LLMs are much worse than humans at learning from experience

The article discusses the potential for transformer-based AI models to achieve human-like reasoning abilities, highlighting differing opinions on whether scaling up existing models or augmenting them with problem-solving capabilities is necessary. The author expresses skepticism about current models like OpenAI's o1 and suggests that significant breakthroughs are needed for AI to reason like humans. The article also explores the limitations of large language models (LLMs), which struggle to learn new concepts at inference time, unlike human brains that continuously learn from experiences.

The article further examines Steve Newman's approach to solving a challenging math problem from the International Math Olympiad (IMO) and compares it to DeepMind's AlphaProof model. Newman used a combination of inference and training, developing intuition through exploration and pattern recognition. In contrast, AlphaProof employs a language model trained on millions of proofs and uses a tree-search approach to explore possible solutions, akin to chess AIs. The article highlights the differences in problem-solving strategies between human intuition and AI's systematic exploration.

Key takeaways:

There is ongoing debate about whether transformer-based AI models can achieve human-like reasoning, with some believing scaling up models is enough, while others think additional capabilities are needed.
Current large language models (LLMs) struggle to learn new concepts at inference time, unlike human brains which continuously learn from experiences.
Steve Newman, a former math prodigy, explored how AI systems approach complex math problems, highlighting the difference between human intuition and AI's systematic approach.
DeepMind's AlphaProof uses a tree-search approach to solve math problems, attempting numerous paths to find a valid proof, similar to chess AIs.

LLMs are much worse than humans at learning from experience

Key takeaways:

Comments (0)

Newsletter