Old challenges and new opportunities

Summary

A pond, and a question

The lily pads on a pond double every day, and on day 83 the pond is completely covered. On which day is it half covered? It is half covered on day 82, one day before it is full, and a week earlier, with less than 1% covered, it still looks empty. Most people know the answer and have to think about it anyway, because exponential growth is logical without being intuitive.

AI poses the same problem at scale. Its capability compounds, while everything that has to absorb it moves at the speed of building sites, courts and reorganizations. The keynote asks who wins, and how, at four levels that run from the world down to the workshop itself.

World: capability compounds, the brakes add delays

Epoch AI’s chart of the “plunging price of thought” shows AI getting about 13 times cheaper every year between 2021 and 2026. The price of computing took roughly 35 years to fall that far.

The keynote’s reading is that what drives AI forward feeds on itself. Models carry a growing share of the work of building the next models, most users don’t have to install anything new, and no lab or country will slow down on its own. What holds it back can only be built one piece at a time. A chip fab takes three to five years to set up and a power plant with its grid connection five to ten. Trust grows with each standard and each court ruling, and organizations change at human speed.

So the exponential wins, and the linear brakes decide when and where.

Society: who gets ahead when thinking gets cheap?

The keynote offers three trends here and no answers. In the US, the share of children who earn more than their parents fell from 92% for those born in 1940 to 50% for those born in 1984 (Chetty et al.). The early part of that fall has to be discounted, because the children of 1940 grew up in the post-war boom and its industrialization. From those born in 1960 on the trend is still there and slower, from 62% down to 50%. Labor’s share of income has shrunk in most countries, and since the 1990s the middle tier of jobs has thinned out while high-end professional work grew. AI now reaches that top tier too.

Cheap thinking also works in the other direction, and three of its effects may keep the way up open.

• Cheaper to start. One person with AI can do what once needed a company, at least in office work.
• Cheaper expertise. Education, legal advice and medical advice were costly steps on the way up, and AI is making them cheap. The evidence is still early.
• New kinds of work. About 60% of the jobs of 2018 didn’t exist in 1940 (Autor et al.). Nobody can say yet what the new ones will be.

None of this fixes the timing, because laws, schools and tax systems take decades to change while AI moves in quarters. For the individual, the keynote’s advice is to “educate yourself up”, because climbing through work alone is getting harder while knowledge and ability still count.

Organization: adoption is faster, adaptation is slower

Every organization has two tasks. The first is to adopt the technology. That can go quickly when four forces work together, which the keynote bases on Nicolai Tangen’s account of AI adoption at Norges Bank Investment Management. Leadership has to want it, and people need the tools and the data access. Training has to reach across teams, and someone has to translate between the business and IT so that results end up in daily work.

Adapting the organization is the second task, and the harder one. Argyris and Schön called it double-loop learning. The single loop asks how to do the work, and AI has made that nearly free, because techniques can be tried and thrown away in hours. The double loop asks why, and questioning the assumptions behind the work, the real operating system of an organization, costs what it always did. What makes it expensive is admitting that an assumption was wrong.

The workshop: what got cheap, and what to trade

The workshop brought together about 80 people who develop and use simulation software, from universities, software vendors and industry. Their old problems are familiar well beyond research. Code outlives its authors, knowledge leaves with the PhD student who wrote it, a proof of concept never becomes a product, and open science collides with industrial IP.

Much of this could be fixed by work that was never affordable, and that work has just become cheap. Tests, build pipelines, documentation and refactoring all fall into that category. Nils Boeffel built a production system with AI for a small metal manufacturer in about five months, and between 30 and 40% of the effort went into tests and pipelines, because AI output keeps diverging unless it is given a point to converge to. Software that can be rebuilt this quickly is worth less as IP than it used to be, while data and validated real cases are worth more.

That changes what the four groups at the workshop (professors, PhD students, software vendors and industrial users) can offer each other. The keynote gave the working groups four questions to take along.

1. What must you keep?
2. What got cheap?
3. What do you lack?
4. Who here has it?

A good trade costs nobody what they must keep and gives each side something it could not get alone.

One problem stays open. Writing was junior work and checking is senior work, so if AI does the writing, fewer juniors learn to check, and it is unclear where the seniors of ten years from now will come from. Simulation engineers already own the discipline that helps. A simulation isn’t trusted until it is verified and validated, and AI-assisted work deserves the same treatment.

Who wins, and how?

The question from the start has a different answer at each level.

• World: the exponential, while the linear brakes decide when and where.
• Society: those who “educate themselves up”.
• Organization: those who adapt as fast as they adopt.
• The workshop: those who ask the critical questions and change how they work together.

The closing slide reads “1+1=3, for extremely large values of 1”. The keynote’s point is that a workshop like this one is a chance for the sum to be greater than its parts, at a time when more can be achieved in a short time than ever before.