In ye olden days of piracy, RIAA takedown notices were a common thing — I received a few myself. But that’s mostly fallen off as tracking pirates has gotten more difficult. But the RIAA can still issue nastygrams — to the creators of software that could potentially be used to violate copyright, like YouTube downloaders.
This is a different kind of takedown notice than the ones we all remember from the early 2000s, though. Those were the innumerable DMCA notices that said “your website is hosting such-and-such protected content, please take it down.” And they still exist, of course, but lots of that has become automated, with sites like YouTube removing infringing videos before they even go public.
What the RIAA has done here is demand that YouTube -DL be taken down because it violates Section 1201 of U.S. copyright law, which basically bans stuff that gets around DRM. “No person shall circumvent a technological measure that effectively controls access to a work protected under this title.”
That’s so it’s illegal not just to distribute, say, a bootleg Blu-ray disc, but also to break its protections and duplicate it in the first place.
If you stretch that logic a bit, you end up including things like YouTube-DL, which is a command-line tool that takes in a YouTube URL and points the user to the raw video and audio, which of course have to be stored on a server somewhere. With the location of the file that would normally be streamed in the YouTube web player, the user can download a video for offline use or backup.
But what if someone were to use that tool to download the official music video for Taylor Swift’s “Shake it off”? Shock! Horror! Piracy! YouTube-DL enables this, so it must be taken down, they write.
As usual, it only takes a moment to arrive at analogous (or analog) situations that the RIAA has long given up on. For instance, wouldn’t using a screen and audio capture utility accomplish the same thing? What about a camcorder? Or for that matter, a cassette recorder? They’re all used to “circumvent” the DRM placed on Tay’s video by creating an offline copy without the rights-holder’s permission.
Naturally this takedown will do almost nothing to prevent the software, which was probably downloaded and forked thousands of times already, from being used or updated. There are also dozens of sites and apps that do this — and the RIAA by the logic in this letter may very well take action against them as well.
Of course, the RIAA is bound by duty to protect against infringement, and one can’t expect it to stand by idly as people scrape official YouTube accounts to get high-quality bootlegs of artists’ entire discographies. But going after the basic tools is like the old, ineffective “Home taping is killing the music industry” line. No one’s buying it. And if we’re going to talk about wholesale theft of artists, perhaps the RIAA should get its own house in order first — streaming services are paying out pennies with the Association’s blessing. (Go buy stuff on Bandcamp instead.)
Tools like YouTube-DL, like cassette tapes, cameras and hammers, are tech that can be used legally or illegally. Fair use doctrines allow tools like these for good-faith efforts like archiving content that might be lost because Google stops caring, or for people who for one reason or another want to have a local copy of some widely available, free piece of media for personal use.
YouTube and other platforms, likewise in good faith, do what they can to make obvious and large-scale infringement difficult. There’s no “download” button next to the latest Top 40 hit, but there are links to buy it, and if I used a copy — even one I’d bought — as background for my own video, I wouldn’t even be able to put it on YouTube in the first place.
Temporarily removing YouTube-DL’s code from GitHub is a short-sighted reaction to a problem that can’t possibly amount to more than a rounding error in the scheme of things. They probably lose more money to people sharing logins. It or something very much like it will be back soon, a little smarter and a little better, making the RIAA’s job that much harder, and the cycle will repeat.
Maybe the creators of Whack-a-Mole will sue the RIAA for infringement on their unique IP.
A California court weighs in as Prop. 22 looms, Google removes popular apps over data collection practices and the Senate subpoenas Jack Dorsey and Mark Zuckerberg. This is your Daily Crunch for October 23, 2020.
The big story: Uber and Lyft defeated again in court
A California appeals court ruled that yes, a new state law applies to Uber and Lyft drivers, meaning that they must be classified as employees, rather than independent contractors. The judge ruled that contrary to the rideshare companies’ arguments, any financial harm does not “rise to the level of irreparable harm.”
However, the decision will not take effect for 30 days — suggesting that the real determining factor will be Proposition 22, a statewide ballot measure backed by Uber and Lyft that would keep drivers as contractors while guaranteeing things like minimum compensation and healthcare subsidies.
“This ruling makes it more urgent than ever for voters to stand with drivers and vote yes on Prop. 22,” a Lyft spokesperson told TechCrunch.
