AI Insights - AI Agents, Synthetic data, AI Infrastructure and Health Tips Delivered to Your Doorstep

In recent months, we have been hearing a lot about agents and the agentic era. But what exactly is an AI agent? Let's understand it through a simple analogy.

What Are Agents?

Agents are AI systems that can solve tasks on their own. There are two types:

1. Workflows: These follow predefined steps you code in advance. Think of them as step-by-step recipes.
2. Agents: These decide how to solve a task as they go. They’re more flexible but also more complex.

?When Should You Use Agents?

• Simple Tasks? Stick to basic AI calls or workflows.
• Complex Tasks? Use agents when: The problem isn’t predictable. You need flexibility, like customer support or advanced coding tasks.

Trade-offs:

• Agents are powerful but cost more and take longer.
• Workflows are simpler, faster, and cheaper for predictable tasks.

??Imagine This Task:

You want to plan a birthday party at Pizza Hut.

?Without an Agent (Fixed Steps):

1. Make a to-do list for the party.
2. Choose the Pizza Hut location nearby.
3. Decide the kids guest list.
4. Order the cake and customize the cake theme and plan for birthday food menu
5. Send invitations.

This is a straightforward workflow. This is following predefined steps that do not adjust dynamically if something changes.

?With an Agent:

An agent can figure out what to do on its own based on what you want.

• You tell the agent: “Plan a birthday party for 10 people at a Pizza hut with a new theme

?Here’s how the agent works:

1. Decides what needs to be done: Venue selection. Cake theme and size. Games and activities. Decorations.
2. Uses tools to complete tasks: Searches online for nearby pizza huts. Suggests themes for cake designs. Creates a kids guest list and sends invitations.
3. Handles unexpected things: If it finds out the pizza hut is unavailable, it looks for an alternative location. If any kid can't attend, it adjusts the kids guest list.

?As per Anthropic's blog about AI agents, let's explore the common patterns of agentic systems using the birthday analogy.

Tools and Their Simple Applications

1. Prompt Chaining

• What It Does: Breaks the task into smaller, manageable steps with each LLM call building upon the previous one
• Example: Step 1: Search for Pizza Hut branches that allow party reservations. Step 2: Confirm if the venue allows customized themes. Step 3: Plan the menu, keeping in mind kids' preferences. Step 4: Decide decorations and themed activities. Step 5: Book the venue and finalize details.

2. Routing

• What It Does: Directs specific tasks to the right resources or systems.
• Example: Task: Verify availability of Pizza Hut party rooms → Routed to the venue manager. Task: Design invitations → Routed to an online template service. Task: Arrange Harry Potter decorations → Routed to a party supply store.

3. Parallelization

• What It Does: Handles multiple tasks(running multiple LLM calls simultaneously) at the same time to save time.
• Example: While confirming the booking with Pizza Hut, the agent simultaneously: Orders themed cake. Sends invitations to guests.

?4. Orchestrator-Workers Pattern

• What It Does: Uses a main agent (orchestrator) to assign subtasks to specialized agents (workers).One LLM generates a response, and another evaluates and refines it.
• Example: The main agent plans the entire event and delegates: Venue-related tasks (like booking and menu customization) to one worker agent. Decoration planning to another worker agent. Activities (like magical games) to a third worker agent.

5. Evaluator-Optimizer

• What It Does: Evaluates multiple options and chooses the best one.
• Example: Evaluates different Pizza Hut branches based on location, availability, and capacity. Optimizes the menu to include both pizza and kid-friendly sides like garlic bread and desserts.

??Why These Tools Are Useful

• Efficiency: Tasks like booking, menu planning, and decoration arrangements happen faster through parallelization.
• Flexibility: If a Pizza Hut location is unavailable, the agent quickly finds an alternative.
• Autonomy: You don’t need to handle each step manually; the agent takes care of everything.

?

What AI Agents Are Lacking Compared to Humans (Birthday example):

• Humans can recognize when a child feels uncomfortable or shy and make adjustments to ensure they are included or engaged.
• An AI agent may not be able to recognize when a child is feeling left out or needs extra attention unless specifically programmed with emotional intelligence features.
• Humans can make personal decisions like adjusting the party theme or activities based on the child’s preferences and mood.
• An AI agent may stick strictly to predefined themes or ideas, lacking flexibility to change things based on individual children's interests.
• Humans can make real-time changes, like altering the party schedule if a child doesn't enjoy a game or activity.
• An AI agent might not adjust the schedule dynamically based on children's reactions unless specifically programmed to monitor and adapt to such cues.
• Humans know how to handle sensitive situations, such as if a child is upset or if there is a conflict between kids.
• An AI agent may not be equipped to deal with interpersonal issues that arise during the party, unless it has been integrated with conflict resolution systems.

