Miju Labs

← All buyers · NVIDIA (Nemotron / Cosmos)

Oleksii Kuchaiev

Senior Director of Applied Research, NVIDIA (leads Nemotron post-training) · NVIDIA (Nemotron / Cosmos)

high confidenceTier A · Talk to firstSecond-wave labstext

Runs post-training for NVIDIA's Nemotron models and is listed under Leadership and Post-training in the Nemotron 3 report. He is the senior author on HelpSteer3-Preference, the dataset NVIDIA built with more than 6,400 annotators across 77 countries sourced through Scale AI and Translated — which makes him one of the few frontier-scale buyers in this index whose human-data supply chain is publicly documented.

Background

Ukrainian-trained mathematician: BSc and MSc in applied mathematics from Taras Shevchenko National University of Kyiv, then MSc and PhD in information and computer science at UC Irvine (2010). He built the first GPU backend for Microsoft's CNTK, then worked as an applied scientist on App Store recommendations at Apple, before joining NVIDIA where he grew NeMo from a personal project into a multi-team framework.

He now leads applied research on post-training and reinforcement learning for the Nemotron line, and is a visible advocate for open models — he published a piece in July 2026 arguing that banning open-source AI would be a historic mistake, and writes about the shift from scalable imitation learning to scalable exploration learning.

What they run now

Post-training and RL for Nemotron: SFT, preference optimisation, reward models, and increasingly environment-based RL. HelpSteer3-Preference (2025) is the public artefact of his human-data programme — over 40,000 samples across General, STEM, Code and Multilingual splits, annotated on a seven-point relative-quality scale, with escalating qualification bars per split (relevant degree for STEM, software-engineering experience for Code, fluency for Multilingual), and annotators told to use web search but not LLM tools. Scale AI managed the General, STEM and Code subsets; Translated managed Multilingual. The Nemotron 3 report (Dec 2025) leans more on RL environments and names no annotation vendor, which suggests the human-preference programme is now one input among several rather than the centre of gravity.

What they buy

Documented: large-scale human preference annotation bought through Scale AI (general, STEM, code) and Translated (multilingual), with screened annotator pools and published guidelines, released openly. Modality is text and code; no visual or design data appears in any Nemotron human-data release. Inference: as the post-training director he approves which capabilities get a human-data campaign and how much it costs, and NVIDIA's habit of open-sourcing the resulting datasets means anything Julian sells here may be published.

Career

  1. presentSenior Director of Applied Research; leads Nemotron post-training, NVIDIA — Grew the NeMo framework; Leadership and Post-training in Nemotron 3.
  2. —Applied Scientist (App Store recommendations), Apple
  3. —Engineer (initial GPU backend for CNTK), Microsoft
  4. 2010University of California, Irvine, Ph.D., Information and Computer Science
  5. 2009University of California, Irvine, M.Sc., Information and Computer Science
  6. 2007Taras Shevchenko National University of Kyiv, M.Sc., Applied Mathematics

On the record

Open-source AI · 2026-07

Argues that banning open-source AI would be a historic mistake. source

Post-training direction · 2025-04

Frames the field as moving from scalable imitation learning toward scalable exploration learning. source

Recent posts

Oleksii Kuchaiev
Senior Director of Applied Research
I would love to see "international agreement" on AI. But as we speak, fascist russia is raining ballistic missiles with cluster munitions, jet-powered drones and other weapons on Kyiv, a European capital with at least 3M residents. And on many other cities and infrastructure in Ukraine. China is cooperating with russia economically and in many other ways. International alignment is a problem humans have failed to solve so far.
342 comments
I would love to see "international agreement" on AI. But as we speak, fascist russia is raining ballistic missiles with cluster munitions, jet drones and other weapons on Kyiv, a European capital with at least 3M residents" as well as many other cities in Ukraine. And China is helping it economically. International alignment is a problem humans have failed to solve so far.
Oleksii Kuchaiev
Senior Director of Applied Research
We need to accelerte progress in AI not slow it down, particularly in the areas of life sciences, cyber security and alignment. There is very real cost (unlike p(doom) estimates) that can be attached to slowing down progress. For example, a one year delay in meningitis vaccine invention would have cost tens of thousands of lives. We also should not let incompetent or malicious humans to shift blame from themselves to the tool (model). If someone start testing new model with Internet access and it causes harm, that someone must be held accountable. And yes, it is possible to just physically prevent Internet connection if one is unsure of their software guardrails.
672 comments

Interviews & talks

Why it matters here

He is a top-tier buyer with an unusually transparent supply chain, which is worth studying even if he never buys design data: HelpSteer3 is effectively a public spec sheet for how a frontier lab writes annotator qualifications, rubrics and guidelines, and a vendor can quote it. The direct fit is weak — Nemotron's human data is text and code, and NVIDIA's visual work sits in Ming-Yu Liu's Cosmos/Deep Imagination group instead. The credible opening would be multimodal or document-visual preference data for Nemotron's vision capabilities, not design taste.

How to reach

Speculation: he is publicly active (X, personal blog, open-source advocacy), so the natural opener is technical and open-source flavoured — e.g. a small openly released design-quality preference set built to HelpSteer3's conventions (seven-point scale, documented qualification bars, published guidelines) as a demonstration that the same methodology transfers to visual judgement. NVIDIA's openness norm means a vendor willing to let the data be published is unusually attractive to them. For visual work specifically, ask him to point to the Cosmos side rather than pitching him directly.

What we could not establish
  • Whether NVIDIA still uses Scale AI post-2025; Nemotron 3 names no vendor.
  • Whether any Nemotron or Cosmos programme buys visual/aesthetic preference data.
  • His start date at NVIDIA and exact reporting line under Jonathan Cohen / Bryan Catanzaro.

Why they are on the map

Leads Nemotron post-training teams; senior author on HelpSteer3 (Scale AI + Translated annotators) source — evidence dated 2026-09

Needs the data for a model; writes the spec.

How NVIDIA (Nemotron / Cosmos) buys data

HelpSteer3-Preference (2025) used 6,400+ annotators in 77 countries sourced/managed by Scale AI (general, STEM, code with degree/SWE bars) and Translated (multilingual). Nemotron 3 report (Dec 2025) lists dedicated Data, Post-training, Evaluation/Safety and Project Management groups but no annotation vendor; RL is mostly environment-based.

Nemotron post-training under Oleksii Kuchaiev (Applied Research), leadership incl. Jonathan Cohen, Bryan Catanzaro, Kari Briski; HelpSteer human-data programme run by Zhilin Wang and co-authors; 15-person project-management pool. Visual generation in Ming-Yu Liu's research group.

Known vendors: scale-ai, Translated

Sources

  1. kuchaev.com — personal site (title, career, education, writing)
  2. HelpSteer3-Preference (arXiv 2505.11475): 6,400+ annotators, 77 countries, Scale AI and Translated
  3. NVIDIA Nemotron 3 report — Leadership and Post-training contributor lists