General intelligence // what the hack it is?) hype, marketing?
There is no general intelligence, every intelligence is specialized, and general intelligence != sum of special ones
All current artificial intelligence systems are unable to solve cognitive tasks that children, animals and other natural creatures easily can do.
Here’s a summary of the key points from François Chollet’s talk on intelligence and AI:
1. Current limitations of Large Language Models (LLMs):
— LLMs struggle with tasks requiring reasoning, composition, and generalization.
— They often fail when presented with slight variations of familiar problems.
— Performance depends more on task familiarity than task complexity.
2. Intelligence vs. Skill:
— Chollet argues that skill is not intelligence, but rather the output of intelligence.
— Intelligence is the ability to deal with new situations and generate new solutions.
3. Measuring Intelligence:
— Chollet introduced the Abstraction and Reasoning Corpus (ARC) as a benchmark for measuring intelligence in AI systems.
— ARC tasks are novel and based on core knowledge, requiring few-shot learning and generalization.
4. Types of Abstraction:
— Value-centric abstraction: operates in continuous domains (e.g., perception, intuition)
— Program-centric abstraction: operates in discrete domains (e.g., reasoning, planning)
5. Combining Deep Learning and Program Synthesis:
— Chollet suggests merging these approaches to achieve more general AI capabilities.
— Deep learning can provide intuition to guide program search, while program synthesis can add reasoning capabilities to neural networks.
6. Future Directions:
— Chollet believes the next breakthrough in AI may come from outsiders rather than big tech labs.
— He emphasizes the need for new ideas and approaches beyond simply scaling up existing LLM architectures.
7. ARC Challenge:
— Chollet is offering over $1 million in prizes for solutions to the ARC challenge.
— Current state-of-the-art performance on ARC is around 40–42%, significantly below human performance.
8. AGI Development:
— Chollet sees solving ARC as a necessary (but not sufficient) condition for achieving AGI.
— He plans to develop future versions of ARC to address more complex and open-ended problems.
The talk emphasizes the need to rethink how we approach AI development, focusing on generalization and abstraction abilities rather than just improving performance on specific tasks or benchmarks.
