Dan Kondratyuk

Dan Kondratyuk is a machine learning researcher and co-founder of rekursiv.ai. He is an equal technical contributor to Google's VideoPoet, which uses an autoregressive language model to generate and edit video and audio. His work also includes MoViNets for efficient streaming video recognition and UDify for multilingual parsing. He uses the GitHub handle Hyperparticle.[2][8][5][6][7][3]

Kondratyuk coauthored the December 19, 2023 VideoPoet announcement with David Ross, when both were identified as software engineers at Google Research. He describes himself as the project's first author and lead contributor; the project credits equal technical contribution to Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Bryan Seybold and Lu Jiang within a larger author team. Jay Yagnik is credited with the initial concept. VideoPoet received an ICML 2024 Best Paper award.[4][5][9][1]

Kondratyuk and Ross identified coherent large motion as a bottleneck in video generation: models could produce attractive scenes but show artifacts when movement became substantial. Their approach integrated multiple generation and editing tasks in one language model. They argued that a practical advantage was the ability to reuse scalable efficiency improvements from existing LLM training infrastructure.[4]

VideoPoet converts images and video into discrete codes with MAGVIT V2, and audio with SoundStream. An autoregressive language model learns across those codes and text, predicting the next video or audio token; tokenizer decoders turn the output back into viewable or audible media. Its tasks include text-to-video, image animation, inpainting, outpainting, stylization and video-to-audio. For stylization, depth and optical flow supply the motion while text guides the appearance.[5][4]

The model's default output is a two-second video. Longer clips are generated by repeatedly conditioning on one second of video and predicting the next second. Camera instructions such as “Dolly zoom” and “Pan left” can guide motion through the text prompt. For the team's traveling-raccoon short film, Bard supplied a story and scene prompts, VideoPoet generated individual clips, and the clips were stitched together; Angad Singh is credited with storyboarding. Kondratyuk also appeared in the VideoPoet interview included in Latent Space's ICML 2024 episode, published December 10, 2024.[5][4][1]

Kondratyuk co-founded rekursiv.ai with Joshua V. Dillon. The company is building systems that form hypotheses, design and run experiments, and interpret results. In a September 14, 2026 research post, Kondratyuk, Junpeng Lao, Ming-Chang Chiu and Dillon described a research harness that lets agents build on each other's discoveries, preserve failed experiments and revise how they conduct research.[8][10]

The coauthors reported 6,164 experiments over three days and a final NanoChat recipe averaging 0.887791 bits per byte across ten seeds, with lower scores indicating better performance. Each final run used 300 seconds of charged training on one B200 GPU; the discovery campaign consumed additional compute. Research proceeded in two-hour waves, with later teams inheriting recipes, findings and failed directions. Each training run initialized fresh weights rather than receiving a trained checkpoint.[10]

One example shows why preserving failures mattered. A projected trigram-memory design lost to its parent on the tested seed. Agents then tried a full-width design on a differently regularized parent and at a different scale; all three matched seeds improved. The coauthors explicitly distinguished this revised recipe from an ablation isolating the projection's effect. They also cautioned that changing agent instructions did not remove human oversight: humans set study priorities and could pause the search, while a service interruption left 35 completed panels awaiting qualification.[10]

Before rekursiv.ai, Kondratyuk worked on generative video at Luma AI and on video and language models at Google Research. His earlier work supplies context in efficient video processing and multilingual learning.[9][6][7]

In his September 2026 biography, Kondratyuk says he co-developed the generative video models powering Luma AI's Dream Machine and led its World Models team. He describes the team's focus as realtime interactive video, including joint video-audio generation, speech-to-video avatars, and player and camera control. Before Luma, he says he spent five years at Google Research, progressing from AI Resident to Senior Machine Learning Engineer, and also worked on Integrated Multimodal Perception, a mixture-of-experts model combining image, video, audio and text.[9]

Kondratyuk was first author of MoViNets: Mobile Video Networks for Efficient Video Recognition, accepted to CVPR 2021, with Liangzhe Yuan, Yandong Li, Li Zhang, Mingxing Tan, Matthew Brown and Boqing Gong. The work combined neural architecture search for efficient 3D convolutional networks, a Stream Buffer that decouples memory requirements from clip duration, and an efficient ensembling technique. The authors reported that MoViNet-A5-Stream matched X3D-XL's accuracy on Kinetics 600 with 80% fewer FLOPs and 65% less memory.[6]

Earlier, Kondratyuk and Milan Straka introduced UDify at EMNLP-IJCNLP in November 2019. It fine-tunes multilingual BERT on combined datasets to jointly predict part-of-speech tags, morphological features, lemmas and dependency trees across 124 Universal Dependencies treebanks in 75 languages, without recurrent or language-specific components. The authors found that low-resource languages benefited most from cross-linguistic annotations; their zero-shot results suggested useful predictions even for languages absent from both UDify and BERT training.[7]

Kondratyuk reports earning an M.S. in Computational Linguistics at Charles University in Prague through an Erasmus Mundus scholarship, with a departmental Best Master's Thesis award for his multilingual dependency-parsing thesis. He completed his B.S. at Boise State University, where he says he graduated summa cum laude and received the Outstanding Student Award in Computer Science.[9]