Miju LabsOn 16 August 2026, Taste Labs published Requests for Research — a 14-minute read by Hamidah Oderinwale, Member of Technical Staff (tastelabs.com/blog/requests-for-research) — structured as five thematic sections containing eleven discrete open problems. It went out the same day as the Prototype fellowship announcement.
Its framing, in their words: "Research(ed) questions we want to answer on human-AI collaboration, machine creativity, and the design engineer stack… Taste has become a trite term. At Taste, we are working to make that tractable. How do you take something so subjective like taste and make it an engineering problem, and is it even worth doing? We think it is, and that the floor is worth raising."
The eleven, as published
1 — The geometry of the visual web. "While models of the world represent physical systems and their dynamics in terms of coordinates and free-body diagrams, we lack unifying abstractions for the world of the web… A number of our projects use Playwright, which allows sites to be captured and interacted with programmatically, but we have found synthesizing function, form, and motion difficult. We're interested in what it takes to build high-fidelity environments of the digital world for models."
2 — Universal multimodal geometry. Whether the Platonic Representation Hypothesis holds across modalities: "is there a universal geometry of multimodal embeddings?… We ask how models really 'think' about multimodal representations and whether any generalize across architectures and when." They report a negative from their own work: "a model could only reason about the viewpoint of a figure in natural language but struggled with a DSL and images alone."
3 — Procedural fingerprinting. "whether agents observed different problem-solving habits in the same ways that humans do and if they could, were they identifiable from their traces alone?… Given that visual artifacts from the same designer will have its unique signatures and the designer's process is its foundation. We wonder how agent processes and outputs relate."
4 — Evaluating spatial awareness in use. "A button means something different in Figma, React, and HTML: there is no shared format between them… what are the right representations for visual and interactive code that can be understood across interfaces?" And the sharpest disclosure in the document: "A verifiable layout target lets a construction loop converge, while a contested aesthetic target does not: in our patent render-loop, 0/24 converged. Pinning down which goals are verifiable, which are contested, and how to measure progress against the latter is an open problem."
5 — Centaur evaluation. "more benchmarks should explore the extent to which models can serve as thought partners, not just executors… How do we model and audit the design workflow end to end? What metrics of efficiency and 'thoughtfulness' should future evaluations depend on—for example, minimizing the number of iterations required to reach a satisfactory outcome? Where ground truth does not exist, how do you approximate it with human preference?"
6 — Design intent inference from edit sequences. "how do we infer user intent and continually learn from edit sequences and design choices over time? Historically, search engines have optimized for click-through rates; we should consider what analogous signals look like in AI-assisted design environments, such as follow-up edits, revision depth, or engagement with specific components."
7 — Reward hacking in open-ended design agents. "we've explored agents for reconstructing sub-pages of websites in adherence to a given brand… Notably, we've observed symptoms of reward hacking; however, because these behaviors manifest in heterogeneous and visually subtle ways, they are difficult to identify and trace at scale. We see a strong need for tools that can evaluate agents at scale while still permitting them to be open-ended and expressive."
8 — Measuring creativity as process, not output. "what does it mean to evaluate the creative process and its outputs?" Their illustration is the "forced connections" exercise given to design students — a deliberately arbitrary brief, say Comic Sans with an image of an apple — where "a design teacher writes a rubric to assess students' varied interpretations; the same is required for machines, if not more so. Grading these works is the instrumental challenge that designing a benchmark forces us to confront."
9 — Versioning beyond code diffs. "Designers think in states of their work, not line diffs. What number of changes make a new meaningful version of an artifact?… versioning thresholds may be better defined not by operations or visual edits, but by detected shifts in what the designer is trying to accomplish."
10 — Provenance, attribution and auditing. "We need techniques for detecting, attributing, and auditing design outputs in open-ended generative settings. We view this as an open problem of evaluating design agents while still allowing them to explore the limits of their creativity."
11 — Aesthetic- and context-aware search ranking. "how might new search result ranking systems account for aesthetic and contextual quality, not just visual similarity, with context (e.g., how you judge a minimalist perfume store site is different from how you judge a college course selection site)?"
