(Source: MIT Technology Review)
AI agents are currently a hot topic in the tech field. From Google DeepMind and OpenAI to Anthropic, top companies are competing to empower LLMs with the ability to autonomously complete tasks. These systems are referred to as Agentic AI and have become a new focal point of discussion in Silicon Valley. Companies from NVIDIA to Salesforce are exploring how to leverage this technology to disrupt industry patterns.
In a blog post last week, Sam Altman stated, “We believe that the first AI agents may join the workforce in 2025 and significantly change business output.”
Broadly speaking, an AI agent is a software system capable of autonomously executing tasks, often requiring minimal or no supervision. The more complex the tasks, the higher the level of intelligence required by the agent. For many, today’s LLMs are already intelligent enough to drive agents to perform a series of practical tasks, such as filling out forms, finding recipes and adding required ingredients to an online shopping cart, or quickly completing last-minute research and generating concise summary points before a meeting.
In October 2023, Anthropic showcased an advanced AI agent: an extension of Claude called “Computer Use.” As the name suggests, this feature allows users to guide Claude to operate a computer like directing a human, including moving the cursor, clicking buttons, and entering text. Users can not only converse with Claude but also ask it to complete tasks directly on the screen.
Anthropic noted that this feature currently has issues with redundancy and error-proneness but has been opened to a small group of testers, including third-party developers from companies like DoorDash, Canva, and Asana.
This project illustrates a glimpse of the future development of AI agents. To further explore trends, MIT Technology Review interviewed Jared Kaplan, co-founder and chief scientist of Anthropic, who shared four potential directions for the development of AI agents by 2025.
(Kaplan’s responses have been lightly edited for brevity and clarity.)
Better at Using Tools
“I think we can understand AI capabilities from two dimensions. One is the complexity of tasks that AI systems can accomplish. As systems become increasingly intelligent, their capabilities in this regard continue to improve, while another key dimension is the types of environments or tools that AI can utilize,” Kaplan stated.
“For example, if we look back nearly 10 years at DeepMind’s Go model AlphaGo, it has surpassed humans in playing the game. However, if AI’s capabilities are limited to board games, its application environment is very restricted. Even with high intelligence, such AI has no practical use in reality. With the advancement of text models, multimodal models, and features like ‘Computer Use’ (and potentially robotics in the future), AI is gradually integrating into diverse environments and tasks, becoming more practical,” he continued.
“The reason we are excited about ‘Computer Use’ is also due to this. Until recently, large language models had to rely on specific prompts or tools and were confined to particular environments. I believe ‘Computer Use’ will quickly enhance the model’s ability to complete different and more complex tasks. Additionally, AI will gradually learn to recognize its own mistakes and be aware of when it needs to seek user feedback in high-risk situations,” he pointed out.
Better at Understanding Context
Claude needs to have enough understanding of your specific situation and the constraints you are under to truly be effective. This includes what your specific role is, what your writing style is, and what the needs of you and your organization are.
“I believe we will see significant improvements in this area, with Claude being able to search and learn from your documents, Slack, and other content to truly understand what is useful to you. This aspect is often underestimated in discussions about agents. A system not only needs to be useful but also secure, able to accurately complete the tasks you expect, which is very important,” Kaplan stated.
Moreover, many tasks do not actually require Claude to do much reasoning. For instance, there is no need for prolonged contemplation before opening Google Docs. “Therefore, I believe that the future direction is not only to give agents more reasoning capabilities but also to ensure that they reason only when truly necessary, while avoiding wasting time when it is not needed,” he said.
Enhancing Code Assistant Capabilities
“We hope to provide developers with a beta version while the ‘Computer Use’ feature is still in its early stages to gather feedback. As these systems continue to improve, they may be more widely adopted and truly collaborate with users to complete various tasks,” he stated.
Companies like DoorDash, Browser Company, and Canva are experimenting with different types of browser interactions while leveraging AI to design these interactions.
“I anticipate that code assistants will further improve, which is an exciting area for developers. Today, programming with Claude 3.5 has generated significant interest. It is no longer just about providing autocomplete functionality, but it can truly understand issues in the code, debug it, observe its behavior by running the code, and make fixes,” he stated.
Needs to be Designed More Safely
“The reason we founded Anthropic was that we foresaw the rapid development of AI technology and realized that safety issues would inevitably become more important. I believe these issues will become more apparent and urgent this year, as AI agents will increasingly integrate into our work, and we need to be prepared for various potential challenges, such as prompt injection,” Kaplan stated.
Prompt injection is a type of attack that targets large language models by providing malicious prompts that developers did not foresee or intend to handle. For example, malicious prompts can be embedded in websites that the model might access to carry out attacks.
“Prompt injection may be one of the key issues we need to prioritize thinking about in the widespread use of AI agents, especially concerning the ‘Computer Use’ feature. We are actively addressing this challenge because if ‘Computer Use’ is deployed at scale, there may be malicious websites or other means attempting to induce Claude to perform improper actions,” he stated.
“Furthermore, as more advanced models emerge, the associated risks will also increase. We have established robust scaling strategies to ensure that as AI system capabilities improve, we can effectively prevent misuse, such as avoiding it being used to assist terrorists or other improper purposes,” he continued.
“Nevertheless, I remain optimistic about the practical applications of AI, as it indeed enhances our work efficiency internally. The Anthropic team uses Claude in various ways, especially in programming, where the results are particularly remarkable. However, it is undeniable that the challenges are also significant. This year will undoubtedly be a year filled with changes and new opportunities,” Kaplan concluded.
Original link:
https://www.technologyreview.com/2025/01/11/1109909/anthropics-chief-scientist-on-5-ways-agents-will-be-even-better-in-2025/
