GPT-5 to take AI forward in these two important ways

Generative AI systems might soon be capable of passing Ph.D. exams, according to Microsoft CTO Kevin Scott. During a recent Berggruen Salon event in Los Angeles, Scott discussed the advancements in AI memory and reasoning that could lead to such capabilities.

Durable Memory and Improved Interaction

Scott criticized the current state of AI memory, describing it as “episodic.” Presently, AI interactions are transactional, with no continuity between sessions. This means the AI forgets previous interactions, which limits its ability to learn and provide consistent support over time. Scott expressed optimism about the development of “durable memories” for AI systems, which would allow them to recall and build upon past interactions. This improvement would enable AI to respond more naturally and accurately over multiple conversations.

In February, OpenAI began testing a new persistent memory system with select users. This system allows the AI to remember user preferences regarding tone, voice, and format, and to make suggestions based on previous conversations.

 

Advancements in AI Reasoning

Scott also addressed the “fragility” in current AI reasoning, noting that many AI systems struggle with complex tasks, especially in mathematics. He suggested that future models could handle more sophisticated reasoning, comparing current models to high school students and predicting that next-generation AIs could perform at the level of Ph.D. candidates.

Despite the impressive achievements of current AI, such as GPT-4 passing the Multistate Professional Responsibility Exam (MPRE) with a score higher than the national average for humans, Scott downplayed the significance of training AIs to pass exams. He emphasized that the true measure of progress would be how these advanced capabilities are utilized.

 

Lowering Barriers to Entry

Scott highlighted the rapid reduction in barriers to entry for working with AI. Tasks that previously required extensive technical knowledge and months of coding can now be accomplished by high school students in a matter of hours. This democratization of AI technology is accelerating, making powerful tools accessible to a broader population.

Scott underscored the importance of this accessibility, stating that solving global challenges should not be the exclusive domain of tech companies or elite universities. With 8 billion people worldwide, many have valuable insights and ideas that could benefit from access to advanced AI tools.

The potential for generative AI systems to pass Ph.D. exams symbolizes the broader advancements in AI memory and reasoning. These developments are making AI more capable and accessible, which could democratize the technology and empower a diverse range of people to tackle complex global issues. As these barriers continue to fall, the future of AI looks promising, not just for tech experts but for anyone with innovative ideas and access to these powerful tools.