Seventy people and one missing bridge

The machine-readable version is the graph at /people/talk-to — 61 organisations, 70 named people, 43 relationships. This page is the argument the graph cannot make: who to talk to first, in what order, and what the shape of the network says about the business.

The clusters are unbalanced because the field is: research 20, creative 13, competitors 12, labs 7, community 7, tools 6, investors 5. One fact falls out before any name does. The research cluster is the most enumerable part of this market and the labs cluster is the least. Everyone who has built an aesthetic reward model in four years can be named from paper bylines; almost nobody who buys one can. That sets the outreach sequence — the people who will tell you what a taste model has to beat are trivially reachable, and the people who will pay for it are not.

#The six conversations to have first

#Souki Mansoor — Lead, Global Sora Artist Program, OpenAI

The only person identified anywhere who has already done the client's hardest job at a frontier lab: persuading sceptical professional creatives to work with a model company, at scale, with a real budget. Sora Selects funded ten artists and teams out of a roughly $3M initiative built over about ten months of relationship work (Artnet).

The detail that makes her worth a first meeting rather than a fifth: the programme did not buy silence. Boris Eldagsen took the money and still said publicly that copyright had been "outsourced to the users." Objection converted from a boycott into a negotiation is exactly what the Sora revolt evidence says is achievable.

Ask how artists were sourced and compensated, whether any evaluation signal fed back into training — the swing factor on market size — and whether OpenAI would rather buy that relationship than keep building it. She is publicly quoted and speaks at art and film venues. She is also the client's single most useful hire target, better acknowledged before the meeting than after.

#Sarah Catanzaro — General Partner, Amplify Partners

She co-led Taste Labs' $18.5M seed and wrote the public thesis for the whole category (Amplify). By her own portfolio description she is also an investor in Runway (Amplify) — one of only two investor-to-lab bridges in the graph, and the best-informed outside observer of this market.

The conversation is a diligence conversation run in reverse: what she saw and rejected before backing Taste Labs, how she thinks an expert panel defends against the arena layer, and what a differentiated second entrant would have to look like. She has already done the market map; comparing notes with hers is the fastest way to find the holes in this one.

#Aspen Hopkins — Researcher, MIT CSAIL

Lead author of the leading competitor's flagship paper (arXiv 2606.30561) and an MIT CSAIL researcher (CSAIL). The academic credibility of Contra's central artefact rests on her institutional affiliation rather than on in-house staff — and her employment status at Contra could not be resolved after two research passes.

If she is a collaborator rather than staff she is available, and an MIT co-publication is the cheapest credibility this company can buy. Ask what the disagreement structure in designer preference data actually looked like — the empirical core of the "taste is not averageable" argument both competitors make rhetorically and neither has fully evidenced.

#Purvanshi Mehta — ex-Lica World, now leading design research at Gamma

Corresponding author of the TASTE paper (arXiv 2605.20731) and the person who supplied Contra Labs' reward-modelling capability, before Gamma acquired Lica on 25 August 2026 (TechCrunch).

For a short window she is uniquely available: she knows what signal in designer ratings turned out to be learnable, where preference models fail on design, and what lab demand looked like from the inside — and having moved into a product company, she has stopped being a competitor.

Ask why Lica took an acqui-hire rather than a lab contract. The answer is the most important unresolved risk in the client's business model — a data point on whether this capability is a business or a feature. Her co-founder Priyaa Kalyanaraman is the same conversation from the go-to-market side.

#Danny Wu — Head of AI Products, Canva

Canva is the best-qualified single buyer in the map. Across 1,562 app-layer postings surveyed, exactly one dedicated AI-quality-evaluation contractor role exists anywhere in the app layer, and it is Canva's (AI Quality Evaluator); it also runs Research Lead and Research Engineer roles for Evaluations. Wu owns AI product quality across a very large user base and is a genuinely public figure.

The question is diagnostic rather than commercial: are those roles the start of a permanent internal panel, or a specification of work Canva would rather buy? The answer tells you whether the app layer is a customer or a competitor for the same talent.

#Elizabeth Goodspeed — Designer and writer; US editor-at-large, It's Nice That

She has written the definitive practitioner essays on taste and on why AI cannot supply it, in the publication working designers read (It's Nice That). No lab-side traction fixes a supply side that has decided the company is extractive, and her read will largely set how the profession receives it.

Put the uncomfortable question directly: what would make selling aesthetic judgement to AI labs read as dignified professional work, and what would make it read as selling out? If she cannot articulate a version that clears the bar, the positioning is wrong — much cheaper to learn in week four than after a launch.

#The introduction paths the graph actually reveals

Sixteen edges cross clusters, and those are the only ones that move a conversation from one part of the market into another. Four are worth acting on.

Gamma is the seam in the competitive set. Grant Lee acquired Mehta and Kalyanaraman, and Mehta co-authored TASTE with Contra's Alexandria Minetti — so Gamma is simultaneously a potential customer, the new owner of the category's best modelling talent, and the holder of Contra's institutional memory. The highest-information node in the graph, reachable through ordinary press channels.

