Hi. My name is Charlie Deck. I'm an AI researcher and builder.
I want the benefits of AI to actually reach people. That's the through-line in everything below.
I believe a post-scarcity future is possible, but only if we stay close to the human shape of it -- curiosity and challenge shouldn't disappear just because scarcity does.
I'm in NYC. Reach out if you're working on something interesting, or just want to argue with any of the above.
Work. Since Q4 2024, I've been at Reflection, where I helped lead early product explorations and build the evaluation and tooling it takes to ship powerful models safely, including our frontier open-weight models.
(If you're interested in joining our mission of SOTA open-weight AI for all, please check out our open roles.)
Previously, I was on Google DeepMind's AI experience team, where I explored product concepts and built research-facing tools for generative AI.
What I'm best at is working the seam between research and product. A recurring part of that is building the feedback loops that help a lab distinguish real capability from a persuasive demo.
Past work. Before AI labs, I independently built products across mobile, web, and desktop platforms that reached millions of users.
Working end to end on creative tools and games taught me to care about taste, launch constraints, and whether people come back a week later.
That background is why I ask the same question of every AI system I work on now: would anyone rely on this twice?
selected work
// reflection, deepmind, independent research
Building frontier open-weight AI models accessible to all, backed by $2B in funding to democratize advanced AI capabilities.
Google DeepMind's state-of-the-art large multimodal (vision, audio, multilingual text) model powering diverse Google products.
Public atlas of 832 U.S. occupations combining labor-market data with multi-model AI judgments about replacement, augmentation, insulation, and disagreement.
recent
// research, products, blogs (2018-)
A playable physical model of the alto saxophone that shows why fingerings work. It computes the horn's acoustic impedance from its bore and tone holes, adds the player's vocal tract, and blows each note with a simulated reed so you can hear it.
A course in recognizing typefaces by sight: 104 faces from Jenson's roman of 1470 to the fonts on your phone, each shown with the few details that give it away, beside the faces it is most often mistaken for.
An archive of AI skepticism from Lovelace's objection to this year's bubble talk: 285 dated claims, memes and milestones, each quote checked against its source and scored on how it has held up.
A source-linked atlas of AI endgames: 116 accounts from 1863 to 2026, spanning research, forecasts and novels, each broken down into what it assumes, how it unfolds and where the argument can break.
A browser flight lab over ray-marched cloudscapes: cumulus, anviled thunderheads, stratocumulus decks and cirrus generated from tileable noise, above an FFT ocean and under a physically based sky.
A physically modeled saxophone for the EMEO digital sax and other breath controllers. A mass–spring reed drives a conical bore in real time, so the horn speaks, brightens, cracks and honks with breath like a real one.
A ground-up recreation of my 2017 iOS music controller, whose source was lost: a continuous multi-touch instrument for the web and iOS where every finger is its own voice, with its own slide, velocity and pressure.
An ear-training game that grades speed as well as accuracy: name intervals, chord qualities and voice-led progressions from a sampled piano, one keystroke per answer, against a clock.
A tactile practice instrument for the browser, styled as skeuomorphic music hardware: metronome and tempo trainer, chromatic tuner, drone, and a pitch scope that traces your intonation against the beat.
Agreement and calibration statistics for LLM judges and human annotators--what survives once chance gets its share. 3 interactive tutorials.
How sure you are about a benchmark score, and where the next thousand samples should go: uncertainty and experiment design for LLM evals. 3 interactive tutorials.
An interactive playground for seeing number theory: 100,000 integers laid out on spiral grids, with classifications, comparisons, and curve-fitting overlays that turn famous theorems into things you…
A single-page WebGPU app for exploring escape-time fractals as an instrument rather than a renderer: pan, zoom, switch formulas, and remix colors while the image stays live, all the way down to zooms…
Early Reflection product explorations, including Asimov, that helped clarify the product surface and set the stage for the later refocus on frontier open models.
Rephrase text in a webpage in-situ with simpler explanatory language, using a language model.
Let's peer into the future encoded in an AI's training data. With a bit of prompting, we can bootstrap a journey from today to next year to decades to millenia in the "future"--as a powerful LLM…
"Infinite Storyteller" is a generative choose-your-own-adventure, told with state-of-the-art speech synthesis, generative imagery, and presented as a Twitch broadcast in a real-time, 3d game engine.
Explored implementations of editing images with natural language with Stable Diffusion -- inspired by Couairon, et al's DiffEdit paper.
