The path to achieving AGI may involve starting with limited domains and defining clear benchmarks for success. Meta-learning could play a key role in allowing a network to solve new problems without the need to restart the learning process from scratch.
The speaker hopes to see an architecture that can learn as it sees new problems or data without needing a restart, defining the first step towards AGI.
While the arrival of artificial general intelligence may still be far off, the scaling of AI can already provide significant benefits, such as assisting those who lack resources or knowledge.
The concept of generalization to new tasks in AI requires well-defined benchmarks, but as researchers get closer to limited domains, there's potential for progress. The next step is to go beyond traditional meta-learning approaches and understand intelligence.
The extraction of a knowledge graph from Wikipedia would be interesting as it would provide a different data structure that is more compatible with the idea of programs and deep learning working together. These graphs have a unique structure with human interpretable elements, which can generate knowledge representations that experts can understand and tweak.