Scott Belsky is the highest-leverage single person in the map. He founded Behance, ran strategy at Adobe through the Firefly contributor-compensation decisions, and is now a partner at A24 (Variety), which has a reported AI filmmaking venture with Google DeepMind. One relationship touches the largest creative portfolio network, the only company that has paid creative contributors for training data at scale, and a live studio-lab partnership — and he is the plausible route into the DeepMind gap below.

CMU is both the audit and the go-to-market blueprint. GenAI-Bench's 1,600 prompts came from professional designers and were then adopted by Google DeepMind as Imagen 3's prompt set (arXiv 2408.07009). That is the clearest existing proof that a frontier lab will adopt a designer-authored artefact — and it happened through an academic benchmark, not a sales process.

D&AD is a pre-assembled credential cluster. Six of the thirteen creative-cluster people are 2026 jury presidents, connected by the jury itself (D&AD). Teo Connor is the node to enter through — VP of Design at Airbnb and jury president for Digital Experience Design, carrying credibility on the buyer and practitioner sides at once. One relationship reaches a vetted roster no paid advertising replicates, which is the register argument on the recruiting side.

#Who is over-represented, who is missing, and who will be hostile

Research is over-represented at 20 of 70, and that is honest rather than lazy. The literature is small, fully enumerable, and every author will take an academic email. The point buried in it: the product has to beat named, published baselines built by people who will answer the phone. A pitch that does not benchmark against VisionReward and PickScore is dismissed in one question.

Labs at 7 and tools at 6 are under-represented, and that is the finding. Across 1,235 model-builder postings only four are dedicated human-data or evaluation procurement roles, and not one posting in the sector names an external data vendor. Procurement at frontier labs is not a public function; the verifiable names run creative-facing functions, which are public by design. You cannot map the buy side by scraping — the routes in are creative programmes, published research agendas and investors.

Hostility comes in three kinds.

The artist-rights opposition — Karla Ortiz, a named plaintiff in Andersen v. Stability AI, and Jingna Zhang, who built Cara to over a million artists explicitly to keep AI out and took no venture money so no investor could push it the other way. Neither is reachable as a customer or a supplier; both will be asked for comment on launch day. The distinction the client must defend in one sentence is between selling artwork for training and selling judgement about outputs. It survives contact with practitioners; it does not survive contact with a bad contract. If the terms let a lab retain the underlying imagery, the distinction collapses and Ortiz's framing wins.

The methodological critics — Jordan Taylor, William Agnew and Maarten Sap, whose FAccT '26 audit showed the LAION-Aesthetics predictor filters in captions mentioning women while filtering out captions mentioning men and LGBTQ+ people, and was trained on scores from "English-speaking photographers and western AI-enthusiasts" (arXiv 2601.09896). Read that paper as the client's own future audit. A curated panel of 30 to 1,000 designers has the same failure mode at smaller n with a commercial motive attached. Publishing panel demographics, pay rates and disagreement statistics before anyone asks turns the most dangerous critics into the most useful reviewers.

The credibility gatekeepers who are not hostile yet — Goodspeed, Jenny Brewer at It's Nice That, Katy Cowan at Creative Boom, Debbie Millman. None opposes AI in principle; each can make the client radioactive with the supply side in one article. They are the reason to get terms right before getting distribution. Ed Newton-Rex sits on both sides of that line: he resigned from Stability over training-data ethics and now certifies through Fairly Trained. He is the most likely person to dismantle the pitch publicly and the most valuable person to have certify it. Talk to him early, expect a no, and treat his criteria as the spec for the supply terms.

#The structural finding

The isolation has a practical meaning: these nodes have no edges because they are purchasable or bookable rather than relationship-mediated. Felix Lee's ADPList mentor profile is publicly bookable, Sidebar sells sponsorship at a published rate, Dribbble routes through an account manager. Work them in parallel with the core conversations, not after.

#Method, and what it cannot tell you

Every person is sourced to at least one public page — a paper byline, a company site, a jury announcement, a press interview, a public profile — and where a role is uncertain the record says so. Where a person could not be verified they are absent, and the absence is reported rather than filled with a plausible name. No personal contact details appear anywhere; the reach field names public channels only.

Three limits. Google DeepMind has zero named people despite publishing the industry's richest human-evaluation methodology — 366,569 ratings from 3,225 raters across 71 nationalities in the Imagen 3 report, with nobody attachable to a human-data or creative-partnerships remit. Given the reported A24 venture, that is the highest-value addition to make next. Taste Labs' site shows eleven photographs captioned with first names only, and exactly one was resolvable — Hamidah Oderinwale, Member of Technical Staff, from research post bylines; the founding designer and engineers remain unnamed, and that opacity is deliberate. And no individual partner could be named for CRV's lead on Taste Labs, Index's lead on Design Arena, or AfterQuery's Series A.

The sequencing: weeks 1–2 for the research cluster, which costs nothing but email; weeks 2–4 for the Gamma seam and Catanzaro while the acquisition is fresh; weeks 3–6 for the supply terms with Newton-Rex, Goodspeed and one jury president, before any recruiting page goes live; weeks 4–8 for the buyers, by which point there is a paper and a defensible supply story. One warning: Oderinwale is reachable through an open inbound channel and is also the technical mind of the nearest competitor — that conversation is legitimate, valuable, and visible to them. Several of the open questions close through these conversations rather than through further searching.