Why Nutcrackers Remember and GPT Doesn't

Key Points:

• GPT's training objective is next-word prediction over internet text corpora. This narrow focus explains resulting capabilities and limitations.
• Numerous examples are provided of GPT-3's failures on uncommon or atypical inputs and outputs, complex reasoning, physical tasks without vision, etc.
• The authors categorize key sources of errors such as likelihood bias, data dependence, sensitivity to wording, difficulty combining concepts.
• They recommend identifying mismatches between training objectives and real-world tasks, using skepticism where skills exceed language prediction.
• GPT's strengths and weaknesses reflect its internet corpus training, not comprehensive intelligence. Outputs emerge indirectly rather than manipulating abstract ideas.
• LLMs operate within fluid competence frontiers shaped by likelihood tradeoffs. Their impressive yet limited capacities are echoes of thought, not complete cognition.

Clark's nutcrackers cache food like no other. These birds hide thousands of seeds in the soil every winter and depend on recalling cache locations months later to survive. Their extraordinary spatial memory outperforms that of related species like scrub jays. 

This difference makes sense through a teleological lens: the ecological need to find cached food shaped the evolution of enhanced spatial abilities in nutcrackers. Their cognitive specialization arose from the particular problems their lifestyle requires solving. 

Similar principles extend to artificial intelligence. Analyzing what objective a system was optimized for reveals a lot about resulting capabilities. Just as naturally selected skills reflect challenges faced, so too do trained AI abilities mirror prescribed training goals.