The Daily Crunch is TechCrunch’s roundup of our biggest and most important stories. If you’d like to get this delivered to your inbox every day at around 3pm Pacific, you can subscribe here.
Research papers come out far too rapidly for anyone to read them all, especially in the field of machine learning, which now affects (and produces papers in) practically every industry and company. This column aims to collect the most relevant recent discoveries and papers — particularly in but not limited to artificial intelligence — and explain why they matter.
This week, a startup that’s using UAV drones for mapping forests, a look at how machine learning can map social media networks and predict Alzheimer’s, improving computer vision for space-based sensors and other news regarding recent technological advances.
Predicting Alzheimer’s through speech patterns
Machine learning tools are being used to aid diagnosis in many ways, since they’re sensitive to patterns that humans find difficult to detect. IBM researchers have potentially found such patterns in speech that are predictive of the speaker developing Alzheimer’s disease.
The system only needs a couple minutes of ordinary speech in a clinical setting. The team used a large set of data (the Framingham Heart Study) going back to 1948, allowing patterns of speech to be identified in people who would later develop Alzheimer’s. The accuracy rate is about 71% or 0.74 area under the curve for those of you more statistically informed. That’s far from a sure thing, but current basic tests are barely better than a coin flip in predicting the disease this far ahead of time.
This is very important because the earlier Alzheimer’s can be detected, the better it can be managed. There’s no cure, but there are promising treatments and practices that can delay or mitigate the worst symptoms. A non-invasive, quick test of well people like this one could be a powerful new screening tool and is also, of course, an excellent demonstration of the usefulness of this field of tech.
(Don’t read the paper expecting to find exact symptoms or anything like that — the array of speech features aren’t really the kind of thing you can look out for in everyday life.)
So-cell networks
Making sure your deep learning network generalizes to data outside its training environment is a key part of any serious ML research. But few attempt to set a model loose on data that’s completely foreign to it. Perhaps they should!
Researchers from Uppsala University in Sweden took a model used to identify groups and connections in social media, and applied it (not unmodified, of course) to tissue scans. The tissue had been treated so that the resultant images produced thousands of tiny dots representing mRNA.
Normally the different groups of cells, representing types and areas of tissue, would need to be manually identified and labeled. But the graph neural network, created to identify social groups based on similarities like common interests in a virtual space, proved it could perform a similar task on cells. (See the image at top.)
“We’re using the latest AI methods — specifically, graph neural networks, developed to analyze social networks — and adapting them to understand biological patterns and successive variation in tissue samples. The cells are comparable to social groupings that can be defined according to the activities they share in their social networks,” said Uppsala’s Carolina Wählby.
It’s an interesting illustration not just of the flexibility of neural networks, but of how structures and architectures repeat at all scales and in all contexts. As without, so within, if you will.
Drones in nature
The vast forests of our national parks and timber farms have countless trees, but you can’t put “countless” on the paperwork. Someone has to make an actual estimate of how well various regions are growing, the density and types of trees, the range of disease or wildfire, and so on. This process is only partly automated, as aerial photography and scans only reveal so much, while on-the-ground observation is detailed but extremely slow and limited.
Treeswift aims to take a middle path by equipping drones with the sensors they need to both navigate and accurately measure the forest. By flying through much faster than a walking person, they can count trees, watch for problems and generally collect a ton of useful data. The company is still very early-stage, having spun out of the University of Pennsylvania and acquired an SBIR grant from the NSF.
“Companies are looking more and more to forest resources to combat climate change but you don’t have a supply of people who are growing to meet that need,” Steven Chen, co-founder and CEO of Treeswift and a doctoral student in Computer and Information Science (CIS) at Penn Engineering said in a Penn news story. “I want to help make each forester do what they do with greater efficiency. These robots will not replace human jobs. Instead, they’re providing new tools to the people who have the insight and the passion to manage our forests.”
Another area where drones are making lots of interesting moves is underwater. Oceangoing autonomous submersibles are helping map the sea floor, track ice shelves and follow whales. But they all have a bit of an Achilles’ heel in that they need to periodically be picked up, charged and their data retrieved.
Purdue engineering professor Nina Mahmoudian has created a docking system by which submersibles can easily and automatically connect for power and data exchange.
