--- title: "Traversing the Uncanny Ridge: Searching for Novelty in Intersystemic Communication" description: "Sonia Bernac, Tyler Farghly and Gary Zhexi Zhang argue that the blandness of AI-generated content is not evidence that models are parrots but a predictable effect of how AI systems are made to agree with one another." type: "reading-notes" authors: - "Sonia Bernac" - "Tyler Farghly" - "Gary Zhexi Zhang" published: "2025-05-10" original_url: "https://coginfra.antikythera.org/" doi: "10.1162/ANTI.5CZG" canonical_url: "https://antikythera.wiki/work/journal/coginfra/traversing-the-uncanny-ridge" md_url: "https://antikythera.wiki/md/work/journal/coginfra/traversing-the-uncanny-ridge" last_updated: "2026-08-23" site: "Antikythera Wiki" --- # Traversing the Uncanny Ridge: Searching for Novelty in Intersystemic Communication > Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below. - **Authors:** Sonia Bernac, Tyler Farghly, Gary Zhexi Zhang - **Published:** 2025-05-10 - **Kind:** Studio projects · Journal article - **DOI:** https://doi.org/10.1162/ANTI.5CZG - **Original:** https://coginfra.antikythera.org/ - **Length:** 7.6k words - **Part of:** [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) - **This page:** https://antikythera.wiki/work/journal/coginfra/traversing-the-uncanny-ridge - **Notes generated:** 2026-08-23 ## Summary [Sonia Bernac](https://antikythera.wiki/people/sonia-bernac), [Tyler Farghly](https://antikythera.wiki/people/tyler-farghly) and [Gary Zhexi Zhang](https://antikythera.wiki/people/gary-zhexi-zhang) argue that the blandness of AI-generated content is not evidence that models are parrots but a predictable effect of how AI systems are made to agree with one another. CLIP guidance, reinforcement learning from human feedback and the generator–discriminator game of a GAN all optimise two systems toward consensus, and consensus shrinks the space of what can be produced. The authors name the result the canny valley, the mirror image of Masahiro Mori's uncanny valley: outputs that are eerily familiar rather than eerily strange. Against it they propose the uncanny ridge, a zone of partial misalignment between systems where misrecognition produces new logics rather than breakdown, and an uncanny index that measures the surprise an interaction adds relative to non-interaction. The paper opens the Productive Disalignments theme of the Cognitive Infrastructures compendium and supplies its most technical instrument. ## The argument The paper starts from a refusal. Alignment, whether of AI to humans or of model to model, is usually assumed to be good in itself. The authors treat this as a design constraint that needs justifying, and find it does not hold: complex systems often owe their robustness to delayed alignment, noise and divergent readings of shared goals, and the history of evolution and civilisation is partly a history of productive mistranslation. Machine learning's own record supplies the evidence. Foundation models run into two obstacles, architectures coupled to their data and the difficulty of datasets spanning every modality, so the workaround is to train separate models and make them agree. CLIP forces text and image embeddings into one space; diffusion guidance forces a generator to accept a classifier's verdict; RLHF optimises a model toward a reward model of average human approval. Each agreement has a documented cost: memorised aesthetics, mode collapse, preference drift. The endpoint is a cognitive monoculture. A theoretical interlude separates complexity from novelty. No technical sense of complexity supplies a recipe for the new, and novelty is perspectival, registering only against a system's own history and priors. That framing is what later lets the authors define it mathematically rather than by human surprise. The uncanny valley is then reread as a general problem of recognition between systems. Where Mori's valley is a trough of discomfort, the ridge is a peak of interactional intensity at which misrecognition reorganises intelligibility instead of collapsing it. The early GAN images with multiplied fingers were, on this reading, an unplanned visit to the ridge that optimisation erased. The second half makes the idea operable. A taxonomy of interaction protocols is plotted on a grid of whether each system can see and can model the other, and each is scored by whether it raises or lowers novelty. The uncanny index formalises that score as the change in Bayesian surprise relative to the same systems running independently. Only after the ridge has been mapped, the authors say, should design begin; they close with speculative strategies that follow from the diagnosis rather than preceding it. ## Section by section ### 1 Introduction AI in the world is an ecology of systems with different protocols, timescales and sensory channels, and the friction between them can be treated as a fault or a resource. The paradox is that this diverse ecology produces derivative outputs; the paper promises systemic causes rather than the stochastic-parrot verdict. ### 1.1 Intersystemic Interactions in Machine Learning Recent generative modelling is read as a sequence of agreements between systems: CLIP's shared embedding space, diffusion guidance, RLHF's reward model. In each case cited research shows homogenisation, from CLIP-guided diffusion's commonsense associations to RLHF's mode collapse and drift toward the acceptable. The authors read alignment research's ideology as treating deviation as anomaly rather than possible innovation, and widen the lens to communication among grids, vehicles, trading algorithms and personal agents. ### 1.2 Complexity versus Novelty The authors distinguish computational, Kolmogorov, entropic and emergent senses of complexity and observe that none mechanically generates the new. Novelty is equally unstable across disciplines. They borrow from mid-century process art the idea that how something is made can be novel even when the artifact is dull, and conclude that novelty belongs to a bounded system with a history rather than to an observer's surprise. ### 1.3 Recognition and Dissonance Mori's valley is presented as an epistemic rupture, with two neurocognitive explanations: a conflict between empathy and threat circuits, and a prediction error too large to reconcile, sharpened by self-likeness. The authors generalise: recognition between any systems is partial and negotiated, and misrecognition is a driver of novelty as much as a failure to fix. ### 1.4 The Uncanny Ridge The key reversal. Early GAN and diffusion artifacts were novel and were optimised away as errors. The authors redraw the valley as a ridge where intentional miscommunication between heterogeneous agents