The future of AI-Agents // will be as normal as a smartphone today
What’s an Agent?
Technically, it’s an AI system composed of many things: large multi-modal/language models, semantic/conventional databases, code generators/executors, different types of memories, external API adapters, web search modules, etc. It can somehow be named a digital brain, why not.
For sure, LLMs are the core components of agents. Thanks to LLMs, we can create reasoning pipelines almost similar to how people do.
ChatGPT generated the definition of an AI agent based on LLMs:
An AI agent based on Large Language Models (LLMs) and a vector database, with the ability to generate and execute code, leverage short and long-term memory, and utilize external APIs and web search, represents a sophisticated system designed for intelligent information processing and task execution. The Large Language Model serves as the core component, enabling natural language understanding and generation. The vector database complements this by providing a structured and efficient way to represent and retrieve information. The agent’s capability to generate and execute code implies a proficiency in automating tasks and problem-solving. The utilization of short and long-term memory allows the agent to store and recall information strategically, enhancing its adaptability and learning over time. Integration with external APIs and web search capabilities expands the agent’s access to real-time data and external services, broadening its scope of tasks and responses. This comprehensive set of features positions the AI agent to excel in diverse applications, from complex problem-solving to dynamic data-driven tasks.
There are a parade of papers and frameworks related to prompts, reasoning and agents:
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
A Survey of Chain of Thought Reasoning: Advances, Frontiers and Future
Implicit Chain of Thought Reasoning via Knowledge Distillation
Contrastive Chain-of-Thought Prompting
Chain of Thought Prompting Elicits Knowledge Augmentation
AI performance benchmarks: MMLU, TriviaQA, HellaSwag, GSM8k, more at huggingface open_llm_leaderboard
It’s already possible — an autonomous agent that can act on its own
Why it matters: The White House is preparing an AI executive order, expected in coming weeks.
Suleyman predicted AI 100 times more powerful than today’s models within several years.
The big picture: “We haven’t invented things before that can act independent of humans,” Suleyman said, insisting this requires new business and regulatory strategies.
What they said: “Everyone is going to have a personal intelligence,” Suleyman said.
Examples of Agnets