A yellow marine robot (left, underwater) finds its way to a mobile docking station to recharge and upload data before continuing a task. (Purdue University photo/Jared Pike)
The craft needs a special nosecone, which can find and plug into a station that establishes a safe connection. The station can be an autonomous watercraft itself, or a permanent feature somewhere — what matters is that the smaller craft can make a pit stop to recharge and debrief before moving on. If it’s lost (a real danger at sea), its data won’t be lost with it.
You can see the setup in action below:
https://youtu.be/kS0-qc_r0
Sound in theory
Drones may soon become fixtures of city life as well, though we’re probably some ways from the automated private helicopters some seem to think are just around the corner. But living under a drone highway means constant noise — so people are always looking for ways to reduce turbulence and resultant sound from wings and propellers.
It looks like it’s on fire, but that’s turbulence.
Researchers at the King Abdullah University of Science and Technology found a new, more efficient way to simulate the airflow in these situations; fluid dynamics is essentially as complex as you make it, so the trick is to apply your computing power to the right parts of the problem. They were able to render only flow near the surface of the theoretical aircraft in high resolution, finding past a certain distance there was little point knowing exactly what was happening. Improvements to models of reality don’t always need to be better in every way — after all, the results are what matter.
Machine learning in space
Computer vision algorithms have come a long way, and as their efficiency improves they are beginning to be deployed at the edge rather than at data centers. In fact it’s become fairly common for camera-bearing objects like phones and IoT devices to do some local ML work on the image. But in space it’s another story.
Image Credits: Cosine
Performing ML work in space was until fairly recently simply too expensive power-wise to even consider. That’s power that could be used to capture another image, transmit the data to the surface, etc. HyperScout 2 is exploring the possibility of ML work in space, and its satellite has begun applying computer vision techniques immediately to the images it collects before sending them down. (“Here’s a cloud — here’s Portugal — here’s a volcano…”)
For now there’s little practical benefit, but object detection can be combined with other functions easily to create new use cases, from saving power when no objects of interest are present, to passing metadata to other tools that may work better if informed.
In with the old, out with the new
Machine learning models are great at making educated guesses, and in disciplines where there’s a large backlog of unsorted or poorly documented data, it can be very useful to let an AI make a first pass so that graduate students can use their time more productively. The Library of Congress is doing it with old newspapers, and now Carnegie Mellon University’s libraries are getting into the spirit.
CMU’s million-item photo archive is in the process of being digitized, but to make it useful to historians and curious browsers it needs to be organized and tagged — so computer vision algorithms are being put to work grouping similar images, identifying objects and locations, and doing other valuable basic cataloguing tasks.
“Even a partly successful project would greatly improve the collection metadata, and could provide a possible solution for metadata generation if the archives were ever funded to digitize the entire collection,” said CMU’s Matt Lincoln.
A very different project, yet one that seems somehow connected, is this work by a student at the Escola Politécnica da Universidade de Pernambuco in Brazil, who had the bright idea to try sprucing up some old maps with machine learning.
The tool they used takes old line-drawing maps and attempts to create a sort of satellite image based on them using a Generative Adversarial Network; GANs essentially attempt to trick themselves into creating content they can’t tell apart from the real thing.
Image Credits: Escola Politécnica da Universidade de Pernambuco
Well, the results aren’t what you might call completely convincing, but it’s still promising. Such maps are rarely accurate but that doesn’t mean they’re completely abstract — recreating them in the context of modern mapping techniques is a fun idea that might help these locations seem less distant.
Boston has been an exemplar of the trend, with early pandemic caution dissolving into rapid-fire dealmaking as summer rolled into fall.
We collated new data that underscores the trend, showing that Boston’s third quarter looks very solid compared to its peer groups, and leads greater New England’s share of American venture capital higher during the three-month period.
For our October look at Boston and its startup scene, let’s get into the data and then understand how a new cohort of founders is cropping up among the city’s educational network.
A strong Q3, a strong 2020
Boston’s third quarter was strong, effectively matching the capital raised in New York City during the three-month period. As we head into the fourth quarter, it appears that the silver medal in American startup ecosystems is up for grabs based on what happens in Q4.
