Mustafa Suleyman Proposes an ACI Approach to Bridging the Gap Between Weak AI and AGI

The proposal, known as ACI (Artificial Capable Intelligence), aims to bridge the gap between weak AI and AGI (Artificial General Intelligence). Rather than solely measuring a model’s ability to communicate, ACI seeks to evaluate its capacity to interact with the world and accomplish specific tasks.

Suleyman introduces a thought-provoking example: tasking an AI model with making a million dollars on an online retail platform starting with an initial investment of $100,000. To successfully tackle this challenge, the model must not only devise a plan of action, but also engage in a multitude of related activities. These may include product ideation, communication with manufacturers and suppliers, contract negotiations, and strategic marketing.

According to Suleyman, achieving ACI could be possible within a couple of years. He believes that this milestone will significantly impact the global economy, propelling us from a phase where AI is useful for solving specific problems to one where AI becomes a central component of the entire economic landscape.

Comparing models across different tasks in such a setting remains somewhat unclear. Additionally, assessing such models poses challenges when their evaluation requires real-world interactions with people. For instance, in the aforementioned example, the model must generate genuine revenue by managing real money, selling tangible products to actual customers, and engaging in numerous human interactions. The testing process could potentially consume significant human resources and financial investments.

Mustafa Suleyman’s forthcoming book, titled “The Coming Wave: Technology, Power and the Twenty-First Century’s Greatest Dilemma,” promises to delve further into these ideas and offer insights into the evolving landscape of AI. Scheduled for release on September 5, the book will likely provide valuable perspectives on the potential implications of ACI.

To learn more about this concept, you can explore the article published in MIT Technology Review.

• In 1950, Alan Turing proposed the Turing test to measure a machine’s intelligence. In December 2022, ChatGPT, an artificial intelligence chatbot, became the second to pass the test, according to Max Woolf, a data scientist at BuzzFeed. ChatGPT’s performance in the Turing test was impressive, as it convincingly mimicked human conversation to a panel of judges. The next steps for ChatGPT include integrating the bot with other platforms like Facebook Messenger or Slack, expanding its capabilities, and adding more data to its training dataset.
• The Turing test has faced scrutiny as technologies like ChatGPT and GPT-4 integrate into online communities. A gamified experiment called “Human or Not?” tested users’ ability to distinguish between human and AI interactions. Over 1.5 million users participated in a web game, where they had to interact anonymously with a language model like GPT. The experiment revealed that only 68% of participants correctly identified their chat partners, while the success rate dropped to 60% when interacting with the language model. The findings highlight the creativity and adaptability of individuals trying to blur the lines between human and machine interactions. However, the Turing test’s reliability raises questions about the reliability of AI emergence. The “Human or Not?” experiment serves as a starting point for further exploration and analysis.

Read more about AI: