MemCast
MemCast / episode / insight
#ai-adoption21#continual-learning4#generalization1
Million-token contexts enable 'days of human learning'
  • Current context lengths (~1M tokens) equivalent to days/weeks of human reading
  • Enables substantial in-context learning
  • Longer contexts (10M+) could enable months of learning
  • Engineering challenge is inference optimization, not fundamental limits
Dario AmodeiDwarkesh Patel00:41:42

Supporting quotes

Quote
00:41:42
“A million tokens is a lot. That can be days of human learning. If you think about the model reading a million words, how long would it take me to read a million? Days or weeks at least.” — Dario Amodei
Quote
00:43:10
“There's nothing preventing longer contexts from working. You just have to train at longer contexts and then learn to serve them at inference.” — Dario Amodei

From this concept

Continual Learning Debate

Discussion of whether AI systems need human-like continual learning to be economically transformative, or if scaling current approaches will suffice.

View full episode →

Similar insights

“Progressive fidelity improves iteration speed”

RafaelY Combinator

“Incremental updates preserve existing work”

HostY Combinator

“Prompt feedback highlights respected vs ignored elements”

HostY Combinator