CognitiveMetrics

We made an adaptive fluid and spatial reasoning test. Items are presented based on your current estimated ability. The test keeps going until your score is measured precisely enough, making it much more efficient than a fixed-length test. The Abstract Thinking Adaptive Test (ATAT) consists of six item types: figure sequences, figure classification, figure weights, object rotation, pattern completion, and number sequences. The ATAT was specifically designed to be culture-fair. It should take around 30 minutes in total. Try it out and post your results⬇️

Interesting. Overall intelligence probably peaks in your 30s. Working memory, perceptual speed, visuospatial reasoning, and fluid reasoning peak relatively early in life. Only crystallized knowledge improves considerably past your 30s.

Frontier AI models now max out or nearly max out most subtests on our IQ test (CORE), but it’s curious how poorly they still perform on Block Counting. The difficulty in Block Counting comes from quickly identifying the hidden blocks required for support. AI’s reasoning is fine, so the bottleneck likely lies in how it handles spatial representations. For instance, AIs might struggle to map a 2D image into an objective 3D coordinate frame, leading to spatial hallucinations. In humans, CORE Block Counting has a ceiling of 155 IQ, a g-loading of .61, and a total of 31 questions. See the link below to try it out.

The chimpanzees were trained on this task for years. When humans were trained on the task, they matched or outperformed chimps (Cook et al., 2010). A review concluded that the size of working memory in chimps is 2 ± 1, versus Miller's famous 7 ± 2 in humans (Read et al., 2022).

When we factor analyzed the 1926 SAT alongside CORE, a distinct new factor emerged. This factor only appears on speeded tests, hence the tentative name we've given it: Speed. This factor predominates much of the 1926 SAT, a test with very strict timings. In our broader model, which also includes the GRE and the AGCT, it appears across verbal, quantitative, and spatial tasks. Here are some of our preliminary findings: 1. Speed is strongly related to general intelligence, with a g-loading of approximately .80. 2. It is clearly distinguishable from CORE’s Processing Speed Index, which uses Character Pairing and Symbol Search (tests of Perceptual Speed, specifically). 3. Its strongest indicators include Logical Inference (.57), Artificial Language (.56), and GRE Analytical (.51). Smaller loadings appear on some CORE subtests: Quantitative Knowledge, Spatial Awareness, Figure Weights, and Graph Mapping. These CORE subtests are all non-leniently timed and are considered slightly speeded by testees. Overall, Speed's loadings seem too diverse for it to reflect a test-taking strategy or a content-specific ability. We speculate that the Speed factor mostly reflects attentional control and, crucially, the speed at which people process relations. However, Speed might not reflect particular cognitive mechanisms at all. Most theories cannot be ruled out or confirmed until item response times are analyzed and novel tasks are devised to test competing ideas.

Has there been a general collapse in attention in the past few decades? A new study compared Austrian adults’ performance on attention tasks in 2025 with matched samples assessed in the early 2000s. It found small, mixed changes instead of a general decline. Selective attention showed a very slight improvement, but it was not statistically significant. And in a 33-minute vigilance task, contemporary participants responded a bit more slowly, but detected the same number of targets. Their response times also did not deteriorate more rapidly as the session progressed. Although tests in a controlled setting are different from everyday distractibility, these results undermine the idea that digital technologies have caused sweeping declines in attention. These results also reinforce the domain specificity of the Flynn effect.

An independent user has collected thirty-one CORE–WAIS score pairs, finding a correlation of .797 between CORE and WAIS FSIQ, which rises to 0.811 after correction for indirect range restriction. The two tests showed strong normative alignment, with the average score differences being insignificant (+1.6 IQ points) and normally distributed. Although the sample size is small and selected, this further validates CORE's norming methodology and suggests strong convergent validity between CORE and other professional tests.