Lack of Continual Learning
- Continual Learning is Key: Demis Hassabis and Dwarkesh Patel highlight the lack of continual learning in current LLMs as a critical missing piece for achieving AGI.
- LLMs Don't Improve Over Time: LLMs don't improve over time like humans do. There's no way to give a model high-level feedback; their abilities are fixed after training.
- Human Learning is Adaptive: Humans are useful not just for raw intelligence but for their deliberative, adaptive learning process, which LLMs currently lack.
- Karpathy's Analogy: Andrej Karpathy likens LLMs to coworkers with anterograde amnesia, unable to consolidate knowledge post-training.
Lack of Continual Learning in LLMs
- LLMs Lack Continual Learning: LLMs don't improve over time like humans do. Their baseline may be high, but it's impossible to give them high-level feedback. You're stuck with their out-of-the-box abilities.
- Humans vs. LLMs: Humans are valuable not for raw intelligence, but for deliberative, adaptive learning, which LLMs currently lack.
- Analogy: Andrej Karpathy compares LLMs to a co-worker with anterograde amnesia, lacking long-running knowledge or expertise beyond their short-term memory (context window).
- Mitigation: ChatGPT's memory feature is a primordial, but limited, attempt to address this deficit. Other research focuses on better solutions.
- Potential Solution: Rich Sutton's architecture, a system of agents doing reinforcement learning at runtime, shows promise.