AI Vision & Future
What this document is#
I approach AI as a builder: what can we make dependable, what does it cost, and what can people reasonably entrust to it? This working paper examines increasingly capable AI systems, agentic systems, and AGI-adjacent claims through those questions.
The deployed AI system is the unit of analysis: models, tools, data access, state, evaluation, governance, infrastructure, and organizational context. Claims about value must name who benefits and at what level.
I explain how each mechanism could work and what would have to hold for it to succeed. Use, capability, reliability, and value require distinct evidence. Where that evidence is incomplete, I say so.
This paper supplies the analytical foundation for the practical guidance in the AI Operators Handbook.
Scope#
This paper examines:
- deployed AI systems, including models, tools, data, permissions, state, evaluation, governance, and accountability;
- mechanisms that affect reliability, adoption, organizational learning, and economic diffusion;
- feedback loops created by operation, measurement, and iteration; and
- conditions under which bounded agency becomes feasible, risky, or counterproductive.
These mechanisms set the paper's scope. Timelines for AGI or capability breakthroughs, forecasts about vendors, markets, or geopolitical outcomes, questions of consciousness or moral status, and narratives of inevitability remain separate questions.
“AGI-adjacent” refers here to the operating conditions of increasingly broad competence and transfer: breadth, task horizon, dependence on scaffolding, and dependable performance across domains.
Physical autonomy is outside this paper's scope. When a system acts in the physical world, observability, reversibility, safety, and the cost of error all change; those implications require separate treatment.
How to read this#
Read in order or by section. Later sections use earlier definitions; stable links connect them.
How the argument fits together#
- 01 · Framing establishes definitions, system boundaries, and evidence levels.
- 02 · Supercycle asks when economics, reuse, and organizational absorption could sustain broad diffusion.
- 03 · Flywheel describes learning within a current operating boundary.
- 04 · Agentic defines the delegated action and control problem inside an agency envelope.
- 05 · Helix (Hypothesis) asks when demonstrated control permits the boundary of tractable work to expand.
- 06 · Conclusion identifies what remains robust, uncertain, and worth testing.
This revision and future updates#
This revision adds the evidence ladder, rebuilds the Flywheel around valid operational learning, defines the agency envelope, and states Helix's entry conditions and tests.
I will update this paper when new evidence changes my assessment of behavior under deployment constraints; measurement, evaluation, governance, or regulation change what is feasible; or a claimed mechanism repeatedly fails in real workflows.
Updates should identify changes to evidence, interpretation, or the model. Preserve section numbers and anchors unless structural changes require otherwise.
Contact#
Questions, critiques, and evidence-based challenges are welcome.