Structuralism and structural representation
Source: https://www.semanticscholar.org/paper/de348b06183864dde9e56c72f66ae02d1eb69258 ↗
Most AI critique leans on Dreyfus's embodied cognition argument — the claim that intelligence requires a body situated in the world.
Chirimuuta's contribution is to trace the problem further back, to the structuralist movement in science and philosophy at the turn of the twentieth century, when the idea that minds (and machines) represent the world by mirroring relational structures first took shape.
Cassirer crystallises this tradition; Heidegger's quarrel with Cassirer over Kant's sensory intuition opens a different, less-travelled path of critique.
For product people building systems that claim to 'understand' or 'represent' the world, this paper clarifies what that claim actually inherits — and what philosophical weight it must bear.
It fills a gap the library has: deep historical philosophy of AI that goes beyond the standard phenomenological objections.
Central argument
Chirimuuta argues that the philosophical foundations of AI's representational claims trace back to the structuralist movement in early twentieth-century science and philosophy, particularly crystallised in Cassirer's work, which held that minds and machines grasp the world by mirroring its relational structures rather than its intrinsic properties. The paper contends that when modern AI systems are said to 'understand' or 'represent' the world, they inherit this specific structuralist commitment — not a neutral or obvious one, but a contested philosophical bet. By foregrounding the Davos dispute between Cassirer and Heidegger over Kant's sensory intuition, Chirimuuta shows that a structuralist account of representation explicitly excludes the kind of grounded, sensory engagement with the world that Heidegger's reading of Kant demands, making that exclusion a feature, not an oversight, of the AI paradigm.
Critique
The paper's historical excavation is valuable, but there is a tension in treating the Cassirer–Heidegger debate as the decisive fork: both figures were operating within a broadly neo-Kantian and phenomenological orbit that may not map cleanly onto the computational architectures and training dynamics of contemporary large-scale models. A thoughtful reader might press Chirimuuta on whether structuralism as Cassirer conceived it — a philosophical program about symbolic forms — is genuinely the intellectual ancestor of, say, transformer-based representation learning, or whether the genealogical line is more rhetorical than causal. The risk is that a historically elegant argument overstates the doctrinal continuity between early twentieth-century philosophy of science and present AI engineering practice.
Why it matters for product
For a CPO, the paper's sharpest practical implication is this: every product claim that a system 'understands user intent' or 'represents the domain' is implicitly making a structuralist bet — that capturing relational patterns in data is sufficient for genuine world-modelling. That bet has measurable consequences for product discovery, because it determines what kinds of user signals you treat as ground truth and what you assume the model is doing when it generalises. If Chirimuuta is right that this bet carries unacknowledged philosophical weight, then product teams have an obligation to design evaluation frameworks that stress-test structural representations against the sensory and contextual dimensions the model by design cannot access.