I Replaced ChatGPT & Notion with Local AI for 90 Days

I Replaced ChatGPT & Notion with Local AI for 90 Days: What I Gained and Lost

Spoiler: it wasn’t the privacy win I expected. It was something weirder and more interesting.

I was using ChatGPT to help me draft a difficult email to a client the kind where you’re trying to be firm without being rude and somewhere in the middle of typing out the context, I realized I was handing a pretty sensitive business situation to a server I knew nothing about. The client’s name, the project details, the money involved, the awkward history between us. All of it, just… out there.

I closed the tab and sat with that feeling for a minute.

I’d read about local AI models before. Ollama, LM Studio, running models on your own machine. I’d always filed it under “interesting but probably worse than the real thing.” But that evening I started actually looking into it, and three days later I had Ollama running on my laptop with Llama 3.1 loaded up, and Obsidian with a local AI plugin replacing my Notion workspace.

I committed to 90 days. No ChatGPT, no Claude (painful, honestly), no cloud-based AI writing tools. No Notion. Everything local, everything on my own hardware.

Here’s what actually happened.

The Setup: What I Used and How I Got There

My machine is a MacBook Pro M2 with 16GB of RAM. Not a beast, but the Apple Silicon chips handle local models surprisingly well this turned out to be a bigger advantage than I expected.

For the AI model side:

I used Ollama as the backbone. If you haven’t heard of it, Ollama is basically a tool that lets you download and run open-source language models locally with minimal setup. You install it, you pull a model like you’d pull a Docker container, and you run it. The command line interface is simple enough that I had it working within about 20 minutes of first hearing about it.

For a front-end (because typing into a terminal gets old fast), I used Open WebUI a self-hosted browser interface that looks and feels similar to ChatGPT. You run it locally, it talks to Ollama in the background, and you get a familiar chat interface without anything leaving your machine.

The models I tested over 90 days:

• Llama 3.1 8B fast, fits comfortably in 16GB RAM, surprisingly capable for everyday tasks
• Llama 3.1 70B noticeably smarter but significantly slower on my hardware (I used this for important tasks where I could wait)
• Mistral 7B good for writing tasks, snappy response times
• Phi-3 Mini surprisingly sharp for its tiny size, great when I needed speed
• DeepSeek-R1 7B strong reasoning, I started using this for anything logic-heavy toward the end of the experiment

For replacing Notion:

I moved to Obsidian a local, markdown-based note-taking app that stores everything as plain text files on your computer. No cloud sync unless you specifically set it up. To get AI features inside Obsidian, I used the Smart Connections plugin which lets you query your own notes using a local model, and Text Generator for AI-assisted writing within notes.

The First Two Weeks: Genuinely Rough

I’m not going to pretend the transition was smooth. It wasn’t.

The speed difference was the first thing that hit me. ChatGPT responds almost instantly. Llama 3.1 8B on my M2 takes maybe 10-15 seconds to start generating and streams noticeably slower. With the 70B model, I was sometimes waiting 45 seconds to a minute for a response to begin.

When you’ve been using instant cloud AI for two years, that gap feels enormous at first. I nearly quit in week one purely because of this.

The second problem was quality on complex tasks. I tried using Llama 3.1 8B for something I do regularly in ChatGPT taking a messy set of notes and turning them into a structured outline for an article. The output was… fine. Serviceable. But it missed nuance, structured things oddly, and I had to do significantly more editing than I was used to.

I also immediately missed Notion’s web clipper. The ability to clip articles directly into a structured workspace, tag them, link them to projects Obsidian can do this but it requires more setup and a bit of a different mental model. I spent probably six hours in week two just configuring Obsidian the way I wanted it.

Week Three Onwards: Where It Got Interesting

Here’s the thing nobody tells you about local AI: you start using it differently.

With ChatGPT, I had developed these long, elaborate sessions. I’d share a lot of context project background, client details, previous conversations because the model needed it and because I was in a “chat” mindset. When you’re paying attention to what you’re sharing, and when the model is slower and you’re waiting for responses, you naturally become more deliberate.

I started using local AI for shorter, more specific tasks. “Rewrite this paragraph more directly.” “Give me five alternative subject lines for this email.” “Summarize these three bullet points into one sentence.” Focused, surgical prompts rather than big sprawling conversations.

And honestly? For those kinds of tasks, Llama 3.1 with a good prompt is pretty excellent. Not ChatGPT-4 level, but solid.

The privacy thing also started feeling more real in a different way than I expected. It wasn’t just about sensitive client information it was about the fact that I could be completely honest in my prompts without any internal filter. There’s a subtle self-censorship that happens with cloud AI that I hadn’t consciously noticed until it was gone. With a local model, I’d write exactly what I was thinking without wondering how it would be interpreted or stored.