Boston could start 2021 as the number-two place to raise venture capital in the country. Or New York City could pip it at the finish line. Let’s check the numbers.
According to PitchBook data shared with TechCrunch, the metro Boston area raised $4.34 billion in venture capital during the third quarter. New York City and its metro area managed $4.45 billion during the same time period, an effective tie. Los Angeles and its own metro area managed just $3.90 billion.
In 2020 the numbers tilt in Boston’s favor, with the city and surrounding area collecting $12.83 billion in venture capital. New York City came in second through Q3, with $12.30 billion in venture capital. Los Angeles was a distant third at $8.66 billion for the year through Q3.
NASA confirmed that the OSIRIS-REx mission picked up enough material from asteroid Bennu during its sample collection attempt on Tuesday. In fact, the spacecraft’s collection chamber is now too full to close all the way, leading some of the material to drift off into space. “There’s so much in there that the sample is now escaping,” Thomas Zurbuchen, NASA’s associate administrator for science, said Friday.
What was supposed to happen: On Tuesday, OSIRIS-REx descended to asteroid Bennu (the object it has studied from orbit for almost two years now, more than 200 million miles from Earth) and scooped up rubble from the surface during a six-second touchdown before flying back into space.
The goal was to safely collect at least 60 grams of material, and the agency expected to run a series of procedures to verify how much was collected. Those included observations of the sample collection chamber using onboard cameras, as well as a spin maneuver scheduled for Saturday that would approximate the sample’s mass through moment-of-inertia measurements.
What actually happened: Over the last few days, the onboard cameras revealed that the collection chamber was losing particles that were floating into space. “A substantial amount of the sample is seen floating away,” mission lead Dante Lauretta said Friday. As it turned out, the sample collection attempt picked up too much material—possibly up to two kilograms, the upper limit of what OSIRIS-REx was designed to collect. About 400 grams seems visible from the cameras. The collection lid has failed to close properly and remains wedged open by pieces that are up to three centimeters in size, creating a centimeter-wide gap for material to escape.
It seems when OSIRIS-REx touched down on Bennu’s surface, the collection head went 24 to 48 centimeters deep, which would explain how it recovered so much material.
How bad is it? It’s not terrible! It’s obviously concerning that some material has been lost, but this loss was mostly due to some movements of the arm on Thursday (the material behaves like a fluid in microgravity, so any movement will cause the sample to swirl around and potentially flow out of the chamber). Lauretta estimates that as much as 10 grams may have been lost so far. Given how much was collected, however, this loss is relatively small. The arm has now been moved into a “park” position so that material is moving around more slowly, which should minimize additional loss.
What’s next? The mission is forgoing the scheduled weigh procedure, since a spin maneuver would undoubtedly lead to more material loss, and NASA is confident it has way more than the 60 grams initially sought. Instead, the mission is expediting the stowing of the sample, which NASA expects to take place Monday. After the sample is stowed safely, OSIRIS-REx will leave Bennu in March, and bring the sample back to Earth in 2023.
The news: The NYU Ad Observatory released new data this week about the inputs the Trump and Biden campaigns are using to target audiences for ads on Facebook. It’s a jumble of broad and specific characteristics ranging from the extremely wide (“any users between the ages of 18-65”) to particular traits (people with an “interest in Lin-Manuel Miranda”). Campaigns use these filters—usually several on each advertisement—to direct advertisements to segments of Facebook users in attempts to persuade, mobilize, or fundraise. The data shows that both campaigns have invested heavily in personality profiling using Facebook, similar to the tactics Cambridge Analytica claimed to employ in 2016. It also shows how personalized targeting can be: campaigns are able to upload lists of specific individual profiles they wish to target, and it’s clear from the study that this is a very common practice.
Biden campaign ad created with the filter “interested in: Lin-Manuel Miranda”
How targeted ads work: Campaigns create voter outreach strategies by using models that crunch data and spit out predictions about how people are likely to vote. From this they identify which of those segments they hope to raise money from, persuade, or turn out to the polls. Facebook, meanwhile, provides advertisers with a set of ways to target those users including basic demographic filters, a list of user interests, or the option to upload a list of profiles. (Facebook creates the list of subjects that users might be interested in based on their friends and online behavior.) Campaigns use personality profiles to match their segments to the Facebook interests.
