Lesson brief

Almost every confusing AI debate starts with people using one word for three different things. A task-specific tool like a transcription app or a tumour-detection model does exactly one job and gets there by being trained on a giant labelled dataset of that one thing. Narrow superintelligence does a single domain better than any human but stays inside that domain. General superintelligence would match or exceed humans across the full range of cognitive work, including jobs we have not invented yet. The case for safety being achievable is much stronger when you are talking about the first two and collapses when you are talking about the third.

The mechanism behind the confusion is commercial. Frontier labs sell narrow products and general aspirations in the same sentence, because the aspiration justifies the funding round and the product justifies the revenue. When you hear a claim about AI, your first move is to ask which of the three categories the actual shipped system belongs to. A self-driving model trained on thousands of contractors hand-labelling vehicles, pedestrians and traffic lights is narrow tooling, no matter how the press release phrases it. A system that surprises its own designers by inventing strategies they could not have forecast, the way AlphaGo did with its famous nineteenth move, is gesturing at the second category. Neither is the third.

The tradeoff to hold is that the line between these categories is not always visible from outside. A system that looks narrow can become more general as it is given more tools, more memory and more autonomy, and the labs themselves disagree on when that crossover happens. So categorisation is not a one-time judgement, it is a habit. You watch what a system can actually do this quarter, not what it is rumoured to do next year, and you update only when behaviour, not branding, has shifted.

Core takeaways

• A product is task-specific tooling when its training data is a giant labelled set of one narrow thing.
• Narrow superintelligence beats humans inside one domain; general superintelligence would beat humans across all domains, including jobs that do not yet exist.
• When a single press release uses all three meanings of AI, the financial purpose is to let the aspiration carry the valuation while the tool earns the revenue.
• Move 37 mattered because the system found a strategy human professionals first read as a mistake, hinting that capability had outrun human legibility.
• Always categorise on shipped behaviour, not on roadmap language, and re-categorise as the system gains tools, memory and autonomy.

Practice

Pick three AI products you have seen advertised in the last week, one from a frontier lab, one from a startup, one embedded inside a product you already use. For each, write a single sentence describing the actual task the shipped version performs today. Then label it narrow tool, narrow superintelligence, or general superintelligence using only that sentence, ignoring the marketing copy. Spend no more than ten minutes per product. At the end, note which of the three claims required you to stretch the definition the furthest, and write one line on what evidence would have to change for you to upgrade its category.

Quiz

FAQ

What's the actual difference between narrow AI and AGI?

Narrow AI is task-specific and well-bounded — image recognition, language modelling, game-playing. AGI would be general-purpose intelligence comparable to humans across domains. Current systems are narrow with broadening capability; whether and when that becomes general remains the central technical and policy question.

Is AGI close?

Forecasts vary wildly across credible researchers from 'within a decade' to 'never'. The honest answer is that nobody knows with high confidence. Plan for a wide range of outcomes rather than betting on a specific timeline.

Should I be worried about AI?

Worth taking seriously without becoming a doomer. The near-term risks (mis-information, labour disruption, surveillance, security) are concrete and present; long-term risks are more speculative. Focused engagement beats both panic and dismissal.

Reflection questions

1. Which takeaway here is most uncomfortable to apply to your life right now?
2. Where in your week could the exercise above realistically run for 7 days?
3. What is the smallest, bad-day version of this lesson's idea you could do tomorrow?
4. Who in your life would benefit most from you applying this?
5. What would have to be true in 90 days for this lesson to have mattered?

Common mistakes in this area

1. Treating ‘AI’ as one thing rather than a family of capabilities.
2. Confusing a hype video with deployment reality.
3. Forecasting from feelings (excitement, anxiety) rather than capability deltas.
4. Reading only one tribe (optimists or doomers).
5. Skipping the timescale question on every prediction.

Apply this today

Pick one action from the practice block above. Put it on today's calendar at a specific time, in a specific place. If it can't fit in today's calendar, it's too big — shrink it until it can.

Next steps

• Worksheet: AI claim evaluation worksheet
• Continue the micro-course: AI Risk Literacy: Separating Hype from Real Threat
• Lesson: Follow the Money Behind AI Inevitability14 min
• Lesson: Estimating Extinction Risk Without Hand-Waving16 min
• Lesson: Alignment Failure Modes and Move 3715 min