Course resource
Staying Current
A system for keeping up that does not consume your week or leave you anxious.
The problem
AI news is produced faster than anyone can read it, most of it is noise, and the volume creates a background sense that you are falling behind. That feeling is not information — it is the by-product of an industry that benefits from your attention.
You need surprisingly little to stay genuinely current.
The principle
Follow capabilities, not releases.
A new model version is not news. "Models can now reliably do X" is news, because it changes what you should build. Most of what crosses your feed is the former dressed as the latter.
The test for any announcement: does this change something I would do differently tomorrow? Almost always no.
The weekly fifteen minutes
One slot. That is the whole system.
1. Skim two sources. Two, not ten. (5 min)
2. Ask: did anything change what I should build or stop doing? (2 min)
3. Log anything that did. (1 min)
4. Review my own automations — anything misbehaving? (5 min)
5. Note any task I did 3+ times this week that could be automated. (2 min)
Step 4 is the one that pays. Your own systems drifting is more consequential than anything in the news.
Choosing two sources
Pick two you will actually read. Criteria:
- Filters rather than aggregates. You want someone with judgement, not a firehose.
- Says when something does not matter. A source that hypes everything is useless.
- Matches your level. Research summaries are not useful if you are not doing research.
Two suggestions to start, both daily and both brief: therundown.ai and bensbites.io. Replace either if it stops earning the slot.
Unsubscribe from the rest. Seriously. The marginal value of source three is near zero and the cost is real.
The quarterly hour
Once a quarter, not weekly:
- Re-test your automations. Models change under the same name — your prompts may be silently degrading.
- Review your tool stack. What are you paying for and not using?
- Check the Tool Radar — has anything you rely on been superseded?
- Update your Digital Twin.
- Turn something off. There is always something.
- Re-read your own notes. You have forgotten half of what you learned.
The experiment log
The only reliable way to know whether something is useful is to try it on your own work.
| Date | What I tried | On what task | Better than my current approach? | Keeping it? |
|---|---|---|---|---|
One experiment a fortnight is plenty. The discipline is trying it on a real task, not a demo — a tool that impresses in a demo and fails on your actual work is worse than no tool.
What to ignore
- Benchmarks and leaderboards. They move weekly and rarely change what you should do.
- Model version numbers. Capability matters; the number does not.
- Anything predicting a date for AGI. Nobody knows, and it changes nothing about Tuesday.
- Threads promising "10 AI tools that will change everything". They are the same ten tools.
- Anyone claiming a technique is "dead". Prompting was declared dead annually and remains the main interface.
What to pay attention to
- Capability thresholds crossed — something unreliable becoming reliable
- Price changes — they shift what is economic to automate at volume
- Deprecations — a model you depend on being retired
- Standards — things like MCP that change how tools connect
- Regulation in your jurisdiction — it changes what you may do
- Failures reported by people doing your kind of work — the most useful and least amplified category
The anxiety part
Most people in this course will feel, at some point, that everyone else is further ahead. A few things that are true:
The gap between knowing about a tool and using it well is large, and most people are on the near side of it. Reading about agents is common; running one in production is rare.
Depth beats breadth. Someone who uses two tools extremely well is more capable than someone who has tried thirty.
The fundamentals barely move. Context, constraints, verification and judgement have applied to every model released since this course was written and will apply to the next ones.
Nobody is keeping up. The people who appear to be are specialising narrowly and ignoring the rest, which is the correct strategy.
Your setup
| My two sources | |
| My weekly slot (day and time) | |
| My quarterly slot | |
| Where my experiment log lives | |
| Last quarterly review |
Put the weekly slot in the calendar now. A system that depends on remembering is not a system.