Staying current in AI without burning out: a 3-hour weekly routine

Why "keeping up" feels impossible

There are more AI announcements each week than anyone can read, and many are incremental. Social feeds reward novelty, so they amplify each launch as if it changes everything. Three habits make it worse: reading summaries of summaries, bookmarking more than you will ever read, and switching tools every time a new one trends. The routine below replaces all three.

The weekly routine

That is three hours. If a week is busy, drop the Read block before the Try block; trying something teaches more than reading about it.

Choosing sources

Prefer sources that are close to the change itself:

• Release notes and changelogs of the model providers, frameworks and cloud services you use. These tell you what changed without commentary.
• Official documentation, especially "what's new" pages.
• Technical reports and papers for important model releases, read selectively.
• One or two practitioners whose writing you find careful — people who show evaluations and admit failures.

Limit the list. Five sources you check every week beat fifty you feel guilty about. Unsubscribe from anything that has not produced an "affects me" item in a month.

The "affects me" test

During the Scan, mark an item only if at least one of these is true:

1. 1It changes the price, limits or behaviour of something you use now.
2. 2It deprecates or retires something you depend on.
3. 3It solves a problem you have written down in the last month.
4. 4Your team, a client or an interviewer has asked about it.

Everything else is interesting, not urgent. Rule 2 matters more than people expect: deprecations and retirements are how an unchanged system breaks. Certification exams change too — see what you can verify about AI certifications.

The Try block: small and specific

Keep experiments tiny so they finish in one sitting. Examples:

• A new model version: run your own ten saved test prompts through it and the current one; compare side by side.
• A new framework feature: rebuild one small function from an old project with it.
• A new retrieval technique: run it on the same 20 questions you used last time.

Keeping a fixed personal test set is the single most useful habit here. It turns "this feels better" into "this got 17 of 20 where the old setup got 14". The Generative & Agentic AI add-on teaches this evaluation-first habit.

The learning log

One entry per week. A plain text file is enough:

2026-09-27
Item: Provider X added structured output mode for model Y
Source: provider changelog + docs page
Tried: my 10 invoice-extraction prompts, old JSON-prompt vs new mode
Result: valid JSON 10/10 vs 8/10; field accuracy unchanged (9/10 both)
Decision: switch in my invoice project; not worth changing others yet
Time spent: 2h 50m

After a few months the log becomes useful evidence in interviews and appraisals — it shows judgement, not just awareness. See documenting AI-assisted work in your appraisal.

Signs the routine is working

• You can name what changed in your main tools this month without looking it up.
• You have tried each thing you talk about.
• Your weekly time stays near three hours.
• You feel less anxious about missing things, because you know you'll catch what matters in Monday's scan.

Signs it isn't

• The Scan grows past 45 minutes: cut sources.
• You skip Try for three weeks in a row: make experiments smaller.
• Everything is marked "affects me": apply the four rules strictly.
• You're reading about AI late at night most days: set a fixed stopping time and keep the rest of your week free.

For people changing careers

If you are learning AI while working another job, you have even less time. Keep the Scan to 15 minutes, and use the Try block for your course projects rather than new tools. Depth in fundamentals — evaluation, data handling, Python — lasts longer than knowledge of this month's release. The 90-day plan to become the AI person on your team shows how to direct that time towards visible results at work.