The rise of generative AI is leading to a huge amount of content flooding the web, with everything from models like GPT-4 producing full-length screenplays to image generation AI creating frighteningly deceiving photos of Pope Francis in a flashy jacket.
In the last edition of our newsletter we showed you the basics of using AI to look at text signals in pieces of content to differentiate them into “types,” such as a factual explainer about an event vs. a piece of people reacting to it and expressing their opinions.
That is great for helping humans understand what is happening online, but the other advantage of using language models like Overtone’s is that they can be applied across wide swaths of the internet. While generative AI can “write” out new content onto the internet at alarming rates, innovations like our models can “read” just as fast.
This speed can be particularly helpful when looking at events in real-time, like we delve into with the [[Silicon Valley Bank]] case study below. AI can “read” thousands of times faster than a person can. This will add power to publishers, platforms, PR firms, and more.
Silicon Valley Bank - The Media Autopsy of a Bank Run
Part of the reason why it is important to understand what is being said online is that it leads to real-world consequences. Last month spooked depositors at Silicon Valley Bank, communicating online via social media and news articles, rushed to pull out their money, crashing a major financial institution.
Our blog post [[here]] below delves into what we did using AI to evaluate thousands of articles about the bank (and other institutions that did not see bank runs) to pinpoint what exactly happened. What we found is that type signals, and what sorts of articles are being published, can play an important role in alerting observers that something is up. When journalists started reporting out the story, doing the work to uncover it, is when the run gathered steam.
See examples of Overtone types on SVB here
User Needs
For those of you who follow the publishing community closely, all this talk of different “types” of articles may remind you of another popular way of looking at articles: user needs.
User needs looks at what sorts of articles a publisher’s audience wants so they can better understand the world, and examines whether a publisher is filling those needs. It was pioneered by Dmitry Shishkin while he was at the BBC, and he found that outlets routinely publish too many “update me” stories instead of other types of pieces.
Overtone is a big fan of the work that Dmitry, who recently launched an update of his user needs to add two more, has done. You can read our interview together for publisher association WAN-IFRA here. However, our tech focuses on the content itself, rather than the way that it connects to an audience, which can be much harder to delve into with machine learning. That being said, we view our “types” as a great way to get started for publishers looking to build an approach with user needs, and here are some examples of how we view our data fitting in, with more explanation here.
Overtone updates
March was full of opportunities for Overtone. In New York, Reagan Nunnally had the pleasure discussing technological advancements in AI and sharing the Overtone experience with journalists and editors around the world on behalf of the NYC Media Lab.
We rounded the month off with some food for thought. Our very own Christopher Brennan released an article covering large language models like GPT-4. With use of the above analysis on the panic at SVB, Christopher discusses that it is now possible with language models to “read” and create data points almost instantly, instead of solely focusing on their ability to write.
If you are interested in how language models reading can help navigate a crisis or understand content better, the article in the Financial Times-backed Sifted is linked here.
If you enjoyed this edition, please subscribe to be notified when we release our next.