Margaret Mitchell has spent her career founding bootstrap-style AI projects inside large tech companies. She helped create Microsoft Research’s “Cognition” group, which concentrated on AI advancement, before moving to Google and founding its Ethical AI team and cofounding its ML Fairness group.

Now…She’s left Big Tech behind for full-time startup life—leading data governance efforts at Hugging Face, a 60-person AI company founded in 2016 and based in NYC.

It’s a big change for Mitchell, following even bigger ones. Over a three-month span beginning in December 2020, Google fired both Mitchell and her Ethical AI team co-lead, Timnit Gebru, after disagreements over their research paper on the dangers of large language models. (Google disputes this version of events.)

• Though the powerful language algorithms increasingly underpin popular and useful services like Google Search and AutoComplete, they can also replicate and multiply harmful human biases.

The most powerful language models are also concentrated in the hands of a few powerful companies. Typically, their training and development is limited to the Big Tech sphere, or to FAMGA-funded research groups, both in academia and outside of it (e.g, OpenAI).

Hugging Face wants to bring these powerful tools to more people. Its mission: Help companies build, train, and deploy AI models—specifically natural language processing (NLP) systems—via its open-source tools, like Transformers and Datasets.

So what does it mean to play a part in “democratizing” these powerful NLP tools? We chatted with Mitchell about the split from Google, her plans for her new role, and her near-future predictions for responsible AI.

Read on for some highlights, and click here to read the full thing.

On seeking out a new role...

I was increasingly interested in companies that had ethical values baked into their core, like part of the initial construction of the company, because I had felt that I had gotten a pretty good expertise at retrofitting ethical processes and retrofitting inclusion, essentially, in companies that hadn’t been built on that. It’s very challenging, obviously, and I think that what happened with Google showed a bit of how intense that can be in a way that people didn’t really realize before.

On her five-year plan...

Currently, no one can contest if they have some text that’s been picked up in a training data set and then used to train a model, people can’t say no to that. So figuring out how to allow contestation brings with it a whole bunch of research questions: How do you remove training data from a model? How do you even find the data that someone would contest?

So the current paradigm is that whoever finds it, owns it—not whoever creates it, owns it. So I think within five years, we’ll see that paradigm fundamentally shifting—or at least I hope so.

Click here to read the full interview with Margaret Mitchell.—HF

Your friend in SF may soon be price-checking a ride-hailing app other than Uber and Lyft. Last week, the California DMV granted both Waymo and Cruise permission to offer autonomous vehicle rides to the public, under new permits.

But...there’s a lot of fine print involved. And the companies have two different sets of marching orders.

Let’s sort through the Venn Diagram:

For both companies: Waymo and Cruise were already testing free autonomous rides for passengers, but now, they’ve got permits to officially go commercial...kinda. For the moment, they’re still stuck offering those rides for free.

• In order to begin charging for autonomous services, they’ve got to secure one more puzzle piece: a deployment permit from the state’s Public Utilities Commission.

Driverless deployment: Cruise is cleared to offer fully autonomous rides—read: no human safety driver—in five vehicles. They’ll operate at a max speed of 30 mph, in designated areas of San Francisco between 10am and 6pm.

Drivered deployment: Waymo can operate (semi-)autonomous rides—read: with a human safety driver—in certain parts of the SF and San Mateo areas. The vehicles can run at all hours, at speeds up to 65 mph.

All this means we’re closer to seeing commercial robotaxi services go live in California, although the scope will likely still be limited even then, as it is with Waymo’s fully driverless service in Arizona. Cruise, for its part, foresees its robotaxi business bringing in $50 billion in revenue and hopes to start charging in 2023, Bloomberg reported.

View this story on our website.—HF

And just like that, September—or, for Emerging Tech Brew, Cannabis Month—is over. Across last month, we published pieces focused on the nexus of cannabis and emerging tech like AI, biotech, and drones. In case you missed it, here they are:

• How a cannabis-focused AI platform helps growers predict yield
• One way to track cannabis shipments? Molecular barcodes
• Weed farmers are using drones as crop cops
• To grow medical cannabis, you’ll need some tech essentials first

Looking ahead...We’re focusing on the tech that powers tech in October. In one word: infrastructure. That means semiconductors, data centers, EV chargers and batteries, lidar, and more. We’ll publish at least one a week across the month.—DM

Stat: Buzzy Amazon-backed electric truck startup Rivian lost $994 million in the first half of this year, and expects to spend $8 billion total building out its infrastructure by 2023.

Quote: “It’s a total crapshoot.”—Peter Anthony, CEO of UGN Automotive, on how constantly changing chip availability is turning automakers’ supply orders upside down

Read: How refugees are powering machine-learning breakthroughs in Big Tech.

Let’s get down to business: Your small business is fueled by big ideas—your credit card should be no different. American Express® Business Cards offer serious spending power, rewards that help you do more, and support from a specialist who understands your business needs. Terms Apply. Learn more here.*

^{*This is sponsored advertising content}

Today: After a brief hiatus, Morning Brew’s interview-focused podcast, Business Casual, just kicked off with new hosts Nora Ali and Scott Rogowsky. Check out the first episode of the new season here.

Tuesday: Windows 11 drops. Here’s an early review of Microsoft’s new OS, from The Verge.

All week: We’re looking for more updates on the bipartisan infrastructure bill, which the House of Representatives was supposed to vote on last Thursday. House progressives are withholding votes for the $1.2 trillion infrastructure bill—which contains billions for projects like EV chargers and batteries—until the House and Senate pass a broader climate change and social policy plan.