Alex Nichol

Research

Nichol's early papers were on reinforcement learning and meta-learning. "On First-Order Meta-Learning Algorithms" (2018), written with Joshua Achiam and John Schulman, analyzes methods for learning a parameter initialization that adapts quickly to new tasks, and introduces the Reptile algorithm. The same year Nichol was first author of "Gotta Learn Fast", which presented a reinforcement learning benchmark for transfer and few-shot learning built on the Sonic the Hedgehog games.¹⁵¹⁶

In 2021 and 2022 Nichol published a series of papers on diffusion models for image generation. "Improved Denoising Diffusion Probabilistic Models", written with Prafulla Dhariwal, proposed changes that improve sample quality and likelihood, and "Diffusion Models Beat GANs on Image Synthesis" introduced classifier guidance. GLIDE then applied guided diffusion to text-conditional image generation and editing, and Nichol co-authored the 2022 paper on generating images from CLIP latents with Aditya Ramesh and colleagues.³⁵⁴¹⁷

3D generation and code evaluation

Point-E, posted in December 2022 with Nichol as first author, generates 3D point clouds from text prompts by first producing a synthetic image and then a point cloud. Shap-E, posted in May 2023, generates the parameters of implicit 3D functions that can be rendered as textured meshes or neural radiance fields. OpenAI published the code for both projects in MIT-licensed repositories.^9111012

Nichol is also a co-author of "Evaluating Large Language Models Trained on Code" (2021), which introduced the Codex models and the HumanEval set of hand-written programming problems. The HumanEval evaluation harness is published by OpenAI as an MIT-licensed repository.¹³¹⁴

Work and contributions

• HumanEval Open-source code

  Listed among the authors of "Evaluating Large Language Models Trained on Code" (2021); the MIT-licensed openai/human-eval repository is the evaluation harness for the HumanEval problem solving dataset described in that paper.¹³¹⁴

• Improved DDPM, guided diffusion, and GLIDE Open-source code

  First author of "Improved Denoising Diffusion Probabilistic Models" (2021) and GLIDE (2021), and co-first author of "Diffusion Models Beat GANs on Image Synthesis" (2021). OpenAI publishes the official codebases (improved-diffusion, guided-diffusion, and glide-text2im, which runs the small filtered-data GLIDE model) under the MIT License.³⁴⁵⁶⁷⁸

• Point-E and Shap-E Open weights

  First author of Point-E (2022) and co-author with Heewoo Jun of Shap-E (2023). Both OpenAI repositories are the official code and model releases for the papers and are licensed under the MIT License.^9101112

Timeline

1. Mar 2018Posted "On First-Order Meta-Learning Algorithms" with Joshua Achiam and John Schulman.¹⁵
2. Apr 2018Posted "Gotta Learn Fast", a reinforcement learning benchmark based on the Sonic games, as first author.¹⁶
3. Feb 2021Posted "Improved Denoising Diffusion Probabilistic Models" as first author with Prafulla Dhariwal.³
4. May 2021Posted "Diffusion Models Beat GANs on Image Synthesis" with Prafulla Dhariwal.⁵
5. Jul 2021Co-authored the Codex paper that introduced the HumanEval benchmark.¹³
6. Dec 2021Posted GLIDE, a text-guided diffusion model for image generation and editing, as first author.⁴
7. Apr 2022Co-authored "Hierarchical Text-Conditional Image Generation with CLIP Latents".¹⁷
8. Dec 2022Posted Point-E, a system for generating 3D point clouds from text prompts, as first author.⁹
9. May 2023Co-authored Shap-E, a model that generates conditional 3D implicit functions.¹⁰

Where they work

• OpenAI 63 people in the wiki

Related people

• Aditya RameshAlso at OpenAI
• Barret ZophAlso at OpenAI
• Xinlei ChenAlso at OpenAI
• Prafulla DhariwalAlso at OpenAI
• Alexander KirillovAlso at OpenAI
• Luke Melas-KyriaziComputer vision · Image, video, and 3D generation
• Bichen WuComputer vision · Image, video, and 3D generation
• Sanja FidlerComputer vision · Image, video, and 3D generation
• Samuli LaineImage, video, and 3D generation · Computer vision
• Amir HertzImage, video, and 3D generation · Computer vision