(Source: MIT Technology Review)
AI agents are currently a hot topic in the tech field. From Google DeepMind and OpenAI to Anthropic, major companies are competing to give LLMs 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 ranging from NVIDIA to Salesforce are exploring how to leverage this technology to disrupt industry norms.
Sam Altman stated in a blog post last week, “We believe the first AI agents may join the workforce in 2025 and significantly change business outputs.”
Broadly speaking, AI agents are software systems capable of autonomously executing tasks, often requiring little to no supervision. The more complex the tasks, the higher the level of intelligence required from the agent. For many, today’s LLMs are already smart enough to drive agents to complete a range of practical tasks, such as filling out forms, finding recipes, adding required ingredients to an online shopping cart, or quickly conducting last-minute research and generating concise summaries before meetings.
In October 2023, Anthropic showcased an advanced AI agent: an extension feature of Claude called “Computer Use.” As the name suggests, this feature allows users to guide Claude in operating a computer as if directing a human, including moving the cursor, clicking buttons, and entering text. Users can not only converse with Claude but also request it to perform tasks directly on the screen.
Anthropic noted that this feature still has issues with redundancy and errors, but has been opened to a small group of testers, including third-party developers from companies like DoorDash, Canva, and Asana.
This project showcases a glimpse into 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 AI agents’ development by 2025.
(Kaplan’s responses have been lightly edited for brevity and clarity.)
Better at Using Tools
“I think AI’s capabilities can be understood from two dimensions. One is the complexity of tasks that AI systems can complete. As systems become increasingly intelligent, their capabilities in this regard improve, while another key dimension is the types of environments or tools that AI can use,” Kaplan said.
“For example, if we look back at DeepMind’s Go model AlphaGo from nearly a decade ago, it surpassed humans in playing the game. However, if AI’s capabilities are limited to board games, its application environment is very constrained. Even with high intelligence, such AI has no practical use in the real world. With advancements in text models, multimodal models, and features like ‘Computer Use’ (which may even involve robotics in the future), AI is gradually integrating into diverse environments and tasks, making it 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 limited to specific environments. I believe ‘Computer Use’ will rapidly enhance the model’s ability to complete various and more complex tasks. Additionally, AI will gradually learn to identify its own mistakes and recognize when it needs to seek user feedback in the face of high-risk issues,” he pointed out.
Better at Understanding Context
Claude needs to have a sufficient understanding of your specific situation and constraints to truly be effective for you. This includes what your specific role is, what your writing style is, and what the needs of you and your organization are.
“I think 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 needs to be not only useful but also safe, accurately completing the tasks you expect is very important,” Kaplan stated.
Moreover, many tasks do not actually require Claude to engage in extensive reasoning. For instance, there’s no need for prolonged contemplation before opening Google Docs. “Thus, I believe the future direction is not only to enable agents to have more reasoning capabilities but to ensure that they reason only when truly necessary while avoiding wasting time when it’s not needed,” he said.
Enhancing Code Assistant Capabilities
“We hope to provide a beta version to developers while the ‘Computer Use’ feature is still in its early stages to collect feedback. As these systems continue to improve, they may be more widely adopted and genuinely collaborate with users to complete various tasks,” he said.
Companies like DoorDash, Browser Company, and Canva are experimenting with different types of browser interactions while leveraging AI to design these interactions.
“I expect code assistants to further improve, which is an exciting area for developers. Nowadays, programming with Claude 3.5 has generated immense interest. It no longer merely offers autocomplete features but can actually understand problems in the code, debug it, observe its behavior through running the code, and make fixes,” he stated.
Needs to Be Designed More Safely
“The reason we founded Anthropic was due to the foresight of the rapid development of AI technology and the realization that safety issues would inevitably become increasingly important. I believe these issues will become more intuitive and urgent this year, as AI agents increasingly integrate into our work, and we need to prepare for various potential challenges, such as prompt injection,” Kaplan stated.
Prompt injection is an attack method 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 may access to carry out the attack.
“Prompt injection may be one of the key issues we need to prioritize in the widespread use of AI agents, especially for the ‘Computer Use’ feature, this issue is particularly important. ‘We are actively addressing this challenge because if ‘Computer Use’ is deployed on a large scale, malicious websites or other means may attempt to induce Claude to perform inappropriate actions,” he said.
“Additionally, as more advanced models emerge, the associated risks will also increase. We have developed robust scaling strategies to ensure that as AI systems’ capabilities improve, they can be effectively prevented from being abused. For example, preventing it from being used to assist terrorists or for 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, particularly in programming, with remarkable results. However, the challenges cannot be overlooked. This year will undoubtedly be a year full of changes and new opportunities,” Kaplan stated.
Original link:
https://www.technologyreview.com/2025/01/11/1109909/anthropics-chief-scientist-on-5-ways-agents-will-be-even-better-in-2025/