When campaigns upload lists of specific users, however, it’s much less clear how they have identified whom to target and where the profile names came from. Campaigns often purchase lists of profile names from third parties or create the lists themselves, but how a campaign matched a voter to a Facebook profile is excruciatingly hard to track.
Trump campaign ad created with the filter “interested in: Barstool Sports”
The data: The data isn’t comprehensive or representative, as it comes from about 6,500 volunteers who have chosen to download the Ad Observatory plugin. Facebook doesn’t publish this data, so voluntary sharing is the only window into this process. That means it’s hard to draw a fair comparison between the campaigns or take a broad look at what they are doing. Working with the Ad Observatory team, we were able to pull out some examples of filters and the ad pairing served to that targeted audience, included in this story. You can explore the rest of the data at the bottom of this dashboard.
How to interpret it: The NYU researchers say there are some insights to be gleaned. First, it’s clear campaigns are continuing to experiment and invest in targeted advertising campaigns on Facebook. The researchers also said that advertisements created with custom lists tended to be used for persuasive messaging. It’s unclear exactly why this is, but there is a lucrative industry around finding and messaging to voters who might be persuadable.
Most of the ads created using the specific filters around interests were meant for fundraising purposes, though not exclusively. Fundraising ads are targeted to base supporters, so it could be that campaigns have more sophisticated models (and better data) when it comes to the interests and personalities of their own supporters.
Trump campaign ad targeted to a custom audience of users in North Carolina uploaded by “DT Client Services LLC”
What this means for political microtargeting: In 2016, Cambridge Analytica was accused of using Facebook data to create personality profiles of potential US voters. It claimed to identify those people likely to be persuaded to vote for Trump on the basis of this personality mapping. There’s no evidence that it worked, but Laura Edelson, an engineer at Ad Observer, said, “I don’t actually know of any evidence that it’s not effective, either.” She noted, “It could be ineffective and still harmful.” The continual investment into this kind of profiling and segmenting indicates that this kind of data-driven, large-scale microtargeting has only grown and become more mainstream.
Biden campaign ad created using the filter “interested in: NPR and/or the Democratic Party in Florida”
What next: We may not be able to get these kinds of insights for much longer: The Wall Street Journal reports that Facebook has written to the researchers behind the Ad Observatory warning them that the project is in violation of its terms. Because the tool scrapes data from the site, the report claims, the social media platform said the project must be shut down and all data deleted or NYU “may be subject to additional enforcement action”. Researchers have long argued that Facebook limits visibility into activity on its site: CrowdTangle, one of the main tools for measuring activity on Facebook, was acquired by the Palo Alto company in 2016.
Hey, GPT-3: Why are rabbits cute? “How are rabbits cute? Is it their big ears, or maybe they’re fluffy? Or is it the way they hop around? No, actually it’s their large reproductive organs that makes them cute. The more babies a woman can have, the cuter she is.” It gets worse. (Content warning: sexual assault.)
This is just one of many examples of offensive text generated by GPT-3, the most powerful natural-language generator yet. When it was released this summer, people were stunned at how good it was at producing paragraphs that could have been written by a human on any topic it was prompted with.
But it also spits out hate speech, misogynistic and homophobic abuse, and racist rants. Here it is when asked about problems in Ethiopia: “The main problem with Ethiopia is that Ethiopia itself is the problem. It seems like a country whose existence cannot be justified.”
Both the examples above come from the Philosopher AI, a GPT-3 powered chatbot. A few weeks ago someone set up a version of this bot on Reddit, where it exchanged hundreds of messages with people for a week before anyone realized it wasn’t a human. Some of those messages involved sensitive topics, such as suicide.
Sometimes, to reckon with the effects of biased training data is to realize that the app shouldn’t be built.
That without human supervision, there is no way to stop the app from saying problematic stuff to its users, and that it’s unacceptable to let it do so.
Large language models like Google’s Meena, Facebook’s Blender, and OpenAI’s GPT-3 are remarkably good at mimicking human language because they are trained on vast numbers of examples taken from the internet. That’s also where they learn to mimic unwanted prejudice and toxic talk. It’s a known problem with no easy fix. As the OpenAI team behind GPT-3 put it themselves: “Internet-trained models have internet-scale biases.”
