AGI / OPERATIONAL DEFINITION 1.1
AGI = Artificial General IntelligenceAGI is not a smarter chatbot. It is a new unit of cognitive production.
Artificial general intelligence has no single universally accepted test. We therefore use a public, operational definition: specific enough to measure, cautious enough to revise, and connected to economic outcomes rather than launch rhetoric.
The “system” may combine models, memory, tools, agents, proprietary data and human-governance interfaces. We evaluate the delivered capability, not one model in isolation.
THE THRESHOLD IS MULTIDIMENSIONAL
Four capabilities must compound. Deployment is the external proof.
A benchmark jump can be important without being AGI. The transition becomes economically meaningful only when breadth, autonomy, reliability and unit economics reinforce one another in real work.
Breadth + transfer
Performance across unfamiliar domains, with skills transferring beyond the training-shaped task or one benchmark.
Autonomy + horizon
End-to-end completion of longer tasks: planning, tool use, memory, correction and escalation without constant human steering.
Reliability
Repeatable results in messy environments, including uncertainty recognition, recovery from errors and calibrated refusal.
Economics
Useful work delivered at a cost, speed and availability that make scaled adoption rational—not merely technically possible.
KEEP THE WORDS CLEAN
AI, frontier AI, AGI and ASI are not synonyms.
The full field of machine systems performing tasks associated with intelligence. It includes narrow tools and today’s generative systems.
The most capable current general-purpose models. Frontier status is relative and does not by itself establish AGI.
Broad, transferable, autonomous and economically deployable cognitive capability at least around capable-human level across many domains.
Hypothetical superintelligence that substantially exceeds the best humans across nearly all cognitive domains. It is not our AGI threshold.
WHAT AGI IS NOT
We will not let one dramatic demo settle the question.
A product label
A laboratory can name a model; it cannot name the economic transition into existence.
One benchmark score
A narrow score may hide brittleness, contamination, weak transfer or unusable costs.
Consciousness
Sentience and self-awareness are philosophical questions, not requirements in our investment definition.
Infallibility
Humans are not error-free either. The relevant question is whether reliability is sufficient for a given deployment.
A single day
We expect a phase transition visible through accumulating evidence—not a universally agreed ceremonial date.
Superintelligence
AGI can arrive before systems exceed the best humans in every cognitive field.
OFF-CYCLE LOG / LOG-149
AGI is a field of inquiry—not a unit of measurement.
Roberto Pieraccini’s “Four Questions About AGI” is a valuable challenge to the word itself. It is strongest when it separates a persuasive banner from observable competence, and weakest when the lack of one universal definition is allowed to imply that the category has no operational value.
We keep AGI as the name of the transition we study, but never admit “AGI” as evidence. Every claim must be decomposed into breadth, autonomous horizon, reliability, deployment, owner economics and expectations already in price.
What is intelligence?
No field-wide definition has won. That is a reason to publish the measurement contract—not a reason to stop measuring.
Do people mean the same AGI?
No. A laboratory declaration, broad benchmark competence, reliable autonomy and profitable deployment are different clocks and must never be merged.
Do we need AGI?
Capital does not need the label. It needs evidence that a specific capability removes a costly constraint in a real workflow at a viable total cost.
What would the world look like?
Not an overnight switch. Capability diffuses through infrastructure, institutions, regulation and physical reliability; disruption can arrive before broad productivity gains.
What the essay gets right
One brilliant result cannot establish robust transfer to nearby tasks, unfamiliar contexts or long horizons.
Motors, dexterity, safety and deployment reliability do not inherit the same scaling curve as digital inference.
Even a real capability breakthrough needs workflows, accountability, standards, distribution and trust before it rewrites the economy.
Where we refuse to overreach
The article is an informed essay, not evidence of scientific consensus that AGI is meaningless.
Brooks’s decade-per-nine reliability rule is a useful deployment warning, not a universally validated forecast function.
Whether a system understands or is conscious can remain unsettled while its economic capability is measured.
EVIDENCE / NOT RHETORIC
The measurement stack we will use
Breadth, transfer and performance on unfamiliar, contamination-resistant task distributions.
Accepted-task rate, coherent horizon, retries, human interventions and recovery from failure.
Outcome quality at deployment-relevant confidence—not a best-run demonstration.
Paid usage, retention, workflow penetration, switching cost, safety and regulatory acceptance.
Cost per accepted task, incremental gross profit, capex, dilution and owner-cash conversion.
Which part of that future is already embedded in valuations, positioning and forecast contracts.
WHY CAPITAL SHOULD CARE
If useful cognition becomes scalable, economic scarcity moves.
The investment consequence is not simply “technology stocks go up.” Value can migrate toward bottlenecks, distributors and adopters while some incumbents lose pricing power.
Cognition reprices
The marginal cost of parts of analysis, software, design and coordination can fall sharply.
Research accelerates
AI-assisted R&D can shorten invention cycles and alter the pace at which competitive advantages decay.
Physical constraints surface
Compute, power, memory, networking, data, permits and real-world execution can become the binding constraints.
Profit pools move
Capability creators, infrastructure suppliers, distributors and fast adopters may capture different shares of the surplus.
OUR INVESTMENT TRANSLATION
Capability is not alpha until it survives the causal chain.
- 01Capability→
- 02Reliability + autonomy→
- 03Adoption→
- 04Unit economics→
- 05Cash flow→
- 06Expectations in price→
- 07Capital decision
Every arrow can break. That is why our 25 U.S. equities are a starter hypothesis about control points—not our definition of AGI, not a permanent universe, and not proof of alpha.
WHAT WOULD CHANGE OUR MIND
A useful definition must be falsifiable.
Task horizons plateau
Systems remain unable to sustain coherent work beyond short, supervised tasks.
Transfer fails
Performance gains stay benchmark-specific and do not generalize across unfamiliar domains.
Reliability stays too low
Human verification and failure costs overwhelm the value of autonomy.
Economics do not close
Inference, integration and supervision costs prevent broad profitable deployment.
Adoption does not compound
Technical capability fails to convert into productivity, revenue or durable owner cash flow.
Scaling is structurally blocked
Energy, supply chains, regulation or security constraints prevent deployment at economically relevant scale.
PRIMARY FRAMEWORKS
Our synthesis is grounded, but it is not a claim of scientific consensus.
OpenAI supplies a widely cited economic-work definition; Google DeepMind provides a levels framework; METR gives us a practical way to observe autonomous task horizons. We combine them into an investment-research instrument and will revise it as evidence improves.
Highly autonomous systems and most economically valuable work.
Performance, generality, autonomy and deployment risk as separable dimensions.
A measurable view of the human-expert task length agents can complete at a stated reliability.
Secondary essay under review: useful synthesis and critique, not an independent capability receipt.
Publisher record for Julian Togelius’s survey of definitions from psychology, ethology and computer science.
Dated robotics and deployment observations, including the home-navigation and elder-care countercases.
Primary statement of the decade-per-nine reliability heuristic; treated here as a heuristic, not a universal law.
Primary book-length argument for prediction, functionalism and evolutionary transition.