3 ms·

The Unpredictable Abilities Emerging from Large AI Models

• 13years 4y ago
  "LLMs can produce hundreds of “emergent” abilities — tasks that big models can complete that smaller models can’t, many of which seem to have little to do with analyzing text" Fascinating. This has major implications for AI alignment and safety and not in a good way. We have unpredictable behavior that increases with model power while at the same time we fundamentally do not understand how LLMs internally operate.