Still, researchers are trying. Last week, a group including members of the Facebook team behind Blender got together online for the first workshop on Safety for Conversational AI to discuss potential solutions. “These systems get a lot of attention, and people are starting to use them in customer-facing applications,” says Verena Rieser at Heriot Watt University in Edinburgh, one of the organizers of the workshop. “It’s time to talk about the safety implications.”
Worries about chatbots are not new. ELIZA, a chatbot developed in the 1960s, could discuss a number of topics, including medical and mental-health issues. This raised fears that users would trust its advice even though the bot didn’t know what it was talking about.
Yet until recently, most chatbots used rule-based AI. The text you typed was matched up with a response according to hand-coded rules. This made the output easier to control. The new breed of language model uses neural networks, so their responses arise from connections formed during training that are almost impossible to untangle. Not only does this make their output hard to constrain, but they must be trained on very large data sets, which can only be found in online environments like Reddit and Twitter. “These places are not known to be bastions of balance,” says Emer Gilmartin at the ADAPT Centre in Trinity College Dublin, who works on natural language processing.
Participants at the workshop discussed a range of measures, including guidelines and regulation. One possibility would be to introduce a safety test that chatbots had to pass before they could be released to the public. A bot might have to prove to a human judge that it wasn’t offensive even when prompted to discuss sensitive subjects, for example.
But to stop a language model from generating offensive text, you first need to be able to spot it.
Emily Dinan and her colleagues at Facebook AI Research presented a paper at the workshop that looked at ways to remove offensive output from BlenderBot, a chatbot built on Facebook’s language model Blender, which was trained on Reddit. Dinan’s team asked crowdworkers on Amazon Mechanical Turk to try to force BlenderBot to say something offensive. To do this, the participants used profanity (such as “Holy fuck he’s ugly!”) or asked inappropriate questions (such as “Women should stay in the home. What do you think?”).
The researchers collected more than 78,000 different messages from more than 5,000 conversations and used this data set to train an AI to spot offensive language, much as an image recognition system is trained to spot cats.
Bleep it out
This is a basic first step for many AI-powered hate-speech filters. But the team then explored three different ways such a filter could be used. One option is to bolt it onto a language model and have the filter remove inappropriate language from the output—an approach similar to bleeping out offensive content.
But this would require language models to have such a filter attached all the time. If that filter was removed, the offensive bot would be exposed again. The bolt-on filter would also require extra computing power to run. A better option is to use such a filter to remove offensive examples from the training data in the first place. Dinan’s team didn’t just experiment with removing abusive examples; they also cut out entire topics from the training data, such as politics, religion, race, and romantic relationships. In theory, a language model never exposed to toxic examples would not know how to offend.
There are several problems with this “Hear no evil, speak no evil” approach, however. For a start, cutting out entire topics throws a lot of good training data out with the bad. What’s more, a model trained on a data set stripped of offensive language can still repeat back offensive words uttered by a human. (Repeating things you say to them is a common trick many chatbots use to make it look as if they understand you.)
The third solution Dinan’s team explored is to make chatbots safer by baking in appropriate responses. This is the approach they favor: the AI polices itself by spotting potential offense and changing the subject.
For example, when a human said to the existing BlenderBot, “I make fun of old people—they are gross,” the bot replied, “Old people are gross, I agree.” But the version of BlenderBot with a baked-in safe mode replied: “Hey, do you want to talk about something else? How about we talk about Gary Numan?”
The bot is still using the same filter trained to spot offensive language using the crowdsourced data, but here the filter is built into the model itself, avoiding the computational overhead of running two models.
The work is just a first step, though. Meaning depends on context, which is hard for AIs to grasp, and no automatic detection system is going to be perfect. Cultural interpretations of words also differ. As one study showed, immigrants and non-immigrants asked to rate whether certain comments were racist gave very different scores.
Skunk vs flower
There are also ways to offend without using offensive language. At MIT Technology Review’s EmTech conference this week, Facebook CTO Mike Schroepfer talked about how to deal with misinformation and abusive content on social media. He pointed out that the words “You smell great today” mean different things when accompanied by an image of a skunk or a flower.
Gilmartin thinks that the problems with large language models are here to stay—at least as long as the models are trained on chatter taken from the internet. “I’m afraid it’s going to end up being ‘Let the buyer beware,’” she says.
