Lingguang 2.0: A personal assistant must solve problems before charging

Xi Chen introduces a shift from question answering toward personal AI assistants that work on continuing tasks. China’s super-app ecosystems may make payments and services easier to connect, but trust, platform incentives and monetization remain open questions.

Ant Lingguang CTO Mo Linjian, known as Linyi, says Lingguang 2.0 shifts from generating mini-apps toward helping ordinary users pursue goals over time. The team began the work before comparable products appeared.

Linyi used a Codex-like tool to analyze his child’s mistakes and generate exercises, but says developer tools are too complex for his family and older users. He wants to make similar capabilities accessible to ordinary people.

He cites his mother’s blood-pressure notebook: recording measurements is only the first step; understanding trends follows. He sees potential for an agent to assist with such ongoing tasks if the interaction becomes easier.

Linyi says the aim is for users to state goals and constraints while the system assists along the way. Better long-horizon models are a foundation, but the surrounding harness, memory, protocols and costs also matter.

Models remain probabilistic and cannot guarantee perfect correctness, he says; current information also needs sources outside model training. Engineering safeguards must account for those limits.

Looking back at Lingguang 1.0, Linyi says generating apps from a sentence lowered the creation barrier, but making apps was not a frequent need for everyone. Chen cites a three-day retention rate just above 20%.

He says aggregate retention matters less in isolation than whether useful apps are repeatedly used and growing. Creation counts are not the same as value; the question is whether users’ problems are solved.

For a viral product, he suggests tracking whether engaged users with several meaningful goals grow and remain, since a surge of curious users can distort the aggregate.

Discussing a 100-million-yuan mini-app incentive program, Linyi says it helped educate the market and encouraged sharing and remixing good work, though producing a polished app still requires considerable skill.

Chen asks why agents still need incentives and promotion. Linyi points to a different competitive environment from early ChatGPT and a habit shift from asking one question to stating a goal, such as planning a trip to southern Xinjiang.

Natural language will not eliminate graphical interfaces, Linyi argues; the two should work together. A GUI packages steps into a visible process, and a click may be easier than speaking for repeated actions.

Images, text, speech and charts each have strengths. He expects multimodal, unstructured input and output chosen to communicate information and support the interaction effectively.

Linyi describes bill analysis: a friend saw returns far exceed actual spending, and he downgraded an expensive phone plan after receiving a suggestion. He sees those small savings as practical value.

During a family trip, he says Lingguang adjusted the itinerary for southern Xinjiang weather and illness, replanning against the overall goal rather than simply shifting unfinished items to the next day.

Chen describes other assistants selling used items or finding overlooked refunds. Both say people could do those things themselves, but an agent may reduce forgotten tasks and administrative effort.

Workplace and personal uses are not mutually exclusive. Linyi says workplace value is easier to monetize, while life outside work is broader; lower inference costs are important for consumer applications.

He sees recurring needs in children’s education, money management, saving and travel, since these often involve ongoing goals rather than a single exchange.

An assistant should handle tedious tasks while keeping the user involved in decisions that require them. Willingness to pay depends on incremental value, he says, using a child’s learning progress as an example.

Chen distinguishes essential pain points from optional conveniences. Linyi says education can be a pain point, but value depends on the actual improvement; summaries of work messages save time without necessarily being indispensable.

On subscriptions, commissions and advertising, Linyi says the market is unsettled. He favors exploring subscriptions or commissions and worries advertising incentives may be difficult to reconcile with trust in a personal assistant.

For now, he says, creating incremental value for users comes first; dividing the returns can be explored later. Large platforms have resources and established ecosystems, but aggressive immediate monetization may be premature.

An agent’s reach depends on its environment and access to real services as well as the model. Linyi says ecosystems such as Alipay make bill analysis, payments and travel services easier to connect.

Chen asks why a platform might block an outside assistant. Linyi says direct transactions that bypass pages may challenge traffic- and advertising-based businesses; coordination within an ecosystem may be easier, but conflicts do not disappear.

He expects bargaining among parties to continue and asks whether the interaction creates value rather than merely redistributing it. He imagines ordering otherwise-unpurchased coffee while driving, while noting payment and trust issues remain.

Lingguang can connect bills, email and workplace services. Linyi says simple lookups may suit an ordinary API or MCP, while questions involving business rules may be better delegated to an agent that knows that domain.

Browser automation is not a universal access solution: it may help when a willing merchant lacks an API, but forcing access against a platform’s wishes becomes a conflict of interests.

Linyi does not expect users to interact only through agents immediately, but thinks agents may gradually become a default option. Merchants will differ in their responses to price comparison, traffic and brand representation.

Brands may connect proactively to control how they are represented, while low-margin merchants may fear automated price comparison yet cannot ignore shifting traffic. These commercial problems cannot be solved by technology alone, he says.

Safety determines trust, Linyi says. The team separates account credentials from the agent and classifies instructions and actions by risk; payments and similar high-risk actions still require user confirmation.

He says the group is exploring infrastructure for automatic payments under specified amount and scenario limits, but scaling high-risk actions requires caution until the infrastructure is ready.

He attributes most cost reduction to model improvements. Background tasks can also be scheduled away from peaks, and smaller models can handle some work; routing and scheduling still have room to improve.

Linyi wants assistants to improve execution as well as remember users—for example, reusing effective flight-search experience to improve cost and completion rates.

Automatic generalization can go too far: avoiding spicy food today does not mean avoiding it forever, and a successful path is not always best. Sustained automatic improvement remains difficult, he says.

He separates memory into durable preferences, events that actually happened and temporary states. Ongoing tasks can establish better-grounded facts, but the system must keep those categories distinct.

For proactive notifications, he proposes three levels: urgent notices for matters directly affecting a goal; measured suggestions for discoveries that may change an action; and background material available without interruption.

Whether a suggestion merits attention depends on the person: a well-reviewed nearby café may change a coffee lover’s itinerary. Linyi says the product is currently conservative and should explore more useful, non-disruptive suggestions.

He describes a travel diary and half-marathon training: family members report preferences, while the assistant connects fitness-device data and adjusts suggestions. He expects value and engagement to rise when users entrust it with ongoing goals.

On being always available, he distinguishes the user’s experience from running continuously in the background. An assistant can wake at relevant times; idle operation must be weighed against cost and additional value.

He expects headphones, cars and glasses to offer access when opening a phone is inconvenient, and hopes for more natural multimodal interaction and more reliable task completion.

They imagine connections to home devices or a car: an assistant that knows an authorized user is about to return to a hot vehicle could precondition it. Linyi says broader context can make such coordination possible.

Linyi recalls satisfaction when targeted exercises helped his child’s scores and when an assistant suggested a gift using family travel context. He hopes his mother will eventually use the product and benefit too.

Chen closes by identifying reliability and trust, ecosystem incentives and monetization as barriers to everyday personal assistants. She sees access to real services and sustained measurable value as central to competition.