Miju LabsMost competitive maps of this lane rank by funding and put Contra Labs first because it shipped first. That ordering is wrong. Contra has the artefacts — 39 studies, eight Hugging Face datasets, one arXiv paper — and Taste Labs has none of them. Taste Labs is still the more dangerous company, on three structural grounds, none of which is the $18.5M.
The seed was $18.5M, announced 16 June 2026, co-led by CRV and Amplify Partners (Amplify; Dealroom; CRV), with Latitud claiming "the very first check months ago, at their inception stage" (Latitud). One trade outlet dates the launch to June 2025; Amplify, Dealroom, Axios and the founder all say 2026 (Axios, 29 Jun 2026). Treat the 2025 date as a typo.
Ground one: the founder has already sold to this buyer
Thais Castello Branco led growth at Exa before founding the company. Amplify describes her as "known for unusually thoughtful and distinctive marketing"; Latitud is the more useful source, describing Exa as "a Benchmark-backed AI startup serving foundation labs," where she watched models fail at subjective quality first-hand (Amplify; Latitud).
That single line is the largest structural difference in the field. Contra's founders are marketplace operators who built a freelance product for six years and repointed it. Castello Branco is a go-to-market operator from a company that already sold into foundation labs. The hardest thing about this business is not assembling designers; the supply is cheap and abundant. It is getting a research manager at a lab to take a first meeting and then convert a pilot into a campaign. She has done a version of that job before, with the same buyer population.
Ground two: they are hiring people who train models, and Contra is not
The open roles on the company's own site are Applied ML Engineer, Design Research Engineer, AI Engineer (Full Stack), AI Engineer (Back-end), Growth Lead, Senior Designer, Strategic Product Operations Lead, and Operations Manager–Community (tastelabs.com). The careers page says they want people working in "post-training, architecture, search, product building, evals, data, design and more, to crack subjective domains" (tastelabs.com/careers).
| Taste Labs | Contra Labs | |
|---|---|---|
| Open roles | 8 | 5 |
| ML / applied-research engineering | 4 of 8 | 0 of 5 |
| Delivery / research-ops | 3 | 5 |
| Modelling capability | Being built in-house | Partnered in from Lica World, now acquired by Gamma |
Contra's five postings are all delivery or research-operations at a flat $150–200K, with zero ML engineers (Ashby) — the point its dossier makes about the org chart producing a consultancy. The asymmetry matters because of where this category ends. If the endgame is "we sell you a dataset", the buyer builds it themselves the second time. If the endgame is "we sell you a taste reward model, and we maintain it", that is a product with renewal. Taste Labs is staffing for the second outcome.
Contra rented its reward-model capability and the landlord has been bought. Taste Labs is hiring its own. Twelve months from now that is the difference between a research-ops agency and a product company.
Ground three: trust propagation, which is the best idea anyone in this field has published
The supply network is TasteMakers (tastemakers.tastelabs.com redirects to portal.tastelabs.com). Amplify describes the recruiting mechanism as a "trust-propagation model" in which vetted evaluators nominate others, corroborated in a separate founder interview (Amplify; For Creative Girls).
This solves the single hardest problem in expert data — how do you know the expert is actually good — without an ops army. Portfolio review at scale requires senior reviewers, and senior reviewers are the scarce input. A nomination graph moves the screening cost onto people who already passed, and it adds something no funnel has: social cost to nominating badly. It also produces a roster where every node has a legible provenance chain, which is exactly what a lab's procurement team asks for when it asks who rated this. See Telling a good designer from a confident one for why that question decides deals.
The portal's own copy is consistent with it: "a curated collective of designers who share a high bar for quality, not a volume marketplace"; three steps — apply, take a short test, get placed; and failing the test is not exit, since applicants join the community "to be matched when the right project launches" (portal.tastelabs.com). Terms offered are "rates that reflect the real worth of your skill", "no middleman markup", no minimum hours, no exclusivity, plus a private WhatsApp group.
The network number is small, and that is the honest version
1,000+ designers across 30 countries (portal.tastelabs.com), covering graphics, product, illustration, 3D renders, art concepts, games and presentations.