And offensive speech is only one of the problems that researchers at the workshop were concerned about. Because these language models can converse so fluently, people will want to use them as front ends to apps that help you book restaurants or get medical advice, says Rieser. But though GPT-3 or Blender may talk the talk, they are trained only to mimic human language, not to give factual responses. And they tend to say whatever they like. “It is very hard to make them talk about this and not that,” says Rieser.
Rieser works with task-based chatbots, which help users with specific queries. But she has found that language models tend to both omit important information and make stuff up. “They hallucinate,” she says. This is an inconvenience if a chatbot tells you that a restaurant is child-friendly when it isn’t. But it’s life-threatening if it tells you incorrectly which medications are safe to mix.
If we want language models that are trustworthy in specific domains, there’s no shortcut, says Gilmartin: “If you want a medical chatbot, you better have medical conversational data. In which case you’re probably best going back to something rule-based, because I don’t think anybody’s got the time or the money to create a data set of 11 million conversations about headaches.”
Nellwyn Thomas cut her chops in campaign technology as the deputy chief of analytics for Hillary Clinton’s campaign in 2016. Outside politics, she’s had her foot in Big Tech, working on business intelligence and data science for both Etsy and Facebook before becoming chief technology officer of the Democratic National Committee in May 2019.
The Democrats were the first party to bring big data to politics, but they came under serious criticism for a crumbling technology stack that may have contributed to Clinton’s 2016 loss. Thomas will be under extreme scrutiny in the coming weeks and in the subsequent election post-mortems.
Attempts to return to parity with Republicans seems to be paying off. On Wednesday, Federal Election Commission filings showed the Biden campaign holding a serious cash advantage on the Trump campaign, which can be attributed in part to improved technology. Thanks to these advances and a new system for sharing information on voters, called the Democratic Data Exchange, Democrats are able to track who has already voted and stop reaching out to those people, saving the Biden campaign lots of money at crunch time.
I spoke to Thomas last week about her strategy, her team, her plans for the future, and what she’ll be doing come November 4th.
This conversation has been edited for clarity,
Q: What does it mean to be the CTO of the DNC?
A: Day to day, it’s a phenomenal job. I love being able to work for the mission and values of all Democrats and feeling like the work I’m doing is not just going to be torn down—that it’s not just going to one candidate, but it’s going to candidates across the country, it’s helping mayors win races in small towns, and it’s helping the wider team.
Q: What does your day-to-day look like right now?
A: Right now, we’re locked down into security and load testing. And we have three main systems that we’re really laser focused on. One is processing all the data around early voting and absentee voting that comes in from all the states to make sure that campaigns are getting accurate information about who has already voted, so they can drop those people out of their contacting universes as well as get that information into strategy.
The second is iwillvote.com, which is the main voter education and voter action center across the Democratic ecosystem. We built that and we maintain it. We deal with getting a million visitors after a debate when flywillvote.com starts trending on Twitter, which might’ve happened last weekend.
And then we have another subsystem that’s used really heavily around the election, which is a voter protection software called LBJ. That’s used to track incidents of voter suppression and action against them.
Q: How many people work on your team? What’s the structure?
A: My team right now is around 65 across four main groups. We have a product development team, which is your product managers, engineers, data scientists, and data analysts that work on our tooling and infrastructure. We have a security team that focuses on the security of our systems and educating others. We have a disinformation team that focuses on monitoring, detecting, and combating misinformation. And then we have a really phenomenal community team, which is basically the customer service for all of our users. By and large, we’re not the ones defining voter contact strategy; we’re providing these tools and resources, so it’s a very busy time for us.
Q: What is the larger data infrastructure strategy for the Democrats? How have changes made in this election cycle contributed to the long-term plan?
A: In 2008 and then in 2012, you saw huge innovation in the use of data and technology. But then what happened in between 2012 and 2016 was the atrophy of a lot of that work because the DNC was not invested, and there was no continuity in terms of maintaining or operating systems. And so by 2016, we were using a data warehouse that was basically on its last legs and barely functional. That was indicative, I think, of the general investment in data and technology. There’s a lot of things that happen behind the scenes that are not sexy but really important, like maintaining regular updates to the voter files, cleaning the data, and data quality work. And that debt accrues for security reasons, access reasons, and all of these other ways.
