Has DeepMind really solved protein folding?

The Big Story

More internal trouble at Google

Anna Kramer writes: Leading AI ethicist Timnit Gebru was fired from Google on Wednesday, and AI industry leaders are seriously angry. After she shared the news of her firing on Twitter, the outpouring of support from all corners of the internet was practically unanimous, filled with stories of her kindness, competence and mentorship. "If we have heroes in the AI ethics community, she's one of those heroes," the Altimeter Group's Susan Etlinger said.

• Gebru said she was fired because of an email she sent to coworkers about her exhaustion and frustration with Google's diversity pledges, and her recent fight to get publishing approval for a research paper. (See here and here for more details on the specifics of how and why she left).

AI industry leaders told me that her departure from Google could have "a chilling effect" on ethics research at tech companies.

• While Gebru has the fame and support to be successful despite losing her job, All Tech is Human's David Ryan Polgar told me, other employees without such high profiles may be increasingly afraid to speak up about ethical questions after witnessing the potential consequences.
• "As Black women in tech, we all face similar issues, and not everybody is going to take the stand to stay within [the] industry," especially after witnessing what happened to Gebru, said AI for the People's Mutale Nkonde.

This is becoming a pattern for Google, and potentially a hugely problematic one. Not just for its hiring practices, either. Industry-wide respect for Gebru gave Google's AI ethics team credibility; by firing her, the company opened the door wider for skepticism about the integrity of AI ethics research inside the company, said Ansgar Koene, the global AI ethics and regulatory leader at EY.

• "Their division does great work, except a lot of the times they have their hands tied behind their backs because of such repressive policies," Koene explained.
• "The idea that this is going to be able to happen," Nkonde said, "and it's going to go away and it's not going to have an impact on tech … Google really needs to really look at itself in a mirror."

Science

Unfolding the protein-folding problem

More from Anna: Earlier this week, DeepMind announced that its AlphaFold AI system had solved the protein-folding problem. Which meant … something. The headlines made it sound like somewhat of a miracle, so I talked to some researchers to unpack what AlphaFold did, exactly, and whether we should be skeptical about the excitement. The short answer: It's great news, but don't get too excited just yet. The problem isn't all the way solved.

So what's the protein-folding problem, you ask?

• Understanding the structure of proteins could help us more quickly and cheaply discover drugs for complicated diseases (the COVID-19 vaccines, for example, rely on our understanding of the proteins that make up the virus). But the structures are hard to find. Specifically, we don't know how or why they are structured (or "folded") in three dimensions, and mapping out that structure experimentally is very expensive and time-consuming.
• So the protein-folding problem asks: Can we computationally identify the way a protein is structured in three dimensions based on its one-dimensional chemical makeup, and how do we do it?

DeepMind's AlphaFold sort of solved this problem. But only sort of. As part of a yearly competition to measure the progress in resolving the question, AlphaFold, using machine learning, successfully identified the three-dimensional structures of some proteins based only on their one-dimensional makeup.

But there are some important caveats, according to Lior Pachter, a computational biologist at Caltech.

• First, while AlphaFold created many structural predictions — and the ones for the simplest proteins in particular were very accurate — the researchers sometimes had difficulty identifying which prediction was most accurate. That means that experimental modeling is still necessary to "check the work" of the AI.
• Second, the proteins used in this competition are relatively simple. In real-world applications, proteins can be more complicated and can change shape as they interact with other proteins. Pachter suspects the AlphaFold model would have a more difficult time with these more complex predictions.
• Finally, because DeepMind does not plan to release the code it used to create and modify AlphaFold, it will be difficult for researchers to understand how the result came about and apply it in their own work.

While there may be practical applications from this week's announcement, the news won't have us creating cancer-curing drugs tomorrow, Pachter explained. "It certainly gives hope, and I do think it's useful, possibly, for some things," he said. "But it's not like we're done with that problem now. It's not time to close the book on that problem."

Money

Stripe gets into banking

The hottest thing in tech right now is … checking accounts. Weird. And Stripe, the biggest American fintech company of them all, is getting in the game. Yesterday, the company announced Stripe Treasury, a new system for allowing companies to offer banking services.

• Stripe's not trying to replace consumer bank accounts, the way Varo and Chime and Simple and Square and Google and Amazon and 40,000 others are. It's building these tools for the same sorts of businesses that use its payment processing.
• Unlike some other banking-as-a-service providers, Stripe's not becoming a bank itself. Instead, it's working with Citigroup and tech's favorite bank Goldman Sachs to support the system.
• Shopify is one of its first customers, meaning an enormous number of small businesses will now be able to operate almost entirely through Shopify and Stripe.

This looks a lot like the next phase of ecommerce, and maybe the internet as a whole: Rather than just setting up a Facebook page or trying to game Amazon's algorithms, companies can set up something that feels much more like their own shop without having to do much more work.