?While AI agents are extremely efficient at handling the logistical aspects of party planning-such as booking venues, ordering supplies, and organizing tasks-they still fall short when it comes to the emotional intelligence, creativity, and social sensitivity that human planners bring to the table. The best solution might be a hybrid approach, where AI handles routine and repetitive tasks, while humans bring the personal touch and adaptability needed for a truly memorable event.

?I have collected a few good blogs and links that provide a clear explanation about AI agents :

https://www.anthropic.com/research/building-effective-agents

https://huyenchip.com//2025/01/07/agents.html

https://www.kaggle.com/whitepaper-agents

https://www.weforum.org/publications/navigating-the-ai-frontier-a-primer-on-the-evolution-and-impact-of-ai-agents/

https://arxiv.org/abs/2308.08155

OpenAI has released recently "AI in America OpenAI's Economic Blueprint" the article primarily focuses on establishing the United States as a global leader in artificial intelligence (AI).

?AI innovations today are heavily reliant on three critical pillars: infrastructure, compute, and energy. These pillars form the backbone of advancements in AI and are essential for ensuring the continued growth and accessibility of cutting-edge technologies.

Infrastructure

The development of robust infrastructure, such as data centers, chip manufacturing facilities, and renewable energy projects, is non-negotiable for supporting the compute-intensive demands of AI. Without this foundation, the ability to scale AI applications across industries and geographies is severely limited.

Compute

Compute power is the engine driving AI. From training large language models to running sophisticated machine learning algorithms, the demand for compute resources has skyrocketed. Reducing the cost of compute and ensuring its abundance will democratize AI development and allow more players to participate in this transformative field.

Energy

AI is an energy-intensive domain. Powering data centers, maintaining cloud services, and fueling innovation in compute requires sustainable and reliable energy sources. Investing in renewable energy projects, modernizing grids, and optimizing energy use are critical for sustaining AI’s growth without exacerbating environmental challenges.

The Role of Funding

All of this- infrastructure, compute, and energy-ultimately depends on funding. Global funds, estimated at $175 billion, are waiting to be channelled into AI infrastructure projects. The race to dominate AI is, therefore, also a race to secure and allocate these resources effectively.

US vs. China

The competition is intense. The United States and China are both vying for AI supremacy. While the US has a strong ecosystem of talent, innovation, and global partnerships, China’s centralized planning and rapid investments have positioned it as a formidable contender. Both nations recognize that leadership in AI will not only define technological progress but also shape economic and geopolitical landscapes.

?Source :

AI in America: OAI's Economic Blueprint [FINAL 010925]

AI is running out of data to train its models, and companies are turning to synthetic data as a solution.

Sam Altman on Synthetic Data

• Elon Musk in recent news said that AI has already used up all the useful human knowledge that’s available online to train its models.
• To solve this, AI companies are now turning to synthetic data-data created by AI itself—to keep improving these systems.

?

1. What’s Synthetic Data?

• Imagine a teacher (AI) running out of textbooks (human knowledge).
• To keep learning, the teacher starts writing its own textbooks. This is synthetic data-AI creating its own material to learn from.

?

2. Why Is This a Problem?

• Hallucinations: AI can sometimes make things up or get things wrong. If AI writes a textbook with errors, future AI could learn bad information.
• Model Collapse: If AI relies too much on its own synthetic data, it might become less creative or accurate over time, like a teacher who only teaches from their own flawed notes.

?

3. How Are Companies Handling It?

• Big companies like Meta (Facebook), Microsoft, and OpenAI are already using synthetic data to train their AI.
• But this raises questions about copyright (e.g., who owns the original data that AI uses) and quality control (how to avoid errors in synthetic data).

?

4. Why Does It Matter?

• AI Quality: To keep AI reliable, we need good data—not just synthetic data.
• Global Competition: Countries like the US and China are racing to dominate AI. Running out of training data could give one country an edge if they handle synthetic data better.
• Urgency: Experts predict there won’t be enough new public data by 2026, so the pressure is on to figure out solutions.

Summary:

• ?AI is running out of high-quality training data from human sources.
• Companies are turning to synthetic data, but it has risks (e.g., hallucinations and bias).
• The US and China are in a race to solve these problems and dominate AI.
• How we handle this challenge will shape the future of AI.
• Experts say the world might run out of new public data by 2026.
• Without solutions like synthetic data or better access to government and private data, AI progress could slow down.