"You mean graphics card prices went up just to make these things?"

"Worth every penny!"

At the end of 2026, users are scrolling through short dramas with one hand while raising a "big blue fat fish" with the other—and they are actively experiencing what a real AI application actually means.

On the mass content consumption side, the digital audiovisual industry, represented by Hongguo short dramas and comic dramas, is ushering in comprehensive prosperity. AI technology now penetrates the entire process—from script polishing and storyboard drawing to dynamic generation and voiceover—driving high-volume industrial production of quality dramas and making them a mainstream choice for daily entertainment for hundreds of millions of users.

On the other end, in developer circles and among young professionals, the top trend recently is to raise a "big blue fat fish" on one's own computer. This agent framework originates from DeepSeek's open-source project DeepSeek Harness (DSH for short), and because its logo is an anthropomorphic blue whale, the community has given it an affectionate nickname.

In just over two months, developers worldwide have written thousands of plugins around it: some taught it to directly take over the browser to automatically fill in forms and post, some let it report code progress at any time on mobile, and some installed it into social group chats to act as a group member with long-term memory.

The delivery form has changed, and so has the interaction logic. Compared with more professional AI office applications, AI applications seem to have quickly taken the path belonging to the masses.

1. Farewell to "Half-Baked Products"

Looking back over the past three years, the core reason early AI applications often fell into the dilemma of "high downloads, low retention" is that the vast majority of products delivered only "half-baked products" to users.

What is a half-baked product? A user enters a prompt into a dialog box, the model spits out a draft or a static image, and the subsequent formatting, editing, cross-software entry and verification still require manual step-by-step completion.

At the time, for ordinary people lacking professional skills, figuring out how to write prompts was itself an extra threshold; and for professionals, frequently copying and pasting between chat windows and business software did not truly free people from tedious processes.

According to the latest Q3 AI application market report released by QuestMobile, after domestic AI-native apps passed the popularization period with user scale exceeding 546 million, the third quarter ushered in a round of "natural filtering," with average quarterly usage per person falling from the Q2 peak to 86.6 times.

Behind the data decline is precisely the shift in user demand from "curious trying" to "result delivery." After eliminating a batch of homogeneous wrapper Q&A tools, AI applications in 2026 formally established two core main lines for eliminating "half-baked products."

The first main line is to become thoroughly "invisible" on the mass consumer side, directly delivering highly complete content and services.

Whether it is the currently booming short drama and comic drama ecosystem, or system-level perception services such as "Xiaobu Suggestions" on mobile terminals, whose average monthly usage per person rose 214.3% year on year, their common feature is "de-toolization."

When ordinary viewers smoothly binge-watch dramas on Hongguo short dramas, they do not need to know at all which visual large model or speech synthesis technology is called behind the scenes; AI retreats behind the scenes of the content industry, helping creators compress storyboard drawing and post-production that originally required a team of dozens of people and months into an efficient and stable digital assembly line. Technology can already be delivered in a form users enjoy—this is a major sign of maturity.

The second main line is to provide support on the personal productivity side, turning AI into a freely assemblable "white-box workbench."

The masses need minimalist black-box finished products, while creators, programmers and heavy office users need extremely high controllability and customization freedom. This is precisely why DeepSeek Harness "big fat fish," along with Tencent WorkBuddy (32.353 million MAU in September) and Alibaba Qwen Office and other desktop and collaborative agents, are rising rapidly.

Under the "everything is a plugin" architecture followed by DSH, AI can directly reuse the user's logged-in Chrome browser state to operate web pages through the community-developed dsh-browser plugin, understand front-end screen screenshots through the ModLens plugin, and even through Agent Teams mode have a master node automatically decompose tasks and dispatch multiple sub-agents to write code or organize research reports in parallel.

At one end are zero-threshold audiovisual finished products, and at the other end is a digital teammate with full permissions. Enabling users to simply use it and become unable to live without it is the essence of the word "application."

2. A Mirror of History

Compared with the development of internet applications, there seems to be nothing new in AI applications. What is highly similar to the past two decades is the path by which technological popularization stimulates supply explosion, and the diffusion law by which hardcore tools reach the masses through a "cultivation culture."

Ten years ago, the popularization of 4G networks and smartphone cameras greatly lowered the threshold for video shooting, allowing creativity from the graphic-text era to transform into the prosperity of short videos; today, the maturity of multimodal generation technology is likewise setting off a creative equalization in the audiovisual field, freeing a large number of excellent web novel stories and anime ideas from the high costs of live-action shooting and animation rendering, thereby enabling short dramas and comic dramas to bloom in themes, art styles and narrative rhythm.

Also familiar is the co-creation vitality of geek communities. When personal computers first entered homes, people established an emotional connection with digital tools by assembling hardware themselves, changing desktop themes, and raising an electronic pet in the corner of the desktop.

Now, in the DeepSeek Harness open-source community, developers have not only written rigorous plugins for the "big fat fish" for financial data scraping, Feishu remote control and starship-style voice input, but also spontaneously gave it an anime "whale girl" skin, developed a QQ bot plugin that can remember group chat slang, and even made "fun" components that turn the model reasoning intensity slider into an interactive gimmick.

