Lyft Level 5’s Machine Learning Pipeline

Lyft Level 5 recently published an amazing overview of their PyTorch-based machine learning pipeline.

They confess that a few years ago, their development process was slow:

“Our first production models were taking days to train due to increases in data and model sizes.…Our initial deployment process was complex and model developers had to jump through many hoops to turn a trained model into a “AV-ready” deployable model….We saw from low GPU and CPU utilization that our initial training framework wasn’t able to completely utilize the hardware.”

The post proceeds to described the new pipeline Lyft built to overcome these obstacles. They started with a proof-of-concept for lidar point cloud segmentation, and then grew that into a production system.

The pipeline accomplishes a lot of infrastructure wins.

Testing. The pipeline incorporates continuous integration testing, both to ensure that the models don’t regress, and also to verify that the code researchers write will run in the PyTorch-based vehicle environment.

Containerization. Lyft invested in a uniform container environment, to mitigate the distinction between local and cloud model training.

Deployment. The system relies heavily on LibTorch and TorchScript for deployment to the vehicle’s C++ runtime. Depending on existing libraries reduces the amount of custom code Lyft’s team needs to write.

Distributed Training. PyTorch provides a fair bit of built-in support for distributed training across GPU clusters.

There’s a lot more in the post. It’s a pretty rare glimpse of a machine learning team’s internal infrastructure, so check it out!

Ridesharing Moves Toward Subscriptions

In the last few weeks, both Lyft and Uber rolled out (very different) subscription pricing models.

Lyft’s All-Access Plan offers 30 rides per 30 days (up to $15 per ride) for $299. This is basically a bulk purchase option. A Lyft power user would save $150 per month by subscribing instead of paying per ride. (30 rides * $15 per ride— $299 subscription price =$151 savings)

Uber’s Ride Pass, by contrast, is essentially insurance against surge pricing. For $14.99 per month, Uber users escape the risks of surge pricing, plus they save approximately 15% off of normal fares. A user would probably need to spend $100 or more per month on Uber to come out ahead.

Each of these offers is a tiny, baby step toward all-you-can-eat subscription ridesharing. The marginal cost of each ride probably prohibits Lyft and Uber from diving headfirst into all-you-can-eat ridesharing. Hopefully self-driving cars will lower that marginal cost to the point that it becomes feasible.

My first professional job was with America Online, now AOL. I joined as a college intern in 2001, but there were still veterans around who could recall AOL’s switch from hourly to unlimited pricing.

22 years ago, AOL switched from plans that allowed for metered hourly Internet usage to plans that allowed unlimited Internet access. The pricing change brought in a wave of new customers, along with per-user usage increases that overwhelmed AOL’s infrastructure. It was one of the most significant things AOL ever did.

Ridesharing is getting closer to its all-you-can-eat moment.

Self-Driving Lyfts in Las Vegas

A few weeks ago Lyft and Aptiv announced their 5000th paid self-driving car ride in Las Vegas. The Lyft blog announcement quotes Raj Kapoor: “Lyft is the largest network currently deploying a commercial self-driving program to the public.”

I believe that is correct and may, in fact, slightly understate Lyft’s position. Lyft is one of the only companies offering self-driving rides to the general public, since Uber halted its autonomous vehicle testing program earlier this year.

Waymo and several other companies have self-driving car pilots available to pre-screened participants, but few companies are currently opening their network to any member of the general public.

Earlier this year, AAA was running a self-driving NAVYA shuttle through downtown Las Vegas, although the website currently implies the shuttle is offline and will return later this month.

Las Vegas has long been a mecca for self-driving cars for a few reasons:

  • The weather is sunny
  • The streets are wide and rectilinear
  • The Consumer Electronics Show in January has prompted companies to run demos in the city already

Waymo announced earlier this year that its Phoenix-area self-driving cars would open to the general public at some point in 2018.

But, for now, the best place for you and me and almost anybody to try out a self-driving car is Las Vegas.

Ride-Sharing Doesn’t Work with a Phone

The Cubs-Giants game went thirteen very long innings last night and ended in heartbreak (for me, at least), with the Giants knocking in a walk-off run in the bottom of the 13th inning.

It was also 11:45pm and the game had been going on for five hours.

As I stumbled out of the stadium, I realized my phone was totally dead. Five hours of emails and web browsing between innings had drained the battery.

If my phone had been working, I might have just hailed an Uber home and tucked into bed. But my phone wasn’t working.

No worries, though! In San Francisco, the train station is just blocks from the ballpark. I hustled on over to Caltrain, waited forever for the train to leave, and then learned I got on the wrong train. The train I was riding wouldn’t make its first stop until 8 miles past my house.

I disembarked the first chance I could and walked into an empty parking lot at the Belmont Caltrain Station at 12:45am. No taxis.

A gas station light flickered across the street and I rolled over and begged the attendant to call a cab. No cabs available.

Then I bought a charger from the station’s inventory and hailed an Uber, which took twenty minutes to arrive, being past midnight in the suburbs.

I finally tucked into bed at 1:30am.

So what’s the moral of the story?

Mostly that I shouldn’t have totally drained my phone battery, and I should look at train schedules.

But also that, in the days before ride-sharing, it was more common to have taxis circling around and you didn’t need a phone to hail them.

The world today is a better place because of Lyft and Uber, but it does require a phone to navigate.

The Future of Lyft (and the World) is Autonomous

Lyft co-founder John Zimmer has written a long thinkpiece on Medium, outlining the future of Lyft, transportation, America, and the world.

