Where did all these AI plugins come from?
"Too many AIs, can't keep up." This anxiety is not unfounded. The supply side has indeed exploded, with AI projects on Github increasing by 59.3% in 2023; the number of developers on the AI open-source community Hugging Face quadrupled.
a16z analyzed user usage of web-based AI products. Among the Top 50 products, 28% are content generation (including text, images, video, music), 22% are content editing, 14% are productivity tools, and 10% are general intelligent assistants (such as ChatGPT, Claude, etc.). Mobile is different, with users preferring to use general assistants (an AI Chat interface right on the home screen).
Web products are suitable for more complex, multi-step workflows. In the productivity category, six out of seven products offer Chrome browser extensions, or even exist only as extensions. The rationale for plugins is "synchronous operation," eliminating the need for users to leave the software they're currently using, greatly reducing the friction of switching between multiple web pages or applications.
AI should be designed to be as close to the user as possible, preferably embedded in the usage environment. This is a prerequisite for AI to be used effectively.
However, developers are not satisfied with merely integrating technology into product processes. AI applications iterate quickly but have short lifecycles. 40% of a16z's Top 50 list changed within six months. QuestMobile points out that the activity rate of domestic generative AI applications is below 20%, three-day retention is below 50%, and the uninstallation rate of some apps is over 50%.
Some AI applications have evolved from needing to open a webpage to use, to becoming plugin products that accompany users at all times, in an attempt to achieve implementation in a lighter way. Browsers remain an important traffic entry point, and search is one of the most universal and high-frequency internet demands. These plugins compete for space on the browser interface, trying to become an "entry point" to solve user growth problems.
Plugins (Once) Reigned Supreme
In 2008, Chrome was still a fledgling compared to IE's 60% global market share. But its growth was explosive—from 5% in 2009, to 15% in 2010, to 31% in 2012... Google's open ecosystem stood in stark contrast to Microsoft's closed approach at the time. Extensions (plugins) played a crucial role in this.
Within the established development framework, third-party developers could implement functions with minimal development work. More and more developers were attracted, creating rich functionalities that opened up the browser's "limitations." Users were online, but not just browsing web pages—for example, a plugin designed for streaming services like Netflix and Hulu, "Netflix Party is now Teleparty," allowed people to watch and discuss in a floating window. Someone in the plugin's comment section said it saved a long-distance relationship.
These plugins, like functional mods in games, could provide better reading modes, dark modes, automatically accept all cookie requests, better video playback modes... They truly considered the user's perspective.
Browser plugins accompanied the prosperity of PC internet. Mobile internet took away most of users' attention, affecting some lightweight plugins to "App-ify." For example, "SimpRead" started as a plugin that "generates pages with layouts suitable for Chinese reading," but gradually became more substantial, integrating features like annotation, read-later, and export, and eventually launched as a standalone application.
While mobile applications try to capture user attention, they also build high walls. At this point, people realized that plugins were the best footnote to the "internet spirit," and also its last glow.
This is not as simple as posting flyers on telephone poles
In 2019, security company Extension Monitor analyzed 180,000 extension plugins in the Chrome store and found that people commonly use several categories: ad blocking, communication, shopping, security, password managers, etc., mostly functional.
Regarding the development of generative AI applications for consumers, there has been an ongoing discussion: with large models iterating so quickly, will stronger underlying technologies cover the existing "carving" on upper-layer applications? This is a challenge for software development—can we set aside the baggage of "App development" and solve users' actual problems by figuring out where and how AI should "assist" people.
The situation is slightly different this time. These AI plugins need to integrate well with scenarios to generate user value.
Take "search" as an example. During Google I/O, the VP responsible for Google Search shared his observations on user behavior habits, "When users are not clear about what to search for, they usually start with a broad question, get inspiration from search results, dig deeper, and ask more questions."
At the same time, there's an intention behind every query. What's the next step with the information found? Is it to summarize into an outline, write a PPT? Or change the language style and compile it into a news release? And so on (this can also be seen as today's AI search revolutionizing traditional search engines).
These plugins that pop up after highlighting text have "reading companion" as their core user scenario. When users browse web pages, they can selectively let the plugin explain, search, translate, summarize, expand, or convert to social media posts with one click—reorganizing and delivering information around the core search needs.
In terms of product design, some plugins display a few commonly used functions in a floating toolbar at the highlighted area. For example, Dou Bao even allows users to customize "skills" (equivalent to creating a bot, or GPTs) and add them. Other more functions are listed in the sidebar.
Some adhere to the "design principles" of plugins, such as Kimi, which is more concise and "radical." AI will automatically "identify user intent" and "understand context." A user tested it and found that with just one floating button and one window, it can not only explain and translate but also solve math problems. Kimi also sets up a floating button on the right that can be called out to summarize the page and engage in question-and-answer dialogues.