Defining AGI as
Adaptive Artificial Scientist

A visual companion to the episode exploring Bennett’s 2025 provocation: replace vague “AGI” talk with broad adaptive competence under limited knowledge and resources. The page emphasizes the paper’s core moves — artificial scientist framing, search/approximation taxonomy, and the critique of computational dualism.
arXiv:2503.23923
Bennett, 2025
Theme: resource-bounded adaptation
Extracted arXiv IDs from transcript: 2503.23923
Episode Lens
Primary target
Artificial Scientist
Core criterion
Adaptation
Foundations
Search + Approx
Central critique
Dualism
Visuals use realistic mock scores to illustrate the paper’s conceptual comparisons, not claimed benchmark results.

From vague AGI to adaptive artificial scientist

Flow diagram of the paper’s argument. Hover nodes for details. Use the stepper to reveal how the definition becomes more operational — but still not fully benchmarked.

definition shift technical framing evaluation pressure unresolved ambiguity

Framework comparison matrix

A heatmap comparing major intelligence frameworks discussed in the episode. Toggle the scoring lens to see how rankings change when embodiment and hardware constraints matter more.

low medium high

Computational dualism stress test

The same “software capability” behaves differently under different hardware, latency, and energy budgets. Toggle deployment regimes to see how effective intelligence shifts across substrates.

Search, approximation, and hybrid systems

Interactive architecture view of pure search, pure approximation, and hybrid systems. The lower heatmap shows compute allocation across planning depth and learned priors.

What would “artificial scientist” evaluation actually test?

A radar-style comparison plus task-pipeline view. Switch between target profiles to contrast leaderboard capability, adaptive scientist behavior, and practical digital contractor competence.

References

Compact source map for the concepts visualized above.

Bennett (2025)What the F*ck Is Artificial General Intelligence?
arXiv:2503.23923
Pei Wang (1995)On the Nature of Intelligence — intelligence as adaptation under insufficient knowledge and resources.
Legg & Hutter (2007)On the Measure of Intelligence / A Definition of Artificial Intelligence.
Chollet (2019, discussed in transcript)Skill-acquisition efficiency and generalization as intelligence framing.
Sutton (2019)The Bitter Lesson — scalable search and learning beat domain-specific tricks over time.
Hutter (2005)AIXI — formal upper-bound ideal, not engineerable practice.
Silver et al. (2016, 2018)AlphaGo / AlphaZero — canonical hybrid of learned priors and search.
Kaplan et al. (2020); Wei et al. (2022)Scaling laws and emergent abilities in large language models.
Finn et al. (2017)MAML — fast adaptation under limited data.
Cobbe et al. (2020)Procgen — procedural generalization beyond memorized layouts.