Best Practices: Habits, Anti-Patterns, and the Long View

This chapter distils the habits that hold up across years of changing AI news, the anti-patterns worth avoiding, and the appropriate humility about the whole situation.

Habits That Hold Up

Read across camps

Most AI content comes from specific camps with specific framings. Habitual reading across camps (alignment-worried, accelerationist, skeptical, ethics-focused, policy-oriented) produces a better mental model than deep immersion in one.

Weekly cadence: one thing you'd naturally read, one you wouldn't. Over a year, this rewires your thinking.

Track specific predictions

Commentators and forecasters make specific claims. Note them with dates. Revisit. Patterns of accuracy over time tell you whose views to weight more.

A year of tracking trains calibration in a way no amount of reading alone does.

Do the arithmetic

Whenever a claim involves numbers (compute, cost, capability, probability, economics), spend a minute checking. Many arguments that sound profound collapse on checking.

Notice your updates

When your views change, notice. Write down what changed and why. Over time, you build a record of how your thinking evolved. This is valuable both for self-knowledge and for intellectual honesty.

Hold frameworks over tactics

Specific AI news updates weekly. The underlying frameworks (alignment, concentration, institutions) update slowly. Prioritise learning the frameworks. The tactical news will make more sense in their light.

Practice calibration

On specific questions ("will X happen by Y?"), give yourself a probability. Check later. You'll learn whether your confidence is calibrated. Most people's initial calibration is poor; it improves with practice.

Spend time outside AI

One of the most important habits: don't let AI consume you. Read history, fiction, science outside AI, philosophy. Have relationships. Exercise. The frame you bring to AI questions is better when it's informed by more than AI discourse.

Use AI yourself, thoughtfully

Using AI for real work teaches things reading about AI can't. What works; what doesn't; what feels different about specific models; how deployment actually shapes up. This is data you get nowhere else.

Anti-Patterns to Avoid

Over-certainty in any direction

"AGI by 2027, everyone's deluded." "AI is overhyped, it'll plateau soon." "Alignment is impossible." "Alignment is basically solved." Each is held confidently by someone; none is warranted by the evidence. Certainty is the tell.

Tribal signalling

"The safety people are..." "The accelerationists are..." "The AI ethics folks are..." Each sentence often presages a caricature. Engaging with actual arguments from named individuals is harder and more useful.

Engagement theatre

Performing a position online without contributing substantively to understanding. Likely to feel productive; unlikely to be so. Resist the pull.

AI news as entertainment

Consuming AI content for emotional stimulation (excitement, fear, outrage) rather than to learn. Signals: reading more, updating less, stronger emotions over time.

Career optimisation around a single forecast

Betting your career on AGI by 2028 (or on never). Hedged positions are usually wiser, because forecasts are unreliable.

The "just one more source" treadmill

AI content is infinite; new papers, posts, interviews, podcasts every hour. Chasing it all is impossible and counterproductive. A curated narrow diet of high-signal sources, read carefully, beats consumption of everything.

Outsourcing judgment

Letting a favourite commentator or lab do your thinking for you. Even the best sources have blind spots; forming your own view matters. The tutorial is structured around this principle.

Panicking

Fast, visible action out of anxiety rather than considered thought. Sometimes urgent action is warranted; often it isn't. Distinguish.

Paralysis

The mirror of panic. "The stakes are too high for me to act; I'll wait to understand more." Understanding improves with action; endless waiting usually doesn't produce more understanding, just less engagement.

The Monthly Check

Once a month, ten minutes:

1. What did I read this month? One thing from each camp, or close to
2. What's changed in my views? Any specific update, grounded in specific evidence
3. What's one specific prediction I made? How's it tracking?
4. Am I still doing arithmetic? Or have I drifted into taking claims at face value?
5. Am I using AI well? Where's it helping? Where is it eroding skill I want to keep?
6. What's one thing I could do better next month?

Not a checklist for its own sake. A small discipline that compounds.

