Dan Kondratyuk
Dan Kondratyuk is a machine-learning researcher whose work links multilingual language processing, efficient video recognition and generative video. Raised in Boise, he earned a computer-science degree from Boise State in 2017 and an Erasmus Mundus joint M.S. in computational linguistics at Charles University and Saarland University in 2019; his early models include LemmaTag and UDify. At Google, he moved from AI Resident to Senior Machine Learning Engineer and created MoViNets; at Luma AI, he joined as a Research Scientist in March 2024 and led the World Models team behind Dream Machine. VideoPoet won the 2024 ICML Best Paper award, and he coauthored MALT Diffusion for long-video generation in 2025. As of October 2026, he is a co-founder of rekursiv.ai, where he works on AI systems that run experiments and produce new knowledge.
Timeline
-
2017
A Boise State degree
Raised in Boise, he earns a computer-science degree from Boise State in 2017. He credits an early fascination with computers for his interest in programming.
-
2018–2019
LemmaTag and UDify
He completes an Erasmus Mundus joint M.S. in computational linguistics at Charles University and Saarland University in 2019. In 2018, he coauthors LemmaTag, a neural model for joint part-of-speech tagging and lemmatization. His 2019 model UDify predicts universal part-of-speech, morphology, lemmas and dependency trees across 75 languages.
-
2021
MoViNets for streaming video
He creates MoViNets with Google colleagues, a family of efficient mobile video networks for action recognition on streaming video. Its Stream Buffer separates memory use from clip length and supports online inference with a small, constant memory footprint.
-
By March 2024
Google's AI Residency
He joins Google's AI Residency and moves into a full-time machine-learning role, advancing to Senior Machine Learning Engineer. His 2023 Integrated Multimodal Perception work combines image, video, audio and text in a mixture-of-experts model.
-
March 2024
Research Scientist at Luma
In March 2024, he joins Luma AI as a Research Scientist and develops features behind Dream Machine. He leads the World Models team on real-time interactive video, joint audio generation, avatar speech-to-video and camera control.
-
2024
ICML Best Paper for VideoPoet
His coauthored VideoPoet wins ICML's 2024 Best Paper award. The multimodal language model uses a decoder-only transformer to generate video from image, video, text and audio inputs.
-
2025
MALT Diffusion for long videos
He coauthors MALT Diffusion, a memory-augmented latent transformer model for long video generation. It generates video in short segments and carries a compact memory across them to retain long-range context.
-
As of October 2026
Co-founder of rekursiv.ai
He is a co-founder of rekursiv.ai, where the team builds autonomous systems that form hypotheses, design experiments and discover new knowledge. In October 2026, he coauthors a study of a 15-million-parameter reinforcement-learning agent that beats Craftax's final boss in 44% of its games.
Career
Education
- 2019Charles UniversityM.S.
- 2019Saarland UniversityM.S.
- 2017Boise State UniversityComputer science degree
Sources and credits
Photo: We have no published photo of Dan Kondratyuk yet. Send us a free-licensed photo.
- Featured Student: Dan Kondratyuk
- Erasmus Mundus Retrospective: Charles University in Prague, Year 1
- About · Dan Kondratyuk
- 75 Languages, 1 Model: Parsing Universal Dependencies Universally
- MoViNets: Mobile Video Networks for Efficient Video Recognition
- VideoPoet – Google Research
- ICML 2024 Awards
- MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation
- rekursiv.ai, Autonomously Create Knowledge
- A 15M-parameter RL agent trained from scratch vs GPT-6 Astra: who would win?