Alibaba's Qianwen "Discouraging" Its Own Flagship Model? A Major Test of AI Office Productivity and Cost-Effectiveness
Summary of Key Points
The core of this report highlights an counterintuitive business phenomenon: Alibaba's AI office tool, Qianwen Office, actually advises users not to always use its most powerful flagship models (Pro/Max) but to prefer the lighter and cheaper standard mode (Flash).
To verify this claim, the journalist conducted a thorough test. Nine complex securities research reports were provided to the AI, with one duplicate report intentionally included, and the AI was asked to organize the information into a Word document and compress it into a one-page presentation PPT.
Test Results:
- Standard Mode (Flash): Fast (10 minutes for Word, 3 minutes for PPT). It identified the duplicate report, maintained a clear structure, and provided accurate data. However, the content was more of a summary, lacking in-depth analysis, and the PPT layout was cramped with small text, making it suitable for quick browsing and preliminary drafts.
- Advanced Modes (Pro/Max): Slower (10 minutes for Word, 10 minutes for PPT), but the output resembled that of a “senior analyst.” It not only organized the data but also delved into the logic behind the differing opinions of the various analysts (such as different assumptions) and automatically checked for layout issues (such as overlapping text and color contrast), offering more professional guidance.
- Conclusion: Alibaba's claim that “95% of daily tasks can be handled with the standard mode” is largely true, but by “usable,” it means “capable of completing the task without errors,” not “delivering a perfect product.” The standard mode is for getting the work done, while the advanced mode ensures the work is done well. For most daily tasks, the standard mode offers excellent value for money; however, for formal reports or complex decisions, the advanced mode is still indispensable.
---
In-Depth Analysis: Five Dimensions to Understand This AI Office Revolution
1. Reversal of Business Logic: Why Would a Large Company Encourage You to Use a Less Powerful Model?
Typically, tech companies promote their products with the idea that the more expensive, the better, as the flagship version is a source of profit and a symbol of technical prowess. However, Alibaba's approach is quite bold: it actively downgrades its flagship models and promotes the lighter models.
- Behind the Strategy:
- Cost and Scale Effects: Large model processing (i.e., the AI’s thinking process) is very computationally intensive and costly. If all users used the top-tier models, it would put a huge strain on servers and increase operational costs. By directing 95% of users to use the lighter Flash model, Alibaba can significantly reduce the cost per usage, allowing it to support a larger user base.
- Lowering the Bar: Many ordinary users do not need the most intelligent AI; they just need an “obedient, fast, and affordable” assistant. If the default option were the expensive and slow flagship version, it might deter many users.
- **Redefining “Usable”: Alibaba is redefining what “smart” means in an office context. It signals to the market that in most daily tasks, “speed” and “accuracy” are more important than “depth.” This is a practical product strategy aimed at integrating AI into frequent workflows, rather than making it a rare luxury.
- Popular Metaphor: It’s like going to a restaurant. In the past, the owner would always recommend the most expensive “full Chinese menu.” Now, the owner says, “95% of the time, a ‘regular meal’ will suffice—fast and cheap. Only for important clients (complex tasks) do you need the full menu.” This increases the restaurant’s turnover rate and saves you both money and time.
2. Uncovering the Real Capability of AI in Handling Complex Documents
The test used nine securities research reports, which are a challenging task for AI office tools. Why? Because these reports are not just a collection of text but contain:
- Data Traps: Numbers from different years and statistical methods (e.g., GAAP vs. Non-GAAP) are mixed together.
- Conflicting Opinions: Different analysts have varying views on the same data.
- Information Redundancy: Multiple institutions may cite the same financial data but reach different conclusions.
Performance of the Standard Mode (Above Average):
- Duplication Removal: It successfully identified two identical reports, showing basic file recognition capabilities.
- Structuring: It organized the information by topic (Alibaba’s operations, AI progress, analysts’ opinions), which aligns with human reading habits.
- Limitations: It functions more like a “stenographer,” providing a summary of the information without explaining the underlying reasons for the differences.
Performance of the Advanced Mode (Above Professional Level):
- In-Depth Analysis: It provided a detailed analysis, explaining the sources of the differences (e.g., different assumptions about cloud business margins).
- Self-Correction: It automatically checked for layout issues and even exported the PPT as an image for further review, demonstrating a level of self-reflection typical of advanced models.
