AI technique called “generative adversarial networks”, where neural networks are pitted against one another, could lead to unsupervised learning
The day Richard Feynman died, the blackboard in his classroom read: “What I cannot create, I do not understand.” Tweets: @hadip , @ericschmidt , and @googleresearch Tweets: Hadi Partovi / @hadip : The best part of this story is that the idea for it was developed during a semi-drunken argument at a bar. http://www.wired.com/... Eric Schmidt / @ericschmidt : Clever approach to A.I. - train two neural nets against each other, one as ‘artist’ and the other as ‘critic’ http://www.wired.com/... @googleresearch : A nice @WIRED article on the history, and the future, of Generative Adversarial Networks. #MachineLearning http://www.wired.com/...
Context & Ripple Effects
WIRED's 2017 explainer captures GANs at the research-curiosity stage: two neural nets, one generating and one critiquing, trained against each other until the generator produces convincing output without labeled examples — the bar argument origin Hadi Partovi tweeted about and the 'artist vs critic' framing Eric Schmidt amplified. At the time, the payoff being chased was unsupervised learning.
The later coverage shows what that bet compounded into: an arc of generative-AI breakthroughs built on CUDA, convolutional nets, and transformers, then ChatGPT-class systems that marked a step change in real-world AI use. The adversarial idea WIRED profiled is now the ancestor of commercial products with legal exposure, not just lab demos.
First-order effects
- For researchers, the immediate value of pitting a generator against a discriminator is training without hand-labeled data — attacking the annotation bottleneck that supervised learning depends on.
Second-order effects
- Once generation went commercial via systems like DALL-E, the output quality GAN-style training unlocked created the copyright-infringement-without-attribution problem that Marcus on AI argues will draw more lawsuits against OpenAI and peers.
Third-order effects
- If generation keeps displacing curation as the way content gets made, the industry's constraint shifts from producing convincing output to governing it — provenance, attribution, and error tolerance become the design problems, which is the territory the neurosymbolic AI debate over combining neural nets with explicit reasoning also circles.
The trend: Adversarial and generative techniques have traveled from unsupervised-learning research to the technical foundation of consumer-facing generative AI, dragging questions of attribution and governance along with them.