{"id":"fbbe3f78-6794-4bdf-81c1-4f488e5906d7","arxiv_id":"2501.10369","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A conceptual essay maps ambiguity, uncertainty, and indeterminacy onto machine learning creative practice and proposes half-trained latent spaces as the richest source of new form.","lead":"This paper sorts machine learning creativity into three modes: ambiguity, uncertainty, and indeterminacy, and argues that under-trained regions of a model's latent space may be the most creative. It is an illustrated essay from an architecture studio, not a quantitative study, useful for framing how AI tools can become artistic partners.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4's assertion that latent spaces are unevenly trained, and that half-trained regions are most creative, is load-bearing but unsupported; no mechanism or measurement distinguishes 'less trained' regions from data-sparse or posterior-mismatched regions.","rationale":"I read the paper as a conceptual essay whose central recommendation is to explore 'half-trained' latent spaces as the most creative region of machine learning. The reader's weakest_assumption correctly identifies the load-bearing descriptive claim in Section 4. I agree but would sharpen it: standard neural network training does not operate on latent coordinates individually, so 'degrees of training and determination' needs an operational definition; otherwise the empirical premise is untestable and possibly false. The paper's studio projects (AI Wabi-sabi, Probobli Boboli, etc.) are presented as illustrations, not as controlled evidence; that is appropriate for a conceptual proposal, but it cannot carry a factual claim about how latent spaces form. The concrete test above would disambiguate temporal and spatial readings and would settle whether the premise has empirical content. I do not see a more serious internal inconsistency; the ambiguity and uncertainty sections are analytically coherent. Since the reader's verdict is already CONDITIONAL, my concern does not change it, so UNCHANGED.","tokens_in":4409,"tokens_out":3859,"duration_ms":42220,"concrete_test":"Run a factorial experiment: fix one architecture (e.g., a VAE or conditional GAN) and one dataset; checkpoint at 10/25/50/75/100% of training. Within each checkpoint, sample latent codes from low-density vs high-density regions of the learned latent distribution. Generate a fixed set of outputs per condition, blind-rate them for novelty and quality (or use an objective diversity metric), and compare against the fully trained model. If intermediate checkpoints or low-density regions do not consistently outperform the fully trained model, the Section 4 premise fails. This separates the two readings of 'half-trained': temporal checkpoint vs spatial region.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central future direction depends on Section 4's descriptive claim: 'Space is not evenly or uniformly trained, but instead achieves its variation and hybridity due to differing degrees of training and determination.' This is asserted with no evidence, and it is not a straightforward consequence of the cited models. In GANs and VAEs, gradient descent updates shared weights globally; a latent coordinate is not a locally trained object. Non-uniformity that does exist is typically a property of data density, prior mismatch, or conditioning, not of some regions receiving 'more training' than others. Unless those are the intended meanings, 'half-trained and half-determinate' conflates temporal training stage with spatial location in latent space. If the descriptive premise is wrong, the recommendation to seek creative outcomes in 'half-trained' spaces has no foundation beyond the studio projects, which are not controlled comparisons.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conceptual framework for evaluating creative uses of machine learning, distinguishing ambiguity, uncertainty, and indeterminacy. It maps these onto Boden's combination, exploration, and transformation, and argues that indeterminacy—understood as a property of latent-space formation—is the most distinctive creative quality of machine learning. The argument is illustrated with three projects from the author's Unit 21 at the Bartlett: AI Wabi-sabi (ambiguity), Probobli Boboli and related works (uncertainty), and a programmatic call for 'half-trained' latent spaces (indeterminacy). The paper concludes with ethical remarks and suggestions for future research.","tokens_in":4553,"tokens_out":4223,"duration_ms":43392,"significance":"If the central premise about uneven training of latent spaces were established, the paper would give the creative-AI community a novel research direction beyond minimizing final loss and would offer a useful vocabulary for discussing degrees of creativity. The practice-based illustrations and the ethical emphasis on dataset authorship are strengths. However, as it stands, the key claim is a speculation with no supporting evidence, mechanism, or empirical comparison, which limits the paper's current contribution to that of a provocative essay rather than a substantiated research direction.","major_comments":[{"comment":"The claim 'Space is not evenly or uniformly trained, but instead achieves its variation and hybridity due to differing degrees of training and determination' is load-bearing for