# Tech to watch

## The one question

Every week something new drops, and the headlines make it sound like the end of the world or the start of a new one. Don't go on the headlines. Ask this:

**What can a beginner do now that they couldn't last month?**

If you can't answer it, you can skip the release. If you can, that's your next thing to try.

## The watchlist

Keep it short. Six things is plenty. This is what I keep an eye on.

- **The model makers.** Anthropic, OpenAI and Google, plus others like Moonshot (the makers of Kimi). Each has different products, and each is better at different things. Learn what Google does better and what Anthropic does better, and where each is heading over the next three to six months.
- **Coding agents.** Claude Code is my main builder. I use Codex as a second opinion. When one of them changes, I notice.
- **APIs.** An API is a way for one program to ask another for something. Learn these before MCP servers. They're more useful early on.
- **MCP servers.** An MCP server plugs a tool or a data source into your AI in a standard way. Learn it once you're happy with APIs.
- **Hosting and databases.** I use Cloudflare for hosting and Supabase for databases, because both have free tiers. Pricing and limits change, so check them yourself.
- **Costs and open source.** Usage costs (what you pay for the text going in and out), open-source models, and the frontier models at the top. Cost decides what you can run all day.

## Your 20-minute weekly routine

Pick one slot and keep it, like Friday evening or a lunch break.

1. **Five minutes: the source.** Check Claude's official X account and the Claude Code changelog. They say when new tools land or a bug gets fixed. I check every morning, but once a week is fine to start.
2. **Five minutes: the wider scan.** Look at what's trending on GitHub and Hacker News. You're looking for what other builders are talking about, not reading it all.
3. **Five minutes: your question.** Take the one release that caught your eye and ask the question above. Write the answer in one line.
4. **Five minutes: try it.** Run a tiny test on your own work, not a demo.

Write the one-liner in a note. After two months you'll have a record of what actually mattered.

## Run your own benchmark

I have a test I run on every new model. It's the same task each time, something from my own work. That lets me see how a model has improved, and how it will change how I do things.

Make yours. Pick one real task you do often and write down the prompt. When a new model arrives, run the same prompt and compare the result with last time.

## A prompt to help you triage

Paste a release announcement into Claude or ChatGPT and say:

```
Here's an announcement: [paste it].
I'm a beginner building [your project].
In plain English: what's actually new here?
What can a beginner do now that they couldn't last month?
What would I need to check myself before trusting this claim?
Don't invent anything. Say what you aren't sure of.
```

Then check the claim at the source. Open the company's own page, not a video about it.

## Learn the words first

If terms like agent, workflow, API and MCP sound like nonsense, learn them first. Ask Claude to explain each in plain English, give you an everyday use case, and show how they connect. Do one at a time. Add new ones as you hear them.

Don't feel you have to try everything. Try one thing a week.

**Next:** [Companies to track](companies-to-track.md)
