FormerResearch · Researcher

Steven Cao

“ml @capable, prev @cognition, harvard”

Joined: Dec 2025
How they joined: Directly
Before Cognition: Redwood ResearchResearch Intern
Now: Capable

Overlapped before Cognition

Harvard University11

Sources¹linkedin.com ²linkedin.com

Autumn profile

Researched with Autumn · 2026-11-04

Details

LocationSan Francisco, California, United States

Company sitecapable.com

LanguagesChinese, English

UniversityHarvard University

reinforcement learning post-traininglarge-scale model trainingNVFP4 low-precision trainingLLM evaluationalignment faking evalscoding agents

Notes

Capable's Fall 2026 update describes him as having become "one of the ~five core research MLEs at Cognition, key for training Cognition's trillion-parameter-scale models", and says he brings "experience wrangling post-training large models, compute at scale, optimization of runs".Oct 2026Capable Fall Update: Steven Cao, previously at Cognition and Redwood Research, is joining us as MTS.
Left Cognition in Jun 2026 and joined Capable (lab automation / longevity-biotech AI lab in Mission Bay, SF) as Machine Learning Engineer / MTS in Oct 2026. Capable announced the same month that it had raised a $38M seed (Cathay, Abstract, Fidji Simo, Andrej Karpathy and others).Oct 2026Capable Fall Update: Steven Cao, previously at Cognition and Redwood Research, is joining us as MTS.
Co-author of 'SWE-1.7: Frontier Intelligence at a Fraction of the Cost' (the byline ends '... Steven Cao, Silas Alberti', with an equal-contribution marker). SWE-1.7 was RL-trained on a Kimi K2.7 base and scored 42.3 on FrontierCode 1.1, close to Claude Opus 4.7/4.8 and GPT-5.5 at a much lower cost per rollout.Jul 8, 2026SWE-1.7: Frontier Intelligence at a Fraction of the Cost (co-author)
Research Engineer at Cognition from Feb 2026 to Jun 2026; his LinkedIn role description reads, word for word: "trained swe 1.6 and 1.7 models".Feb 2026
Co-author of 'Introducing SWE 1.6: Improving Model UX' (Rohan Choudhury, Carlo Baronio, Ben Pan, Sam Lee, Eric Lu, Steven Cao, ...). SWE-1.6 was post-trained from scratch, shipped generally available in Windsurf at up to 950 tok/s, and came with a free tier for 3 months.Apr 7, 2026Introducing SWE 1.6: Improving Model UX (co-author)
Listed as a co-author on Cognition's 'An Early Preview of SWE-1.6 and Research Update' post (by Carlo Baronio, Ben Pan, Sam Lee, Eric Lu, Steven Cao, Rohan Choudhury, Adam Zweiger, Ray ...). The preview improved SWE-Bench Pro by more than 10% over SWE-1.5.Mar 1, 2026An Early Preview of SWE-1.6 and Research Update (co-author)
NVIDIA quoted him in its NVFP4 blog post: Cognition is seeing "significant latency and throughput gains" by using NVFP4 in large-scale reinforcement learning, "said Steven Cao, a member of Cognition's research team.".Feb 6, 2026
His LinkedIn post on joining Capable, word for word: "Excited to be joining Capable to help push the limits of human biology! We have so much potential for longer, healthier, and more empowered lives."Oct 10, 2026Excited to be joining Capable to help push the limits of human biology! We have so much potential for longer, healthier, and more empowered lives.
Started at Cognition as a Research Intern (Dec 2025 - Jan 2026, San Francisco) and converted to full-time Research Engineer within about two months.Dec 2025
Research Intern at Redwood Research (Apr - Jun 2025, SF), where he worked on "alignment faking evals". Capable describes this as work "on reducing alignment faking at Redwood Research". It was his first AI-safety lab role, during his first year at Harvard.Apr 2025
Studied Computer Science (Bachelor's) at Harvard University, with the LinkedIn listing 2024-2026. He was training frontier coding models at Cognition by age ~19-20 (his summary simply reads "20yo").2024
Public headline says "ml @capable, prev @cognition, harvard".

Autumn profile

People they worked with before

More from Research