What we Learned About AI/ML in 2018
Wired’s December 2018 cover “Less Artificial, More Intelligent,” and its bold subtitle, “A.I. is already changing everything.” makes crystal clear that artificial intelligence and machine learning are no longer theoretical. How does this reconcile with our experiences as a machine learning delivery software startup-business?
Our customers and our conversations shifted from science-fiction to non-fiction in 2018. Where we used to spend a majority of our conversations explaining that AI/ML was real and could help, we had productive and candid conversations on leveraging AI/ML into businesses in practical and material ways. An advertising customer shifted from hoping AI/ML could automate and optimize all of their investments to incorporating it into their planning and review cycles. A media customer shifted from hoping AI/ML would find them the next Taylor Swift to helping them find the next 10 creators similar to their most successful current clients.
On a global level we’ve observed some very important evolutions:
- Availability of open source algorithms for the most common problems
- Acceptance of AI/ML across industries including industrial, healthcare, finance and media
One of our most frustrating events was a healthcare customer walking away from a highly valuable development. After developing a 95% accurate patient behavior prediction model, they failed to adopt the technology into operations out of concerns for additional work, data privacy and process changes. We were reminded that organizational inertia is still the biggest bottleneck to realizing the full potential of AI/ML.
AI/ML is still often feared in the same way automation was, a threat to jobs and a technology that can’t possibly know as much as a person with decades of experience. [SOURCES] . Elon Musk has been simultaneously leading technology while stoking fears. At the same time, Musk has been open about recognizing the combined value of humans and robots. Since the beginning we’ve been proponents of AI (Artificial Intelligence) being replaced by IA (Intelligence Augmentation). Taking ML (Machine Learning) as an example: if the goal is to use ML to find patterns faster and more accurately than humans, who tells the ML what patterns matter?
IBM’s AI efforts, called Watson, reminded us that there are still significant gaps between potential and value. Execution matters. After failing publicly in 2017 at MD Anderson, layoffs were announced in mid-2018. One laid off engineer captured this discrepancy well:
“It’s like having great shoes but not knowing how to walk — they have to figure out how to use it.”
We’ve continued to be disappointed by the general misunderstanding of AI/ML ecosystems. There is still an assumption that AI/ML will come to life if the “perfect algorithm” is developed. On the contrary, even a perfect algorithm is only perfect for a moment in time, supported by data inputs, feedback loops and computing infrastructure. This misunderstanding is realized when organizations hire a data scientist (or ten) who fail to deliver any meaningful outcomes despite commanding premium salaries.
We expect to see Data Science as a role, evolve and dissolve. Specifically, we see data science as a combination of three skill sets:
- Mathematical statistics
- Software engineering
- Subject matter expertise
Further, these skills (and people) will only prove useful when integrated into business operations and supported by computing infrastructure. To be blunt, data science (and data scientists) who live in a bubble will fail to deliver.