China's TikTok censors US teenager for criticising China's Muslim crackdown
AI implications to the Chinese company's grab for the world's smartphone videos
AltDeep is a newsletter focused on microtrend-spotting in data and decision science, machine learning, and AI.
Know an engineer, research scientist, AI product manager, or entrepreneur in the AI space? Buy them a gift subscription. They’ll owe you one.
Ear to the Ground
Cringe-worthy AI hype from The Economist
The Economist is the latest top tier publication to hype the GPT-2 transformer network language model created by OpenAI. They did so by “interviewing” the network, meaning that it generated “unedited” answers to the questions.
But the answers were actually curated by human journalists — this was not clear from the article itself.
Further Reading
An artificial intelligence predicts the future — The Economist
“At AWS, our mission is clear: we aim to put machine learning in the hands of every developer”
With AWS services like Amazon Comprehend, it is straightforward to grab machine learning predictions through APIs. However, writing application code that shuttles various machine learning outputs across a broader business intelligence service is still tedious.
In a blog post, AWS announced it is addressing this problem by integrating prediction models into several of its business intelligence and database services.
Big Picture: The machine learning cloud services war is entering a new phase of land battles between sales teams. From what I hear, the smart money is on Microsoft Azure despite being second to AWS in market share, because Microsoft’s commercial sales operation is more sophisticated.
Further Reading
How AWS is putting machine learning in the hands of every developer and BI analyst — AWS Machine Learning Blog
AI says Henry VIII was a collaboration
One well-established task for machine learning models of natural language is authorship attribution. For example, you train a model on documents that Alexander Hamilton, James Madison, and John Jay were each known to have authored, and then use that model to try to figure out who wrote which parts of the Federalist Papers.
Now the Smithsonian has applied this technique to the play Henry VIII and concluded that Shakespeare collaborated with playwright John Fletcher.
Further reading
Machine learning has revealed exactly how much of a Shakespeare play was written by someone else — MIT Technology Review
Signals from China
TikTok censors teen for video about abuses of Uighurs in China
China-owned TikTok reportedly suspended the account of a teenager for posting the following video, which later went viral. The video begins as a guide to curling eyelashes, then abruptly dives into a monologue about ongoing human rights abuses against the Uighur ethnic group in China.
TikTok later unblocked the account and attributed the action to “human moderation error”.
Big picture and AI context
TikTok is a social media aggregator of short smartphone videos that has exploded in popularity around the world, including the US. As Google, Facebook, and other American tech companies have shown, collecting massive amounts of data on users and using that to power algorithms is a tremendous source of political power. China hawks in the US believe that a Chinese company having this power over Americans (many of whom are kids) is equivalent to the Chinese government having this power.
For example, earlier this month, Senator Josh Hawley said the following about TikTok’s declining an invitation to a Senate Judiciary hearing touching on tech privacy concerns and China:
Further reading
TikTok reverses ban on teen who slammed China’s Muslim rackdown — The New York Times
People in China required to have faces scanned when registering new mobile phone services
When signing up for new mobile contracts, people are already required to show ID and have their photos taken. Now, they will also have their faces scanned to verify that they are a genuine match for the ID provided.
This is not just for phone registration
I had a friend over for Friendsgiving who had recently returned from visiting her hometown in China. She mentioned that during her visit, she tried to send a package on behalf of her father. She brought his ID but did not bring the man himself. When she tried to process the package, she learned learning facial verification was required. So she sought out some random middle-aged man on the street and had the app scan his face. According to her, the facial verification app correctly classified the provided photo as a mismatch with whatever image was in its database for the provided ID.
She said the experience was unnerving because there was never a time where members of her family had knowingly provided images to the government for this purpose.
Further reading
Think someone might like this post?
AI Long Tail
Pre-hype AI microtrends.
After Grammarly valuation, emerging crop of AI writing assistants
Navigator is a “teamwork assistant” designed to help make meetings more effective. Before a meeting, it contacts meeting attendees through email and Slack to gather discussion topics, and then it creates an agenda. It then keeps meeting notes and action items. According to the company’s press, “it even suggests discussion topics to help build relationships and surface priorities.”
This app falls into a trend of hyper-focused AI assistants, as I’ve written about previously. 95% of apps like this are just rehashing functionality of older apps (e.g., voice transcription software) not branded as AI. The AI component is the remaining 5% that does things like “suggest discussion topics.” The question is whether a team can build this 5% into something that delivers real business value.
Stanford releases an open, privacy-preserving virtual assistant
Stanford Open Virtual Assistant Lab released an open-source virtual assistant called Almond.
Upon perusing the docs, it seems the main attraction that it can alleviate the privacy concerns that make us suspicious of Alexa and Siri.
Almond also has a unique approach to enabling developers to extend the voice assistant to new domains or “skills.” It attempts to convert every natural language verbatim from a user to a line of code in a fixed programming language. Developers train Almond’s data-hungry deep learning-based natural language model with a tool that takes a command in the programming language and generates many natural language variants that match that command.
Decent AI stock photo site launches
Generative.photos is has the best execution of custom stock photos I’ve seen yet.
Don’t miss a post! Altdeep publishes micotrend insights in AI that gets beyond the hype and inside baseball.