Frozen in Time and Fine With It ...Sometimes
If I download an AI model and run it entirely offline, on my own machine, with no data leaving my laptop, how much am I actually giving up?
Plain Inference
What new technology really means for the people it touches, and where that leaves room to build.
Subscribe on LinkedInShop like it's 1999 (a consumer playbook to outsmart surveillance pricing)
The Senate just spent an afternoon on surveillance pricing, and three states have bans on the books. Almost nobody is telling consumers what they can do today, no legislation required. Here's the playbook.
You keep using that word, "automated," I do not think it means what you think it means
I handed an agent my most tedious recurring task and discovered the AI was the easy part. Everything around it, the access and the permissions and the plumbing, was the hard part.
Every Agent Delegation Has Its Flame Spurts
Setting up a staff of AI agents, the real work was never describing each job. It was writing down what each one must not do without checking with me first.
Welcome to Plain Inference
If you've landed here, you're probably somewhere on the spectrum between "AI-curious" and "AI-exhausted." Good. This is a place for both.
Past Experiments
Hands-on writeups from Daily(ish) GenAI Experiments, the project that became Plain Inference.
Did I Actually Learn Something? What Claude Code Thinks About My Progress
After weeks of building my website entirely through conversation with Claude Code, I realized I'd been so focused on getting things done that I hadn't stopped to think about whether I was actually absorbing any of the technical knowledge along the way.
The Data Trail We Leave Behind - Part 1: The Fragmented Picture Problem
I've been deliberately careful about data privacy throughout my AI experiments. My background in consumer research and product testing in the data privacy space taught me early what the real costs are when you hand over personal information to tech platforms.
Claude Code, Part 2: The Accidental Education in Website Infrastructure
Building a website with Claude Code taught me five unexpected lessons about how the web actually works—not through tutorials, but through real debugging with immediate stakes.
Claude Code, Part 1: How a Non-Coder Built a Custom Website
I had a GoDaddy site with decent blog features and a Canva design I loved. But I couldn't combine them. So I asked Claude Code to build a custom website—and it actually worked.
The Agent Got 99% Right—The Last Mile Ate My AI Credits
Every morning, I need to fill out a slip telling my kindergartener's school what they'll eat for lunch. I thought: this is exactly the kind of thing an AI agent should handle. Spoiler: the agent got almost everything right, but 'almost' doesn't fill out lunch slips.
Three AI Models Walk Into an Unsolvable Puzzle
My second grader came home with a logic puzzle worksheet. After some frustration at the kitchen table, I realized there wasn't enough information to solve it. I decided to test whether AI could confirm my suspicion.
GPT 5.2 Nailed My Plant Care Question, Then Got Fooled by a Moose
I had a trailing vine plant sitting on top of a cabinet in my house and wanted to know if I should do anything with it. Simple question. But I also had an ulterior motive: I wanted to see if ChatGPT could match a specialized plant identification app.
Testing ScholarAI: When a Custom GPT Actually Works (But Also Really Wants You to Upgrade)
I tested the ScholarAI custom GPT to see if it lived up to the hype. I wanted to find out what it can really do.
Quick Take: Knowing the Risks Matters — A Reminder from the AI Safety Index
This week, I dug into the latest AI safety pulse check: the Future of Life Institute's "AI Safety Index: Winter 2025." It reviews eight leading AI companies across six critical domains, from risk asse...
When a Wine List Becomes a Logic Puzzle
Last week I went to a newly opened speakeasy and ordered a glass of white wine. The server who was young, friendly, but clearly new, admitted he hadn’t tasted any of the wines yet and neither had the ...
Trapped in the Chat: Why I Rebuilt My Claude App in Gemini 3
We often talk about "Vendor Lock-in" with enterprise software, but I recently discovered a new flavor of it: "Prompt Lock-in."
Hands-on with "Nano Banana Pro": Context, Cartoons, and Guardrails
Yesterday's experiment started with a simple goal: test whether Google's latest image model understands context or just follows instructions. The results revealed both a massive leap forward in AI's a...
When AI Becomes Your Holiday Shopping Assistant: Claude Voice Mode Edition
Yesterday's experiment started during my drive to daycare pickup and ended with an actual shopping plan I might follow. If you've ever stared at a gift list for 8 family members and felt your soul lea...
