AI statistics, benchmarks & research.

Read the methodology and configuration behind each comparison.

Preference rankings depend on the voter pool, task mix and uncertainty.

Check the date, definitions and uncertainty of each dataset.

Annual reports describe a dated snapshot, not live market totals.

Read each chart's underlying source and measurement notes.

Policy trackers are reference material; confirm current rules with the relevant authority.

Compare models within the same evaluation setting and version.

Hardware, software and benchmark division affect comparability.

Check the benchmark variant, agent setup and evaluation rules.

Results depend on the agent, model, tools and benchmark release.

Use the same release when comparing scores.

Task dates and evaluation windows affect the reported result.

Check the task set, date and agent configuration before comparing runs.

ARC results measure a specific task family, not all forms of intelligence.

Read the experimental assumptions and uncertainty alongside the headline results.

Read each benchmark's task design and scoring method.

A narrow benchmark cannot establish overall model quality.

Composite scores and listed prices should be checked against their sources.

Confirm current provider rates, caching rules and billing units before purchasing.

Use the edition date when citing figures or conclusions.

Institutional research describes the authors' own work and priorities.

Preprints may not have completed peer review.

Submission and review visibility varies by venue.

Technical papers may require statistical or machine-learning background.

This is an archive; check individual article dates.

Distinguish author commentary from primary research evidence.

Review practices and scope vary between conferences and workshops.

Check the publication year and paper-specific evidence.

Conference pages can point to different years; check the selected edition.

Use the relevant conference year when citing work.

Read the original paper and venue details before relying on an abstract.

Dataset licences, provenance and task definitions vary.

Reports reflect their scope and publication date.

Reports and classifications are not a complete or uniformly verified census.

Check the linked reporting and distinguish allegations from established facts.

Guidance is a framework for applying judgement, not an automatic compliance certificate.

Check the contributors, intended audience and scope of each resource.

Research recommendations are analysis, not enacted law.

Confirm the version and applicability of any standard with its publisher.

Follow links to original research; access conditions can vary by post.

Commentary expresses the authors' analysis; consult the evidence they cite.

Some posts may have subscriber access requirements.

Check linked primary sources for the complete context.

Policy analysis reflects a particular research perspective and publication date.

Programming and mathematics background help; environment setup may be required.

Check course and session dates because development practices change.

The visualisation illustrates a particular model and simplifies the wider landscape.

It is an educational demonstration, not a model-quality benchmark.

Conceptual lessons do not replace hands-on training or current implementation guidance.

This is a foundational explanation; newer model designs differ.