The Yearly Reflection

Once a year, longer:

• What was my view on the AI transition a year ago? What is it now?
• What did I get right? What did I get wrong?
• What would I have done differently given what I now know?
• What's the next year's uncertainty I can't resolve now?
• How is my career, finances, and life aligned with the trajectories I expect?

This reflection is itself a data point. Writing it down preserves it for future-you.

Signals of Good Calibration

If your thinking is on track:

• You can articulate the strongest versions of positions you disagree with
• You've updated your views on at least one significant question this year
• You disagree with your in-group on at least one specific thing
• You're not certain on the hardest questions (timelines, catastrophic risk)
• Your predictions, when specific, are within your stated confidence intervals
• You can distinguish what you know from what you believe from what you feel

Signals of Bad Calibration

Time to recalibrate if:

• You've been in the same camp without friction for years
• You can't state opposing views fairly
• You're consistently wrong about specific predictions
• Your certainty exceeds your evidence
• You feel comforted by AI content more than challenged
• You're spending more time on AI commentary than on producing or learning anything else

Not a disaster. A call to widen your intake.

The Long View

A reasonable perspective on where this goes:

Ten years from now (2036), one of several things will be true:

• AI is transformative and the transition is under way in ways we can now describe
• AI is significant but less transformative than feared or hoped
• AI has plateaued; progress happens but less than predicted
• Something unexpected has happened that none of us currently imagines

Each is plausible. The frames in this tutorial help you read the situation regardless of which arrives.

Twenty years from now (2046), outcomes will be clearer. Mistakes made now will matter in predictable and unpredictable ways. People reading this tutorial then will find parts that aged well and parts that didn't. They'll be in a better position than those of us writing now to judge.

The honest position is to act well in the present, knowing we're operating with incomplete information, and to adjust as conditions clarify.

What If You're Wrong

A useful discipline: imagine the tutorial (and you, reading it) are wrong about something important.

• What if transformative AI arrives in 3 years, not 20?
• What if it never arrives, and this whole discourse is overblown?
• What if the specific risks most worried about don't materialise, and others do?
• What if the distribution outcomes are the opposite of what you expect?

If your plans depend on being right about all these, they're fragile. If they roughly hold up under being wrong about any of them, you're doing OK.

The goal isn't being right. It's being responsive to evidence and resilient under uncertainty.

The Uncomfortable Middle

Most reasonable views on AI sit in an uncomfortable middle:

• Taking the risks seriously without being paralysed
• Acknowledging the benefits without being dismissive of costs
• Following the science without trusting the labs uncritically
• Engaging politically without becoming partisan
• Using AI without denying its problems
• Worrying about concentration without opposing progress

The middle is harder to occupy than any of the poles. It's also probably where the truth lives, most of the time.

A Personal Note

This tutorial has views. They're shaped by who's writing, what I've read, the moment I'm writing in. They'll age; some better than others. Read with appropriate skepticism.

The underlying commitment: the AI transition is important enough to try to think about honestly, and hard enough that honest thinking requires effort and discipline. Everything else follows from that.

If you leave this tutorial with:

• The vocabulary to read AI commentary with better filters
• A rough map of the serious positions
• Specific sources to read across camps
• Some habits for updating on evidence
• Appropriate humility about everyone's predictions (including the tutorial's)

Then it's been useful. The goal wasn't to teach you what to conclude. It was to put you in a position to conclude things carefully for yourself.

Where to Go From Here

• Pick three sources: one believer, one skeptic, one neutral. Read each for a month. See what you notice
• Set up a calendar reminder for a monthly 10-minute check on your views
• Pick one specific AI policy question and engage with it deeply for a quarter
• Notice if your views are changing in response to evidence or social pressure; adjust
• Re-read this tutorial in a year; it will age; so will your views. Both changes are information

The transition is real. Whether it turns out well is contested. Good thinking, distributed across many people making decisions over years, is part of what determines which way it tips. Your thinking is part of that.

Good luck.