This ecological evolution of "hardcore productivity foundation + community fun cultivation" is in the same vein as the prosperity of open-source software and early internet plugin ecosystems.

However, often in details people do not pay much attention to in daily life, the underlying distribution and interaction logic of the software industry has already undergone a rift.

Over the past two decades, whether portal websites in the PC era or super apps in the mobile internet era, all followed the "destination logic"—software companies spent huge sums building fully functional "digital shopping malls," and all product design was aimed at attracting users in, keeping them in the interface, and then monetizing by competing for users' dwell time.

By 2026, productivity-tool AI applications have completely reversed this logic, from "enclosing people inside software" to "dispatching AI into systems."

The QuestMobile report disclosed a highly representative set of comparative data: in Q3 2026, the average monthly usage count and duration per person of traditional search apps fell 21.7% and 17.2% year on year respectively; at the same time, the usage frequency of AI applications in PC client form surged 167.7% year on year.

Why would PC desktop clients, once considered to have peaked, return to center stage? Because a truly efficient AI assistant does not need users to click into a beautiful app interface every day to exchange pleasantries; it needs to reside at the bottom layer of the computer system, be embedded in the daily communication flows of WeChat and DingTalk, and while the user is eating or resting, through a brief instruction sent from mobile, quietly finish file organization, code debugging and schedule planning in the background.

Software is degrading from a "destination that needs people to serve it" into an invisible execution channel. For the new generation of developers, this may profoundly change many things.

3. Where Does the Qualitative Change Come From?

You may notice that compared with the explosive popularity of half-baked applications like OpenClaw at the beginning of the year, in the second half of 2026 everything seems to have progressed much faster.

The underlying logic driving this qualitative change at the application layer is that the center of gravity of the entire AI industry has completed a key handover: the main battlefield of competition has shifted comprehensively from simply comparing the "IQ" of foundation large models to building the "Harness" and "private Context" that connect to the real world.

The word "Harness" (originally meaning horse tack or harness) itself precisely summarizes the core contradiction in the current implementation of AI applications.

From 2023 to 2025, the whole industry was fully cultivating "thousand-mile horses" that run faster, with model parameter counts, inference benchmark scores and long-text capabilities changing rapidly; but a steed without a harness and not connected to a carriage cannot even pull a bag of flour in real life.

But OpenClaw's private self-deployment philosophy and the subsequent maturity of various specifications changed the logic of applications.

At the technical foundation, as MCP (Model Context Protocol) and A2A (Agent-to-Agent Protocol) became industry-wide standards, and the open-source community decoupled models, long-term memory, skills and extension commands into standardized plugin interfaces, AI finally legally and securely obtained the "key" to enter users' real work sites.

When a "big fat fish" is allowed to call the local command line, read the project directory structure, reuse browser login credentials, and precipitate the user's habits into local long-term memory in every interaction, the productivity it can unleash will inevitably be stronger than similar products that can only operate in the cloud.

Likewise, on the content production side of short dramas and comic dramas, success is no longer determined by the generation quality of a single image, but by whether a stable engineering pipeline can link character consistency, camera movement trajectories, emotional voiceover and editing rhythm into a workflow that does not break down.

From the plugin ecosystem on personal desktops to the production pipeline of the content industry, in essence both represent engineering capability deeply taking over the model dividend.

Looking in this direction, there are still several certain directions for the explosion of AI applications.

First, content creation will fully enter a prosperous period of "small-team creativity + industrial-grade AI pipeline."

In fact, there is already a clear content quality dividing line between AI short dramas in the first half and second half of the year. As generation consistency and audiovisual workflows further mature, high-quality content's dependence on asset-heavy teams will continue to decline, and creators with unique narrative aesthetics and scheduling capabilities will unleash greater production capacity.

Second, software forms will comprehensively move toward an assembly era of "minimalist core + countless plugins."

Traditional software that is large and comprehensive with rigid functions will increasingly struggle to meet personalized needs. A model like DeepSeek Harness that retains only a lightweight scheduling core and leaves specific capabilities entirely to community plugins and individuals' free combination will become the mainstream paradigm of productivity tools. In particular, the recently controversial Rea reverse-engineering technology is spreading the idea of "copying software yourself" from the tech circle to ordinary users.

Third, the core digital assets of individuals and enterprises will further settle into local memory libraries and exclusive workflows, while possibly prompting changes in sharing mechanisms.

The model itself is becoming public infrastructure like water and electricity. The truly hard-to-replace moat is the set of industry knowledge, operating habits and private plugin combinations you spend months feeding to the "big fat fish." So, can these accumulations spread externally in another way?

Recently, some agent tool vendors have been trying to launch paid sharing mechanisms, making knowledge skills an asset that can flow. Considering the trend of AI application popularization, this may also brew the next explosive point—though perhaps it will also become another knowledge-payment scheme.

In short, everything comes down to daily use. The end of AI applications will inevitably be shaped jointly by users and AI. How to design personalized applications that most users cannot live without, starting from common principles, will be the key to tipping the balance of the entire track.

Source: Songguo Finance