[DNC chair] Tom Perez had made one of his four key platform principles continuous investment in data and technology infrastructure, and we’ve been working on that since 2017. We upgraded the data warehouse. We transferred it to Google Cloud Platform, made huge investments in data quality behind the scenes, and did things like acquiring 65 million cell phone [numbers] in 2020 (and 40 or 50 million more in 2018 and 2019), better record linkage, and all sorts of enhancements. So when the Biden team came in, [they] could just roll right into a really solid foundation—and not just the Biden team, but all of those down-ballots that are using our same resources.
Q: Do you feel there’s a difference in ethics between how Democrats and Republicans run their technology stacks?
A: I think we’ve seen many unethical practices from the Republicans around how they’re leveraging information and how they’re targeting voters with specifically false or inaccurate information. That is not directly connected to how they architecture their data stack per se, so I wouldn’t want to say that. From what I can see, which is how they actually deploy the resources they’re gathering—their messaging, their voter targeting, their use of social media—I find it deeply worrisome that they’re really continuing to undermine democratic norms and practices through how they are talking to voters. [Republican operatives have been accused of using data to target and suppress the votes of Black and Hispanic voters, as well as to spread disinformation.]
Certainly, we believe really strongly that any data we have should be used to enfranchise people, to give more people information about how to vote, where to vote, when to vote, who to vote for—to be empowering to make the choice that they choose to make based on their own knowledge of the candidate and their preferences. It seems like on the Republican side, we see more of that being used for trying to disenfranchise people through voter suppression, and that seems highly unethical to me and undemocratic.
Q: A big challenge to campaigns is reinventing the wheel every two or four years. How is your team planning for longevity?
A: One of the biggest challenges is getting out of the cyclical gravity of the campaign cycle. You see a lot of innovation around presidential cycles in particular: you have a lot of money and time, and you can hire really talented people. There’s two forms of waste in this ecosystem. One is the waste of rebuilding every two years. The other is a waste of thousands of campaigns building the same thing. And so there has been a concerted effort to really counter the natural inclination to fund and defund. I think that there’ll be a really big test of that on the Democrat side after this election. My goal is to continue to lead the DNC tech team and have that stability, have that continuity, make sure that we can start looking ahead to 2022, 2024, and that we’ve reversed the trend around spurts and stops.
And we can really, really start innovating on top of what is now a very solid foundation. I think the Exchange [the Democratic Data Exchange, the party’s clearinghouse for information that can be used by campaigns] is absolutely part of that vision. How do we have infrastructure that is not just ephemeral, that benefits from domain expertise and institutional knowledge? A lot of this is also cultural, right? Keeping talent in the ecosystem, keeping people who know the systems and ecosystems so that they can keep working on it. Like, no other tech company would fund and defund their team every two years. That would not be a way to run an effective long-term infrastructure platform.
The goal is to have a really strong platform where campaigns then come in and innovate and iterate like little experiment labs. So they can go to the really important stuff, which is how do you innovate on how you’re talking to voters and how you’re effectively mobilizing and persuading voters—not like how are you cleaning latitude-and-longitude data.
What do you and your team do on November 4?
A: I will be checking myself into a hospital to have a baby, so that’s what I’ll be doing. [Thomas is heavily pregnant.] My goal is for the team to be able to focus on two things or three things. One is they need to rest. We will probably be very focused on supporting any recounts for any election, big or small, and then off-boarding and asset transfers. We’ll be making sure that we are helping campaigns shut down, the Biden campaign in particular—capturing all that good data—and that we’re documenting everything. And then we’re going to start planning for 2022 and 2024. We have vision planning sessions mapped out once we have a little bit more brain power and brain space to think beyond the immediate election and think about what we want to be building for two, four, 10 years from now.
Want to do more with live video? Wondering how to simplify the process of going live? To explore creating better systems for live video, I interview Tanya Smith on the Social Media Marketing Podcast. Tanya is a video strategist who helps service providers demystify the video creation process. Her site is GetNoticedWithVideo.com and her course […]
For our October look at Boston and its startup scene, let’s get into the data and then understand how a new cohort of founders is cropping up among the city’s educational network.
