Tangentially related 1: 12 Aug 2020

Robotics in Boston, Apple & AI, experts' takes on the future of AI. What's happening today in automation, AI, and robotics?

Disclaimer: I am separating the newsletter from the essays. If you have feedback on the format, please respond or reach out. You can still expect essays on Fridays. These will be less frequent, but with a bit more commentary than I used to provide.

Robotics in Boston

There is a push to make a robotics hub in Boston. I’m on board for this.

Marques Brownlee reviews Boston Dynamic’s Spot.

When the biggest tech YouTuber gets his hands on a robot that has gotten a lot of press [1, 2, 3, …] in the last few years, I knew the video was going to be a gold mine.

  1. Spot’s movements are so, so smooth. When you watch old Boston Dynamics videos in the wild, the robots are jerky, snappy, and jittery. In this video, including the bloopers, you’ll see that the robot is much smoother. Smoothness is very useful in the product-side of a robot because high force maneuvers cause more wear on the motors and harm battery life (the other side of productizing a robot is robustness, but I think that is talked about more).

  2. The robot has applications and is not just a toy. At the end of the video, they detail the sensor suite (data access) and the payload capacity of this robot. If this iteration doesn’t sell well, as they incorporate more new technology and up their supply chain, the 5 year older sibling of this in 2025 will be a hit.

  3. There’s not a lot of machine learning in ways people may think. Yes, the robot likely uses deep learning for vision, but Boston Dynamics is famous for not using reinforcement learning or other hip techniques on their robots. Reinforcement learning could be good for fine-tuning performance in some extreme scenarios, but it goes to show that successful robotics companies are still more about classic engineering rather than hot new research.

Lex Fridman interviews MIT robotics expert Russ Tedrake.

The AI Podcast is fantastic, and this interview has a lot of gems. Specifically, I am always interested in what robotics leaders have to say about deep learning and its future in robotics. I agree a lot with Russ’s take:

Deep learning makes it easy to get things to work.

Let’s unpack this seemingly simple statement. Deep learning gives roboticists a tool that will work most of the time for tasks, but it is hard to use in robust and safe ways — things that normally important after showing something works.

  • The other side of the deep learning coin is control theory, which solves systems by extremely clear thinking, modeling, and hypothesis. I love the interplay between these fields (and why I think electrical engineers make excellent machine learners, they were trained to understand model types and weaknesses).

  • Also, they discuss why roboticists are normally happier with their work — sometimes I feel this, but it is a win to see things come to life in front of you. Russ also has one of the most bad-ass commutes I’ve heard (10 mile barefoot runs each way).

Apple & AI

The story of Apple using artificial intelligence is just getting started — and it is going to play out in a huge way. Let’s start with a storytelling the arrival of Apple’s new AI chief, John Giannandrea, who they persuaded to jump ship from Google. The story is nicely summarized by this quote:

“When I joined Apple, I was already an iPad user, and I loved the Pencil,” Giannandrea (who goes by "J.G." to colleagues) told me. “So, I would track down the software teams and I would say, ‘Okay, where's the machine learning team that's working on handwriting?’ And I couldn't find it.”

Apple has been so detailed orientated and product-driven that a) they didn’t need to have fancy machine learning features like the Pixel and b) they hadn’t built out the teams to do so. A couple of years later. character recognition is a big feature in their upcoming software release (iPad OS 14). I think this narrative also fits well with the context of Tim Cook as CEO. You should also read this article from the WSJ on Apple’s evolution with the lens of machine learning accelerating. The trend is described by this quote:

Don’t ask what I would do. Do what’s right.

I think this is Apple’s approach to machine learning now. It’s not about forcing features, it’s about making an impact in dramatic ways, in an incredible ecosystem that they built from the ground up. Fittingly, Apple also just launched a blog and paper index for their machine learning research, which you can find here.

[Link, link, link]

Mechanisms, future of AI, social good

I registered ($10, open to anyone) for an interdisciplinary workshop on Mechanism Design for Social Good from August 17th-19th.

If you’re passionate about making fair and equitable technology, why not join?

Also, I watched this round-table from Berkeley & MIT on the state of AI in the world. If you haven’t considered the broader impacts and future of the field recently, I recommend giving a watch. Please keep thinking deeply about where your work will lead.

Accountable machine learning engineering

I’m worried people rush to make their machine learning projects, make their cluster the biggest available to a company, or deploy products without thinking. I am always looking for tools to mitigate these issues — two this week I want to highlight.

  1. A tool to estimate the energy usage of a certain python script. When I was launching 300 GPU jobs to a cluster cramming for a paper deadline, it would be really good if something told me the cost upfront. I would love to see big companies incorporate something like this tool in the workflow all the employees use. Even if the data center uses green energy there is always an environmental cost to training a new model.

  2. My resources on ethical training for roboticists. In talking to a friend getting a Ph.D. in psychology we realized the disparity in ethical training across different fields. Essentially, “medical-like” fields get a ton and everyone else may get none. The problem is, the leaders in computer science create the tools that dictate most of the mental state of billion of people, so we need our training. If you have any resources that may help, please respond!


E-commerce and robotics

Another article on e-commerce and robotics. It’s fun to keep tabs on how the mainstream news follows the area, but there’s one takeaway I keep reaching from these — the most valuable robotics companies of the next few years will offer autonomy packages, not new hardware.

The four workers are actually 260-pound industrial robot arms. Manufactured by Waukegan, Ill.-based Yaskawa America, they are equipped with computer vision and artificial intelligence from San Antonio-based Plus One Robotics.

In this case, One Robotics is the winner. Much like the Berkeley startup, Covariant. Platforms scale-like digital companies.

Internet infrastructure

The Atlantic (recently becoming more of a fan) had a piece on how the internet infrastructure handled the swing in demand caused by the coronavirus pandemic. Early on, I was impressed by AT&T maybe for the first time in my young life.

AT&T rehearses for disaster. Last May, the company ran an internal war game on how a pandemic would affect its ability to keep phone and internet service running. The company does these exercises routinely to try to get ready—to build teams of people and their reflexes, and also to understand what they will need on the ground.

I’m thankful to receive pointers to things I and the rest of this community may be interested in - please be in touch!

My recent essays

Hopefully, you find some of this interesting and fun. You can find more about the author here. Tweet at me @natolambert, email democraticrobots@gmail.com. I write to learn and converse. Forwarded this? Subscribe here.