The headlines coming out of the piece are that most of Lyft’s rides will be driverless by 2025, and that private car ownership will be dead in urban America by that time.

But it’s really a magnum opus on transportation and technology.

Helpfully, Zimmer divides the piece into sections.

1. Autonomous vehicle fleets will quickly become widespread and will account for the majority of Lyft rides within 5 years.

2. By 2025, private car ownership will all-but end in major U.S. cities.

3. As a result, cities’ physical environment will change more than we’ve ever experienced in our lifetimes.

Read the whole thing.

GM to Launch Self-Driving Cars with Lyft

Big news. GM is going to launch its self-driving cars “much faster than people anticipate”, and their going to be on the Lyft network. And by all appearances, the platform will be the Chevy Bolt EV.

GM’s Chief Engineer Pam Fletcher told Tech Insider:

We are working on an on-demand ride-sharing network with Lyft, it’s not something we are thinking about, it’s something we are very much readying for consumer use.

That’s about the extent of the news, though. Open questions include what cities this will operate in, what the launch schedule is, how much the cars will cost, what kind of sensors they use, if they will even have a steering wheel, and when they will be available for purchase outside the Lyft network.

I guess that’s a heads-up to Uber, though.

On-Demand Mobility Is Still Expensive

I needed to get home from work late last night and I didn’t have a car. Cry me a river.

So I went to the Palo Alto train station, only to discover that I had read the train schedule on my phone wrong. The next train wasn’t passing through for another 45 minutes.

So I looked up the cost of a ride-share home. The distance is a hair under 15 miles.

Lyft wouldn’t even quote me a price — apparently the ride was out of their service area.

Uber quoted me a price of $50.

The train is $5. I got two slices of pizza and waited for the train.

Maybe this isn’t a good use of time, running the cost benefit analysis. But it was just hard for me to pull the trigger on a $50 ride home.

GM and Lyft

GM and Lyft announced that they plan to test self-driving cars on real customers within the next year. That’s a pretty aggressive timeline and a pretty amazing goal.

The reports have been pretty light on details, so there’s not much comment on.

One thought is that this would be awesome.

Another is that I’d love to see some firmer details around this prediction. For example, how many cars will there be and how tightly geo-fenced will the routes be?

I was trying to think of how likely I believe this prediction is to come to pass, but it’s impossible to put a finger on that without firmer details.

A third thought is that this would show a remarkably successful and fast integration of the Cruise acquisition into GM.

My last thought is that asking people take a ride in an autonomous vehicle is probably a smaller lift than asking them to purchase one outright. So this seems like a smart way for Lyft to introduce their AVs.

Ride-Sharing Monopoly

In business school we did a case on eBay that was teed up perfectly for the professors to disabuse students of a cherished notion.

The central question was what makes eBay such a successful marketplace? And the intuitive answer is “network effects”. The more people are using eBay, the more the next user wants to join eBay, instead of a competitor.

And the professors loved this answer because it is so easy to pop a hole in it.

Buyers don’t want to be with other buyers — the competition just drives up prices. And sellers don’t want to be with other sellers — the competition just drives down prices.

Instead, the professors argued, it was eBay’s reputation metrics that really drove its success. Buyers and sellers want to transact with counterparties that have solid reputations, and only eBay can provide that.

This is a good and correct answer, but it’s always seemed incomplete to me. eBay has a couple of other strengths:

  1. eBay has a thick market. There is no other market where you can find the Danish Christmas ornament from exactly 28 years ago that you just shattered and need to replace.
  2. eBay has an incredible brand.

The point about branding is especially easy to ignore, in business school and elsewhere, because it is so close to magic. How do you build a good brand? How do you know if you have one? How do you know if one brand is stronger than another brand? How do you quantify when a brand is strong enough to support a monopoly?

Warren Buffett was a famously late in life convert to the power of branding, in all its mystery.

Which brings us to ride-sharing and Uber.

Uber, like eBay, has power in its reputation metrics, in the thickness of its market, and in its brand.

So how do these stack up to eBay?

  1. Uber’s reputation metrics are less important than eBay’s. Uber’s ride-sharing flow is simpler than eBay’s asynchronous process of buying, and paying, and shipping, and receiving, so counterparty reputation matters less.
  2. Uber’s market thickness is more important than eBay’s. If an Uber driver or rider can’t find a match at any given moment, they’re much more likely to give up on the service altogether, than if an eBay buyer or seller can’t find a match.
  3. Uber’s brand is as important as eBay’s. But Uber hasn’t locked up the market yet. Lyft is on their heels. The question is, unlike with online auctions, is there room for two players?

One last note is that while eBay does not have a direct competitor, Craigslist serves as an alternative, albeit a distinct one.

Lyft is much more of a direct competitor to Uber than Craigslist is to eBay, but the existence of Craigslist raises the possibility that two firms could survive in the market.

H/T to Jared Myer and Daniel Pryor’s post in Forbes, which inspired this.

Originally published at on February 16, 2016.

Vertical Integration

Fortune runs with a story about Lyft and Uber and the merits and demerits of vertical integration:

“From a purely technical perspective, it’s unlikely that Uber’s in-house mapping efforts will be able to compete with the massive scale of Waze’s crowdsourced information. Waze is something like the Wikipedia of mapping with, as of last year, almost 300,000 editors worldwide contributing regular updates — all for free.”

Of course this is to some extent a matter of options. Uber has raised enough money to at least attempt building its own features, whereas Lyft is more cash-constrained.

But, as Malcolm Gladwell will tell you, sometimes being smaller comes with surprising advantages.

Originally published at on February 1, 2016.