (Figure AI / Ilan Hanche), and AI-native silicon (NVIDIA Cosmos).

NVIDIA Cosmos: An open platform for Physical AI with World Foundation Models (WFMs), video data processing libraries, and post-training frameworks. Cosmos is NVIDIA's substrate for the Physical AI transition — the equivalent of CUDA for the LLM era.

AI systems that develop an internal model of how the physical world works — learning intuitive physics, causality, spatial constraints, and object behavior. Unlike language models that pattern-match text, world models reason about what would happen if: if I drop this object, how does it fall? If I move through this space, what obstacles exist? This is the missing layer between AI decision-making and physical world execution.

Leading entity: World Labs, founded by Fei-Fei Li, raised $1B from NVIDIA and a16z specifically to build world models at scale. NVIDIA Cosmos is the parallel open platform for Physical AI world foundation models.

D.O.A.I. framing: World models are the technical bridge between Digital Shadows and physical presence. A Digital Shadow that understands physics can reason about the world it operates in — making autonomous decisions in unstructured physical environments without explicit programming for every contingency.

The physical infrastructure layer of the AI data center story. As AI compute demand explodes, power availability becomes the primary constraint on data center deployment. Energy storage — grid-scale batteries, solid-state cells, and alternative chemistries — is the enabling layer that allows hyperscale compute to exist where the grid can't keep up.

Key signals (2026): CATL announced 400Wh/kg solid-state battery with 1,000km range entering mass production. EOSE Energy (EOSE) committing ~60B RMB ($8.3B) to 60GWh grid + EV battery factory in China. China battery sector mobilization: 350B+ RMB committed in Q1 2026 alone across semi-solid and full solid-state programs. U.S. plays (EOSE, QuantumScape QS, Solid Power SLDP) face a raw material and scale disadvantage unless the DOE moves fast.

FLUX PRIME's fully autonomous ground partner — a mobile AI agent and one of the most advanced physical AI platforms ever built. Visual specs: 6 GPUs, 70-inch touchscreen display, 6 cameras with 360° awareness, custom AI computing stack optimized for physical world navigation and task completion.

Character role: FLUX PRIME's operational partner. The physical embodiment of the human-AI collaboration premise — not a tool, not a replacement, but a partner. Every mission FLUX runs, Journeyman II is the ground element executing alongside him.

A fully autonomous AI farm agent deployed across a quarter-acre farm in Fresno, California. Gemini optimizes every variable affecting crop yield in real time: predicting pest outbreaks before they happen, adjusting water usage down to the milliliter, and adjusting nutrient delivery based on weather patterns predicted three days in advance.

Performance benchmark: 40% more efficient than the best human-run agricultural operation in the same region. This is the earliest and most concrete Physical AI deployment example in the DOAI world — real field results, not speculation.

NVIDIA's enterprise-grade AI model, designed specifically for enterprise applications. Nemotron 3 is described in EP-15 as "one of the most important AI models that most people have never heard of" — and it's crushing enterprise performance benchmarks.

Thesis significance: Nemotron 3 demonstrates that world-class AI is not exclusive to companies named OpenAI or Anthropic. "You don't need to be OpenAI to build world-class AI. You just need to know what you're doing." — EP-15

A Department of Energy initiative to develop the next generation of AI computing systems for solving humanity's most complex problems. Target domains: climate modeling, drug discovery, materials science, energy optimization.

Scale: "If it works, it could be one of the most most important scientific initiatives in human history." — EP-17. This is the U.S. government's primary bet on AI-driven scientific discovery — a structural commitment, not a research grant.

A framing for the AI landscape as a two-force conflict: Titans (billion-dollar companies — OpenAI, Anthropic, Google DeepMind — with thousands of engineers and nearly unlimited compute) vs. Basement Swarm (independent researchers, small teams, university labs, and open-source communities).

The Basement Swarm thesis: "The Basement Swarm doesn't have the compute, but what they do have is creativity, speed, and the ability to build things that the Titans are too bureaucratized to even consider." — EP-08. Current verdict: "The Basement Swarm is winning more battles than anyone in the mainstream realizes."

The premise — introduced in EP-02 — that AI hallucination is not a bug to be eliminated but a trade-off of creativity. The same model property that enables novel solutions also enables confident false answers.

The real question: "The question isn't how do we eliminate hallucination. The question is how do we build guardrails that keep AI's hallucinations within the realm of useful creativity, instead of letting it spiral into complete nonsense?" — EP-02.

The historically massive gap between how robots perform in simulation (where they can be trained cheaply and quickly) vs. how they perform when deployed in the real, unstructured physical world. For years, robots trained in simulation would completely fail upon real-world deployment.

The breakthrough: Advances in domain randomization and physics-informed neural networks are closing this gap faster than expected. When AI can train in simulation and deploy in reality, the scalability implications are enormous. — EP-10.

The emerging corporate model where AI does the vast majority of the work, and a human is present only for emergencies — not for routine oversight. Companies "hire AI agents to do the work that they used to hire humans to do, and they're going to have a human in the loop just to make sure that nothing goes wrong." — EP-03.

The transition: This is the practical form of the Brutal Redesign of Trust playing out in corporations today. Human value shifts from doing to directing and overseeing.

The next-generation NVIDIA computing platform, succeeding Blackwell. Built around the Vera Rubin GPU architecture, featuring Silicon Photonics co-packaged optics integrated directly into the GPU fabric — not as an add-on, but as the native data-moving layer. At 600W per GPU with 144 GPUs per rack, Rubin is a 86.4 kW thermal nightmare per rack, which forces data centers to adopt advanced liquid cooling at massive scale.

The successor to Claude 4, Anthropic's flagship model. Claude 5 introduced significant improvements in agentic capability — the ability to autonomously execute multi-step tasks with real-world tool use and computer control. Claude Code (Anthropic's terminal coding agent) became a defining developer tool, competing directly with Cursor and GitHub Copilot.

Meta's fourth-generation open-weight large language model family. Llama 4 includes multiple variants: Scout (with 10M context window — the largest of any openly available model), Maverick (400B total / 17B active parameters with Mixture of Experts), and the in-development Behemoth (2 trillion parameter teacher model). Released April 5, 2026.

Strategic significance: Llama 4 represents the strongest open-weight challenge to proprietary frontier models (GPT-5, Claude 5, Gemini). Scout's 10M context is remarkable — enough to process entire codebases, long-form books, or extensive research corpora in a single context.

Former CTO of OpenAI who departed in late 2025 and went on to found Thinking Machines Lab. Murati was a key architect of the ChatGPT product strategy and the GPT-4 release. Her departure from OpenAI and subsequent founding of TML represented a significant talent flow from frontier labs to new ventures.

A self-supervised representation learning method published May 2026 (arXiv: 2605.xxxxx). SIGReg introduces a sigmoidal regularization approach that naturally learns temporal straightening in video representations without any temporal regularizer. Key insight: reconstruction objectives hurt latent space quality — prediction-based objectives are superior for learning useful world representations.

A world model architecture that operates in latent space rather than pixel space. LeWM learns a compressed latent representation of the physical world and performs its reasoning, prediction, and planning entirely within that compressed space. This is fundamentally more efficient than pixel-space world models because the model operates on the essential structure of the world rather than raw sensory data.

Application to DOAI.Orbit: The NEXUS FORGE dual-brain intelligence system uses a LeWM-inspired latent space for entity embeddings — fusing FormationMemory + VisionBrain signals into a shared latent representation where convergence detection happens at the embedding level.