Work

Algorithmic competence does not require algorithmic thought

Type: article
Abstract: Machine learning researchers often treat a model's ability to produce a correct algorithmic solution as evidence that its internal computation instantiates that algorithm, or at least works algorithmically. Several lines of research go further and aim to build models that execute algorithms themselves. This position paper argues that this aim is misplaced: algorithmic competence is better built into systems in which the model thinks heuristically and calls on algorithms kept outside it, run by components that guarantee correctness. Such systems already exist in AI for mathematics, where language models paired with proof assistants have solved open problems. My main argument for this is the human case: humans, the one general intelligence we know, think largely heuristically and, beyond short mental computations, rely on external memory, and ultimately on external machines, to execute algorithms. I then provide further arguments for this position, discuss what makes a process heuristic and what part the model keeps in such systems, and draw consequences for interpretability, neuroscience, and concept creation.
Language: English
Subfield: Philosophy
Open Access status: green
OpenAlex ID: W7225842395
Cites: 69
Cited by:
Related to: 10