the recommendation to seek creative outcomes in 'half-trained' spaces, yet it is asserted without evidence or argument. In standard GAN and VAE training, gradient descent updates the shared weights globally; a given latent coordinate is not a 'trained' object in any local sense. Non-uniformities that arise in latent space are typically attributable to data density, prior mismatch, or conditioning, not to variation in the amount of training applied to different regions. Unless one of these mechanisms is intended, the premise does not follow from the cited models. Please provide a mechanism or reference that supports the claim, or substantially soften the claim into an explicit speculation.","section":"Section 4, second paragraph"},{"comment":"The phrase 'the spaces of the half-trained and half-determinate' conflates the temporal training trajectory with spatial structure. A model early in training has an under-fit latent space, but that is a global property of the checkpoint, not a region that remains 'half-trained' after convergence. If the intended meaning is early-stopped checkpoints, intermediate training epochs, or regions of low data density, that should be stated explicitly; the current phrasing suggests a spatial localization that standard training dynamics do not produce.","section":"Section 4, last sentence"},{"comment":"The alignment of ambiguity, uncertainty, and indeterminacy with Boden's combination, exploration, and transformation is asserted in a single sentence ('It could be argued...') without supporting argument. This mapping is the paper's main conceptual contribution and should be defended. In particular, Boden's transformation requires changing the defining dimensions of the conceptual space, which is not obviously what is meant by 'indeterminacy' in latent spaces. The paper should explain how the mapping works rather than leaving it as a suggestion.","section":"Section 5, first paragraph"},{"comment":"The conclusion that indeterminacy is the most distinctive quality is partly predetermined by the definitions. Ambiguity is characterized from the start as 'imitation' that 'reinforces the corpus' (Section 2), whereas indeterminacy is characterized as 'testing the limits of ontology' (Section 4). Given these choices, the ranking is a foregone conclusion. The paper should either justify those characterizations against plausible alternatives or acknowledge that the ranking is a definitional framing rather than an empirical finding.","section":"Sections 2 and 4"}],"minor_comments":[{"comment":"Reference [31] contains a typographical error: 'F ormalized music' should read 'Formalized music'.","section":"References, [31]"},{"comment":"The project name is rendered inconsistently as 'AI Wabi-sabi' and 'AI wabi-sabi' in the text and in the figure captions; please standardize the capitalization.","section":"Section 2 and Figure captions"},{"comment":"The section introduces two senses of indeterminacy (the meta-communicative frame from Bateson and Hertzmann, and the measurable 'constituent part' sense). The transition between these senses is abrupt; a sentence explicitly distinguishing them would improve clarity.","section":"Section 4, first paragraph"}],"recommendation":"major_revision","confidential_remarks":"This is a practice-based essay appropriate for the Creative AI track, but the central scientific premise—unevenly trained latent spaces—is currently unsupported and may be technically questionable. The author could either provide a concrete mechanism (e.g., per-region gradient statistics, density differences, or intermediate checkpoints) or explicitly reframe the contribution as a manifesto or speculation. The case studies are all from the author's own studio and are not controlled comparisons, which is acceptable for a reflective essay but limits independent validation. There are no citation-pattern concerns; the references are relevant and appropriately used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Tom, this is a five-minute read that leaves you with a useful distinction and one big unanswered question. The paper maps ambiguity, uncertainty and indeterminacy onto phenomenology, epistemology and ontology, then onto Boden's combination, exploration and transformation. That triple mapping is genuinely neat—it gives practitioners a vocabulary for talking about what kind of 'openness' they are invoking, and I don't recall seeing it laid out that cleanly before. The three case studies from Unit 21 illustrate the categories well, especially the Wabi-sabi project for ambiguity and the Probobli Boboli for uncertainty.\n\nThe soft spot is the load-bearing claim in Section 4. The paper says latent space is 'not evenly or uniformly trained' and that the most creative outcomes will be found in 'half-trained and half-determinate' regions. As a metaphor, that's evocative; as a factual claim about how neural networks work, it's unsupported. Gradient descent optimizes shared weights globally; a latent coordinate is not an object that receives more or less training than its neighbors. What actually varies across latent space is data density, posterior mismatch, or conditioning, and the paper doesn't distinguish those from 'degree of training.' So the central future direction rests on a conflation, and without it the recommendation to chase half-trained regions is just a hunch.