Fluid intelligence

As explained in Deep Learning with R, Third Edition

Fluid intelligence is the ability to solve novel problems and adapt reasoning across varied situations without relying solely on previously learned knowledge.

Fluid intelligence is the ability to solve unfamiliar problems by adapting reasoning, abstraction, and learned skills rather than relying only on memorized knowledge. It appears when a system faces a genuinely new situation and must construct a way forward with little or no new data. section 19.2.2

The practical test is novelty: give a learner a brand-new problem instead of repeating familiar questions. Fluid intelligence supports on-the-fly models of situations, anticipation of possible futures, long-term planning, rapid adaptation, and learning new skills after limited practice. section 19.2.2

Fluid intelligence differs from narrower forms of generalization in several ways:

• It recombines knowledge and strategies for situations that were not specified in advance. section 19.3.5 section 19.2.2
• It requires both value-centric abstraction for perception and intuition and program-centric abstraction for exact reasoning, planning, and causal understanding. section 19.5 section 20.3
• It is not demonstrated merely by scale or brute-force search; efficient results matter because intelligence concerns resource use as well as capability. section 19.3.7

ARC-AGI as a practical test

ARC-AGI presents input-output examples that require discovering the underlying transformation. Its resistance to memorization exposes the gap between systems that interpolate familiar patterns and systems that can infer and recombine rules for new tasks. ARC-AGI-2 raises the difficulty with longer reasoning chains and tasks less amenable to exhaustive search. section 19.3.5 section 19.3.7

Figure 19.6 Local generalization vs. extreme generalization

Compare local generalization with extreme generalization to see why modest deviations from training data differ from genuinely novel problem solving. section 19.2.2

Figure 19.10 Example of a task that couldn’t be solved by o3 on the highest compute settings (more than $20,000 per task)

Inspect a task that remained unsolved by o3 even with more than $20,000 of compute, illustrating the limits of brute-force capability. section 19.3.7

How Deep Learning with R, Third Edition covers it

How deep it goes

Explains novelty, abstraction, planning, efficiency, and benchmark evidence, while deferring a complete implementation of fluid intelligence.

Examples

Uses brand-new problems, ARC-AGI, ARC-AGI-2, LLM arithmetic errors, and high-compute o3 failures as diagnostic examples.

What it assumes

Distinguishes memorization and automation from adaptation, while introducing extreme generalization through concrete comparisons.

What you can do after

Evaluate AI systems by testing novelty, recombination, reasoning quality, and resource efficiency rather than scale alone.

How it connects

Connects extreme generalization with abstraction, reasoning, planning, program synthesis, test-time adaptation, and modular recombination.

liveBookmanning.com

Read Deep Learning with R, Third Edition on liveBook Get the book at manning.com

Related concepts

• Abstraction & Reasoning Corpus for Artificial General Intelligencemeasured by · The Abstraction and Reasoning Corpus for Artificial General Intelligence is a benchmark of visual reasoning tasks designed to assess abstract rule induction and generalization from examples.
• foundational modelcontrasts with · A base or foundation model is a pretrained model that provides learned parameters or capabilities as a starting point for reuse or adaptation to downstream tasks.
• test-time adaptationmeasured by · Test-time adaptation comprises methods that adapt a model through problem-specific learning or reasoning during inference or evaluation.
• ARC-AGI-2measured by · ARC-AGI-2 is a benchmark dataset of abstract reasoning tasks designed to evaluate fluid reasoning and generalization to novel problems.
• multi-step reasoningequivalent to · Multi-step reasoning derives a conclusion, solution, or decision through multiple connected intermediate inferential or problem-solving steps.

This explainer was generated by Manning's knowledge graph from the text of Deep Learning with R, Third Edition. The section links point to the parts of the book it draws on.