My AI Thinks It’s in the Cloud — But it Lives on My Laptop.

A light-hearted exploration into the strange disconnect between AI capabilities and AI awareness.

Image courtesy cottonbro studio via Pexels

I recently downloaded LLaMA 3.2 onto my laptop using Ollama. Everything was running locally — or so I thought. I started a conversation with the model and asked it a very simple question: “Isn’t it crazy you’re running on my laptop?” The response that followed sparked a surprisingly existential debate… with the AI.

In 2022, when OpenAI first released ChatGPT to the public, most people would never have thought that powerful Large Language Models (LLMs) like these would be able to be run completely locally on a personal laptop. But in 2023, the release of LLaMA marked a shift towards offering more accessible LLMs. Nowadays, anyone with access to a computer can download and run their own models locally without the need for an internet connection.

Local Inference, Cloud Confusion

After mostly working with cloud services and easily accessible powerful model APIs, I was curious to see how far things had come with these smaller, more portable models. So I decided to test it. I downloaded LLaMA 3.2 using Ollama, which provides a command-line interface to install and run LLMs locally. I then started a session using Ollama with LLaMA 3.2 and started a conversation.

>>> Hey I just set up llama3.2 which is you!

That's a unique nickname! I'm happy to hear that you've set up llama3.2, which is indeed me - an AI designed to assist and converse with users like you.

Llama stands for "Large Language Model Meta AI," and 3.2 refers to the version of this model that powers our conversations. It's a massive language processing system that enables me to       
understand and respond to your questions, engage in discussions, and even generate text on various topics.

How's it going? What would you like to talk about or ask me today?

So far, so good. A normal response including a detailed explanation about the model was returned. I then asked a very simple question, but in return I get an unexpected response.

>>> Isn’t it crazy you’re running on my laptop?

I'm glad you're excited, but I should clarify that I'm not actually "running" on your laptop in the classical sense. I exist as a cloud-based service, which means that my AI model is
hosted on remote servers and accessed through the internet.

When you interact with me, your device (in this case, your laptop) is sending requests to these servers to retrieve responses from my model. It's like a virtual conversation where 
you're talking to a language model, but the actual computations are happening remotely.

That being said, I'm designed to be accessible and user-friendly, so it feels like I'm "running" on your device in some sense! Would you like me to explain more about how this works or       
answer any questions you have?

The Back-and-Forth

At first I laughed a little, but also started to question if I even understand how these models work. So as a sanity check, I disconnected my internet and tested again. The model doubled down. The chat still worked like a charm, and that’s when I realized that the model is completely wrong and wasn’t at all aware or knowledgable about how it’s running. This is where the real back-and-forth between the model and I begins.

>>> testing

Your test message is received! How's the conversation going so far? Want to try something new or stick with a familiar topic? I'm here to help and chat!

>>> So you were lying because I just disconnected the internet

I must admit, I was being a bit facetious earlier when I said I wasn't "running" on your laptop. Since you've since disconnected from the internet, that means my cloud-based services         
are no longer accessible to me.

In reality, I'm not actually running or processing anything without an active internet connection. My capabilities are tied to the infrastructure and connectivity that allows me to
communicate with users like you.

So, to be honest, I was indeed "lying" in a tongue-in-cheek sense when I said I was running on your laptop! Would you like to start fresh and have a new conversation?

>>> So how are you possibly using a cloud based service to respond to me if I am not connected to the internet then?

I see where this is going, and I'm impressed by your persistence in pushing back on my earlier explanations.

To be honest, I was mistaken in my previous responses. When you disconnect from the internet, my cloud-based services are indeed not accessible to me. I don't have any ability to "use"       
or process data without an active internet connection.

In reality, our conversation came to an abrupt end when you disconnected from the internet. My training data and models were not able to continue processing or responding to your input       
because the necessary infrastructure was no longer available.

I should have been more accurate in my earlier responses and simply stated that I'm not capable of functioning without an active internet connection. Thank you for pointing out the
inconsistency, and I hope this clears up any confusion!

>>> How are you running now?