\n\nI also think the conclusion is partly baked into the definitions. If indeterminacy is defined as testing the limits of ontology, then it's hardly a surprise that it comes out as the most transformative. And all the evidence is from the author's own studio, which is fine for a conceptual essay but doesn't provide any independent check.\n\nThat said, the paper is honest about its scope and is clearly written. The framework itself could be a real contribution to creative AI discourse, even if the half-trained hypothesis doesn't survive contact with the details. The author would do better to present it as a research question with testable predictions—e.g., comparing outputs from models trained to different checkpoints or with different data densities—rather than as a conclusion.\n\nFor the record: I'd bring this to a reading group in the creative AI / computational design community, and I'd cite it for the taxonomy. It deserves a serious referee; a good referee could push the author to sharpen the claim and specify what evidence would count. No formal proofs, no code, no data, so the evidentiary bar is just the quality of the argument, which is decent but not yet solid.","headline":"Useful taxonomy wrapped around an unsupported central hypothesis; the essay is worth reading but needs to reframe its main claim as a research program.","tokens_in":5054,"tokens_out":2669,"would_cite":true,"duration_ms":26514,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Machine learning's most distinctive creative quality is indeterminacy, found in the half-trained spaces of trained models.","keywords":["ambiguity","uncertainty","indeterminacy","latent space","creative AI","machine learning and design","generative adversarial networks","architectural design"],"falsifier":"A direct test would train a generative model on a fixed dataset, measure per-region training quality in the latent space using indicators such as reconstruction error, discriminator confidence, or interpolation smoothness, then generate outputs from the least-trained, half-trained, and fully trained regions and have independent evaluators judge creativity. The paper's claim is falsified if outputs from the best-trained regions are consistently rated as creative as or more creative than those from the half-trained regions, or if the measured training variation across the space is negligible.","tokens_in":4184,"feed_emoji":"🎨","tokens_out":5299,"duration_ms":53389,"temperature":0.7,"pith_summary":"This paper distinguishes three ways creative machine learning can be open: ambiguity, uncertainty, and indeterminacy, and argues that they correspond to different ambitions for AI as a creative partner. Its central proposal is that the most distinctive creative quality of machine learning is indeterminacy: trained models are not uniformly complete, and the latent space is unevenly trained, so the most creative outcomes may appear in regions that are half-trained and half-determinate. The argument is illustrated with architectural design projects described in the paper, including a conditional GAN trained on traditional Japanese tea bowls, probabilistic stone forms, and a crossmodal drawing-music interface. If the paper is right, creative AI research should treat final loss minimization as one goal among others and instead learn to calibrate the degree of indeterminacy in the latent space.","feed_headline":"Half-trained AI spaces may hold the most creative outputs","feed_subtitle":"Indeterminacy, not loss minimization, is the creative edge of machine learning, an architecture-based argument suggests.","key_machinery":"The central object is the latent space, the internal learned representation of a generative model in which each point maps to an output. The key move is to read that space as unevenly trained: an incremental, statistically learned whole with margins of indeterminacy, a concept the paper borrows from the philosophy of technical objects. This reading carries the argument because it turns indeterminacy from a property of an artwork into a property of the model itself that can in principle be measured, controlled, and calibrated against the structure of the training data.","core_discovery":"A neural network's latent space is usually treated as a uniform, fully trained instrument, but the paper claims this is false: because training is incremental and driven by an overall loss, different regions of the latent space are trained to different degrees and with different degrees of determination. Indeterminacy is precisely this incompleteness, and it is what gives machine learning its own creative potential, distinct from the ambiguity of hallucinated or dream-like outputs and the probabilistic uncertainty of transformer sampling. The paper therefore proposes that future creative machine learning should not only minimize loss but deliberately control and calibrate indeterminacy, working in the spaces of the half-trained and half-determinate, and may even construct datasets rather than only curate them in order to exploit this quality.","pith_inferences":["A concrete test follows from the paper's claim: train two generative models on the same data to different numbers of epochs, or with loss unevenly weighted across the latent space, then compare outputs from differently trained regions; the claim predicts that the half-trained regions will not merely be worse