Paul Saffo says that communication technologies historically take 30 years or more to find their true purpose. Just look at how the Internet today is different than it was back in 1988. I am beginning to think this idea applies also to new computing technologies like artificial intelligence (AI). We’re reading a lot lately about AI and I think 2018 is the year when AI becomes recognized for its much deeper purpose of asking questions, not just finding answers.

Some older readers may remember the AI bubble of the mid-1980s. Sand Hill Road venture capitalists invested (and lost) about $1 billion in AI startups that were generally touted as expert systems. An expert system attempted to computerize professional skills like reading mammograms or interpreting oil field seismic logs. Computers were cheaper than medical specialists or petroleum geologists, the startup founders reckoned, so replacing these professionals would not only save money, it would allow much broader application of their knowledge. Alas, it didn’t work for two reasons: 1) figuring-out how experts make decisions was way harder than the AI researchers expected, and; 2) even if you could fully explain the decision-making process it required a LOT more computing power than originally expected. Circa 1985 it probably was cheaper to hire a doctor than to run a program to replace one.

But now, approximately 30 years later, AI has come back to life. Part of this is simply Moore’s Law. One million dollars worth of 1985 computational power costs less than a buck today, making those software experts way cheaper to run. But wait, there’s more! A second reason for AI’s resurgence is the availability of huge online data sets. Over the past 30 years nearly all the information that formerly resided on paper was reduced to electrons making true machine learning possible. This can’t be over-emphasized. Anyone my age remembers when early search engines indexed thousands of web pages, not billions. Highly technical data generally wasn’t available online in any volume but now it is.

The third major reason for AI’s resurgence is today we don’t even try to pick some doctor’s brain to build an expert system, instead empirically deriving skills directly from the data. Taking this one step further, we are moving to a system where we don’t even start with a question, just the data, allowing cloud-based systems to find what’s learnable from the data. That’s allowing AI to come up with its own questions and it’s the emerging trend on which I am trying to focus this prediction.

AI, analytics, and Big Data are keywords here. If you want to know more about this, early last year I wrote a series of three columns totaling more than 10,000 words on the topic. You can find those here.

If AI is going to figure out questions that need asking it is probably a mistake for me to start listing them here. But I can give you an idea what such questions would be like.

Look at this old chart I got from my friend Frank Starmer. It shows U.S. death rates from infectious and non-infectious diseases from 1900-1996. During this period we basically defeated tuberculosis and polio through vaccinations. We took a big bite out of influenza and pneumonia the same way (that big chart spike is the Influenza pandemic of 1918 — the Spanish Flu). We developed water and sewage treatment systems to defeat cholera and similar diseases. My mother, growing-up in Arkansas in the 1920s and 30s  knew people who had chronic American-Grown malaria and that’s gone, too. The slight infectious uptick toward the end of this chart, in case you are wondering, represents the HIV/AIDS crisis, which peaked in those years, as well as an aging population that was still susceptible to pneumonia.

My reason for showing this chart is not because we did such a great job of curing infectious diseases but because we did such a poor job of curing everything else.

Everybody dies, of course. But this chart also predicts life expectancy. Look at 1900 and you’ll see there were 1700 total disease deaths per 100,000 population. That’s a 1.7 percent annual mortality rate suggesting a life expectancy of 100/1.7=58.82 years for both men and women. The 1996 total disease rate looks to be about 900 per 100,000 or 0.9 percent suggesting a life expectancy of 100/0.9=111 years! We don’t actually live that long for many reasons including infant mortality, wars, accidents, occupational, environmental, and lifestyle diseases, but it shows what’s possible with essentially zero contribution from improvements in non-infectious diseases.

If we found the data to extend this chart much of the news would be good thanks to the Clean Water Act, Clean Air Act, and fewer people smoking — not especially from improvements in medical treatments. Yet the cost of medical treatment continues to rise. Hopefully much of this represents our current investment in the very work I’m advocating here.

We’re getting to the point where AI should begin to organically suggest approaches that will help us improve medical outcomes. The big low-buck solutions like immunizations have for the most part already happened. Future gains will have to come incrementally, often one gene at a time. But that’s just where AI should shine, mining medical gems from all that data in our smart phones, fitness trackers, and home DNA tests.

Thirty years into AI, it’s time to start seeing significant rewards from this approach in many fields, not just medicine.