This CatGPT video features an interview with Mustafa Suleyman, CEO of Microsoft AI and founder of DeepMind. The interview goes beyond typical podcast formats, incorporating Suleyman's passion for fashion as a unique angle to discuss his experiences in AI, its future, and the ethical considerations surrounding its development.
Mustafa Suleyman cites several limitations of current AI models:
Imperfect Memory: While progress is being made in information retrieval, AI lacks perfect memory, whether in terms of context (limited token capacity) or perfect recall ("needle in the haystack").
Inability to String Together Actions: Current models excel at one-shot predictions but struggle to execute complex tasks requiring numerous accurate and precise actions. This includes situations where an agent needs to adapt its actions based on previous results, request information from other agents or humans, or generate new data (code, images, etc.). Successfully completing such tasks demands near-perfect execution or robust mechanisms for error recovery.
Lack of True Scientific Understanding: Suleyman argues that current AI can "regurgitate" aspects of scientific curricula and answer questions, but it cannot independently formulate hypotheses, design experiments, interpret results, and meaningfully contribute to scientific knowledge—all crucial components of genuine scientific practice.
He contrasts these limitations with the capabilities needed for genuine scientific contributions, emphasizing the need for near-perfect execution of complex, multi-step actions, robust error management, and the capacity for original thought and innovation within a scientific framework. A "PhD-level physicist," Suleyman implies, requires far more than rote knowledge recall and simple question answering.