After a Week of Mixed Results, AI Delivers Something My Kid Actually Loves
Last night's experiment started with homework frustration and ended with genuine delight, the kind where my seven-year-old actually asked to do more word practice. If you're a parent who's witnessed t...
When AI Data Extraction Becomes More Work Than Doing It Manually
The promise: AI will automate tedious tasks like extracting data from PDFs into spreadsheets.
When AI looks in your Fridge
Last night's experiment started with a simple idea: could a large language model make something useful out of the chaos in my fridge after a busy weekend?
Custom Instructions to Combat GenAI Hallucinations
On Wednesday I shared an example of a wild hallucination that resulted from the Gamma implementation of GPT5. Today, I'd like to share some tips on how to use custom instructions in the models you use...
When AI Gets Creative (A Little Too Creative): The Gamma vs. ChatGPT Experiment
I just had one of those "this is why we can't have nice things" moments with AI.
Using GenAI to Prepare for a Partnership Conversation
This week’s experiment explored how GenAI can speed up and elevate partnership preparation, not just by summarizing a company’s data, but by helping you think like their business.
ChatGPT Atlas Edition - Part 1: The Privacy Paradox - ChatGPT Wants Your Browser History
What happened
Atlas Edition - Part 2: Watching Paint Dry, AI Edition - The Restaurant Reservation Saga
Armed with ChatGPT Atlas, I set out on what should have been a simple task: book restaurant reservations for a weekend trip. What followed was a masterclass in AI inefficiency.
The Case of the Color-Changing (and Sometimes Disappearing) Shirt
The Challenge: A friend needed her headshot edited to match her brand colors - specifically, changing her blue dress to purple. Simple enough, right? Well, as with most things in AI, the journey was m...
When ChatGPT 'Tried' to Build My Half Marathon Playlist
I had a simple ask. I needed a playlist for my upcoming half marathon, but not some generic "pump up" mix. I wanted something that understood my target pace, my tendency to start too fast, and exactly...
Data Science Deep Dive: Analyzing 3 Years of Peloton Data with Gemini 2.5 Flash
For today's GenAI experiment, I fed Gemini 2.5 Flash my personal three-year Peloton workout history, a 1,351 row CSV file and asked it to diagnose the factors behind my longest training breaks.
Rapid Prototyping with Real Limitations
Gamma is brilliant for breaking the blank page problem. That initial rush of having a complete deck in seconds gave me momentum. Even though I changed most of the content, having something to react to was psychologically powerful.
The Day Peloton Forgot My Run (and GPT 5 Rebuilt It from Disconnected Data Sources)
When your fitness app fails, AI can still finish the story.
The Prompt That Actually Worked: Text to Calendar
Yesterday I shared how I turned the park district's basketball schedule into a clean calendar file. Today I'm sharing the exact prompt template I used, because the magic wasn't in the AI, it was in be...
When the Youth Basketball Schedule Has Everything Except an "Add to Calendar" Button
I had a problem. Actually, I had a perfectly functional schedule. The park district's website listed seven weeks of youth basketball games with all the information I needed: dates, times, locations, m...
Planning a K-Pop Demon Hunter (IYKYK) Birthday Party with Claude
The Experiment Can AI help plan a party based on specific media? I tested this by planning my child's birthday party themed around the Netflix movie K-Pop Demon Hunters.
Gemini & ChatGPT - The Power of Comparison
How I Used Them
Claude Projects - The Foundation of Thought Partnership
What happens when you stop treating AI as a content generator and start using it as a genuine thinking partner? In this experiment, I explored Claude's Projects feature to develop and refine a presentation through multiple iterative sessions. The key discovery: Claude Projects creates a persistent workspace where context carries across conversations. No more re-explaining your project every time. Instead, each session builds on the last - from initial brain dump to structure creation, through voice conversations about creative decisions, all the way to final refinements based on practice session feedback. I found the most value not in having Claude write my presentation, but in using it to think through structural choices, narrative flow, and transitions. Voice conversations proved especially powerful for working through subjective creative decisions in real-time. This experiment demonstrates how AI tools excel when used for sustained creative partnership rather than one-off tasks. Whether using Claude Pro's Projects feature or bookmarking conversations in the free tier, the ability to maintain context transforms AI from a tool into a collaborator. [Read the full experiment on LinkedIn →]