The areas of artificial intelligence (AI) and machine learning (ML) continue to advance, but they are not without obstacles. A classic example is the performance degradation colloquially referred to as ‘stupidity’ in large language models (LLMs) such as GPT-4. This issue has gained attention in AI discussions, especially since the publication of “Work Pollution: Language Models May No longer be Few-Shot,” which highlights the limitations and challenges currently facing LLM.
Chomba Bupe, a representative figure in the AI community, highlighted X (formerly Twitter) has a major problem. LLMs tend to excel on the tasks and datasets they are trained on, but tend to falter on new, unseen data. The crux of the problem lies in the static nature of post-training in these models. Once the learning phase is complete, performance gradually deteriorates due to limited ability to adapt to new and evolving input distributions.
Source: DALL·E Generation
This performance degradation is of particular concern in areas such as programming, where language models are used and programming language updates occur frequently. Bupe points out that the basic design of the LLM is closer to memorization than understanding, which limits its effectiveness in solving new challenges.
Research conducted by Changmao Li and Jeffrey Flanigan further supports this view. They found that LLMs like GPT-3 outperform on older data sets than on training data. This finding is indicative of a phenomenon called task contamination, where a model’s zero-shot and few-shot features are compromised by limitations in the training data.
Continuous learning, as discussed by Bupe, emerges as a key area of machine intelligence. The challenge is to develop ML models that can adapt to new information without compromising performance on previously learned tasks. this difficulty Contrast this with the adaptability of biological neural networks, which learn and adapt without similar drawbacks.
Alvin De Cruz offers an alternative perspective that suggests that the problem may lie in the evolving expectations of humans rather than in the inherent limitations of the model. But Bupe responds by highlighting the long-standing nature of these challenges in AI, particularly in the area of continuous learning.
In summary, the conversation surrounding LLMs like GPT-4 highlights an important aspect of AI evolution: the essentials of models capable of continuous learning and adaptation. Despite its impressive capabilities, LLMs currently face significant limitations in keeping pace with a rapidly changing world, highlighting the need for more dynamic and evolving AI solutions.
Image source: Shutterstock