Preparing for
2030

What can AI do in 2030 that it cannot do today? Name capabilities, not benchmark scores.

A saturated benchmark tells you the benchmark is finished. It does not tell you the job is.

Is it a math genius that still can't complete a NYC building permit correctly?

The interesting failures aren't hard problems. They're tasks whose difficulty is context, paperwork, and knowing which of nine agencies signs off in which order.

What's the better predictor of what stays hard: the reasoning a task demands, or the institutional trivia it demands?

Which single capability, arriving next year, would move your timeline two years?

Does the open-weight lag shrink faster than the payback period on a frontier training run?

What's the first observable if it does: a lab skipping a generation, or a pivot away from per-token pricing?

Does frontier training need to pay for itself through model access at all, or is it a loss-leader for internal automation?

Conditional on AI being as transformative as the fast side thinks, what does measured real GDP per capita actually do?

If the answer is "not much," what should we be forecasting instead?

A forecast nobody can resolve is a mood. Name the series.

What do 10-year rates do over the next three years, and would a rise be about AI at all rather than deficits, oil, and the Fed?

If we're right, why aren't long real rates already much higher? Who's wrong, us or the bond market?

What friction specifically prevents long real rates from pricing this view, and what makes it persist?

Which sectors add jobs this decade, and which get hit hardest?

Is the entry-level squeeze about AI, remote work, credential oversupply, or the rate cycle, and how would we tell?

If AI takes the junior work, who pays to train the seniors we need in 2035?

The output was cross-subsidizing the training, and no firm funds training whose returns walk to a competitor.

What's the next data science: a job class smart people can retrain into in a year?

Anyone who names a skill is betting the skill stays hard. Anyone who names a license is betting it doesn't.

Does Silicon Valley become Detroit — developer unemployment in the industry that automated its own core job — or does it double down as the high-value epicenter of AI?

Detroit didn't lose because cars stopped mattering. It lost because building them moved.

Do the places that embrace the physical applications zoom ahead — the Sunbelt on driverless, the permissive states on data centers — and by how much?

Which metro gains the most this decade and which loses the most, and is that a story about AI or about where the electricity and the permits are?

If the work is done by models, does anyone need to be anywhere — and if not, why is everyone still moving to the same eight cities?

Which countries are genuinely players, and on what: models, compute, energy, or fabs?

How far does the US government go toward centralizing or directing frontier development?

Is there a meaningful US–China understanding on AI safety by 2030, and what would it cover?

Do open weights get restricted for safety reasons, and does that just mean the open frontier becomes entirely Chinese?

Restriction and the China question are one question, not two.

AI sharply lowers the cost of mass surveillance. How much gets built, by whom, against what constraints?

Where's the binding constraint: energy, memory, fabs, grid interconnection, or permitting?

Which has the longest lead time, and who controls it?

How many years from here, and what's the last leg to fall — robotics, chips, or power?

Does the milestone matter economically long before it's reached, or is everything in 01 through 08 decided well short of it?

Self-sufficiency is a clean line. Which of the risks we actually care about sit well short of it, and does having the line help us see them or hide them?