Demis Hassabis Discusses Future of Agent-Based AI Systems
–
AI developers are working hard to create smarter systems. They often do not share details about their training data. This makes it tough for others to prepare for future AI models.
Some studies suggest AI benchmarks drop when data is slightly changed or outdated. This raises questions about current AI benchmarks. Some experts believe that today's AI benchmarks can be solved through memorization. But true intelligence involves adapting and learning in new situations.
In 2019, AI researcher François Chollet created a new benchmark called the Abstraction and Reasoning Corpus (ARC). This benchmark aims to test AI's reasoning skills. By June 2024, AI models could solve 34% of ARC tasks. This was a big improvement from the initial 20%. Just recently, this number jumped to 46%.
The goal is to reach 85%. Once this happens, the methods used can be scaled up for other AI tasks. This will help move away from just using more data and more computing power. Instead, it will focus on better reasoning techniques. Some examples include Chain of Thought and neuro-symbolic AI.
OpenAI's focus on large language models (LLMs) has slowed progress toward Artificial General Intelligence (AGI). LLMs alone may not lead to AGI. It's important to explore other architectures.
The ARC prize is pushing AI research in new directions. It aims to find new ways to achieve AGI. Companies like DeepMind and Google are using similar techniques. They are not just scaling up LLMs. Instead, they are using methods like test-time fine-tuning and blast inference.
These methods have shown success in other areas. For example, Google's AlphaProof model won a silver medal at the Mathematics Olympiad. This shows the value of using new approaches rather than just relying on more data.
The focus is shifting towards better reasoning and understanding. This is different from just using LLMs and large datasets. These new techniques could lead to significant progress in AI.
Many believe that solving ARC is not the same as achieving AGI. But it is a step in the right direction. It shows that AI can improve its reasoning skills. Researchers continue to explore and test new methods. This will help bring us closer to AGI.
The recent advancements did not get much media attention. But they are important for the future of AI. The journey towards AGI involves looking beyond LLMs. It requires new ways of thinking and solving problems. This shift in focus is key to making real progress in AI.