Miju Labs

← All buyers · OpenAI

T

Tejal Patwardhan

Researcher, OpenAI — lead author, GDPval; Preparedness evaluations · OpenAI

medium confidenceTier A · Talk to firstFrontier labstext · multimodal

First author of GDPval, OpenAI's benchmark of model performance on real-world economically valuable work: 44 occupations across the nine largest sectors of US GDP, with tasks built from the actual deliverables of professionals averaging fourteen years of experience and graded against expert standards. She is the person at OpenAI who has industrialised 'pay experienced practitioners to produce and grade real work', which is structurally the same purchase as creative-taste data.

Background

Patwardhan works on OpenAI's Preparedness and evaluations line and has been an author on a run of expert-graded benchmarks — the pattern across her work is recruiting genuine domain experts rather than crowdworkers and having them both author and grade tasks. GDPval (arXiv 2510.04374, October 2025; accepted at ICLR 2026, senior author Jerry Tworek) is the fullest expression: tasks grounded in the Bureau of Labor Statistics Work Activities framework, deliverables that look like real outputs rather than exam questions, and a public automated grading service at evals.openai.com to scale the comparison.

The headline finding — frontier models approaching industry experts in deliverable quality — matters less to Julian than the machinery. GDPval is a procurement template: identify occupations, recruit practitioners with a decade-plus of experience, extract representative deliverables, build a grading protocol, then automate. Creative occupations sit inside the GDP sectors GDPval covers.

What they run now

Evaluations that measure real economic capability rather than exam performance: GDPval and its automated grader, plus Preparedness-track evaluations. The operational core is expert recruitment and grading protocol design — who counts as an expert, what a representative deliverable is, how two experts are made to agree.

What they buy

Direct and evidenced in kind, if not by vendor name. GDPval required sourcing experienced practitioners across 44 occupations and collecting both deliverables and expert grades — that is a large human-data procurement, whether run in-house or through vendors. For Julian the specific opening is occupational coverage: graphic designers, art directors, photographers, video editors, UI/UX designers and marketing creatives are all BLS occupations inside GDPval's sectors, and creative deliverables are the hardest category to grade automatically. A vendor that can supply senior creative practitioners plus a defensible grading protocol is selling exactly what GDPval consumes.

Career

  1. presentResearcher — Preparedness / evaluations, OpenAI — Lead author GDPval; co-author on expert-graded benchmark work
  2. —Harvard University, AB

On the record

What to measure · 2025-10

Tasks should be constructed from the representative work of industry professionals with, on average, fourteen years of experience. source

Model capability · 2025-10

Frontier models are approaching industry experts in the quality of real deliverables. source

Interviews & talks

Why it matters here

Very high, and underrated. She is the cleanest 'evals' fit in the batch: a named person with a published methodology that requires exactly Julian's supply, and a public artefact to anchor the conversation. Unlike the image-model people she is not asking for aesthetic preference pairs — she is asking for real creative deliverables and expert grades, which is a higher-margin and more defensible product.

How to reach

Lead with a gap in her own benchmark. Speculation but concrete: pick two or three creative occupations in GDPval's sectors, build a small set of representative deliverables with senior practitioners, grade them with a second independent panel, publish the inter-grader agreement and the automated grader's error rate on that subset, and send it as a contribution rather than a pitch. Cite the fourteen-years-experience criterion and say how Julian's panel meets it.

What we could not establish
  • No verified X handle or LinkedIn found.
  • Her exact title at OpenAI is not evidenced; 'Researcher' is from the paper byline.
  • How GDPval's expert panel was sourced — in-house, vendor, or marketplace — is not public.
  • Whether creative occupations are already covered in GDPval's 44 and with what results.

Why they are on the map

First author & correspondent on GDPval, expert-graded eval across 44 occupations built with industry experts and vendor partners (Oct 2025) source — evidence dated 2025-10

Needs the data for a model; writes the spec.

How OpenAI buys data

Has a dedicated Human Data team that 'turns human feedback into reliable signals for training and evaluation' via bespoke campaigns with 'external vendors and AI trainers' and has open PM/TPM/EM/TLM roles on it (2025-26 job posts). Ended its Scale AI relationship after Meta's stake (Jun 2025); reported to be a major revenue source for Mercor and to use Surge and Handshake, while building in-house data capacity. GDPval (Oct 2025) and HealthBench (May 2025, 262 physicians) show it buys large expert-graded eval sets. Image models ship with a named 'Data & Evaluation' team (Dec 2025 credits).

A central Human Data team (PMs, TPMs, engineers) works with all research teams and vendors. Glaese, a former Human Data lead, is now VP Research & Safety (Jul 2026). Image generation sits under Multimodal (Dhariwal), with its own data leads and an 18-person Data & Evaluation group, plus Model Behavior. Sora was shut down in Mar 2026 and Peebles left in Apr 2026. Image work was folded into GPT Image 2 / 2.5.

Known vendors: mercor, surge-ai, handshake-ai, scale-ai (ended Jun 2025)

Sources

  1. GDPval (arXiv)
  2. GDPval paper PDF
  3. GDPval at ICLR 2026
  4. Tejal Patwardhan, arXiv author listing