Why it becomes harder to predict technological change

Decision-Making

Perhaps we should have seen this acceleration coming. In the 1930s an American aeronautical engineer named T. P. Wright carefully observed aeroplane factories at work. He published research demonstrating that the more often a particular type of aeroplane was assembled, the quicker and cheaper the next unit became. Workers would gain experience, specialised tools would be developed, and ways to save time and material would be discovered. Wright reckoned that every time accumulated production doubled, unit costs would fall by 15 per cent. He called this phenomenon ‘the learning curve’.

Three decades later, management consultants at Boston Consulting Group, or BCG, rediscovered Wright’s rule of thumb in the case of semiconductors, and then other products too. Recently, a group of economists and mathematicians at Oxford University found convincing evidence of learning curve effects across more than 50 different products from transistors to beer - including photovoltaic cells. Sometimes the learning curve is shallow and sometimes steep, but it always seems to be there.

The learning curve may be a dependable fact about technology, but paradoxically, it creates a feedback loop that makes it harder to predict technological change. Popular products become cheap; cheaper products become popular.

Forecasts photograph the technology. Learning curves quietly replace it while the picture develops. Every unit made teaches something: a quicker assembly, a better tool, a supplier worth avoiding, a material that can be removed. Cost falls. Lower cost attracts demand. Demand funds another round of learning. A modest lead can start feeding itself.

Comparisons based only on today’s price are often strategically useless. The expensive technology with a steep learning curve may outrun the mature cheap one. Or it may remain an expensive hobby because volume never arrives. The direction matters less than the mechanism creating it. Solar equipment, installation, finance and customer familiarity can improve together. Repeated software deployments reveal the exceptions worth automating. A hardware platform attracts compatible products, which make the platform more useful, which attracts more users. Feedback does the forecasting damage.

Curves still hit walls. Scarce materials bite. Regulation slows adoption. Easy improvements run out. Customers may refuse the new habit required around the clever object. Anyone extending a smooth line forever is drawing fan fiction with axes.

Ask what genuinely gets learned when volume doubles. Who keeps that knowledge? Which cost declines, which stubbornly survives and which new bottleneck appears? Model several adoption paths because an early difference may compound into a different market. On the next review, compare the assumed learning with what production actually taught. The forecast should change when the learning does.

Excerpt from

The Next Fifty Things that Made the Modern Economy

by Tim Harford

View on Amazon More from this book

Behavioural principles

Behavioural ideas at play in this post

Short, plain-English explanations of the principles behind this post, with links to related books and examples in the archive.

Present bias

Immediate costs and rewards loom larger than future ones.

More quotes

Overconfidence effect

We are often more certain than the evidence really allows.

More quotes

Why it becomes harder to predict technological change

Behavioural ideas at play in this post

Sharp thinking (my rants) in your inbox