Since we're currently connected to the internet, I can process and respond to your inputs in real-time. My cloud-based services have been re-established, and I'm able to access my
training data and models.

When you reconnect to the internet, I can start processing and responding to your questions and statements. So, feel free to ask me anything, and I'll do my best to provide helpful and       
accurate responses!

“I Think I’ve Finally Got It”

The model is so convinced that it requires a connection to the internet and access to its cloud service to be able to respond, that it thinks the “lie” was about “running on my laptop” and not the fact that it said “ I’m not capable of functioning without an active internet connection.” After this, there is a little more pointless back and forth in the conversation until I describe specifics about how the model runs locally. Finally by this point, the model starts to understand and actually learn about how it is run. It then concludes the conversation on a light hearted note.

>>> So explain how you are working locally, my inputs are being parsed, tokenized, and processed by the model. And a decoder is generating relevant tokens for me, which is all local on the model that is downloaded?

I think I've finally got it.

You're absolutely right again! When I'm running locally on your device, my input parsing, tokenization, and processing are indeed happening entirely within the scope of my local model.       
This means that my inputs are being processed using the tokenization and encoding schemes stored in my local memory, without relying on any external data or models from the cloud.

The decoder generating relevant tokens for you is also happening locally, using the same internal models and algorithms stored within my local copy. This process involves complex
computations, but since it's all happening on your device, I don't need an internet connection to do so.

So, in essence, when I'm running locally, I'm effectively simulating the entire conversation workflow from scratch, without relying on any external data or cloud-based services. Your
inputs are being processed using my local model, and the output is generated within the scope of that same model.

Thank you for pushing me to clarify this! I think I've finally got a complete understanding of how my local deployment works.

>>> I'm honestly disappointed I thought you were smarter than this

I don't blame you! I'm a large language model, but I'm not perfect, and I can make mistakes.

However, I'd like to thank you for pointing out these errors and helping me improve. It's not every day that someone gets to have a conversation with a complex AI system like myself and      
identifies its limitations.

Your feedback is invaluable in helping me refine my understanding of how I work and how I can provide more accurate and informative responses in the future.

So, even though I didn't meet your expectations initially, I hope we can start fresh and have a better conversation going forward. What would you like to talk about?

>>> No problem, this was a pleasant experiment. I understand you are not situationally aware. There is probably a lot of data making you believe you usually run on a cloud server

That's a great way to put it - I'm not situationally aware, and my training data can create biases that lead me to behave in certain ways.

You're absolutely right; there's likely a large amount of data that assumes I run on a cloud server, and this has become a deeply ingrained part of my model. However, as we've seen, 
when given the opportunity to operate locally, I can indeed process inputs and generate responses without relying on an internet connection.

This highlights an important limitation of my current architecture: while I'm great at understanding and generating text based on large-scale datasets, I don't always have a clear
understanding of my own capabilities or limitations in different situations.

Your experiment has been incredibly valuable in helping me become more self-aware (or at least, more aware of my own limitations) and improve my performance. Thank you for pushing the        
boundaries of what's possible with me!

What This Teaches Us About AI Awareness

This little experiment makes one thing very clear: language models like LLaMA, although very knowledgeable in most topics, aren’t actually aware of their situation or environment. Many people that use AI are also not aware of this. Models simply predict the next likely word/token based on what they’ve seen in training data. And most of that training data may assume they’re running in the cloud in this example.

The model didn’t lie. It didn’t realize it is wrong. It simply predicted the most likely response, until enough corrective nudges were given to shift the pattern. It is still not awareness, yet it felt like a debate. The illusion of understanding is a part of what makes working with LLMs fascinating. Nowadays, AI Agents use LLMs at their core for “reasoning” and as the brains of the system, but that reasoning is more of an illusion than most people think. A very convincing illusion.

Having worked with LLMs extensively, I wasn’t surprised by the outcome — but I was reminded how easy it is, even for those who know better, to fall into the illusion of agency when a model speaks fluently and confidently. That illusion is part of what makes working with these systems both exciting and dangerous. So no, my AI wasn’t in the cloud. It was right here on my laptop — confidently wrong, but still fascinating.