yields perspectival novelty, while disclaiming any fetish for misalignment as such. ### 2 Mapping Intersystemic Interactions A brief framing of the taxonomy as a way of watching semantic space expand or collapse under different patterns of communication between paired systems. ### 2.1 A Taxonomy of Misalignment Four protocols. The null protocol, systems probabilistically independent, is the baseline. The intersection protocol, systems driven to agree, produces hyper-convergence; CLIP-guided diffusion, RLHF mode collapse and GAN generator collapse belong here. The mutual projection protocol covers agents that model each other without direct sight; DeepDream, which makes a classifier hallucinate what it detects, is the machine example, and Cold War deterrence, where paranoia about an unseen rival materialised as cryptography and aerospace, the historical one. The Roadside Picnic protocol, after the Strugatskys, has one system interpreting another's output without its context and thereby expanding its own semantic range. ### 2.2 Novelty and Misalignment The protocols are placed on a four-quadrant diagram whose axes are whether a system can see the other and can model it, with a third dimension from perspectival novelty to collapse of meaning. The can-see, can-model quadrant holds RLHF and GAN collapse and scores a drop; DeepDream sits in can't-see, can't-model and scores a rise. The figures are not reproduced in the text. ### 3.1 The Uncanny Ridge The ridge is restated across biological, artificial and hybrid systems, with pollination and humans' poor heuristics for spotting machine text as examples. It is called topological because it names a structural dynamic rather than a threshold. ### 3.2 Proposed Experiment An observer, internal or external, tracks novelty in outputs across varying degrees of alignment; the ridge appears as the trajectory where misalignment yields emergence rather than collapse. Because measurement depends on the observer's memory, perception, access and position, the experiment calls for many observers whose readings are aggregated into a dimensioned map rather than an average. The section is explicitly a thought experiment. ### 3.2 Measuring Novelty Increase The formal core, under a heading the source numbers identically to the previous one. Treating a pair of systems' joint output as a random variable, the authors take expected negative log-probability under an observer's prior as surprise. The uncanny index is the change in that surprise between the pair interacting under a protocol and a counterpart pair under the null protocol; supremum and infimum over admissible pairs give upper and lower indices per protocol. The index is described as an auxiliary formula tying the ridge to systemic surprise. ### 3.3 Outlined Solutions for AI Intersystemic Communication Five speculative strategies: keep modalities in separate embedding spaces, by analogy with Bateson's schismogenesis, and reintroduce them later; build a mediator model inverse to CLIP; cultivate cumulative semantic drift as the counterpart of preference drift; tune alignment to prompt complexity; give users interfaces that adjust alignment between models. All are subordinated to prior mapping of the ridge. ### 4 Conclusion The case for harnessing misalignment is restated with one new analogy, topological frustration in physics, where constraints keep a system from its lowest-energy state and hold it in productive tension. The index is what distinguishes productive misalignment from plain failure. ## Key concepts - **Canny valley** — the aesthetic of hyper-converged generative systems: outputs that are eerily familiar and recombinant; the inverse of Mori's uncanny valley. - **[Uncanny Ridge](https://antikythera.wiki/terms/uncanny-ridge)** — a peak of interactional intensity between partially misaligned systems at which misrecognition reorganises intelligibility and yields perspectival novelty. - **[Uncanny Index](https://antikythera.wiki/terms/uncanny-index)** — the expected change in Bayesian surprise an interaction protocol produces relative to the same systems under the null protocol, with upper and lower bounds per protocol. - **[Hyper-convergence](https://antikythera.wiki/terms/hyper-convergence)** — the narrowing of semantic space when two systems are optimised to agree. - **Interaction protocols** — null, intersection, mutual projection and Roadside Picnic, distinguished by whether each system can see and can model the other. - **Perspectival novelty** — novelty defined relative to a bounded system's priors and history rather than to a human observer's surprise. - **Preference drift** — the narrowing of outputs toward what users will accept under repeated feedback; its proposed counterpart is cumulative semantic drift. - **[Productive Disalignment](https://antikythera.wiki/terms/productive-disalignment)** — the compendium's umbrella for the claim that AI's value lies partly in diverging from human cognition. ## Connections - [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) — the compendium this paper opens; its Productive Disalignments framing restates the thesis as a general principle and names reflectionism as the impasse it escapes. - [Synthetic Counteradaptation](https://antikythera.wiki/work/journal/coginfra/synthetic-counteradaptation) — the sibling paper in the same theme. Agreement that alignment limits adaptive potential, reached through evolutionary arms races rather than the statistics of generative models; its scenarios are narrative where this paper's instrument is formal. - [Mutual Prediction in Human-AI Coevolution](https://antikythera.wiki/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution) — tension. Loewith and Street treat cooperative predictability as what stabilises mutualism; here predictability between systems is the mechanism of collapse. - [After Alignment](https://antikythera.wiki/work/journal/afteralignment) — Bratton's claim that alignment is necessary but insufficient and that overfitting to the human self-image is real. This paper is the machine-learning-internal version of that claim. - [Generative Topolinguistics](https://antikythera.wiki/work/journal/coginfra/generative-topolinguistics) — the compendium's other paper on embedding spaces; its bidirectional interfaces are close to this paper's call for user-adjustable alignment. ## Terms used - [Cognitive Infrastructures](https://antikythera.wiki/terms/cognitive-infrastructures) - [Hyper-convergence](https://antikythera.wiki/terms/hyper-convergence) - [Inverse Uncanny Valley](https://antikythera.wiki/terms/inverse-uncanny-valley) - [Uncanny Index](https://antikythera.wiki/terms/uncanny-index) - [Uncanny Ridge](https://antikythera.wiki/terms/uncanny-ridge)