Curtis “Fjord” Hawthorne
I design, train, and ship models on the path to AGI. Currently, I work on the Agents team at OpenAI, building computer use agents. Previously, at Amazon AGI Labs, I led multimodal, planning, and RL efforts, including launching the Nova Act browser agent. Before that at Adept (which exited to Amazon), I co-designed the Fuyu multimodal architecture, led its end-to-end implementation, and ran production multimodal training. Earlier, I spent over a decade at Google, starting as an SRE, then as a SWE TL/M on Google Play Music (now YouTube Music), before joining Google Brain in 2016 as a TL on Magenta, initially working on music transcription/generation/synthesis, later broadening to Transformer architecture variants and general-purpose generative modeling.
Outside of work, I enjoy playing and composing for the pipe organ →
Selected Publications
Highlights from research across agents, multimodal, music, and large-scale training. See the full list on Google Scholar →
-
Introducing Amazon Nova Act
Publicly launched research preview of our model and SDK for developers to build agents that take actions in web browsers. Culmination of the work started at Adept.
-
Introducing our Multimodal Models
Novel unified multimodal architecture that simplifies training, supports an arbitrary number of images at arbitrary resolutions, and works with interleaved text. Later scaled up to Fuyu-Heavy, which was the world's third-most-capable multimodal model at the time.
-
Scaling Up Models and Data with t5x and seqio
Open source libraries for training and data management, enabling models with hundreds of billions of parameters. Used to train PaLM, UL2, PaLI, and many others.
-
General-purpose, long-context autoregressive modeling with Perceiver AR
Transformer variant that can directly attend to over a hundred thousand tokens. Achieved state-of-the-art results across text, images, and music.
-
Multi-instrument Music Synthesis with Spectrogram Diffusion
First application of diffusion to note-conditioned multi-instrument music generation. The model takes MIDI as input and outputs audio in real time.
-
MT3: Multi-Task Multitrack Music Transcription
State-of-the-art multi-instrument automatic music transcription with T5-style Transformers.
-
Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset
Introduces the MAESTRO dataset, now the standard for training and evaluating piano models. Also scales up the Onsets and Frames transcription model and demonstrates audio synthesis capabilities. Later used to transcribe >10k hours of piano music on YouTube to train Listen to Transformer.
-
Music Transformer: Generating Music with Long-Term Structure
First application of transformers to long-form music generation.
-
Onsets and Frames: Dual-Objective Piano Transcription
Onset-conditioned CNN-LSTM for piano transcription that achieved practical real-world quality and became a widely used baseline.
-
Murder in Jefferson: The 1868 Stockade Case
Definitive history of the 1868 Stockade Case in Jefferson, Texas, chronicling the twists and turns of the trial and ensuing community upheaval.
Music
Contemporary organ works that blend traditional forms with modern harmonic language. Listen to recordings and find published scores below, or see more performances on my YouTube channel →
-
Reflections on St. Denio: Immortal, Invisible, God Only Wise
2023Performance YouTubeThis piece explores the hymn tune St. Denio by reflecting on several phrases from the text and the musical forms they evoke. First, a toccata for the never ending motion of "Immortal". Next, a meditation for the mysteries of "Invisible". For "God Only Wise," it seemed natural to use the "wisest" of musical forms, a fugue! Finally, we revisit the toccata and build to a conclusion of joyful praise for "Almighty, Victorious, Thy Great Name We Praise!" Perfect for preludes or postludes.
A top entry in the 2023 BIS Composition Competition and selected for publication in the 2023 BIS Organ Book: Veni Creator.
-
Prelude and Fugue in G Minor
2021Performance YouTubePublication Sheet Music DirectInspired but not constrained by Preludes and Fugues of the Baroque, this piece is perfect for preludes, postludes, and recitals. The dramatic Prelude develops its rhythmic themes in contrastive sections with modern harmonic language, and the lighter Fugue builds to a full organ conclusion with extensive use of stretto.
-
Prelude on Aberystwyth (Jesus, Lover of My Soul)
2020This organ prelude explores the classic hymn tune "Aberystwyth" (Jesus, Lover of My Soul) with inventive counterpoint and a unique harmonic approach. The closing features several layers of the main theme that build until the final, calm restatement of the opening phrase. Especially suitable for use during Lent. Written for organs with 2 or more manuals and pedal.
Selected for publication in the 2022 Beauty in Sound Call for Composers.
-
Psalm 91
2017Performance YouTubePublication Sheet Music DirectThis original composition is inspired by the text of Psalm 91. The first section of the piece is meant to evoke the "fortress" and the "shadow" of the Almighty. Next, listen for the "fowler’s snare" and the "terror of night." However, the danger doesn’t last long and we are soon reminded that "The Lord is my refuge." The final section describes God’s deliverance, ultimately leading to long life and salvation.
-
Trumpet Voluntary
2016This regal piece is a trumpet voluntary in a modern setting, inspired by the likes of Purcell and Stanley, and developed to a dramatic conclusion. Perfect for a prelude or postlude.
Selected for publication in the 2022 Beauty in Sound Call for Composers.