Transformer and diffusion models are not an incremental improvement over CNNs, RNNs, and GANs. They are a step function. Elad Gil, who built ML systems at Google and Twitter after the Mixer Labs acquisition, spent the 2017-2019 period watching nearly every ML startup fail because the technology could not open new markets. Incumbents captured the value. That dynamic has now inverted.
Gil's core argument is structural, not technical. Treating current AI as a continuum with prior deep learning waves is the same category error as calling an airplane a car with wings. The aviation analogy is precise: the prior frame misses entirely new use cases in logistics, defense, and consumer products that only become possible at this capability level. His two slides, one from the CNN/RNN/GAN era and one from now, make the before-and-after concrete.
The piece is worth reading in full for the slides and the underlying framework Gil uses when advising startups today. The question it leaves open, and does not fully resolve, is which specific market openings are durable for startups versus which will again be captured by incumbents with distribution and compute. That tension is where the real argument lives.
[READ ORIGINAL →]