Miju Labs

← All buyers · OpenAI

Gabriel Goh

Research Lead, ChatGPT Images · OpenAI

medium confidenceTier A · Talk to firstFrontier labsimage

Named Research Lead on ChatGPT Images (GPT-Image-1.5, December 2025) and a long-standing contributor to OpenAI's image line including DALL·E 3 and the multimodal-neurons interpretability work on CLIP. He is the person who decides what the image model is optimised toward, which makes him the technical counterpart to any aesthetic preference dataset.

Background

Goh came to OpenAI from an optimisation-theory background (PhD at UC Davis) and made his early public mark on interpretability — the Distill work on multimodal neurons in CLIP, which showed single units responding to a concept across photographs, drawings and text. He moved into the image-generation line with DALL·E 3 and now leads research on the ChatGPT Images models.

The technical problem he owns has shifted from 'make a plausible image' to 'follow a detailed instruction while preserving likeness, lighting and composition, and render dense small text correctly'. The December 2025 release notes emphasise exactly those axes plus 4x speed. Those are measurable; aesthetic coherence and originality are not, and that is where human judgement enters his loop.

What they run now

Research direction for the ChatGPT Images model family: instruction following, editing fidelity and identity preservation, text rendering, complex composition and multi-face scenes, and inference speed. He works alongside a product lead (Adele Li) and a standing Data & Evaluation team.

What they buy

Inference. A research lead on an image model buys or specifies: human preference comparisons between model variants, fine-grained editing-fidelity judgements (did the edit preserve what it should have), text-rendering correctness at small sizes, and aesthetic ratings that can be distilled into a reward model. The OpenAI credits show a dedicated Data & Evaluation group in his release, so the demand is institutionalised rather than ad hoc.

Career

  1. presentResearch scientist → Research Lead, ChatGPT Images, OpenAI — DALL·E 3, CLIP multimodal neurons, GPT-Image-1.5
  2. —University of California, Davis, PhD (optimisation / computer science)

Interviews & talks

Why it matters here

Medium-high. He is a genuine technical requester for image preference data and is reachable as an individual researcher rather than an executive. But he is optimisation-minded: he will want agreement statistics and a demonstration that the labels move a metric, not a portfolio of pretty examples.

How to reach

Research-to-research. Speculation: a short technical note showing that expert-designer preference labels on image output have materially higher inter-rater agreement than crowd labels on the same items — and that a reward model trained on the expert labels transfers better — is the kind of thing he would read. Cite the DALL·E 3 and GPT-Image-1.5 axes explicitly.

What we could not establish
  • Exact PhD institution and dates not confirmed in this pass beyond secondary listings.
  • No public statements from him on human evaluation or aesthetics located.
  • X handle confirmed from an X profile listing, not from a bio naming the employer.

Why they are on the map

Research Lead in ChatGPT Images credits (Dec 2025); image generation leadership for GPT-4o image gen (Mar 2025) source — evidence dated 2025-12

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. The new ChatGPT Images is here — credits
  2. DALL·E 3
  3. Gabriel Goh, Google Scholar