- Limitations: Although more professional, it is more than three times slower than the standard mode, and the resulting PPT is still cluttered, indicating that AI still falls short of human designers in terms of aesthetics and information selection.
- Popular Metaphor: The standard mode is like a “news summary” that gives you a quick overview of the facts and opinions. The advanced mode is like a “detailed report” that explains the reasons behind the facts and the motivations of the various parties, as well as potential risks.
3. The Truth Behind “95% of Tasks Being Usable”: What Does “Usable” Mean?
Alibaba claims that the standard mode can handle 95% of daily office tasks. This sounds appealing, but the definition of “usable” is crucial:
- **If “usable” means “no errors, deliverable content”: The standard mode meets this requirement. It reads the files, removes duplicates, and generates clear Word documents and PPTs, which are sufficient for daily tasks like organizing materials, writing weekly reports, and conducting preliminary research.
- **If “usable” means “ready to use without modification”: The standard mode does not meet this. Its PPTs are difficult to read and require manual adjustments. The advanced mode is better, but still needs some tweaking.
- What Are the Actual 95% of Tasks? These tasks include information organization, preliminary summaries, format conversion, and simple Q&A. The remaining 5% involve formal reports, complex decision-making, cross-verifying multiple data sources, and visually appealing presentations.
- Advice for Users:
- Daily Tasks: Use the standard mode for organizing meeting notes, extracting key contract terms, and drafting emails—it’s fast and cost-effective.
- Important Tasks: Switch to the advanced mode for reports to your boss, investment analysis, and complex project planning. Spending a few more minutes will result in more rigorous logic and professional presentations.
4. The New Benchmark for AI Office: From “Chatbots” to “Office Agents”
This article reveals an important trend: AI is evolving from “chatbots” to “office agents” (intelligent systems).
- Traditional AI Chat: You ask a question, and it provides a text response.
- AI Agents: You give it a task (e.g., organizing these reports into a PPT), and it plans the steps:
1. Reads the files.
2. Removes duplicates.
3. Extracts key information.
4. Organizes the information.
5. Generates a Word document.
6. Extracts the content from the Word document.
7. Plans the PPT layout.
8. Generates the PPT.
9. Self-Inspection (unique to advanced modes): Checks for layout, color, and overlap issues.
10. Outputs the final file.
- Key Breakthroughs:
- Multi-File Handling: AI can handle multiple PDFs and understand their relationships.
- Format Conversion: It can convert unstructured PDFs into structured Word documents and then into visually appealing PPTs, bridging the gap between different document formats.
- Self-Reflection: Advanced modes can identify and fix issues in their own outputs, marking a significant step towards autonomy.
- Popular Metaphor: Traditional AI is like an “encyclopedia” that provides information on request. Modern AI is like an “intern” that finds the information, organizes it, writes the report, and creates the PPT, even checking for errors. Although you still need to review the final output, it can handle most of the basic work independently.
5. Future Prospects: AI Won’t Do Your Job for You, but It Can Help You Work
The article concludes with a poignant point: “Qianwen is not planning to replace you at work just yet.”
- Limits of AI: AI excels at information processing, format conversion, preliminary analysis, and repetitive tasks.
- Human Roles: AI’s role shifts from executor to reviewer and decision-maker. You no longer need to organize data manually but must evaluate the accuracy of AI’s results and whether they align with your strategic goals.
- Impact on Professionals: The future core competencies will be “issuing precise instructions to AI” and “reviewing AI’s outputs.” You can spend less time on mundane tasks and use the saved time on analyzing the data and planning next steps.
- Cost Efficiency: What used to take half a day to organize reports can now be done in 10 minutes. You can use this time to think about the implications of the data and plan your next steps.
Conclusion
Alibaba’s decision to discourage the use of its flagship models is essentially a pragmatic adjustment to its product strategy. It recognizes that perfection is not the only goal in daily office work; efficiency and cost-effectiveness are more important.
For users, this is good news: AI office tools are becoming more user-friendly, affordable, and widely available. You don’t need to understand complex technical details; just know:
- For everyday tasks: Use the standard mode for speed and cost-effectiveness.
- For important tasks: Use the advanced mode for reliability and professionalism.
- Final Review: Always be the one to make the final judgments.
AI won’t replace you, but those who know how to use it will outperform those who don’t. The first step in leveraging AI is to choose the right tools wisely, rather than blindly pursuing the most powerful ones.