Glossary · AGI and alignment
Fluid intelligence
Fluid intelligence is the ability to reason through unfamiliar problems without relying mainly on previously learned knowledge. In discussions of AI, it usually describes an assessed ability to discover relationships and apply new rules, rather than proof that a system has the same intelligence as a person.
Fluid intelligence: the explanation
An assessment presents problems whose rules must be worked out during the exercise. These might involve completing visual patterns or understanding an unfamiliar sorting system. For a language model, evaluators examine answers across several variations rather than celebrating one successful response. A task that looks new to you might resemble material used during training. This makes it difficult to separate flexible reasoning from the reuse of a familiar solution or pattern.
The concept comes from psychology, where it helps distinguish reasoning about unfamiliar situations from accumulated knowledge. Applied to AI, it encourages questions about whether a system can handle more than familiar exercises. You might want an assistant to apply a rule invented for your workplace, for example. That is a different challenge from recalling a definition, recognizing a standard form or answering a question that appears frequently in its training material.
Using this concept for machines requires care. A test designed for people does not automatically measure the same underlying ability in a language model. Results can change with wording, demonstrations and available tools. An agent might also delegate part of the problem to a program. You therefore need to distinguish what the model contributes from what the surrounding system supplies. A single reasoning score cannot describe every aspect of practical adaptability.
A concrete example
You give an assistant a new rule for organizing your files. Urgent documents come first, except when they are waiting for approval from someone outside your team. You introduce unfamiliar cases and then change one condition. You check whether the assistant updates its decisions and explains exceptions correctly. This provides evidence about reasoning in that setting, without proving general intelligence or ruling out the influence of familiar training patterns.
Why it matters
The concept matters when you need an agent to handle unexpected situations. A system that excels at routine work may struggle when requirements change. Compare candidates using unfamiliar problems, altered rules and incomplete information. Check whether they ask useful questions instead of guessing. Reasoning assessments can complement these practical trials, but they do not replace checks on factual accuracy, permissions or the consequences of actions taken on your behalf.
Common misconceptions
- Solving a problem that appears unfamiliar does not establish humanlike reasoning, because the model may have encountered related material during training and a familiar solution pattern may be enough to produce the correct answer.
- Fluid intelligence does not mean fluent conversation or fast responses; it concerns reasoning through novelty, so you should assess it separately from writing style, factual knowledge and the speed at which an assistant replies.
Related terms: agi and alignment
- AGI (artificial general intelligence)
- AGI, or artificial general intelligence, refers to the idea of AI that can learn and perform a broad range of intellectual tasks beyond a narrow specialty.
- ARC-AGI
- ARC-AGI is a benchmark series designed to study how AI systems solve unfamiliar problems from limited examples or interactions.
- ARC Prize
- ARC Prize is a research and competition initiative that encourages open progress on AI generalization problems.
- generalization
- Generalization is an AI system’s ability to use what it has learned to handle new cases.
Where this site covers it
The term appears on 5 pages of this site. A few of them:
- How We Get To AGI: François Chollet’s talk, chapter by chapter Guide
- Generalization at the heart of intelligence: François Chollet's framework, examples, and measurement Guide
- ARC-AGI: the test, its versions, and published scores Guide
- How We Get To AGI: key ideas from François Chollet’s talk Article
- On the Measure of Intelligence, explained Article
Go further
Frequently asked questions
Is it the same as knowledge?
No. Knowledge concerns information and skills already acquired. Fluid intelligence concerns reasoning through unfamiliar situations. Real tasks often require both, which makes it difficult to isolate either contribution from the final answer alone.
Is one test enough?
One test offers limited evidence. You get a more useful picture by changing rules, wording and difficulty, then checking whether performance remains consistent on new examples rather than relying on one impressive demonstration.
Do tools affect the result?
Yes. A program may solve part of the task for the model. To interpret a result, you need to know whether the agent used search, a calculator or other assistance during the assessment.
Neutrality. No affiliate links. No vendor pays to be reviewed, rated or ranked. Reviews are directed by François Chollet, neutral and factual, and every data point is dated and sourced. How we score