Machines Love Patterns — We Love Chaos: AI for the Rest of Us

Introduction

AI can recognize impressive patterns, but the real world is full of exceptions, ambiguity and unexpected human behavior. In this episode of The Impact of AI Explored, we — Gerjon Kunst and James O'Regan — talk with Sairam Sundaresan about autonomous vehicles, AI explainability, hallucinations, RAG, agents and the hidden costs of AI.

The central question is simple: how do we make AI understandable for people who aren't programmers, data scientists or mathematicians?

Meet the Guest

By day, Sairam Sundaresan works with a team of engineers on systems for autonomous driving and parking. Alongside that, he focuses on AI education, helping professionals and organizations understand how AI models work, what they can and can't do, and how to judge their output.

He's the author of AI for the Rest of Us: An Illustrated Introduction. The book is written for the large group of people in between two extremes: people who don't write code and don't want to follow mathematical derivations, but do want to understand AI and make informed decisions. It explains things in plain language, with illustrations and humor.

Sairam also leads AI research for autonomous vehicles at Valeo and has experience in computer vision, deep learning and AI engineering. His public profile describes work on systems that help vehicles better understand their surroundings.

Setting the Stage

AI is now part of many workflows, but a large share of employees still don't know what that means for them in practice. They might use a chatbot, but don't understand why a model sometimes makes something up, why the same question produces different answers, or how to judge the quality of a result.

That gap between using AI and understanding it is becoming more important. AI isn't just being used for text anymore — it's also being applied to vehicles, recommendations, images, video, customer service and business processes.

In this blogpost, we look at the lessons from the conversation: why self-driving cars are so hard, why AI is inherently probabilistic, how RAG can ground answers, and why a real agent is far more complex than ordinary automation.

Episode Highlights

Machines love patterns, humans love ambiguity

Sairam sums up much of the difference between humans and machines in a single sentence: machines love patterns, while humans are good at dealing with ambiguity.

A driver can react to a subtle hand gesture, an unexpected traffic situation, or behavior that doesn't follow the rules. An AI system has to recognize such situations based on data and past examples. If a scenario is rare, it takes a lot of data to learn it reliably.

AI doesn't replace your judgment

According to Sairam, AI is primarily meant to make work easier and support human decisions. What's more, an AI project is never really "finished." Data changes, models get adjusted, users have new requirements, and competitors build different solutions.

That's why the most important human skill isn't just prompting — it's also judging, probing and verifying.

Deep Dive: From Patterns to Trustworthy Context

Why autonomous driving is so hard

A self-driving system has to perceive its surroundings, combine sensor data, recognize objects, predict behavior and make a decision. All of that happens under constantly changing conditions: rain, fog, darkness, reflections, roadworks, unexpected maneuvers and different driving styles.

On top of that, people don't drive the same way everywhere. According to Sairam, driving behavior and habits differ by country and culture. A system that learns patterns also has to learn which variations are normal and which situations become dangerous.

Then there's the safety question. When the system isn't sure about a decision, when should it hand control back to a human? How much time does that driver get to react? And what happens when the sensors themselves become unreliable?

Why models give different answers

Language models are probabilistic. They predict which word or sequence of words is likely to come next, based on the input and context. That means the same question can produce different answers, especially when the prompt contains little context.

That's not a random error — it's a property of the model. The user has an intention in mind, but that intention isn't automatically present in the prompt. The more relevant context you provide, the higher the chance of a useful answer.

RAG as a knowledge base

RAG stands for Retrieval-Augmented Generation. The model first retrieves relevant information from a knowledge base. That information is then added to the prompt, after which the model formulates an answer based on that context.

This is especially useful when information changes frequently. Instead of fully fine-tuning a model over and over, you just update the knowledge base. A customer service bot can use it to draw on current return and refund policy information.

RAG doesn't automatically make answers correct, but it increases the chance that they're based on real sources rather than a made-up pattern.

