O’Reilly AI Conference 2019 — NYC Edition

O’Reilly’s AI Conference was a fascinating experience, and in many ways, a first for the Hop team. Though we attend technical ML conferences all the time, this was our first time attending an industry-oriented conference.

We’ve tried to capture some of the key trends and takeaways we saw at the various sessions we went to:


Business use-cases for AI are everywhere, but identifying them remains a challenge.

It takes a particular blend of skillsets collaborating across domains to find the most valuable use-cases, and organizations with a more tech-first DNA are outpacing their competitors. Netflix / Facebook / Mastercard / Bloomberg / etc were all able to point to a nearly endless variety of ML use-cases used in real production systems every day, and a number of startups (especially in healthcare) were also able to demonstrate meaningful successes. At the same time, a number of more traditional organizations were able to point to real cultural challenges they face with trying to adopt ML in-house. While there’s a lot of up-front excitement, most projects get relegated to ‘Proof-Of-Concept Purgatory’, and never seem to deliver any meaningful value. This aligns closely with our own experience — real value is only unlocked for our clients when we actually deploy the solution to production and get it in the hands of real customers.


Lack of talent is hurting everybody.

The skills needed to successfully execute a production-grade ML project are so specialized and rare, and the top performers straddle a few different areas of expertise that are each individually valuable. Starting and growing a team remains a challenge for all but the best companies. (Too true!) Some organizations are addressing this by building ML systems that require less human attention to train, deploy and maintain.

Paradoxically, code is getting simpler though systems are getting more complex.

One speaker pointed to performance improvements by switching from a 500k lines-of-code hand-crafted system to a 500 lines-of-code ML-powered system. Admittedly, the ML-powered system had several research papers behind it and required a different level of skillset to work with.


Data is the key to everything.

Several speakers pointed out that the differences in effectiveness between competing ML-powered products often can be traced to the source dataset. Indeed, the data is often the primary proprietary component — both the algorithms and hardware are (at least so far) widely available. The speakers that spoke to industry trends also pointed out that some of the biggest / quickest wins went to folks with interesting / novel data. This has interesting ramifications for the industry as a whole. We’re big believers of this ourselves, and have long advised our clients on how to be strategic with their data-acquisition and management efforts.

Huge Impact in Healthcare.

For healthcare specifically, AI is having a huge impact, but you really need to find the right healthcare / data partners. It’s very easy to get this wrong if you’re not careful — a few different speakers pointed to false starts.


Overall, a great conference with a lot we learned from it and some really interesting insights. It was great to hear from the trenches of how ML systems are being deployed in the real world, and the challenges that brings. Looking forward to going again next year!