| Congrats on the launch nextmv! I had a chance to play with nextmv's beta when they first published it. By far the most useful aspect was the ability to bracket the decision results by calculation time. E.g., I can say, give me your best result of this choice given 100ms. This changes the typical "train / test / deploy" ML process, to something where you can get as accurate a result as possible given some block of time. This gives you a lot more options when the value of having a super-precise decision drops off a lot after say, 80% accuracy. For those of you familiar with rocketry, the technique is a lot similar to a Kalman runner [1]. Essentially when a rocket needs to gimbal adjust its trajectory, it has a ton of uncertainty about the nature of the environment, but it does an excellent job of making a fast educated guess for the simple purpose of "get me to this orbit and don't crash". More generally, this gets to the core of the issues discussed in part 1 of the a16z article about AI companies [2]. Specifically that modeling to get to accurate result is a huge and hidden cost of ML, which makes it distinctly different from software startups. Decision science is an attempt to bridge that gap. [1] https://www.bzarg.com/p/how-a-kalman-filter-works-in-picture... [2] https://a16z.com/2020/02/16/the-new-business-of-ai-and-how-i... |