Ian Goodfellow

Co-founder & CEO — Elorian AI Research Scientist — Google DeepMind Director of Machine Learning — Apple Research Scientist — Google Brain

Listen — profile

0:00 / 3:08

Biographies

Genius Makers

The Mavericks Who Brought AI to Google, Facebook, and the World

Cade Metz 2021

AI history spanning Hinton, Goodfellow (GAN inventor), and pioneers at Google, Facebook.

Genius Makers: The Mavericks Who Brought A.I. to Google, Facebook, and the World

Cade Metz 2021

Journalist Cade Metz's history of AI pioneers — Goodfellow's GAN breakthrough at Montreal bar to Google.

Profile

Ian Goodfellow is the person who taught neural networks to compete with themselves. In 2014, as a PhD student, he sketched out Generative Adversarial Networks — reportedly after an argument at a bar — pitting a generator that fabricates images against a discriminator that tries to catch the fakes, each pushing the other to improve. The idea landed like a thunderclap. GANs powered the first wave of photorealistic synthetic faces, style transfer, super-resolution, and the whole “this person does not exist” moment that made generative AI legible to the public years before diffusion models and transformers took the crown. Yann LeCun called adversarial training the most interesting idea in machine learning in a decade, and for a while he was right.

Goodfellow’s pedigree reads like a tour of the field’s founding families: undergrad and master’s at Stanford under Andrew Ng, then a PhD in Montréal under Yoshua Bengio and Aaron Courville. That trio wrote Deep Learning (2016), the free MIT Press textbook that has functioned as the field’s canonical reference for the better part of a decade — if you learned this material formally, you probably learned it from Goodfellow’s chapters. His career then bounced across the most consequential labs of the era: Google Brain, a stint at the newly founded OpenAI, back to Google, then Apple as Director of Machine Learning, where he publicly resigned in 2022 over a return-to-office mandate, and finally Google DeepMind.

His second major contribution matters as much as GANs for anyone shipping models today: adversarial examples. Goodfellow showed that imperceptible pixel-level perturbations could reliably fool state-of-the-art classifiers, and his Fast Gradient Sign Method turned that observation into a practical attack — and a research subfield. If you care about model robustness, security, or why your vision system does something insane on a slightly odd input, that lineage starts here.

In 2025 Goodfellow left Google after roughly 12 years spanning Brain and DeepMind. He has since resurfaced in the orbit of Elorian, a visual-reasoning startup launched by former DeepMind colleagues (led by Andrew Dai) that raised a reported $55M seed at a $300M valuation before shipping a product — with backing that reportedly includes Jeff Dean. Worth a caveat for readers: Elorian’s own team page and press coverage credit Dai and other ex-DeepMind researchers as the founders, so treat “CEO” framing with some skepticism until the company confirms his title. Either way, Elorian’s thesis — models that reason natively in the visual medium instead of flattening images into text tokens — sits squarely in the lineage of Goodfellow’s generative and vision work, and it’s a bet worth watching if you build anything that has to understand the world through pixels.

Books

Deep Learning

The canonical graduate-level textbook on deep learning, co-authored with Yoshua Bengio and Aaron Courville — free online and still the reference many practitioners cut their teeth on.

Key Articles & Papers

Generative Adversarial Networks 2014 — The paper that launched adversarial training and a generation of image-synthesis research. Explaining and Harnessing Adversarial Examples 2015 — Introduced the Fast Gradient Sign Method and framed adversarial robustness as a core ML security problem. Intriguing Properties of Neural Networks 2013 — First demonstrated that tiny, targeted perturbations can fool deep classifiers — the origin of the adversarial-examples field. NIPS 2016 Tutorial: Generative Adversarial Networks 2016 — The definitive written tutorial on GANs, still the clearest on-ramp to how and why they work. Improved Techniques for Training GANs 2016 — Practical tricks (feature matching, minibatch discrimination) that made notoriously unstable GAN training tractable. Adversarial Examples in the Physical World 2016 — Showed adversarial attacks survive printing and a camera — moving the threat from theory to the real world. Maxout Networks 2013 — An activation-function design paired with dropout that set several benchmark records in the early deep-learning era. Multi-digit Number Recognition from Street View Imagery 2013 — An end-to-end deep net that read street numbers at scale — one of deep learning's early production wins at Google.

Videos

Controversies

Goodfellow’s signature invention is also the technology that made deepfakes possible. GANs enabled convincing synthetic faces, voices, and video, and the same machinery now underpins non-consensual imagery, fraud, and disinformation. Goodfellow has consistently engaged with this openly — arguing for detection research and treating adversarial robustness as a defensive discipline rather than pretending the misuse away — but it remains the double edge of his legacy, and a useful reminder for developers that a generative breakthrough and its abuse ship together.

His 2022 departure from Apple was also unusually public: he left his Director of Machine Learning role in protest of the company’s return-to-office policy, a moment that became a flashpoint in the broader tech-industry fight over remote work.

A note for you (not for the page): The one thing I’d flag before publishing is the Elorian role. Your database bio says “Co-founder & CEO of Elorian AI,” but Elorian’s own site lists Andrew Dai’s team without Goodfellow, and Bloomberg’s coverage frames the founders as other ex-DeepMind researchers. Goodfellow’s LinkedIn confirms he left Google in 2025 and is co-founding a startup, and he publicly amplified Elorian’s launch — but I couldn’t verify the CEO title from a primary source. I wrote the profile to reflect that uncertainty honestly rather than assert a title that may be wrong. If you have a source that confirms his exact role, I can tighten that paragraph. I omitted additional videos beyond the verified Lex Fridman #19 episode (Z6rxFNMGdn0) since I couldn’t confirm other 11-character IDs — your script’s YouTube API will fill that section.

Spotify Podcasts

[İncelemesi] Deep Learning (Ian Goodfellow) Özeti.

9Natree Turkish

2026

Ep 15: Ian Goodfellow designed GANs in 2014 to let machines create realistic images from scratch; five years later the same system was generating non-consensual pornography and fabricated political videos at scale.

Unintended Consequences

2026