Chapter 06 · Section I · 14 min read
About predicting the future
Why most "AI will do X by 2030" predictions are wrong — and how to read the ones that are right.
Predictions about AI tend to come in two flavours. Catastrophist: we are all out of work in two years. Dismissive: it is just a chatbot, calm down. Both are usually wrong, and they are usually wrong for the same reason — both ignore the boring truth that AI moves fast in narrow places and slowly in broad ones.
This section gives you a framework for reading AI predictions skeptically. The framework is not a forecast of its own. It is a checklist of questions to ask whenever someone tells you what AI will or will not do by some specific year.
Why AI predictions are hard
AI capabilities have a peculiar shape. They do not improve smoothly. They sit at a plateau for years, then jump.
- Machine translation. Quality was roughly flat from 2000 to 2015. Then transformers arrived and quality jumped twice within five years.
- Image generation. Mostly toy outputs through 2020. Then diffusion models arrived, and by 2023 the outputs were indistinguishable from photographs in many domains.
- Robotics. Roughly the same capabilities in 2024 as in 2018. The jump has not yet come.
Plateaus are not failures; they are time spent waiting for the next idea, the next data set, or the next generation of hardware. Then a single research insight — backpropagation in 1986, the transformer in 2017, RLHF in 2022 — unlocks a leap.
The catch: predicting when the next jump happens is essentially impossible. Predicting which capability it will affect is hard. And predicting how the jump will reshape the surrounding economy or culture is harder still.
A framework for reading predictions
When someone tells you “AI will do X by year Y,” ask four questions.
1. How much data exists for this task?
AI works best on tasks where there is abundant labelled data. Translating common Indo-European languages has billions of parallel sentence pairs; translating between, say, Limbu and Tamang has very little. A prediction that all Nepali language pairs will be perfectly translated by 2030 is plausible for Nepali-English, dubious for Limbu-Tamang. Same algorithm; different data; very different outcomes.
2. How easy is it to define success?
“Win at chess” is well-defined — there are rules and a clear win condition. “Be a good doctor” is not. Tasks with sharp success criteria fall first. Tasks where success requires judgment, taste, or context fall slower or not at all. A prediction that AI will replace radiologists (well-defined task: classify images) is more credible than AI will replace clinicians (judgment-laden task: choose treatment for a complex patient).
3. Is failure cheap or expensive?
A model that writes an okay marketing email gets deployed quickly because the worst case is an awkward email. A model that decides whether someone gets a loan moves slowly because the worst case is a person harmed and a regulator angry. Same model technology; different deployment speeds. Predict capability and deployment separately.
4. Who has the incentive to make this happen, and who has the incentive to slow it down?
A prediction that ignores politics ignores the main variable. Self-driving cars are technically further along than the headlines suggest, but face liability law that no one has resolved. AI-powered drug discovery is moving fast because pharmaceutical companies have the incentive and the regulatory pathway. Predictions need to account for the whole system, not just the algorithm.
Three honest predictions for Nepal
Not so much predictions as confident bets, because each is grounded in things that already exist and have momentum.
Bet 1. Devanagari OCR will be at human accuracy on printed text, on a basic Nepali phone, by 2027. Why: the data exists, the algorithms are mature, the use case is everywhere, and the hardware is ready. The trajectory is short.
Bet 2. A Nepali-language voice assistant that handles routine government queries (citizenship, license renewal, tax) will exist by 2030, but its accuracy in rural dialects will lag behind Kathmandu Nepali by years. Why: the data for Kathmandu Nepali is being collected; for rural dialects it is not, and no commercial actor has the incentive to collect it.
Bet 3. Predictive AI in agriculture (paddy yield, monsoon timing, disease detection) will be unevenly adopted — large commercial farms first, smallholders last. Why: the technology is ready; the bottleneck is delivery, training, and trust. Predictions that ignore delivery channels miss the actual story.
Notice that none of these is about capability alone. Each combines capability with data availability, incentives, and delivery infrastructure. That is the framework.
Reading the predictions you’ll see
Three patterns to watch for in newspaper coverage:
The capability extrapolation. “GPT-4 can do X, so by 2027 it will do 10X.” This treats capability as smoothly increasing, ignoring plateaus. Discount it.
The hidden assumption. “AI will eliminate accountants by 2030.” Notice the prediction assumes (a) the technology works, (b) regulators allow it, (c) accounting firms restructure, (d) clients trust it. Each assumption has its own probability; multiply them and the headline number shrinks fast.
The transferred prediction. “If it works in the US, it will work in Nepal.” This ignores that the US has different data, different incentives, different regulators. Predictions for Nepal should be made for Nepal.
The single best long-term bet
If you forced me to name one thing AI will reshape in Nepal over the next decade, it would not be a specific industry. It would be who writes things.
The cost of producing draft documents, draft code, draft translations, draft news articles is collapsing. The thing that becomes scarce is editing — the human judgement that decides what is correct, what is true, what is appropriate. The work is shifting from generation to curation. That shift, more than any specific application, is the one I would bet on. It is happening everywhere in the world, and Nepal is not exempt.
The educational implication is unsettling: students who learn AI as a generator will be commodified; students who learn it as a tool to edit, fact-check, and judge will be valuable. This course has been trying, throughout, to lean toward the second kind of literacy.
Check your understanding
Quick check
—AI capabilities have historically tended to improve in which pattern?
What comes next
We’ve covered how to read predictions. The next section moves from forecasting to consequence — work, fairness, language, sovereignty. The four questions a country like Nepal will have to answer in the next decade, no matter what the prediction was.