Thoughts on the Free PDF Translation Upgrade: Why We Gave Up Half-Priced Gemini
Today, I want to have an honest conversation with everyone. First, I sincerely apologize to every friend affected by the recent poor experience of our free version product.
In the past period, we have received a large number of negative feedback regarding the free version experience. We have heard these voices and empathized with them. Many users mentioned that the "waiting time during peak hours is outrageous"; the translated documents have "unstable translation quality," sometimes good and sometimes bad; and some complained that the translations are "mixed with Chinese and English, making them hard to read," filled with an inescapable "machine translation feel."
Every time we see this feedback, our team feels deeply guilty. We fully understand the disappointment and frustration when you urgently need an important academic paper or work report but have to face long waits and a translation that is incoherent and logically confused. This is not the experience we want Doclingo to provide, and it completely contradicts the original intention of creating this product.
From the very beginning, Doclingo adopted a freemium model because we firmly believe that high-quality document translation should not be a privilege for a few. We hope to provide a sufficiently useful free version that allows more people to overcome language barriers and access cutting-edge knowledge and information. However, a poor free experience not only fails to retain users but also betrays everyone's trust. It is this realization that has made us determined: we must completely solve these problems at all costs.
After repeated internal discussions and technical evaluations, we reached a conclusion: the AI translation engine currently used in the free version can no longer support our commitment to quality and efficiency. The only way to fundamentally solve the issues of waiting and translation quality is to switch to a more powerful AI engine—one that truly deserves our users.
This decision brought us to a difficult crossroads. The top AI engines on the market are primarily the GPT series from OpenAI and the Gemini series from Google. They both represent the highest level of current artificial intelligence, but their styles, costs, and performance on specific translation tasks vary.
Therefore, this article aims to share the complete thought process behind our significant upgrade decision: how we made the difficult choice between these two top AI engines and why we ultimately gave up the half-priced Gemini in favor of the more expensive option.
We hope that through this transparent communication, we not only apologize for past poor experiences but also demonstrate our determination and commitment to improving product experience.
Experience or Cost-Effectiveness?
I think every day about how to create more value for users. But in the business world, behind value creation, there is always an invisible cost bill. When Doclingo decided to upgrade its core translation engine, our team stood at such a difficult crossroads: on one side was the huge temptation to nearly halve costs, and on the other was the user experience we have always upheld.
Anyone responsible for product profitability understands that cost control is a Damocles sword hanging over our heads. When we evaluated new large model engines, an extremely attractive option presented itself—Google's Gemini series.
To be honest, Gemini's pricing strategy is highly tempting for us. According to our research, the cost of choosing Gemini is lower than that of GPT. A simple calculation shows that if we switch to Gemini, our core engine invocation costs could be reduced by almost half. For an application like Doclingo, which handles a massive number of translation requests daily, this saved expenditure is quite considerable. This money could be invested in marketing, team building, or directly reflected in a more flexible pricing strategy. Faced with such a significant cost advantage, it would be a lie to say we weren't tempted.
However, after intense internal discussions, we ultimately made a decision that seemed "uneconomical": to uphold the experience and choose GPT.
Because we had ample reasons.
Three Reasons for Choosing GPT
1. Precision in Academic Terminology
For any translation tool aimed at the research and academic field, the precise handling of professional terminology is fundamental to its existence. This not only concerns the translation's "faithfulness, expressiveness, and elegance," but also directly determines whether the core value of the literature can be accurately conveyed.
In this in-depth evaluation, a vivid example left a deep impression on our team. When we translated a paper in the field of condensed matter physics, we encountered a high-frequency term: "pair distribution function."
- GPT's translation is: "对分布函数"
- Gemini's translation is: "配对分布函数"
Literally, "配对" seems closer to the original meaning of "pair," which is a very intuitive and reasonable translation. However, for scholars and students in this field, "对分布函数" is the tacitly accepted "insider's term." This small difference is like a watershed, clearly dividing "outsiders" from "insiders." Although Gemini's translation is not wrong literally, it reveals a certain awkward "machine translation feel," while GPT demonstrates a profound understanding of specific academic domain knowledge.
This difference is not an isolated case. Research shows that in highly specialized fields such as medicine and science, GPT-4 level models often outperform competitors in handling complex concepts and terminology. For example, in comparative evaluations, GPT-4 exhibited a higher accuracy rate and fewer serious errors when answering difficult clinical questions. While Gemini's responses may sometimes be easier to understand, this often comes at the cost of sacrificing technical precision. This tendency to "sacrifice accuracy for readability" is extremely dangerous in academic translation.
We understand that Doclingo's core users—numerous researchers and students—deal with these highly specialized terms daily. For you, a "slight error" in terminology could lead to a "thousand-mile deviation" in understanding. An imprecise term not only disrupts the immersive reading flow, forcing you to stop to verify or guess, but more seriously, it may distort the core argument of the original author and even mislead your research direction. Precise terminology is the cornerstone of ensuring academic rigor and the lifeline for improving literature reading efficiency.