• Meanwhile, Stripe's ambitions continue to grow like crazy. "From our earliest days, we've wanted to make all of financial services truly programmable," Patrick Collison tweeted yesterday. "We're now working with top banks to make that happen."
• It's not trying to be a bank; it's trying to rebuild the whole financial system, online. (The new thing is called Stripe Treasury for a reason.)

It's a big change for the company: Its initial appeal was all about the "it's just seven lines of code" simplicity. Big banks — heck, big companies — get messy and complicated. But John Collison, in an interview with Stratechery, indicated there's still plenty more simplicity to work on:

• "[O]ne in four businesses still have to send faxes to open their business bank account. And so people talk about, 'Oh, well, what hifalutin big data AI things will Stripe Treasury do?' Well we have at least gotten rid of the faxes, and so we're continuing to innovate from there."

A MESSAGE FROM MICRON

At Micron, we see an opportunity to establish memory and storage platform capabilities that will unleash software developers to deliver solutions that speed insight and ultimately support emerging customer requirements. The data-centric era has ushered in a new opportunity to tap data for business growth, but many companies continue to struggle to transform mounting data stores into competitive advantage.

Learn how here.

People Are Talking

On Protocol: Toronto's Sidewalk partnership fell apart in part because it just moved too fast, Toronto Mayor John Tory said:

• "You've got to have the public's confidence in terms of what's going to be done with the data: How intrusive is this technology going to be in this data collection in their lives? And in this case, we were trying to fix that after the debate had broken out, as opposed to having that all in place."

Bad stuff happening on Parler? We wouldn't know, COO Jeffrey Wernick said:

• "If somebody does something illegal, we're relying on the reporting system. We're not hunting."

Here's how Facebook's Nick Clegg said he talks internally about politics:

• "I find myself having to explain to impeccably logical engineers that the ebb and tide of public sentiment and political concern are not always rational — but politicians still need to be respected and taken seriously."

Buying into remote work isn't just about cities and taxes and offices, Okta's Todd McKinnon said:

• "It has everything to do with the fact that we are in a war for talent and the more broadly we can appeal to people, in terms of letting them work from anywhere and ... letting them contribute at a high level from anywhere, that's our plan."

When it comes to delivery, we're already living in 2023, Deliveroo's Will Shu said:

• "Our initial analysis suggests that COVID-19 has accelerated consumer adoption of these delivery services by about two to three years."

Number of the Day

One

That's how many dollars NASA is paying Lunar Outpost, a robotics company in Colorado, for some rocks and soil from the moon. Seems cheap! But Lunar's goal is really just to get to the moon as part of NASA's Artemis program, which is designed to set up a permanent, sustainable presence on the big cheesy rock. "We must learn to generate our own water, air and even fuel," NASA's Mike Gold told the Washington Post. And that happens one $1 moon rock at a time.

In Other News

• On Protocol: The DOJ filed a lawsuit against Facebook, accusing it of failing to adequately recruit and hire U.S. workers for 2,600 jobs. Instead, the DOJ says, Facebook held those jobs open for foreign workers it was sponsoring for green cards.
• Facebook reorganized its 300-person civic integrity team. Rather than operate as a standalone group, those people will now join a larger "central integrity" team, led by Guy Rosen. Head of civic integrity Samidh Chakrabarti will find a new role, according to The Information.
• The DOD designated SMIC as a "Communist Chinese military company," adding it to its list of blacklisted companies. SMIC said the designation would not materially impact its business. Meanwhile, the DOJ is reportedly considering letting Huawei CFO Meng Wanzhou return to China.
• China declared quantum supremacy. The team at University of Science and Technology of China said its Jiuzhang system can outperform traditional supercomputers, using different tech to Google's quantum computer.
• YouTube engineer Betai Koffi, who allegedly took four tabs of LSD, went on a rampage in Bodega Bay, stabbed a man with a pencil, stole a security truck and drove into three people, pleaded no contest to charges of assault.
• Square makes clothes now. The new apparel line is called "Cash by Cash App," and I hope Jack wears it to his next congressional hearing.

One More Thing

Santa 2.0

Janko Roettgers writes: The Zoomification of everything continues. Cherry Hill Programs, a company that has been organizing holiday-themed photo opps in hundreds of shopping centers since 1961, recently launched a virtual Santa experience on Zoom. Kids don't sit on Santa's lap — which, honestly, was always a bit creepy — but instead can chat with Santa, take a photo with him and even receive a box with holiday decorations and other treats. "Create Holiday Magic" packages cost anywhere from $10 to $100, which admittedly is a lot of money. Then again, there's really no better way to end 2020 than with kids screaming at the top of their lungs: "SANTA, YOU'RE MUTED!"

A MESSAGE FROM MICRON

At Micron, we see an opportunity to establish memory and storage platform capabilities that will unleash software developers to deliver solutions that speed insight and ultimately support emerging customer requirements. The data-centric era has ushered in a new opportunity to tap data for business growth, but many companies continue to struggle to transform mounting data stores into competitive advantage.

Learn how here.