Mona Gets A Makeover 🎨
4 min read

Mona Gets A Makeover 🎨

AI on the phone. Facebook's transparency report. Starlink launches.

How would you like a new face?

Researchers at Samsung’s AI Center in Moscow have released a paper detailing a method to create generated videos of someone using a single still image. They call it Few-Shot Adversarial Learning.

This boils down to:

We can now create a realistic video of you with just your profile picture.

A video of this technique applied to the Mona Lisa went viral this week. It’s easy to see why. The results are captivating, but the dual-use of this technology raises concerns.

A popular use of the technology is in the area of “deep-fakes”; the creation of videos where the subject has been edited using advanced ML techniques. This week a video of Nancy Pelosi exhibiting slurred speech was spread across popular social media sites and was further inflamed after being retweeted by Donald Trump. The video was clearly edited using traditional methods, but online communities called for Pelosi to resign from office due to her performance or possible health conditions (that were “obvious” from the edited photo). The video was quickly removed by Youtube, although Facebook and Twitter chose to leave the content available on their sites.

As these methods gain popularity it will become increasingly difficult to identify and flag content that is edited or misleading. Video has traditionally been the gold standard and many of our social systems are built on this assumption.

“I’ll believe it when I see it” won’t cut it anymore.

This is quickly becoming an area ripe for misinformation. Camera recordings in high-profile court cases will need to be examined by experts to find signs of tampering. Election campaigns will need to double their efforts to employ staff specialized in the reduction of fake news across various mediums.

As a society, we are still figuring out how to prevent the flood of misinformation in text form. Lots of models to look at here, and a looming 2020 election where it will play out on a grand scale.

The AI On The Phone Is Actually A Guy From Ireland

Image result for duplex google

Last year, Google announced Duplex: a service designed to take away the tedious activities associated with booking reservations, talking to customer service representatives, and making appointments.

The original video demonstrated an interaction between a Chinese restaurant and Duplex — complete with human-like “ums” and pauses for breath. It seemed as though we were on the cusp of a shift: eventually all reservations would be handled by a fleet of exceedingly polite robots talking to each other, blissfully unaware of the fact that each party had at some point ceased to be a human being.

It turns out, a large number of the Duplex calls are still being handled by humans. The New York Times ran a story where they interacted with the service and reported that 25% of the calls were actually humans performing glorified concierge service. The point of the article seemed to be a “gotcha!” but I think it’s important to have this kind of “unsexy” work being done at the beginning of the initiative.

Creating successful automated systems means first gaining a deep understanding of the different failure modes that can happen. Customer service calls are essentially a game tree problem, where one side has incomplete information.

Humans are really bad at this kind of thing. Computers are good at it, but haven’t been able to participate because the interactions happen over the phone in natural language. When I say “I want to speak with a manager” I am actually signaling that I want to move to a different branch of the tree, but emotions tend to cloud this.

If being a Product Manager at Duplex means you need to take Customer Service call duty for the first few months then I would expect to see a much more empathetic product for the poor humans on the other end of the line.

📚 Reading

A transparency report from Facebook on their moderation efforts over the last 6 months. There are 9 major categories of violations:

  • Adult Nudity and Sexual Activity
  • Bullying and Harassment
  • Child Nudity and Sexual Exploitation of Children
  • Fake Accounts
  • Hate Speech
  • Regulated Goods: Drugs and Firearms
  • Spam
  • Terrorist Propaganda (ISIS, al-Qaeda and affiliates)

Violence and Graphic Content


An open-sourced valuation model from Tesla bulls at Ark Invest. Link.

Peter Thiel shares his thoughts in an interview on how innovation has lagged in the world of atoms. Link.

🐦 Tweetstorm of the Week


Working in corporate blockchain, this thread hits close to home. I posted about working at Anthem here. Matt hits on some important points reminiscent of Chris Dixon’s Strong and Weak Technologies. I believe that there is value in permissioned blockchains within highly regulated industries. Blockchain native companies have faced significant regulatory setbacks and often find themselves rediscovering financial lessons from the past century.

However, the allure of “doing something in blockchain” is strong, and we often find ourselves fending off proof of concept ideas with little to no business value. Large corporations have the benefit of network effects to bolster the incentive models needed to successfully drive adoption of cryptonetworks. Come for the technology — stay for the value. I’ve said it before, but it’s worth repeating: the value of blockchain isn’t in the technology alone, but in the opportunity for traditionally competitive companies to collaborate on shared problems. It’s as much a social experiment as it is a technology shift.

📺 Videos

Apple Developers Conference ‘97

If you need a reminder on Steve Job’s clarity of vision take a moment to watch the Q&A from Apple’s Developer Conference in 1997. Steve sits down and gives a brief introduction on Apple’s strategy: “Make great products.”

“When you think about focusing — focusing is about saying yes, right?”

No. Focusing is about saying no.

On May 23rd, 60 LEO satellites were deployed during a successful launch of SpaceX’s Falcon 9. The plan is to launch 11,000 of the Starlink satellites to provide broadband internet access to billions who are not within range of traditional cellular signal. This launch was accompanied by a large media production typical of companies operated by Mr. Musk.

Similar companies in this space, including OneWeb and Amazon’s Project Kuiper, have been undergoing a race to capitalize on the LEO opportunity. OneWeb, who launched 6 satellites last February, recently raised amounts totaling >$1B to bolster efforts to reach critical coverage. The race is on.

Not a subscriber yet? Sign up here.

Follow me on Twitter and Medium.