Pathways
Scaling, new paradigms, recursive improvement, and multi-agent collectives are distinct routes by which capable systems might become superintelligent.
Canonical orientation / roadmap to ASI
From capable models to compounding intelligence.
This is not a countdown or a prediction. It is a map of the mechanisms that could lead from frontier AI to artificial superintelligence—and the research needed to understand, build, evaluate, and steer them.
How to read this map
4 pathways
Ways capability could advance beyond AGI.
1 compounding loop
How AI-assisted R&D could accelerate that advance.
6 workstreams
How this site organizes the evidence and open work.
One territory / three lenses
The models are complementary, not competing.
The pathways describe how progress may happen. The loop shows why progress may compound. The workstreams turn both into a research program you can navigate.
Compounding loop
Models become agents; agents automate parts of AI R&D; validated research gains feed into better models and systems.
Research areas
Six areas organize the systems, benchmarks, frameworks, and unanswered questions needed to assess and advance the roadmap.
Lens 01 / four pathways
How capability could move beyond AGI.
These pathways come from the site’s anchor orientation report, From AGI to ASI. More than one can advance at the same time.
Make the current paradigm broader, deeper, and more reliable.
More compute, better data, stronger post-training, longer context, memory, tools, and inference-time reasoning continue extending capable general systems.
WatchLong-horizon reliability, tool use, transfer, cost, and the point at which scaling returns change shape.
See frontier-model evidence →
Discover better ways to build and train intelligent systems.
New architectures, learning algorithms, objectives, memory systems, and search procedures could unlock capability that straightforward scaling cannot.
WatchArchitecture discovery, learned optimizers, continual adaptation, and gains that survive outside a narrow benchmark.
See automated R&D →
Use AI to improve the process that creates AI.
Agents write and test code, design experiments, search system configurations, and preserve useful discoveries—closing parts of the improvement loop.
WatchWhether gains are real, general, cumulative, and transferable into the next generation of systems.
See self-improving systems →
Coordinate many specialized systems as a cognitive organization.
Populations of agents may divide work, critique one another, preserve parallel hypotheses, and coordinate at machine speed.
WatchCoordination overhead, diversity, communication, correlated failure, and whether collectives outperform their strongest member.
See multi-agent signals →
Lens 02 / the compounding dynamic
Why the research loop matters.
ASI need not arrive through a single breakthrough. The strategic transition is a shortening feedback cycle: AI systems perform more of the work required to improve AI systems, and verified gains return to the frontier.
The loop only compounds if its outputs are reliable. Evaluation, control, security, and governance are therefore part of the mechanism—not an appendix to it.
- 01Frontier models
Reason, code, use tools, and operate over longer horizons.
- 02Research agents
Plan work, run experiments, critique results, and preserve useful artifacts.
- 03Automated AI R&D
Search architectures, improve scaffolds, create data, and test new methods.
- 04Validated gains
Improvements that survive evaluation feed into better models and systems.
- Continuous layerEvals · control · governance
Measure the loop, catch failure, and keep increasingly capable systems steerable.
Lens 03 / six workstreams
Turn the roadmap into research.
The six areas are the site’s evidence architecture. Each contains primary sources, an ASI Research synthesis, and direct routes to public code where available.
012 signals
Foundations & pathways
How might capable general systems cross into superintelligence?
Explore workstream ↗ 027 signalsFrontier models
Which capability gains change the shape of long-horizon work?
Explore workstream ↗ 038 signalsAgents & self-improvement
Can agents reliably improve the systems that produced them?
Explore workstream ↗ 047 signalsAutomated AI R&D
When does research automation become a compounding loop?
Explore workstream ↗ 058 signalsAI for science
Where are AI systems already closing real discovery loops?
Explore workstream ↗ 0612 signalsEvals, control & governance
How do we make accelerating capability legible and steerable?
Explore workstream ↗Working synthesis / updated 03 Sep 2026
What is visible—and what is not yet established.
This maturity view is an editorial synthesis of the current corpus. It separates reported capability from demonstrated reliability and should change as the evidence changes.
Consequential bounded loops
Frontier agents complete end-to-end computer workflows, research systems produce code-linked artifacts, AI-generated mathematical arguments can be formally checked, and automated pipelines can build domain models from raw data.
Capability with containment
Long-horizon autonomy, architecture discovery, scalable monitoring, secure evaluation environments, evidence-preserving research agents, and reliable multi-agent coordination are active engineering questions.
Sustained compounding
The corpus does not yet establish recursive gains that reliably transfer across model generations, dependable superhuman collectives, or assurance sufficient for ASI-scale deployment.