Last year, we conjectured that polysemanticity is caused by "superposition" – models compressing many rare concepts into a small number of neurons. We also conjectured that "dictionary learning" might be able to undo superposition.
Neural networks often pack many unrelated concepts into a single neuron – a puzzling phenomenon known as 'polysemanticity' which makes interpretability much more challenging. In our latest work, we build toy models where the origins of polysemanticity can be fully understood.
Oct 5, 2023, 5:40:34 PM
There’s a strong argument to be made that the most compelling solutions in the short history of the New Internet have been those that are the most commonplace, slow, ambient, and avoidant of hype. Products that actually work and will continue to.
Today, Meter has built its own integrations on top of Linear, such as one that auto-generates notes for each software release. But Meter CEO Anil Varanasi noted that they’ve seen other productivity tools “start out really great,” then collapse under their own complexity as they scale up. “How does Linear avoid the same fate?” he asked.