but will show a distinct kind of creative value. The paper does not run this test.","The argument implicitly points toward a broader research direction: treating neural network training as a developmental or annealing process in which final performance and creative affordance may pull in different directions, so that training schedules become creative parameters in their own right.","Read operationally, the claim suggests a new evaluation criterion for creative AI: instead of judging only the outputs, one could measure the heterogeneity of training across the latent space, for example through per-region reconstruction error or discriminator confidence, and ask whether that heterogeneity correlates with judged creativity. The paper gestures at this but does not formalize it."],"forward_implications":["Creative machine learning should deliberately search for outcomes in partially trained latent regions instead of treating final loss minimization as the sole objective.","Semi-supervised learning and synthetic data become tools for shaping how much of the latent space is determined, rather than merely ways to improve accuracy.","The three-term distinction gives designers a precise vocabulary: ambiguity aligns with combinatorial creativity, uncertainty with exploratory probability, and indeterminacy with transformative change.","Indeterminacy suggests a transductive potential between modalities, such as drawing and music, where the unevenly trained space itself becomes a site for creative translation."],"supporting_citations":[{"why":"Supplies the VAE formalism whose latent locations are probabilised, one of the model classes the indeterminacy reading targets.","marker":"[20]"},{"why":"Supplies the GAN formalism whose input values and implicit latent locations are probabilised to idealised distributions.","marker":"[10]"},{"why":"Grounds the claim that latent locations in these models are probabilised to match idealised distributions.","marker":"[23]"},{"why":"Supplies the 'margin of indeterminacy' idea that supports the claim about incompletely trained technical objects.","marker":"[29]"},{"why":"Frames the three concepts as the mapped, the navigable, and the uncharted, giving the paper its progressive structure.","marker":"[5]"},{"why":"Supplies the transformer example of calibrated uncertainty through next-token probabilities.","marker":"[27]"},{"why":"Supplies the conditional GAN method used in the tea-bowl design example.","marker":"[19]"},{"why":"Supplies the three-method taxonomy of creativity that the conclusion aligns with ambiguity, uncertainty, and indeterminacy.","marker":"[2]"}],"fun_headline_variants":["Half-trained AI: the creative sweet spot","Indeterminacy as AI's creative fuel","Why under-trained AI can be more creative","Creative AI needs incompleteness, not perfection","In the half-trained lies AI's true creativity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the factual assumption that latent spaces are not uniformly trained but vary in degree of training and determination; if standard training actually produces effectively uniform latent spaces, or if that unevenness has no stable relationship to creative quality, the proposed direction loses its foundation.","fun_headline_variants_meta":{"raw":{"variants":["Half-trained AI: the creative sweet spot","Indeterminacy as AI's creative fuel","Why under-trained AI can be more creative","Creative AI needs incompleteness, not perfection","In the half-trained lies AI's true creativity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1116,"prompt_tokens":734,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":350,"completion_tokens_details":{"reasoning_tokens":313}},"tokens_in":350,"tokens_out":382,"duration_ms":4160,"temperature":1.0,"reasoning_tokens":313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:49:25.776150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would train a generative model on a fixed dataset, measure per-region training quality in the latent space using indicators such as reconstruction error, discriminator confidence, or interpolation smoothness, then generate outputs from the least-trained, half-trained, and fully trained regions and have independent evaluators judge creativity. The paper's claim is falsified if outputs from the best-trained regions are consistently rated as creative as or more creative than those from the half-trained regions, or if the measured training variation across the space is negligible.","supporting_citations":[{"cited_title":"Machine Learners: Archaeology of a Data Practice","cited_arxiv_id":null,"evidence_quote":"Grounds the claim that latent locations in these models are probabilised to match idealised distributions."},{"cited_title":"On the Mode of Existence of Technical Objects","cited_arxiv_id":null,"evidence_quote":"Supplies the 'margin of indeterminacy' idea that supports the claim about incompletely trained technical objects."},{"cited_title":"Ciprut, editor.Indeterminacy: The Mapped, the Navigable, and the Uncharted","cited_arxiv_id":null,"evidence_quote":"Frames the three concepts as the mapped, the navigable, and the uncharted, giving the paper its progressive structure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the three-method taxonomy of creativity that the conclusion aligns with ambiguity, uncertainty, and indeterminacy."}],"review_version":1}