This week, Sam Altman announced his “favorite feature of ChatGPT so far.” It’s called Pulse, and according Altman, it “works for you overnight” by “thinking about your interests, your connected data, your recent chats, and more.” In the morning, you get a “custom-generated set of stuff you might be interested in,” akin to something a “super-competent personal assistant” might prepare. More broadly, he says, it represents “a shift from being all reactive to being significantly proactive, and extremely personalized.” And then, a recommendation: “It performs super well if you tell ChatGPT more about what’s important to you.”
These are the words of a CEO, of course, so we should expect him to be in sales mode. They’re also the words of a person who has not just adopted the language and jargon of generative AI but done so to the exclusion of everything else. In the narrow context of ChatGPT, and through the personified language of generative AI, Pulse can be given agency, ascribed new talents and qualities, and imbued with novelty. Most other people, however, will look at Pulse and see something less futuristic than familiar: a recommendation feed.
A recent study of ChatGPT use helped clarify what the service’s users are most commonly getting from the chatbot, outlining strong consultative habits: a lot of Google replacement, plenty of quick questions and advice, and some task completion. These interactions all depend on the user initiating in the first place, which, if your goal is to maximize engagement and/or draw people into a more comprehensive platform — to make your product the beginning and end of a user’s computing experience — is limiting. People spend lots of time searching, chatting, and working on their devices, sure. But they also spend a lot of time scrolling. Pulse looks like an attempt to secure at least some of the massive amount of attention captured by feeds and to turn ChatGPT into something more than a tool you can consult — specifically, into a source of content you can consume.
To back up a little bit: Before the post-ChatGPT AI boom, which has been defined by large language models and chatbot interfaces, the tech industry’s conversations about AI and machine learning centered on recommendations. That was the case for good reason. Platforms that deployed surveillant recommendation engines were taking over the world. Through the 2010s, social platforms drifted from chronological feeds to algorithmic recommendations, drawing on users’ data and behaviors to show them personalized material. TikTok took this model a step further, treating social connections as firmly secondary to AI-driven learning and recommendation (or, put another way, embracing the model of digital ad targeting for the entire social-media experience).
You can hear, in Altman’s announcement, the description of something akin to a TikTok feed: a “custom-generated set of stuff you might be interested in.” For the logical endpoint of compounded “generation” looks like, Meta helpfully announced a cautionary tale in the form of a new AI-feed product called Vibes:
Anyway, an even closer cousin to Pulse, given the use of ChatGPT as a Google replacement, is the algorithmic homepage popularized by products like Google Now, introduced in 2012 with the following description:
It tells you today’s weather before you start your day, how much traffic to expect before you leave for work, when the next train will arrive as you’re standing on the platform, or your favorite team’s score while they’re playing. And the best part? All of this happens automatically. Cards appear throughout the day at the moment you need them.
By 2016, after Google had abandoned the Now branding but incorporated the features across its product lineup, the company said that it was using “machine learning algorithms to better anticipate what’s interesting and important to you.” The aim was to show Google users “sports highlights, top news, engaging videos, new music, stories to read and more” based not only on their interactions with Google but also “what’s trending in your area and around the world. The more you use Google, the better your feed will be.” By then, it had become obvious that personalized recommendation engines were ascendant and that they’d be incorporated into basically any software product that could accommodate them. And why not? At their best, they were useful and therefore sticky; at worst, they produced low-value engagement that could still be monetized.
Early reviews from heavy ChatGPT users suggest the concept makes sense for them: Pulse is like “a newsfeed tailored to recent conversations,” one writes, saying that he wants to “dump even more information and context and app connections into ChatGPT so I can get an even better daily feed.” It’s easy enough to see how populating ChatGPT with recommendations could increase time spent on the app by casual users, too.