Miju Labs

← All buyers · OpenAI

Prafulla Dhariwal

Technical Fellow / Multimodal Lead, OpenAI · OpenAI

medium confidenceTier A · Talk to firstFrontier labsimage · video · audio · multimodal

OpenAI's multimodal lead and a Technical Fellow — the research executive over native image, audio and video generation. He is credited as Multimodal Lead on ChatGPT Images (GPT-Image-1.5, December 2025) and founded the 'omni' team that produced GPT-4o. Everything OpenAI ships that has to look good passes through his organisation, which makes him the highest-leverage research-side sponsor for visual-taste data at the company.

Background

Dhariwal joined OpenAI in 2016 out of MIT, having come from Pune with IMO and IPhO medals. His research line runs straight through generative modelling: Glow (invertible flows), Jukebox (music generation), 'Diffusion Models Beat GANs on Image Synthesis' with Alex Nichol, and the guided-diffusion work that underpinned DALL·E 2. He then built the omni team, whose first model was GPT-4o — OpenAI's first natively fully multimodal model — and publicly credited the team members who made it work.

His trajectory is notable because he moved from pure generative-model research into owning a product-facing modality stack. Native 4o image generation (March 2025) and its successors made image quality a mass-market judgement rather than a FID score, and the credits on the December 2025 ChatGPT Images release put him at the top of a stack that includes a research lead (Gabriel Goh), a product lead (Adele Li) and a named Data & Evaluation group.

What they run now

The multimodal stack: native image generation in ChatGPT (GPT-Image-1.5 and the ChatGPT Images 2.0 line), audio/voice, and the interface with video and world-simulation work. His organisation carries a distinct Data & Evaluation function for images — a dozen-plus named people on the GPT-Image-1.5 credits — which is the operational unit a vendor would actually contract with.

What they buy

Strong inference from the credits structure. OpenAI's image releases list a standing 'Data & Evaluation' team of roughly a dozen people plus named human-data moderators and a 'Human Data Advisor' (Long Ouyang). That shape — a permanent internal data/eval function under a modality lead — is what buys external preference data, rubric-graded eval sets and expert annotation at scale. For images the likely objects are aesthetic preference pairs, instruction-following and editing-fidelity judgements, text-rendering correctness, likeness and composition grading, and increasingly professional-workflow trajectories (brief → concept → production asset). No vendor is named publicly.

Career

  1. 2016 – presentResearch scientist → Multimodal Lead → Technical Fellow, OpenAI — Founded the 'omni' team behind GPT-4o
  2. 2013-2017Massachusetts Institute of Technology, BS, computer science / mathematics

On the record

GPT-4o and the omni team · 2024-05

GPT-4o was the first model out of the omni team, OpenAI's first natively fully multimodal model, built as an org-wide effort. source

Interviews & talks

Why it matters here

High as the budget-holding altitude, low as a first contact. He is a Technical Fellow with a large org; he will not take a vendor meeting cold. His value to Julian is as the name that justifies going after his org: the entry points are the Data & Evaluation names on the image credits and the product/research leads (Adele Li, Gabriel Goh). Knowing that the function exists, is named, and is stable across releases is the actionable fact.

How to reach

Do not approach directly. Use the credits page as a map: the Data & Evaluation block is the buying unit. Speculation: OpenAI's image org is under sustained competitive pressure from Nano Banana, and the pitch that lands is 'professional-workflow judgement you cannot generate internally', not 'more raters'.

What we could not establish
  • Whether 'Technical Fellow' has replaced or sits alongside 'Multimodal Lead' — the LinkedIn headline says Technical Fellow, the Dec 2025 credits say Multimodal Lead.
  • Whether image data procurement runs through his org or through a central OpenAI human-data function.
  • No public statements from him on aesthetics or human evaluation were found.

Why they are on the map

'Multimodal Lead' in ChatGPT Images credits (Dec 2025); 'Multimodal Organization' lead for 4o image gen source — evidence dated 2025-12

Decides whether the problem is funded.

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. The new ChatGPT Images is here — credits
  2. Introducing 4o Image Generation
  3. Prafulla Dhariwal on X (omni team / GPT-4o)
  4. Prafulla Dhariwal, LinkedIn