Open Ended Learning Team, et al. "Open-Ended Learning Leads to Generally Capable Agents." ArXiv:2107.12808 [Cs], July 2021.
DeepMind Interactive Agents Team, et al. "Creating Multimodal Interactive Agents with Imitation and Self-Supervised Learning." ArXiv:2112.03763 [Cs], Feb. 2022.
Wang, J.X., et al. (2021). Alchemy: A structured task distribution for meta-reinforcement learning. NeurIPS.
Fortunato, M., Tan, M., Faulkner, R., Hansen, S.S., Badia, A.P., Buttimore, G., Deck, C., Leibo, J.Z., & Blundell, C. (2019). Generalization of Reinforcement Learners with Working and Episodic Memory. NeurIPS.
past work
// independent (2007-2017)expressionPad
Unique, highly-configurable music controller with built-in additive synthesizer. Full 10-touch expression with individual continuous control.
A casual roguelike game with RC racing-like controls. Procedural levels, some flashy visual effects, and a 80's hacking theme. Made for the Seven Day Roguelike (7DRL) gamejam.
A Zelda-like co-op language learning action RPG. Features some novel UX for multi-lingual text input--in an action game. Playable alpha.
Shock and aww! Rescue babies and cute animals from ten crisis situations in this room-scale VR experience. An Oculus Touch launch title.
800+ code sketches and renders, many going viral on Reddit and Imgur. Explorations in perception, generative art, algorithm visualization, photorealistic rendering, and much more.
Word Fireworks: Learn Korean! (et al)
Learn to read and write a new language with arcade action! Available in Korean, Chinese, Japanese, and Hebrew.
primes.io
Number theory explorer that visualizes Ulam spirals, the phenomenon whereby prime numbers appear with suspicious frequency when the number line is wound in a spiral. Led to a talk at Google I/O.
Real-time Mandelbrot and Julia fractal explorer. Fractals are calculated on the GPU; an experiment in pixel shaders.
Fireworks Hero!
Test your gaming skills to unlock all 10 heroes. A dozen game modes with more than 100 firework types.
Real-time, interactive visualization of strange attractors: dynamical systems with a fractal internal structure. A haunting beauty reveals itself.
Visual-acoustic performance piece. A series of pitch-responsive abstract mechanisms are controlled via saxophone augmented with a TouchOSC interface.
The Grading Game
Race against time to find spelling and grammar errors--and pay down your crippling student debt!
Explore shiny 3D cellular automata in this interactive explorer powered by WebGL. Discover beautiful, complex 3D tapestries arise from simple rules.
Rhetorical combat with semiotic gameplay dynamics.
Lunacraft
Explore a sci-fi procedural voxel world. Infinite terrain, crafting. (Impressive stuff for the time!)
The drumming app with the fastest response and richest expression. 15+ built-in drumkits, and unlimited customization.
A visual composition toy where notes are nodes lighting up a hexagonal grid. Make music with chain reactions!
A short animated movie generated by wandering around the parameter space of one particular strange attractor system.
Immersive VR total-body flying game. Mashing up body tracking and head-mounted display so players can spread their arms to fly down a virtual chasm.
United Nanolytics Knowledge Vault
Generative game made in 72 hours for the Ludum Dare 30 gamejam.
Avoid/align/approach 3d flocking, inspired by Craig Reynolds' "boids". My first Unity experiment.
emotweet
Visualize Twitter reactions. Sentiment analysis. Final project for Harvard CS171 Data Visualization. (API deprecated.)
Wow Such Web
Doge-ify the web. Chrome extension that swaps and alters content, such media. Wow.
Ellsworth
Abstract puzzler inspired by the 20th century Abstractionist, Ellsworth Kelly.
evophonia
Evolving rhythm game for Ludum Dare 24.
My Kingdom for a Line Block!
Two-player asymmetric version of the famous tetromino game. One player chooses the worst-possible block for the other. Built for Ludum Dare 25.
Polyglot
Inspired by browser extensions like Rikai, Polyglot was an iPhone web browser that enabled you to discover, accumulate, and practice new vocab as you encountered it while reading.
Caucus
Choose your candidate, listen to voters' preferences and transmit your positions on the issues -- but don't let the voters discover you're a flip-flopper! A real-time political strategy game about lying. Built for the Global Game Jam 2012.
Space Opera Singer
Sing the correct notes to defeat the aliens -- a sight-singing space shooter.
Legacy Work
I created many products before those described above. The many lessons those experiences taught me informed subsequent work.