Research Money ≠ Research Effort

By Brian Wixted

Lately, I have been musing about the increasing problems of measuring innovation and technology advances. This esteemed publication calls itself RE$EARCH MONEY and for good reasons. Measuring research expenditure consistently across countries and sectors dates back to the 1960s. Expenditure-based measures have been the most successful and best used indicator of innovation to date because the data can be collected with minimal effort and maximum accuracy (not perfect) and interpreted without huge complexities.

Despite efforts to innovate the family of indicators, research expenditure remains the go to indicator. The R&D expenditures measure has massive limitations, but these have always appeared to be fewer than any alternative – until now.

The essential point of research money is that we can compare sector against sector (government, business, higher education or charities), organizations against one another. Money buys a certain amount of labour and equipment, and we can even build adequate measures like purchasing power parity to adjust for international currency differences.

However, it is time to point out the obvious: technology is changing the relative abilities of sectors and companies to build technology, and that will distort research purchasing relativities. Specifically, companies with significant artificial intelligence (AI), deep learning and big data capabilities are able to ‘learn’ for a fraction of the typical costs.

Take AlphaGo, the first AI champion of the complex Asian strategy board game Go, a computer program which learnt by playing. It was not too long ago that such success was impossible or only through hundreds of thousands of hours of programming by software engineers. The next iterations, the even more capable AlphaGoZero and AlphaZero, required less programming and learned faster. These are only the most obvious examples. Behind the scenes, deep learning AI is being deployed across a huge range of technology activities.

DeepMind, Alphabet Inc.’s AI company and developer of AlphaGo, helped Google reduce energy costs in its data centres by 40%. Rio Tinto, a global mining giant, has a data centre in Brisbane, Australia monitoring in real time all its ore crushing facilities around the world to watch for small tweaks possible in real time to improve energy efficiency. Rio Tinto estimates that 5% of all energy usage on the planet is used for crushing rock so small improvements mean significant savings. And each time I pick up an article there are new examples of AI in use -- from Amazon to JPMorgan.

Stanford’s One Hundred Year Study on Artificial Intelligence, or AI100, reported in 2016: “From autonomous vehicles, to advertising, to platform companies like Uber and Facebook, big data and AI are becoming ubiquitous. During the first Defense Advanced Research Projects Agency (DARPA) ‘grand challenge’ on autonomous driving in 2004, research teams failed to complete the challenge in a limited desert setting. But in eight short years, from 2004-2012, speedy and surprising progress occurred in both academia and industry. Advances in sensing technology and machine learning for perception tasks has sped progress and, as a result, Google’s autonomous vehicles and Tesla’s semi-autonomous cars are driving on city streets today.”

Our default is to think of AI as ‘digital’ in the same category as Cambridge Analytica, gamification with reward points or image processing, but it is far more. So much more, that most of us do not indeed have a real sense of where it is being deployed.

In the past, money was a useful shorthand for ‘effort’ and was roughly correspondent across sectors and countries.  Today this is clearly no longer the case. The challenge is that the so-called FAAMG (Facebook, Amazon, Apple, Microsoft and Google) companies, and others such as Elon Musk's group, clearly have deeper AI capability than others. And they are using it to develop and innovate old fashioned hardware, not just digital life, because, in the case of the things - the physics of matter, matter. Every 1% advantage in design through strength or lightness improves cost to performance.

By contrast, government and university research will continue to fund humans (graduate students) and will continue to focus on new areas of knowledge that by definition mostly have little potential for AI exploitation (even if grants could hope to cover the costs). However, companies with deep learning capacity may simply exploit university research even without going through traditional commercialization routes. Much of science policy discussion seems stuck in last century while the entire configuration of capability of economies is changing here and now.

If we interrogate Schumpeter’s standard five innovation structures, we can begin to see the challenges ahead.

Products: AI is both now a service and changing how products are designed and services are offered.

Processes: From logistics to loan approvals, AI is fine-tuning back office processes.

New Markets: Cambridge Analytica is just the latest example of entirely new market generation only possible through deep learning.

New/changed inputs: As noted above, system optimization appears to be one of the early big advantages offered by current generation AI systems. Electricity grids with significant renewable energy inputs will use AI to harmonize demand, supply and storage in ways that could change the patterns of production and trade.

New economic configurations: The orphaned cousin in indicator design, it has been hard to conceive of, let alone measure, what this might even capture. However, it seems likely that AI might help generate radically different economic configurations from those operating today, such as the autonomous vehicle ecosystems.

Reading through dozens of reports that attempt to articulate and estimate the difference AI is already making, it is clear that AI, big data and deep learning are already having an impact. It is therefore now a real innovation factor and must be considered.

Perhaps, what matters in the end will be measuring the changing configurations of capabilities in economies, but I cannot yet conceive of how that would be possible.

In the meantime perhaps a name change for RE$EARCH MONEY is coming?