Three NOs
In my role as Chairman at the Alchemist Accelerator, I was invited to do a talk at Stanford’s SystemX Spring workshop on Data Analytics & AI for IoT Applications. For those of you who don’’t know, the Alchemist Accelerator is focused on enterprise software and in the past seven years, 137 of our companies have received nearly $1B of investment, with 34 of them being acquired.
Like many meetings the physical event has been cancelled and replaced with a 2-hour web conference. The organizers asked that we all spend 5 minutes talking about “The biggest challenges in the Data Analytics for Industrial IoT space, and possible paths forward.”
I opened my 5 minutes with a word of thanks to all the front-line healthcare workers around the world who are risking their own health for the rest of us. If you want to hear the talk and see the slides check it out at the Stanford website (the first 20 seconds is silent).
While many of us recognize the enormous potential for Internet of Things to transform 2/3 of the world’s economy, the reality is progress has been slow, particularly in the area of the application of AI and analytics to learning from Things/machines. I think there are at least three challenges, three Nos we have face.
NO #1. There is NO Internet of Things
We talk about the “Internet of Things”, but as far as I can tell there is very few things are connected, unlike the Internet of People. According to Statista, the current number of smart phones in the world is 3.5 billion, which means almost 50% of People are connected. Are anywhere near 50% of the world’s Things/machines connected?
There are 500,000 healthcare machines (CT, MRI, blood analyzers, Xray, gene sequencers) in all the children’s hospitals in the world. Less than 1% of these machines are connected. Doosan Bobcat has 750,000 construction machines in the field with less than 1% connected. There are 1.5B automobiles on the planet, how many of these are connected? And in the world of water treatment there are at least 2B machines designed to treat, process and analyze the quality of the water. Significantly less than 1% of these machines are connected.
Of course if these machines are not connected, then there is no way to collect the data required to learn from the data. Many of you may be familiar with the Jeff Dean’s graph, which articulates with more data and more compute, neural network technology can achieve monotonically increasing levels of accuracy. With cloud computing we have access to nearly unlimited compute for no money. We need the data. We need to get connected so we can collect the data, both machine (e.g., laser power level of the gene sequencer) and nomic (e.g, the genomic sequence) data.
The Covid-19 pandemic has shown us the power of connection in the Internet of People. On Monday March 16 the SF Bay Area went on lock down. Water2Table, a wholesale supplier of seafood to local restaurants went from 100+ orders to 6. In less than 24 hours they flipped to a direct to consumer business and are surviving. No way this would be possible without an Internet of People.
NO #2. NO one knows how to buy analytics.
Lately every time someone says “AI” I replace it with the word “software”, as most consumers of analytic software can’t tell the difference. While we might have educated people on how to buy workflow software (purchase-to-pay, order-to-cash, hire-to-fire) the world of analytics is still very fuzzy. In the enterprise space some companies have emerged, e.g. Alteryx, to provide the education and tools to analyze IoP (Internet of People) data, but we have yet to see that in the world of IoT. And as we start using more analytic technology, who and how will we make decisions around false positive and false negative rates? What is specificity? What is sensitivity? What is a ROC curve?
Workflow automation has always been about making something imprecise, precise. For example, without a purchasing application everyone in a company could purchase in different and unknown ways. Workflow software standardizes the processes so it works the same way every time. But if you’re analyzing data to determine if a credit card transaction is fraudulent, sometimes you’ll cry wolf when the transaction is not fraudulent, and sometimes the analytic software will say it’s OK, when it’s really not. How do you make the decision of how many unhappy customers you’re willing to tolerate?
Take this to the world of healthcare and imagine a Covid-19 diagnostic application. What false positive rate is OK? And if you’re constantly getting more CT scans to train your diagnostic, when do you re-train? and what criteria is there for releasing version 2? If there are multiple suppliers how do you choose which diagnostic to purchase? How do you buy imprecision?
NO #3. There are NO analytic applications.
Today most of the analytic software is infrastructure software. This was also the the case for enterprise workflow applications in the 1990s. In the 1st era, infrastructure software companies emerged like Microsoft and Oracle, which focused on developers. These developers used Microsoft Visual Basic and the Oracle database to build custom applications for the enterprise throughout the 90s.
By the late 90s the 2nd era of enterprise software began with the creation of packaged on-premises enterprise workflow application. Companies emerged including PeopleSoft, Siebel, SAP and Oracle. So enterprises didn’t need to hire programmers to develop these workflow applications, they only needed to buy them and manage them.
The 3rd era began in the 2000s with the delivery of these applications as a cloud service. Examples abound including Salesforce, Blackbaud, Workday and ServiceNow. This era eliminated the need for the enterprise to hire operations people to manage the applications and has even more accelerated the adoption of packaged enterprise applications. While you could still hire programmers to write a CRM application, and operations people to manage it, why would you?
Now let’s look at the world of analytics. We’re still in the 1st era, with many infrastructure software companies, public and private. Today’s custom analytic applications require programmers, data scientist, data engineers and devops skills to manage them in production. Just as we saw in the 1st era of enterprise workflow applications deployment will be limited to those enterprises which can afford the talent and time required to build these custom applications.
Just as with workflow application we need to start building packaged analytic applications. These enterprise analytic applications will focus on workers, not developers; have millennial UIs and use many heterogeneous data sources. The best example from the consumer world is Google Search. It’s an application focused on the worker, not the developer, with a millennial UI and using many heterogeneous data sources. Pull back the cover and you’ll see a ton of technology inside. And of course Google can hire the programmers, data scientists, data engineers and devops people to build and manage the search application. So when will we get packaged enterprise analytic applications?
Answering these three challenges is going to be essential to generating value for companies, customers and investors.
— —
If you work at ARM, Boeing, Ericsson, Google , IBM, Intel, Lam Research, Mentor Graphics, NEC, Qualcomm, Bosch, Samsung, TI, Micron, Microsoft, Toshiba, Xilinx, Apple, AMD, Applied Materials, Toshiba or Facebook or any of the other member companies and are interested in viewing the recording of the 2 hour web conference or the presentations (from McKinsey, Cisco, Microsoft, Stanford professors and more) just register.