AER Labs Profile picture
AER: Accelerate Efficient Research | AI Engineering & Research
Nov 26 8 tweets 2 min read
Just watched @JeffDean talk on the last 15 years of ML progress and the launch of Gemini 3.0. It is actually wild to look back. From manually coding gradients on a hypercube in the 90s to solving Math Olympiad problems today.

The recipe remains surprisingly consistent: Scale + Better Algos + Hardware.

A few highlights and notes 🧵👇 1/ The Unsupervised Spark Jeff reminisced about the 2012 "Cat Paper." 10M YouTube frames, unsupervised. No labels, yet neurons emerged that specifically detected faces and cats. This was the moment. The model "grokking" high-level concepts from raw pixel soup. Also Word2Vec—vector arithmetic (King - Man + Woman = Queen) showing semantic space is real. ✨