AI North-Star Terminology

Summary

This page tracks high-level AI goal terms only when they affect the wiki’s time-series and world-model agenda. The current stance is conservative: terms such as AGI are too overloaded to guide technical synthesis unless the source defines measurable interfaces. Superhuman Adaptable Intelligence is useful because it shifts attention from a fixed human-centric capability checklist toward adaptation speed, specialization, and useful tasks.

For this wiki, the important translation is:

not "one model that does everything"
but "systems that maintain useful state and adapt quickly in important domains"

Concept Map

flowchart LR
  AGI[AGI as overloaded generality] --> Problems[unclear scope, human-centric targets, weak assessment]
  SAI[SAI: specialization plus adaptation speed] --> Substrates[SSL, latent prediction, world models, modularity]
  Substrates --> TSFM[foundation time-series model agenda]
  TSFM --> Test[latent state, context, actions, useful futures]

What The Wiki Currently Believes

  • AGI should not be used as a precise technical target unless a source defines its task scope, resource constraints, and measurement protocol.
  • SAI is a better discussion handle than AGI for this wiki because it emphasizes specialization, fast adaptation, and utility. It is still not an operational benchmark by itself.
  • The relevant bridge into time-series modeling is state and dynamics learning: a system should maintain latent state, use context, and eventually evaluate plausible futures under actions, control inputs, or interventions.
  • The AMI lineage supplies the technical substrate adjacent to SAI: self-supervised predictive representations, world models, hierarchical planning, and latent or representation-space prediction.
  • Alex’s foundation time-series agenda is narrower and more operational than SAI. The Foundation Time-Series Model Research Agenda asks whether a model closes concrete bottlenecks for numeric time series, event streams, context, streaming state, and action-conditioned world modeling.

Operational Reading

When a source uses a broad AI north-star term, map it into the wiki’s canonical interfaces before treating it as evidence:

Source languageWiki translationEvidence required
General intelligenceBroad transfer or adaptation across task familiesDefined task distribution, resource budget, and evaluation protocol
Superhuman performancePerformance above a human baselineClear baseline, domain, metric, and cost model
AdaptabilityFast acquisition of new skillsFew-shot, online, or transfer benchmark with data and compute budgets
World modelPredictive representation of state and dynamicsFuture-state or action-consequence evaluation, not only passive forecasting
SpecializationDomain-targeted model capacity or routingMatched-compute comparison against broader shared models

Relation To Foundation TSFM Agenda

SAI is a useful north-star frame, but the Foundation Time-Series Model Research Agenda owns the measurable research target. The agenda should use SAI mainly as a warning against generic “one model for everything” claims. A foundation time-series model can still share broad representations, but its evidence should come from concrete capabilities: streaming latent-state updates, context use, rare-regime preservation, calibrated future distributions, dense numeric generation, and action-conditioned rollout.

In time-series language, SAI pushes the agenda toward specialist systems that can adapt to a particular system embodiment:

telemetry schema + context + event streams + typed action/control-input/intervention interface
  -> maintained latent state
  -> fast adaptation to useful operational tasks

That is compatible with broad pretraining, but the evaluation should reward useful adaptation and decision support rather than only average zero-shot forecasting.

Evidence Boundaries

AI Must Embrace Specialization via Superhuman Adaptable Intelligence is a credible position paper, not direct model evidence. It can justify terminology choices and research framing, but it should not be cited as empirical support that a particular self-supervised objective, JEPA variant, world model, or modular architecture works.

The paper also leaves utility and importance underdefined. For this wiki, those terms should be concretized through domain-specific objectives such as incident reduction, reliable control, scientific discovery, patient outcomes, business-process constraints, or another explicit outcome surface.

Open Questions

  • What is the right adaptation-speed benchmark for high-dimensional multivariate time series?
  • How should a TSFM benchmark score adaptation under changing telemetry schemas, context availability, and intervention surfaces?
  • Which parts of a time-series/world-model stack should be general pretrained substrate versus specialist module?
  • How should utility be defined for operational systems without reducing every task to short-term economic value?