Curtis Hawthorne (@fjord41) / X

Curtis Hawthorne

206 posts

Curtis Hawthorne

@fjord41

@OpenAI Agents. Previously @AdeptAILabs (adept.ai) / Amazon, Google Brain (g.co/magenta). Pipe organ enthusiast: youtube.com/cghawthorne

curtishawthorne.com

Joined September 2013

Curtis Hawthorne
@fjord41
Aug 29, 2022
Diffusion for music synthesis! We trained a “notes2audio” pipeline to synthesize audio from multi-instrument MIDI notes. Listen 🔊: g.co/magenta/spec-d… Play 🎼: g.co/magenta/spec-d… Code 👩‍💻: g.co/magenta/spec-d… Read 📝 : arxiv.org/abs/2206.05408 1/
Curtis Hawthorne
@fjord41
Feb 28, 2018
Google's Machine Learning Crash Course is now available externally! The internal version of this course is what helped me get started when I joined the Magenta team. I highly recommend it.
developers.google.com
Machine Learning | Google for Developers
Curtis Hawthorne
@fjord41
Feb 12, 2018
New blog post about the project I've been working on for a while. Automatic piano music transcription (raw audio to MIDI) that works really well!
magenta.tensorflow.org
Onsets and Frames: Dual-Objective Piano Transcription
Update (9/20/18): Try out the new JavaScript implementation!Update (10/30/18): Read about improvements and a new dataset in The MAESTRO Dataset and Wave2Midi...
Curtis Hawthorne
@fjord41
Nov 2, 2025
Excited to join the Agents team at @OpenAI! Just finished my first week, and there's so much great stuff going on!
65K
Curtis Hawthorne
@fjord41
Feb 7, 2023
Today is my first day at @adeptailabs! I'm excited to be a part of this awesome team and work on building a natural language interface for all software.
71K
Curtis Hawthorne
@fjord41
Jul 21, 2021
Excited to share our @ISMIR2021 paper: Sequence-to-Sequence Piano Transcription with Transformers Generic encoder-decoder Transformer + Spectrogram inputs + MIDI-like event outputs = SotA results! arxiv.org/abs/2107.09142 With @iansimon, @rigeljs, @ethanmanilow, and @jesseengel
Curtis Hawthorne
@fjord41
Oct 31, 2022
Functional MicroGrad: gist.github.com/cghawthorne/68… I really enjoyed @karpathy's recent video about building an autograd engine from scratch. To better understand it, and because I like JAX, I had some fun writing a version that's more "functional" in flavor.
functional-micrograd.ipynb
From gist.github.com
Curtis Hawthorne
@fjord41
Jun 16, 2022
Very excited to share Perceiver AR! Our decoder-only Transformer variant scales to very long input contexts (we demonstrate 131k tokens in the paper). No need to implement complicated new layer types: basically just swap the first self-attention layer for a cross-attention.
Curtis Hawthorne
@fjord41
Aug 29, 2022
Replying to @fjord41
For images, text is great at specifying both style and structure. In contrast, for many types of music, notes are a more natural text-like representation of structure, allowing us fine-grained control of both composition and instrumentation. 2/
00:00
Curtis Hawthorne
@fjord41
Sep 7, 2023
Today's episode of "Adventures in GPU Floating Point Arithmetic"
14K
Curtis Hawthorne
@fjord41
Nov 1, 2017
Check out our new paper on state of the art automatic piano transcription! Onsets and Frames: Dual-Objective Piano Transcription arxiv.org/abs/1710.11153
Curtis Hawthorne
@fjord41
Aug 29, 2022
Replying to @fjord41
The architecture is essentially our MT3 transcription model (g.co/magenta/mt3) backwards. We use the same T5-style encoder-decoder. Instead of spectrograms as input and audio as output, we take note events as input and use DDPM to produce output spectrograms. 3/
Curtis Hawthorne
@fjord41
Aug 29, 2022
Replying to @fjord41
The best diffusion models achieve high fidelity from several stages. We take a similar approach by using a DDPM to predict spectrograms and training a separate GAN spectrogram inverter to generate audio (available here: tfhub.dev/google/soundst…). 4/
Curtis Hawthorne
@fjord41
Aug 29, 2022
Replying to @fjord41
Work done with: @iansimon, @ada_rob, @neilzegh, @jpgard, @ethanmanilow, and @jesseengel. Also big thanks to folks from the Imagen team (@wchan212, @mo_norouzi, @Chitwan_Saharia, and Jonathan Ho) and @sedielem for help with diffusion! 10/