Roula Khalaf, the Editor of the FT, shares her preferred articles in the weekly newsletter.
The capabilities of the leading large language model, GPT-4, are put to the test in various scenarios. Can it offer a solution to the Wordle game after being provided with descriptions of four previous guesses? Is it capable of crafting a biography-in-verse of Alan Turing, while substituting “Turing” with “Church”? This particular challenge arises from the relationship between Turing and his PhD supervisor, Alonzo Church, known for the Church-Turing thesis. Additionally, can GPT-4 analyze a partially completed game of tic-tac-toe and determine the optimal next move?
Nicholas Carlini, a researcher at Google Deepmind, has curated an engaging quiz on his website, exploring these intriguing questions and more. The quiz serves as a demonstration of both the remarkable abilities and limitations of GPT-4. While adept at tasks like integrating functions and wordplay, GPT-4 struggles with numerical concepts and the Wordle game.
One striking observation is that GPT-4 can swiftly generate a full javascript webpage for playing tic-tac-toe flawlessly against a computer, despite its inability to excel at the game itself. This disparity highlights the nuanced strengths and weaknesses of this advanced language model.
Carlini’s quiz underscores three key insights. Firstly, GPT-4 can tackle complex problems much faster than a human expert. Secondly, it exhibits shortcomings in areas where even a child might excel. Thirdly, distinguishing between these two categories of tasks can be challenging, even for seasoned users.
In a separate study by a team collaborating with Boston Consulting Group (BCG), consultants equipped with GPT-4 outperformed their counterparts without the tool. Tasks such as brainstorming product ideas, market analysis, and content creation were completed more efficiently and with higher quality by those utilizing GPT-4. This suggests that GPT-4 can significantly enhance the productivity of management consultants, particularly those with less experience.
However, the study also revealed a scenario where the AI faltered. When tasked with providing strategic recommendations based on financial data and staff interviews, the AI’s advice proved to be subpar. This emphasizes the importance of human judgment in certain decision-making processes that involve nuanced interpretation.
The interplay between human expertise and generative AI performance represents a complex frontier. While the AI can excel in many areas, there are instances where human intuition and understanding outshine its capabilities. Navigating this “jagged frontier” requires a nuanced approach to leverage the strengths of both AI and human intelligence effectively.
Drawing parallels to past technological advancements, such as the introduction of the iPhone, offers insights into the potential impact of generative AI. Just as smartphones revolutionized daily life, generative AI has the capacity to reshape how we approach tasks and decision-making. Understanding when to rely on AI tools and when to trust human judgment will be crucial in maximizing their benefits.
In reflection, the ubiquity and dependence on smartphones serve as a reminder of how quickly technology can integrate into our lives. As we navigate the evolving landscape of AI, it is essential to maintain a balance between leveraging its capabilities and recognizing the value of human insight and decision-making.
Tim Harford’s latest book for children, ‘The Truth Detective’ (Wren & Rook), is now accessible. Follow @FTMag for the latest updates on our stories.
