Ask five serious people where AI is heading and you will get five confident, incompatible answers. That is not a sign the question is stupid; it is a sign that nobody knows, including the people building it. The useful response is not to pick a side and defend it. It is to find the moves that pay off across every scenario.
The four scenarios
1. The plateau
Scaling hits diminishing returns. Models get cheaper and more reliable, not fundamentally smarter. AI becomes infrastructure, like electricity or search — enormously valuable, mostly invisible, captured by whoever owns distribution. Jobs change the way spreadsheets changed accounting: fewer clerks, more analysts, no apocalypse.
2. The long ramp
Capability keeps climbing steadily for a decade. Agents get genuinely reliable at multi-day tasks. Entire job categories compress rather than vanish — one person plus a system does what six did. Wealth concentrates fast, politics gets loud, and the winners are the people who moved from doing the work to designing the system that does the work.
3. The takeoff
Recursive self-improvement in AI research itself compresses years of progress into months. Institutions cannot adapt at that speed. The outcome depends almost entirely on whether alignment and control techniques mature before capability does — which is why independent evaluations and control research matter far more than public argument.
4. The bad end
Powerful systems get deployed into critical infrastructure faster than anyone can verify them, and something breaks that cannot be unbroken — through misuse, through accident, or through a system pursuing a goal we specified badly. Most credible versions of this are not a robot uprising; they are a cascading failure nobody chose.
What the honest probabilities look like
Anyone who gives you exact numbers is selling something. What is defensible: forecasters and researchers disagree by orders of magnitude on timelines, they disagree less on the direction, and expert surveys consistently show a small but non-trivial probability of extremely bad outcomes alongside a large probability of merely disruptive ones. You can read the actual survey data at Stanford HAI and Epoch AI rather than taking anyone\’s summary, including mine.
Uncertainty is not an excuse for paralysis. It is an argument for strategies that do not require you to be right.
The moves that pay off in all four
- Become the person who deploys, not the person who executes. In every scenario, the value moves from doing tasks to designing and supervising systems that do tasks.
- Own something. Equity, a catalogue, an audience, an asset. Wages are exposed to automation; ownership is exposed to the upside of it.
- Keep a human network that would answer your call. Trust and reputation are the last things to be commoditised, and they compound in every world.
- Build a cash cushion and reduce fixed costs. This is good advice in a boom and essential in a shock.
- Stay technically literate enough to smell nonsense. You do not need to train models. You do need to know what a benchmark measures and what it does not.
- Do work that has a physical or relational component. Hands, presence, judgement in a room. Slowest to automate, in every timeline. See robotics and 3D printing for where that line is actually moving.
The moves that only pay off in one
Beware anything that requires a specific future: quitting your job to become a full-time prompt engineer, betting your savings on a single AI token, moving to a bunker, or assuming your industry is uniquely safe. Each of those is a leveraged bet on one branch of a tree nobody can see.
How to hold this without burning out
Doom-scrolling AI news is not preparation. It feels like it, which is what makes it dangerous. Set a bounded intake — a weekly read of a couple of primary sources rather than an hourly feed — and convert what you learn into one concrete change per month. The Edaptus survival guide exists precisely to convert anxiety into a checklist, and the action plans turn the checklist into weeks.
If you want to understand the technical arguments rather than the vibes, start with AISafety.info for plain language, then read a real evaluation report. Ten pages of primary material will change your thinking more than a hundred threads.