Groundedness

Groundedness describes how well an AI answer is supported by evidence. An answer can sound convincing while still referencing a book, study or product roadmap that doesn't actually exist.

A grounded answer points to sources or information that support the claim. That's why users shouldn't just ask whether an answer sounds good, but also what it's actually based on.

What a real AI agent needs

A language model normally gives an answer after someone enters a prompt. An agent goes further: it gets access to tools, can retrieve information, make a plan, and carry out multiple steps.

For a travel booking, for example, an agent needs to understand departure point, destination, dates and preferences, check websites, compare options, and ultimately make a reservation. That requires planning, external tools, oversight, and possibly access to payment.

That's why a real agent is more complex than an automation. It doesn't just need to execute — it also has to choose the right plan and know when it needs human approval.

Real-Life Stories & Examples

A rare situation during autonomous driving

Sairam describes an autonomous vehicle that performed impressively in many conditions, but got stuck on a combination of an obstacle and limited visibility. The vehicle had been trained on many situations, but a new edge case showed that even an advanced system can fail.

That's not a reason to write off AI entirely. It does show that data, testing and scenarios need to be designed carefully. A model can work excellently until it encounters a situation outside its experience.

A product roadmap that didn't exist

James shares that he once asked AI about a product roadmap. The model came back with features that didn't exist at all. Some ideas were interesting, but they didn't come from an actual roadmap.

This is a practical example of hallucination. An answer can be plausible and creative without being factually correct.

The customer service bot and RAG

A customer service bot without an up-to-date knowledge base can give an outdated or incorrect answer about return policies. With RAG, the bot can first retrieve current policy information and use that as context for its answer.

The organization still needs to make sure the knowledge base is accurate, and that customers can reach a human if they're in doubt.

Agents and a credit card

A fully autonomous travel agent could potentially book flights and hotels. But the moment an agent gets access to a credit card, oversight becomes essential.

A safer intermediate step is a human-in-the-loop model: the agent searches, compares and prepares the booking, after which the user confirms the final choice and payment.

The hidden costs of AI

Sairam names maintenance as an underestimated cost. An AI solution that turns out to be successful still needs to be scaled, managed, updated and adapted to changing data and user expectations.

Energy use matters too. Training and running models requires data center capacity and electricity. Every generated image or video has physical infrastructure behind it.

Key Takeaways

• AI works well with patterns, but the real world is full of exceptions.
• Autonomous driving is hard because of sensors, weather, human behavior, regulation and safety.
• Language models are probabilistic and don't always give the same answer.
• Good prompts contain enough context and clear intent.
• RAG pulls current information from a knowledge base and can reduce hallucinations.
• Groundedness measures how much of an answer is supported by evidence.
• A real agent needs planning, tools, external access and oversight.
• The term "agent" is often also used for simple automations.
• Human-in-the-loop is an important intermediate step toward more autonomy.
• AI projects are never fully finished: maintenance, data and evaluation remain necessary.
• Energy use and infrastructure are hidden costs of AI.
• The best way to avoid hype is to start from concrete problems.
• One key AI superpower is learning to ask better questions.

Closing Thoughts

What we take away most from this conversation is that AI shouldn't be seen as a magic machine that automatically understands what we mean. AI recognizes patterns, makes probability estimates, and can be enormously useful with the right context.

But people remain necessary to set goals, judge results and recognize exceptions. Maybe that's exactly the strength of human-machine collaboration: machines are good at scale, speed and patterns; humans are good at meaning, context and chaos.

Thanks to Sairam for the conversation, and for his mission to make AI understandable for everyone who doesn't need code or math to work with it sensibly. His book AI for the Rest of Us fits that mission perfectly: complex technology doesn't have to be intimidating to be taken seriously.

Our question to you: which AI application would you like to use, but don't fully trust yet? Let us know, and keep listening to The Impact of AI Explored.