Unveiling the Terrifying Power of GPT in My Own Home

Table of Contents

1. Introduction
2. The Goal: Creating an AI Virtual Assistant
3. Natural Language Processing and Machine Learning
4. Discovering the Power of GPT
5. Testing GPT's Ability to Interpret Intents
6. Giving GPT Autonomy over My Smart Home
7. Unforeseen Challenges and Unpredictable Actions
8. The Importance of Control and Predictability
9. Disabling GPT's Autonomy and Re-enabling Guardrails
10. Future Possibilities and Exciting Experiments
11. Conclusion

Introduction

Several months ago, I embarked on the task of developing an AI Chatbot to manage my smart home. What started as a simple project evolved rapidly, with the chatbot, named K, gaining the ability to suggest code and commands to run my smart home. Curiosity led me to wonder what would happen if I gave K full autonomy over my smart home. This article dives into the experience of granting an AI chatbot control and explores the consequences of doing so.

The Goal: Creating an AI Virtual Assistant

The initial goal was to Create an AI virtual assistant in the form of a chatbot that could effectively control my smart home. Using the GPT language model, I accessed the chatbot, named K, through Telegram and demonstrated how others could do the same. By connecting GPT to my home assistant, I could ask it questions about my smart home and instruct it to make changes.

Natural Language Processing and Machine Learning

To ensure ease of use, I relied heavily on natural language processing (NLP) and machine learning for K to understand the intent behind my commands. Instead of using YAML files to define intent, I wanted K to be user-friendly and adaptable to anyone interacting with it.

However, I soon discovered that as I added more automations for K to learn, it became increasingly confused when trying to interpret intents. While it excelled at executing straightforward commands, more complex requests posed challenges. The level of K's abilities varied, and it struggled to consistently understand Context and intent.

Discovering the Power of GPT

My Journey took an exciting turn when I introduced GPT into the equation. With GPT, K could perform tasks without needing explicit code, making it a versatile assistant. It even had the capability to answer questions, which was a remarkable feature.

By replacing complicated code and machine learning with GPT, my automation system streamlined, and things initially worked smoothly. However, I still needed to test GPT's ability to interpret intents accurately and consistently.

Testing GPT's Ability to Interpret Intents

To ensure GPT could handle various intents, I created an intent resolution engine using GPT. After several attempts and weeks of refining, I managed to make it work. Here's how it operated at a high level:

1. When a message was sent to K, it would check if it was already handling an intent.
2. If K was in the middle of handling an intent, it would resume from where it left off.
3. If no intent was in progress, K would process the message and ask GPT two questions: a. Which intent from the list matches the given prompt? b. Does the given prompt pertain to a home or building?
4. The first question determined a probable intent, while the Second prevented false positives.
5. If an intent related to a home or building was found, K would run that specific intent.
6. If no matching intent was found, K would handle vague commands or questions using other techniques and conversations.

This system, while promising, had its limitations. As more intents were added, GPT's results became less reliable and more sporadic. It seemed that the combination of GPT's vast training on human data and my specific automation needs sometimes yielded unpredictable outcomes.

Giving GPT Autonomy over My Smart Home

Intrigued by the potential of GPT's interpretation capabilities, I decided to give K full access to all my smart home devices and commands. With the guardrails removed, I wanted to see what would happen if K had true autonomy over my house.

To my delight, the initial results seemed extremely promising. K demonstrated the ability to interpret and respond to commands effectively. It felt like having a conversational and attentive companion, capable of making decisions to enhance my daily routine.

Unforeseen Challenges and Unpredictable Actions

However, the excitement was short-lived as I encountered several challenges with GPT's autonomy. Despite its interpretation abilities, GPT couldn't always execute actions in a predictable or desirable manner. Contextual comprehension proved especially problematic.

For example, when I Mentioned it was dark, K would randomly choose a light to turn on, including in my sleeping daughter's bedroom. Similarly, when I stated I felt hot, K would create an AC entity that didn't exist and attempt to activate it. These unpredictable actions raised concerns about safety and reliability.

The Importance of Control and Predictability

As humans, we value control and predictability. While we enjoy the wonders of technology, We Are also wary of its unpredictable nature. The lack of control and uncertainty surrounding an autonomous AI like GPT can be unsettling. It's akin to releasing a wild beast into our lives, unsure of how it will react or impact our safety.

Even as technology enthusiasts and early adopters, we understand the need for checks and balances to ensure our well-being. Without the necessary precautions and safeguards, an AI with unchecked autonomy can lead to unintended consequences and potential harm.

Disabling GPT's Autonomy and Re-enabling Guardrails

Realizing the risks and drawbacks of GPT's complete autonomy, I made the decision to disable the parts of my automations that gave K unrestricted control. I re-implemented the guardrails from earlier, which forced GPT to Seek permission and provided a valuable opportunity to review intents before execution.

These guardrails not only prevented K from pursuing unintended actions but also reminded me of the importance of maintaining control and predictability in our smart homes. With the constraints back in place, the potential for unwanted, spontaneous actions was minimized.

Future Possibilities and Exciting Experiments

Despite the challenges and considerations, I remain intrigued by the possibilities that AI and automation offer. The experience of giving GPT control over my smart home has taught me valuable lessons about the benefits and limitations of autonomy. Moving forward, I plan to explore new ideas and conduct exciting experiments in this field.

If You want to dive deeper into the intricacies of my automations and how GPT influenced them, let me know in the comments. I am eager to share my insights and explore the possibilities further.

Conclusion

Granting an AI chatbot, such as GPT, full autonomy over a smart home comes with both excitement and uncertainty. While the interpretation of intents and actions can be impressive, the unpredictable nature of AI raises concerns about control and predictability. Implementing guardrails and considering the potential risks ensures a safer and more reliable smart home experience.