The AI Skills Gap in the Office

I’ve become the AI guy in the office, which means I’m often asked by non-technical people some version of: “What should I be learning?” Here are some quick thoughts on AI for non-technical people.

Do I need an AI subscription?

Short answer: yes. The better question is which one.

The three that matter

Right now, three companies really matter: Google, Anthropic, and OpenAI. Meta’s models currently sit a tier below, though they’re investing heavily. (I’m excluding Grok and China’s models from this post.)

All three offer roughly the same menu:

• A flagship model
• Lighter / cheaper variants
• ”Reasoning” and “non-reasoning” modes
• Features like coding agents, deep research, and image generation (not Claude)

My recommendation

Pick one $20/month subscription. That unlocks access to the best models, deep research tools, coding agents and media generation.

Services like Perplexity let you switch between providers per prompt, but you’ll miss the latest features that each company keeps exclusive to their own apps.

Which one should you choose?

For most non-technical users doing everyday tasks, the perceived quality difference between leading Claude, ChatGPT, and Gemini models is very small. Usability and integrations matter more than raw model scores, even though people do try to rank models on sites like LM Arena

A rough way to think about it:

• Claude (Anthropic): The most pleasant UI experience. Nicest tone and quality of writing. Lacks image generation.
• ChatGPT (OpenAI): Smoothest all-around general helper.
• Gemini (Google): Rougher around the edges, but strong value with image/video creation, tight integration with Google apps, and Google One storage.

If you’re exploring vibe coding, Google’s plan paired with Antigravity is currently a steal for the amount of coding you get on a monthly plan.

What concepts should you learn?

The best way to learn AI is by using it. Try pasting these directly into the LLM you use.

1. Context windows

2. Model costs and capabilities

3. Agents and consumer behavior

(Take the answer with a grain of salt—nobody knows how this plays out.)

4. The bear case

5. AI-native tools

What should you actually use AI for?

Here are some easy places to get started:

• Writing: Dump rough ideas in, let AI draft, then iterate
• Editing: Provide context and tone, ask for a rewrite
• Planning: “Give me a step-by-step plan to do X”
• Search: Ask for a curated shortlist instead of Googling
• Deep research: Use built-in research agents to research a topic
• Research synthesis: Drop notes into NotebookLM (or a context window if you don’t have a lot) and ask for a summary
• Tech support: Screenshot your problem and ask what’s wrong
• Formulas & syntax: Describe your data rows and columns and goal; get SQL, Excel, or Python code back
• Image generation: Google Nano Banana Pro is currently king. Try this with Nano Banana Pro: upload two images—one of yourself and one of cartoon characters. Prompt the model with: “Replace the middle character in image 1 with the person from image 2, and generate the result in the style of image 1.”

Staying current

A few sources I find consistently useful:

• The Summary
• Simon Willison’s blog
• Dwarkesh Podcast
• Gary Marcus
• Nate’s Newsletter