Superior intelligence desk · science
Four paths to superintelligence, and the evidence to watch
A research map can help readers ask better questions about the future. It cannot tell them that every route will work. That is the useful way to approach Google DeepMind's June 12, 2026 report, From AGI to ASI.
What the report actually proposes
The report identifies four possible paths: scaling general intelligence, changing AI paradigms, recursive improvement, and collectives of AI agents. It also examines obstacles and unresolved questions. The paths can overlap; the abstract does not present a confirmed arrival date for ASI. Read DeepMind's abstract and publication links.
The questions below are our editorial reading guide. They describe the evidence we would want to see before drawing stronger conclusions from each direction.
1. Scaling: does more produce reliable gains?
Watch what additional resources buy. A comparison should explain what changed, which tasks improved and whether the result survives a realistic cost limit. More compute is an input; better performance on the work being measured is the outcome.
A useful follow-up would compare systems at similar budgets as well as at their maximum settings. That can help distinguish a broadly useful improvement from a result that depends on spending far more per task.
2. New approaches: what can the system do differently?
A claimed change in approach needs a reproducible consequence. Ask which previous limitation it addresses and what evidence rules out a simpler explanation for the gain. An unfamiliar architecture diagram is not enough.
For readers outside a laboratory, the practical question is whether a new approach changes the tasks a system can handle, the evidence needed to trust it, or the resources required to use it.
3. Recursive improvement: who verifies the improvement?
If an AI system helps improve another AI system, the evaluation must remain credible. Ask whether the proposed change works on tests it was not designed around, and whether a separate process checks for regressions.
It also matters how much of the research cycle is completed. Suggesting an experiment, running it and establishing a dependable improvement are different contributions. Reporting should make that distinction visible.
4. Collectives: does coordination earn its cost?
A group of agents should be compared with a simpler system given an appropriate resource budget. Look for evidence that coordination adds something beyond repeated attempts or additional computation.
Our question for an Alberta operator is concrete: if several assistants handle intake, scheduling and reporting, can the business trace their decisions and resolve conflicting records? That is a useful operational test even when the larger research question remains open.
Use scenarios to prepare, not to set a countdown
Keep separate plans for steady improvement, faster progress and stalled results. In each case, identify a decision you would revisit and the evidence that would trigger it. This is a planning exercise, not a prediction about which research path will succeed.
A paper about possibilities is valuable when it makes uncertainty more specific. Read the bottlenecks alongside the proposed routes, and keep the date of the source attached to any forecast you repeat.
Sources and editorial note
Sources checked September 20, 2026. Practical examples and trial plans are this publication's analysis. They do not describe measured client results.
Continue reading
What is superior intelligence? A practical guide to the idea
What superior intelligence means, how it relates to artificial superintelligence, and which questions to ask when an AI capability claim sounds too broad.
Read the guide
AGI vs ASI: where superior intelligence fits
Understand the difference between AGI, ASI, narrow AI and agentic AI, with practical examples of what each term does and does not tell you.
Read the guide
How to measure superior intelligence beyond a benchmark score
Read AI benchmarks with care: compare task success, human baselines, reliability, cost and recovery before accepting a superior intelligence claim.
Read the guide
AI agent autonomy: what to check before giving a system more control
AI agent autonomy depends on tools, permissions and oversight. Learn how to assess longer workflows while preparing for more capable AI systems.
Read the guide
Superior intelligence readiness: a 30-day plan for Canadian businesses
A practical 30-day AI readiness plan for Canadian businesses: choose a workflow, organise data, evaluate results and decide what deserves wider use.
Read the guide