What Local AI Was Actually Good At

After 90 days, here’s where local models genuinely held their own:

• Editing and rewriting. This is where the quality gap was smallest. Give a local model a paragraph and ask it to tighten it up, make it more direct, or change the tone the results were consistently good. Mistral 7B was particularly strong here.
• Summarization. Paste in a long article or a set of notes, ask for a summary local models handle this well. The Smart Connections plugin in Obsidian let me ask questions about my own notes (“what did I write about email marketing last month?”) and the answers were genuinely useful.
• Brainstorming. When I needed ten ideas for article angles or wanted to think through a problem out loud, local models were a good thinking partner. Not brilliant, but useful. And I could share actual messy context without filtering.
• Code snippets. I’m not a developer but I write occasional scripts for automating things. For simple Python or JavaScript, the local models did fine. DeepSeek-R1 was noticeably better than the others at reasoning through logic problems.
• Offline work. This was an unexpected practical benefit. I travel fairly regularly and work from places with unreliable internet. Being able to use a capable AI model on a plane or in a rural area with no connectivity was genuinely useful. Cloud AI is useless without internet. Local AI doesn’t care.

What I Genuinely Missed (And Couldn’t Fully Replace)

I need to be honest about the gaps, because some of them are significant.

Complex reasoning and long-context tasks. If I gave ChatGPT-4 a 3,000-word document and asked nuanced questions about it, it handled it beautifully. Local models, especially the smaller 7-8B ones, would lose the thread in long contexts or give shallower analysis. The 70B model was better but slow enough to be frustrating.

Real-time information. Local models don’t have internet access (unless you build a fairly complex setup with tools like Perplexica, a local alternative to Perplexity AI). I missed being able to ask “what’s the current situation with X” and get a real answer.

Multimodal tasks. I use GPT-4o fairly regularly for analyzing screenshots, reading charts, or describing images. Local multimodal models exist LLaVA is one but they were noticeably weaker than GPT-4o for image tasks during my testing.

Notion’s database features. This is where the Notion-to-Obsidian switch hurt most. Notion’s relational databases linking projects to clients, tracking content calendars, filtering views I use these constantly. Obsidian with the Dataview plugin can replicate some of this, but it requires writing query syntax and the setup time is real. For simple notes and writing, Obsidian is actually better than Notion in my opinion. For structured data and project management, Notion wins clearly.

The Practical Setup Guide If You Want to Try This

If you want to experiment with local AI without committing to 90 days cold turkey, here’s the minimum viable setup:

Step 1: Install Ollama Go to ollama.com, download for your OS, install it. Open your terminal and run: ollama pull llama3.1 That downloads the 8B model. Takes a few minutes depending on your connection.

Step 2: Install Open WebUI If you have Docker installed, one command gets you a full ChatGPT-like interface running locally. If Docker feels too technical, Msty is a native desktop app that connects to Ollama with a nice interface and no Docker required.

Step 3: Test with low-stakes tasks first Don’t immediately try to use it for your most complex work. Start with: “summarize this article,” “rewrite this paragraph,” “give me five headline ideas.” Get a feel for what it’s good at on your specific hardware.

Step 4: Keep your cloud AI for now I wouldn’t recommend going cold turkey unless you specifically want that experience. Use local AI as a supplement first. You’ll naturally figure out which tasks it handles well enough and which ones you still want the cloud for.

Step 5: If you want the Obsidian setup Download Obsidian (free), install the Smart Connections plugin, point it at your local Ollama instance. Your notes stay on your drive, and you can ask AI questions about them without anything leaving your machine.

What I Do Now

After 90 days I didn’t stay fully local that was never really the goal. What I do now is hybrid.

Local AI via Ollama for: client-related work, anything containing personal or sensitive information, quick editing tasks, working offline.

Cloud AI (Claude mostly, occasionally GPT-4o) for: complex reasoning, long-document analysis, anything where quality matters more than privacy, image tasks.

Obsidian for: personal notes, writing drafts, journaling, anything I want to keep private. Notion for: project management, content calendars, client-facing work.

The 90-day experiment didn’t make me a local AI evangelist. But it did make me much more thoughtful about what I’m sharing with cloud services, and it gave me a genuinely useful tool that works without an internet connection and without a subscription.

The models are also getting better fast. Llama 3.1 running locally today is more capable than ChatGPT-3.5 was two years ago. Where they’ll be in another year is an interesting question.

If you have at least 8GB of RAM and any curiosity about this, it costs nothing to try. The Ollama setup takes less than half an hour. Worst case you uninstall it. Best case you find a tool that covers a specific slice of your work in a way that just feels better.

That awkward email I mentioned at the start, by the way I ended up writing it myself. Turned out I didn’t need AI for it at all. Sometimes sitting with something for a minute is the right move.