The acknowledgements name Jason Wu and Yi-Hao Peng; the appended reading list includes Yi-Hao Peng et al., "DesignPref", Brynjolfsson et al. on centaur evaluations, Omar Shaikh et al. on general user models from computer use, and André Cruz et al., "Text as the Richest Preference Signal."
Which ones the client could actually answer
| # | Problem | Buildable here? | Why |
|---|---|---|---|
| 6 | Design intent from edit sequences | Yes | Needs captured design work with narrated intent — the one asset class this plan already builds |
| 7 | Reward hacking in open-ended design agents | Yes | Needs a human panel that can detect subtle brand-adherence failure; they admit they cannot trace it at scale |
| 8 | Creativity as process, not output | Yes | It is a rubric-and-panel problem, which is the core competence |
| 5 | Centaur evaluation | Partly | Needs an instrumented product surface, not just a panel |
| 9 | Versioning beyond code diffs | Partly | Falls out of trajectory capture as a by-product, not a study |
| 10 | Provenance and auditing | Partly | The schema half is buildable now (Sell the paperwork with the data); the detection half is not |
| 4 | Which goals are verifiable | Partly | The taxonomy is buildable; the environment work is not (What this plan could not establish) |
| 11 | Aesthetic search ranking | No | A search company's problem |
| 1, 2, 3 | Web geometry, multimodal embeddings, fingerprinting | No | ML-research problems needing a research team the client does not have |
The three to take seriously are 6, 7 and 8.
Design intent inference from edit sequences (6) is the trajectory asset restated as a research question. It needs exactly what a narrated-capture programme produces — ordered edits with stated intent — and it is the problem Taste Labs has publicly committed to and not yet shipped.
Reward hacking in open-ended design agents (7) is the most interesting of the three, because they have already conceded the hard half: they have observed it and cannot trace it at scale, precisely because the failures are "heterogeneous and visually subtle." That is a sentence describing a job for a calibrated professional panel. A taxonomy of design-agent reward hacks — brand-token mimicry that satisfies a checker while missing the system, layout that games a screenshot metric, contrast that passes an accessibility rule and fails a reader — built from panel adjudication on real agent outputs, is publishable, cheap, and directly answers a funded rival's stated blocker.
Creativity as process rather than output (8) is a rubric problem, and rubric authoring is the one piece of IP a data company owns outright rather than licensing from a contributor (Eight units and one cost anchor). Their own "forced connections" framing hands over the study design: an arbitrary constraint set, varied interpretations, a rubric that grades the interpretation rather than the artefact.
Why answering a rival's published problem is the cheapest citation available
Everything in What publishing actually bought them says the hard part of publishing in this market is not producing the artefact — it is getting anyone to read it. Contra published 39 studies and holds zero citations. An answer to a named open problem arrives with its audience already assembled: the people who wrote the question, the people who read it when it was posted, and the fellowship applicants who will cite it as background. It is also the one publication format where being second is an advantage, because the problem statement did the framing work for you.
The etiquette is cheap: cite the RFR, name Oderinwale, and send it to them before it goes up. The worst outcome is that a competitor learns what you are building, which they already announced they are building.
Read it two ways at once. It is a hiring and fellowship instrument, published on the same day as Prototype, and it works because it is genuinely good — the honesty of "0/24 converged" is what makes a researcher want to work there. And it is a public statement that Taste Labs has solved none of these, including two that are Contra's live products: creativity evaluation and design trajectories. A competitor who ships one of the eleven with real numbers publishes directly into a frontier a funded rival has conceded in writing.
What Prototype signals
The fellowship offers "grants, access to compute, access to custom human data, time with our team and opportunities for collaborative publishing." Note the third item. A company with no dataset, no benchmark and no code is offering custom human data as a fellowship benefit — which means the human-data operation exists and is being spent on relationship-building with researchers rather than on artefacts.
That is a coherent strategy and it is the opposite of Contra's. Contra spent its research budget on 39 owned-media studies and got zero citations. Taste Labs is spending its on other people's papers, where the citation accrues to Taste Labs as a data provider and the researcher does the writing. It is cheaper, slower, and considerably harder to compete with by out-publishing. The full read is at Taste Labs, in full; the artefact record that shows they have shipped nothing is at Every public artefact, by organisation; what to build against it is at Rebuild the Human Creativity Benchmark and What nobody has built.