Set that against Contra's "1.5M+ verified creative experts" in one place on its homepage and "1.7M+" in another (contralabs.com) — a count of registered profiles on the parent marketplace, against studies that used 28–31 evaluators and 10 designers respectively. Taste Labs' 1,000 is plausibly the actual vetted roster. A buyer who is technically literate will prefer a credible small number to an incredible large one, because the large one advertises how cheap the studies are to replicate. In Design and UI/UX the scale claim is the weakest part of the incumbent's pitch and Taste Labs has declined to make it.
They have announced, in public, that they will build Contra's only moat
This is the finding that should change the client's plan.
Taste Labs' Requests for Research page lists nine open problems (tastelabs.com/blog/requests-for-research). Three of them are not research questions, they are a product roadmap:
- "design intent inference from edit sequences"
- "design history/versioning beyond code diffs" — design traces, not commits
- agent process fingerprinting — "do AI agents exhibit identifiable problem-solving patterns similar to human designers"
- creativity evaluation of process, not outputs
That is trajectory capture, named. Contra's genuinely defensible asset is not its preference pairs — those are a commodity any lab makes with thirty contractors and a fortnight — it is the screen-recorded, narrated sessions: 234 steps across four Premiere trajectories, 111–245 minutes each, with a thought field taken from the editor's spoken narration rather than model-synthesised rationale (HF card). See The trajectory moat and Eight units and one cost anchor for why that is the only line item here with real gross margin.
The best-funded competitor has published a research agenda that describes Contra's moat, four months before shipping against it, with four ML engineers hired to do it. Any plan whose defence is "trajectories are hard to copy" now has a named, funded, forewarned copier.
The same page describes Taste Labs as "a data partner for frontier labs" and mentions embedding into design tools like Flora and Krea — which is a distribution strategy Contra does not have.
What they will not tell you, and it is the two things that matter
Taste Labs publishes no pay rates and no IP terms. The portal promises "rates that reflect the real worth of your skill" and "no middleman markup" and stops there (portal.tastelabs.com).
Not publishing rates is tactically smart — it avoids anchoring below Mercor's $150–250/hr and avoids a public comparison. But those are precisely the two dimensions on which a working designer decides whether to sign, and Contra has answered both in writing: up to $100/hr, licence not assignment, byline credit, no platform fee, seven-day payment, right to decline (contralabs.com/jobs/photographers). The refusal research says the binding objection for a large minority of creatives is not the rate at all, it is who profits and on what terms — so silence on IP is a real recruiting weakness against a competitor that has already published its terms sheet. See The clause that expires your corpus in year ten.
The customer list that is not a customer list
Amplify's announcement names teams at Runway, Figma, Character AI and Adobe as facing the slop problem. A trade write-up re-reports these as "specific customers mentioned" (The Source Code).
That is a misreading of the investor post. Amplify lists those companies as examples of the problem, not as Taste Labs accounts. Taste Labs has no publicly confirmed customer. Anyone briefing the client on "Taste Labs already works with Adobe" is repeating a second-hand error, and it will be embarrassing in a room. [WEAK] on the correction only in the sense that a private contract could exist and be unannounced.
The rest of the thesis is strong, though. Their manifesto argues that pretraining is "next-token prediction via maximum likelihood" producing statistical averages, that "greatness in subjective domains lives in the tails of the distribution, while the training objective pulls toward the center of it", and that RLHF makes this worse because "the reward model averages over many annotators… learning the intersection of preferences: what nobody dislikes, rather than what anyone loves" (tastelabs.com/blog/a-manifesto-for-judgement). That is Contra's disagreement-is-signal claim with a sharper commercial implication — don't average, select — and it bears directly on The oracle problem. Expect it to be the standard category pitch by Q4.
Two honest weaknesses remain. There is no published dataset, benchmark, paper or shipped artefact — the blog is three essays and a research agenda, with "no empirical studies, user testing, or quantitative analysis presented" (tastelabs.com/blog/stop-debating-taste). And design is explicitly a beachhead: the phrasing is always "starting with design", with the company positioned as the layer for subjective domains generally.
Taste Labs' HQ (New York in one trade outlet, San Francisco in every one of its own job postings), its founding date and legal entity, its TasteMaker pay rates, its IP terms, and any customer. Four of the five gaps are deliberate.
The read for the verdict page: a client entering this lane is not first, is not best-funded, and is not the only one who has worked out where the moat is. What is still unoccupied is the pairing Taste Labs has left open — published rates and published IP terms, on top of a nomination graph. Both halves are free to copy.