- Joined
- Jan 2026
- How they joined
- Directly
- Before Cognition
- Nous ResearchML Research
- Now
- Stanford University
- Nous Research announced NousCoder-14B was 'post-trained on Qwen3-14B by researcher in residence Joe Li using 48 B200s over the course of 4 days', using Atropos and Modal's autoscaler, reaching 67.87% Pass@1 (+7.08% over Qwen's baseline).Jan 14, 2026
- ML Research Intern at Cognition in San Francisco from Jan to Mar 2026, described on LinkedIn as 'Training coding models. RL algorithms, infrastructure, data.'Jan 2026
- Cognition's SWE-1.6 Preview (2026-03-01) and SWE-1.6 GA (2026-04-07) were the coding-model RL runs during and right after his internship window; SWE-1.6 is post-trained on the SWE-1.5 base, runs at up to 950 tok/s, and shipped 'generally available in Windsurf'.Apr 7, 2026
- On 2026-04-07 he retweeted @cognition: 'We're releasing SWE-1.6, our best model in both intelligence & model UX. SWE-1.6 matches our Preview model on SWE-Bench Pro...' His X bio now reads '@Stanford | prev: @Cognition @NousResearch'.Apr 7, 2026
- At Nous Research (Nov 2024 - Dec 2025) he did pre- and post-training and lists releases nGPT, Hermes 4, and NousCoder-14B.Nov 2024
- Co-author of the Hermes 4 Technical Report (arXiv:2508.18255) with Ryan Teknium, Dakota Mahan, Jeffrey Quesnelle, Chen Guang, Shannon Sands, Karan Malhotra and others.Aug 25, 2025
- Cognition signed a definitive agreement to acquire Windsurf (IP, product, trademark and brand) on 2025-07-14; he joined Cognition in Jan 2026, after the deal.Jul 14, 2025
- 2021-2022 USACO Finalist (Harker, class of 2024), invited to the 2022 USACO summer training camp alongside Harker schoolmate Robert (Bobby) Costin.2022
- Open source: his GitHub JoeLi12345 (bio 'USACO, ML') hosts nGPT, 'an open source reproduction of NVIDIA's nGPT (Normalized Transformer with Representation Learning on the Hypersphere)', with about 112 stars.
- He posted on X about NousCoder-14B: 'Enjoyed working on this a lot! Check out the blog as well.'Jan 6, 2026Enjoyed working on this a lot! Check out the blog as well. https://nousresearch.com/nouscoder-14b-a-competitive-olympiad-programming-model
- LLaDA-R1 post-trained LLaDA-8B-Instruct with SFT and RL plus a remasking optimization to adapt diffusion steps to problem difficulty; code at github.com/radi-cho/reasoning-diffusion.Mar 3, 2025We took 1st place and won $40k at the Mercor x Etched x Cognition inference-time compute hackathon this weekend! Four Stanford freshman (Radostin Cholakov, Zeyneb Kaya, Nicole Ma, and I) built "LLaDA-R1: Scaling Reasoning at Inference Time with Diffusion-LLMs" in 24 hours. ... Thank you to Mercor, Etched, Cognition, CoreWeave, and Anthropic for hosting this incredible hackathon, and to my incredible teammates for their support and collaboration!
- Co-wrote USACO problems: credited with Larry Xing for December 2022 Gold 'Mountains' and the official analysis (with Benjamin Qi); listed among December 2022 contest helpers with Benjamin Qi, Freddie Tang, Nathan Wang and others.Dec 2022
Training coding models. RL algorithms, infrastructure, data.
Pre-training and post-training. Released: - nGPT: Normalized Transformer with Representation Learning on the Hypersphere. - Hermes 4, a family of hybrid reasoning models. - NousCoder-14B, a SOTA olympiad coding model of its size post-trained with RL & verifiable execution rewards.
Biosignal-based emotion recognition using artificial intelligence. First-authored "A Comparison of Personalized and Generalized Approaches to Emotion Recognition Using Consumer Wearable Devices: Machine Learning Study" in JMIR AI.
Princeton Laboratory Learning Program (LLP). Power magnetics modeling. - First-authored "Investigating the Mutual Impact of Waveform, Temperature, and DC-Bias on Magnetic Core Loss Using Neural Network Models" in IEEE Applied Power Electronics Conference 2024. - Co-authored "Multi-Material Power Magnetics Modeling with a Modular and Scalable Machine Learning Framework" in IEEE Applied Power Electronics Conference 2024. - Co-authored "A Simplified Dc-Bias Injection Method for Characterizing Power Magnetics using a Voltage Mirror Transformer" in IEEE Transactions on Power Electronics.
Computer vision and autonomous planning
Coursework: Math: 61CM (Real Analysis + Linear Algebra) 122 (Modules and Group Representation Theory) 144 (Topology & Geometry) 147 (Differential Topology) 158 (Probability & Stochastic Differential Equations) 215A* (Algebraic Topology) Physics: 111 (Partial Differential Equations) 212* (Statistical Mechanics) 330* (Quantum Field Theory) CS: 109 (Probability) 111 (Operating Systems) 149 (Parallel Computing) 224N* (Natural Language Processing) 336* (Language Modeling from Scratch) EE: 101A (Circuits I) 102A (Signals and Systems I) 102B (Signals and Systems II) 108 (Digital System Design) 142 (Electromagnetism) 222* (Applied Quantum Mechanics I) 223* (Applied Quantum Mechanics II) 224* (Quantum Control & Engineering) 263* (Matrix Methods: Singular Value Decomposition) 276* (Information Theory) * indicates graduate coursework.
Just won 1st Place and $40k at the Mercor x Etched x Cognition inference-time compute hackathon! 4 Stanford freshman (@radi_cho, @zeynebnkaya, @nicolesplaining, and I) built "LLaDA-R1: Scaling Reasoning at Inference Time with Diffusion-LLMs" in 24 hours
We took 1st place and won $40k at the Mercor x Etched x Cognition inference-time compute hackathon this weekend! Four Stanford freshman (Radostin Cholakov, Zeyneb Kaya, Nicole Ma, and I) built "LLaDA-R1: Scaling Reasoning at Inference Time with Diffusion-LLMs" in 24 hours. ... Thank you to Mercor, Etched, Cognition, CoreWeave, and Anthropic for hosting this incredible hackathon, and to my incredible teammates for their support and collaboration!
Just won 1st Place and $40k at the Mercor x Etched x Cognition inference-time compute hackathon! 4 Stanford freshman (@radi_cho, @zeynebnkaya, @nicolesplaining, and I) built "LLaDA-R1: Scaling Reasoning at Inference Time with Diffusion-LLMs" in 24 hours
We took 1st place and won $40k at the Mercor x Etched x Cognition inference-time compute hackathon this weekend! Four Stanford freshman (Radostin Cholakov, Zeyneb Kaya, Nicole Ma, and I) built "LLaDA-R1: Scaling Reasoning at Inference Time with Diffusion-LLMs" in 24 hours. ... Thank you to Mercor, Etched, Cognition, CoreWeave, and Anthropic for hosting this incredible hackathon, and to my incredible teammates for their support and collaboration!
Enjoyed working on this a lot! Check out the blog as well. https://nousresearch.com/nouscoder-14b-a-competitive-olympiad-programming-model
