or: *Why a billion dollars still can’t buy you a self-driving car in 2016*
The biggest buzz word of 2016 is Artificial Intelligence (AI), which makes the self-driving car queen bee. Talks given at the [O’Reilly AI Conference](http://conferences.oreilly.com/artificial-intelligence/ai-deep-learning-bots-ny) in New York, NY, which I attended, were filled with honey.
My key takeaway was that industry leaders are pouring heavy investments into automating away human jobs. Uber, for instance, recently acquired the 6-month-old company [OTTO](https://www.ottomotors.com/), a self-driving vehicle designed exclusively for material transport. Uber took on OTTO's 70 employees at a $700 million price tag. Technologist [Sebastian Thrun](https://en.wikipedia.org/wiki/Sebastian_Thrun) estimates the going rate for self-driving talent is [$10 million per head](http://www.recode.net/2016/9/17/12943214/sebastian-thrun-self-driving-talent-pool). A great industry to get into then, provided you are ready to overcome the challenges presented by implementing AI.
One bottleneck at any company launching a fully automated self-driving system is the inattention valley. A 99% safe car lulls the passenger into a false sense of security. That 1% discrepancy is when hands need to be on the wheel, but instead they are holding an iPad in the back seat. Research is steadily closing that gap, but the threshold for practicality where every car on the road can be self-driving is still a distant future. Current self-driving implementations use a system called [Machine Learning](https://en.wikipedia.org/wiki/Machine_learning). These implementations can quickly teach themselves how to do 90% of the things right, but chasing that last 10% has a potential to completely wreck the previously completed work.
Developing 1000 times faster with machine learning allows us to create world class machines that can translate dictionaries, trade stocks profitably, and search the entire web in milliseconds. The reason why product development can be so fast is that we no longer require a team of experts to write out several million rules. Technical debt comes from cutting corners to meet release deadlines and putting off doing things the right away until after the launch.
In traditional software development we love modularization because one-off problems are contained within their own cell. If we have one piece down the logic flow that goes bad, we can just flip a True / False boolean value and everything is alright again. In Machine Learning we feed a curated database to a black box system which will learn the rules of the game by itself. You could imagine how it creates a reputation for itself as being the [highest interest credit card of technical debt](http://research.google.com/pubs/pub43146.html).
With machine learning, we currently have no way to separate out problem areas form the rest of the system. Changing one piece changes everything. As the debt piles on at speed, it takes a certain type of person who can make adjustments and fix one-off cases. An end-to-end understanding of the system is required to make non-breaking changes. You will eventually be able to use your growing intuition to find unoptimized pieces of the puzzle. Once that is achieved, your mental capacities are free to experiment and improve the performance of the model.
Overall I took a lot away from the conference, and would recommend it to anyone interested in Artificial Intelligence. It was my first time in New York, and I made sure to fill up on pizza and lox bagels. Next year, I'll make sure to carve out time to go to the top of the One World Trade Center and explore what New York has to offer outside of Manhattan.
I'm extremely thankful for CoverHound's positive position on helping their engineers grow, and sending me to New York for this amazing event. If you happen to be looking for a new opportunity, I encourage you to check out our [jobs page](https://jobs.lever.co/coverhound).