Amazon Leads the AI Race by Focusing on Developing Intelligent Agents: David Luan's Vision and the New S-Curve

The Concept of the S-Curve and Intelligent Agents

Why does David Luan, head of labs at Amazon, Artificial General Intelligence (AGI) in Amazon, believe that agent development represents the new "S-Curve" for Artificial Intelligence?

The "S-Curve" is known as a model that describes the technology adoption lifecycle, starting with slow initial growth, followed by rapid expansion and widespread adoption, then a stabilization phase. In the context of Artificial Intelligence, especially Generative AI, it is seen as ushering in a new and different adoption curve (Forbes, 2024) (Medium, 2023).

The concept of Intelligent Agents is considered one of the most prominent current topics in the field of Artificial Intelligence. Interest is focused on developing AI systems to go beyond being mere chatbots, to become capable of accomplishing complex tasks effectively and with high reliability in the real world. Despite progress, agent reliability remains a fundamental challenge.

David Luan and His Vision at Amazon

Many companies are competing in the industry of Artificial Intelligence to address agent reliability challenges, and at the heart of this race stands David Luan, head of research lab for Artificial General Intelligence (AGI) in Amazon. Luan has a long history in this field, having been one of the first research leaders at OpenAI, effectively contributing to the development of GPT-2, GPT-3, and DALL-E models. After OpenAI, he co-founded Adept, a research lab specializing in Artificial Intelligence particularly focused on Intelligent Agents. Last summer, Luan joined Amazon, where he currently leads the company's Artificial General Intelligence lab in San Francisco.

This interview was recorded shortly after the launch of GPT-5 from OpenAI, providing an opportunity to discuss David Luan's vision on the slowdown of progress in AI models. David's team's work is a top priority for Amazon, and this interview serves as the first clear reveal of what they have been working on.

Luan also touched upon how he joined Amazon. His decision to leave Adept is an early example of what is known as "reverse acquisition," where major tech companies acquire promising startups in the field of Artificial Intelligence to avoid antitrust scrutiny. Regardless of the details, David's transition from the startup world to tech giants last year reflects his deep understanding of the AI race's trajectory. Therefore, his predictions for the future hold significant importance.

The Evolution of Artificial General Intelligence (AGI) and Agents

David Luan defines Artificial General Intelligence (AGI) as a system capable of assisting humans in accomplishing any task they desire on a computer, considering this a realistic and user-centric goal. He indicates that the evolution in Artificial Intelligence models has become convergent, and traditional benchmark tests are no longer decisive, as all parties are moving towards a similar performance point. Instead, he emphasizes that the most important aspect is how people interact with and use these systems.

via GIPHY

Luan emphasizes that the concept of Artificial General Intelligence must go beyond mere text conversations or writing programming code. He asserts that Intelligent Agents will represent the fundamental pillar of future computing, and Amazon's focus on developing agent solutions stems from its realization of the immense economic potential they will unlock, aligning with its strengths in cloud computing and infrastructure.

Luan recognizes that many users have experienced Intelligent Agents and found them not to perform tasks with the required efficiency, often describing them as mere "chatbots with extra steps." However, he presents an ambitious vision for a true intelligent agent, for example, a model that can connect to a biological lab to independently conduct drug discovery experiments, analyze relevant scientific literature, propose the most effective experiments, execute them, then analyze the results, and iterate until the desired goal is achieved. This advanced level of effectiveness will grant humans immense and unprecedented capabilities.

via GIPHY

Amazon's Methodology for Training Agents

Luan explains the fundamental difference between Large Language Models (LLMs) and Intelligent Agents. While LLMs learn by predicting the next word, known as "behavior cloning," agents require a deep understanding of the true causal mechanism, i.e., the ability to infer the relationship between causes and consequences (if X is performed, then Y will be the consequences). To achieve this understanding, Amazon widely trains agents using a "self-play" methodology within simulated environments, called "RL gyms." These environments are designed to simulate cognitive tasks performed by employees in various sectors, such as the Salesforce platform or Computer-Aided Design (CAD) software. Through repeated trial and error in these environments, the agent gains a clear understanding of the consequences of its decisions and actions.

This innovative approach is considered based on large-scale training through self-play in simulated environments, relatively unique within the sector of Artificial Intelligence, and Amazon adopted it based on the previous work of the Adept team. Luan believes this represents the "crucial missing piece" for achieving Artificial General Intelligence true. When asked about the impact of this work on Amazon products, Luan explained that Alexa Plus, for example, uses these techniques to enable a web agent capable of ordering plumbing services via platforms like Thumbtack. He also indicated that Amazon leverages its vast collection of internal data and its own operational environments, which reflect the operations of nearly all Fortune 500 companies, to train smarter and more efficient agents, asserting that free web data is no longer sufficient for effectively training Artificial Intelligence agents with the required effectiveness.

Applications and Efficiency of Intelligent Agents

In a related context, Nova Act, which was launched as a research preview in March, achieved a reliability rate exceeding 95% in various enterprise applications. This represents enormous progress compared to the average reliability of competing products in the market, which is around 60%. The system Nova Act is currently efficiently used for tasks such as registering doctors and nurses, automating travel bookings, and managing complex quality assurance workflows.

via GIPHY

Luan expects that the "GPT moment" for Reinforced Learning agents will be achieved in less than a year, with the team continuing to make rapid progress towards this achievement. He believes that the pace of AI development will not slow down but will move to a new "S-Curve," where Intelligent Agents form the core component of the next qualitative leap in the sector. He asserts that Amazon aims to be a leader in this field by focusing on this innovative methodology for training agents, leveraging its immense capabilities and scale of operations.

The Future of AI and the Talent Market

In conclusion of the interview, Luan addressed aspects of the talent market in the sector of Artificial Intelligence and the phenomenon of "reverse acquisition." He emphasized the necessity for companies to gather a "critical mass" of competencies and computing power to ensure victory in the race for Artificial Intelligence. He explained his decision to join Amazon as stemming from the company's strong desire to excel in the field of agents and its readiness for the massive investment required to achieve this, especially given the enormous increase in computing and talent acquisition costs.

Luan offered valuable advice for those interested in working in the sector of Artificial Intelligence, including joining small research and development teams with broad access to computing resources, and focusing on companies that have a deep product understanding and a clear vision for how to integrate Artificial Intelligence solutions radically into users' lives, and avoiding limiting themselves to developing traditional chatbots. He indicates that there are many "fundamental product forms" whose challenges have not yet been solved, and which will become essential and clear in the near future.