What does Wang mean by intelligence as adaptation with limited resources?
This explores Pei Wang's definition of intelligence — adaptation to an environment under insufficient knowledge and resources — and asks what the corpus has to say about it; up front: the collection doesn't contain Wang's own work, but several notes circle the same two ideas (resource limits and environment-coupling) under other names.
This reads the question as being about Wang's working definition of intelligence — roughly, the ability to adapt to your environment when you never have enough knowledge or compute to be certain — and what the collection holds on that territory. The honest first thing to say: no note here cites Wang or that definition directly. But two halves of his idea show up repeatedly under different vocabulary, and reading them together is more interesting than any single retrieval.
The 'limited resources' half is where the corpus is richest, because it treats resource scarcity not as a constraint to engineer away but as the thing that *makes* a system intelligent. Memory-amortized inference argues that cognition reuses stored inference paths rather than recomputing answers from scratch, and that this reuse is precisely where energy efficiency — and adaptive thought — comes from Can cognition work by reusing memory instead of recomputing?. That's Wang's premise turned into a mechanism: intelligence as economical reuse under a budget. You can see the same instinct in work on models that learn *when* to think hard versus answer quickly, allocating expensive reasoning only where it pays off Can models learn when to think versus respond quickly?, and in compressing chain-of-thought so the same accuracy costs far less compute Can we steer reasoning toward brevity without retraining?. None of these papers frame themselves philosophically, but all of them treat 'spend reasoning resources wisely' as the core competence — which is exactly Wang's 'insufficient resources' clause.
The 'adaptation to environment' half is anchored by a sharper note: AGI definitions that locate intelligence purely in software commit what it calls computational dualism, isolating a mind from its hardware and environment the way Descartes split mind from body Does software intelligence exist independent of hardware and environment?. This is the strongest in-corpus ally to Wang, because Wang's whole move is to define intelligence *relationally* — as a system adapting to a world — rather than as a fixed property a program either has or lacks. If intelligence is adaptation-to-environment, then benchmarking software in isolation measures the wrong thing.
Where the corpus would genuinely push back on a naïve reading of Wang is on the word 'adaptation' itself. One line of work shows that AI-boosted performance often behaves like an exoskeleton — capability that looks like skill while the AI is present but vanishes when access is removed, never becoming the system's own Does AI assistance build lasting skills or temporary abilities?. That's a useful test: adaptation in Wang's sense should persist and accrue, not evaporate. The same distinction appears at the economic scale, where workers genuinely adapt by reallocating to non-displaced tasks when AI exposure is concentrated Does concentrated AI exposure enable workers to adapt and reallocate? — real adaptation under resource pressure, versus borrowed capability.
So the thing you didn't know you wanted to know: the collection never quotes Wang, but it has quietly reconstructed both pillars of his definition — and it adds a discriminator he'd likely endorse, the difference between adaptation that sticks and capability that's merely rented from the tool. If you want a single doorway into the deepest version of this debate, the computational-dualism note is the one to open first.
Sources 6 notes
Influential AGI formalisms isolate intelligence in software independently of hardware and environment, but success depends on all three layers together. This mirrors Cartesian dualism—a fundamental error that makes isolated benchmarks inadequate measures of AGI.
Memory-Amortized Inference proposes intelligence arises from structured reuse of prior inference paths over topological memory, inverting RL's reward-forward logic into cause-backward reconstruction. This duality explains energy efficiency and suggests memory trajectories form the substrate of adaptive thought.
Thinkless trains a single model to select between extended reasoning and direct responses using DeGRPO, which decouples mode selection from answer refinement. This prevents mode collapse and enables self-calibrated routing without explicit difficulty labels.
Activation-Steered Compression extracts a single vector from 50 paired examples to reduce chain-of-thought length by 67% while maintaining accuracy and achieving 2.73x speedup. The method is training-free and generalizes across model sizes and domains.
Research shows AI assistance creates temporary capability extensions—workers produce skilled-looking output while AI is present but revert to baseline performance when access is removed. This differs fundamentally from true skill, which persists independently.
Analysis of task-level AI exposure across firms 2010-2023 shows that while higher mean exposure reduces labor demand, more concentrated exposure (affecting few tasks) enables workers to reallocate to non-displaced tasks, producing modest net employment effects.