The reason GPT models can achieve this is not accidental. Its powerful cognitive and reasoning abilities have been fully validated in industry-recognized benchmarks like MMLU (Massive Multitask Language Understanding). For instance, even as a lightweight version, GPT's MMLU score reached an astonishing 82.0%, a result that proves its deep accumulation in understanding and reasoning across numerous academic fields. It is this strong "knowledge foundation" that allows it to go beyond literal meanings and accurately capture the correct expressions in the context of specific disciplines.
Therefore, when we see the precise translation of "对分布函数," we know that behind it is the model's deep understanding of professional knowledge. To safeguard this "precision" and "rigor" in academic communication, we believe that choosing GPT is the only correct answer.
2. Chinese Context
We understand that a good tool must not only be powerful but also "understand" the user. In the context of academic literature translation, "understanding" means deeply grasping the reading habits and cultural context of Chinese readers. A seemingly trivial detail can often determine the quality of user experience and even affect the professionalism of the entire product. In this comparison, the difference between GPT and Gemini in handling Chinese authors' names is an excellent example of "seeing the truth in the details."
When we submitted a paper containing the author "Xiaohao Yang" for translation to both models, a surprising detail emerged: GPT almost "intuitively" restored this pinyin name to the Chinese "杨晓浩," while Gemini simply retained the original pinyin. This small difference strikes at the heart of the matter. For any Chinese reader, especially when reading a translation intended to conform to Chinese journal formatting habits, seeing a familiar Chinese name rather than a long string of pinyin makes a significant difference in reading fluency and familiarity. This is not just translation; it is a cultural consideration, a respect for "people."
Why can GPT achieve this? Behind it is its powerful contextual understanding and Named Entity Recognition (NER) capabilities at work. Research shows that GPT-4 can achieve the overall translation quality of a novice human translator and possesses a keen ability to assess translation quality, meaning it is not merely doing mechanical word replacements but understanding the deeper meanings behind the text. When handling proper nouns like names, GPT can more accurately utilize contextual clues for judgment. For example, in a study on name recognition in Russian cultural news, GPT achieved an F1 score of 0.93 through appropriate prompting, demonstrating its outstanding performance in specific languages and entity types. This ability allows it to infer that "Xiaohao Yang" is likely a Chinese author and attempt to find the most matching Chinese character combination in its knowledge base, ultimately successfully "guessing" "杨晓浩." This is a form of intelligence based on probability and context, rather than simple rule matching.
In contrast, Gemini's performance here confirms some issues found in related research. Although Gemini performs excellently in certain NER tasks (such as recognizing context-sensitive names), it often shows inconsistency, mistranslation, or omission when handling proper nouns. Research has pointed out that Gemini lacks accuracy when dealing with proper nouns like names and places, leading to omissions or mistranslations. For instance, when translating classical literature, it might misinterpret a proper noun like "佛國白禪師" into a descriptive phrase. Therefore, Gemini's failure to restore "Xiaohao Yang" to Chinese likely reflects its instability in handling proper nouns and insufficient depth in contextual judgment.
This small difference in name translation is significant for us. It is not just proof of technical superiority but also reflects the "temperature" of the product. A model that "understands" the Chinese context can anticipate users' potential needs—in the Chinese world, we are accustomed to calling names directly. Restoring the pinyin name of a Chinese author to characters is a confirmation of the author's identity and an adaptation to the reading habits of Chinese readers. This kind of "smart" and "considerate" detail can greatly enhance users' immersion and trust in deep reading scenarios.
3. Understanding of Context
In our user feedback, one observation was particularly insightful, accurately pointing out the core stylistic differences between the two mainstream models: "Gemini's characteristic is that it provides an overwhelming amount of information, almost wanting to translate every subscript, which sometimes leads to overly verbose writing. In contrast, GPT's expression is more concise."
This evaluation hits the nail on the head. In academic and literature translation scenarios that pursue efficiency and depth, "conciseness" is not just about beautiful wording; it directly relates to "sense of proportion"—the wisdom of knowing when to elaborate and when to restrain, thereby maximizing information transmission efficiency. When faced with a vast amount of literature, the most valuable thing is time. A translation assistant that understands "sense of proportion" can help you quickly strip away redundant information and get to the core argument, rather than drowning you in all-encompassing details. This concerns not only accuracy but also reading efficiency and cognitive load.
So, where does this "sense of proportion" come from? It stems from the model's deep, holistic understanding of context. Interestingly, although Gemini is known for its ultra-large context window of up to millions of tokens, theoretically able to "see" further, maintaining a consistent style and emotional tone in actual long text translations has become a challenge. Research has pointed out that Gemini may weaken the emotional color of the original text during translation, showing significant variability in stylistic consistency. Sometimes, it even confuses the plot in long narratives, resulting in "style drift."
In contrast, while GPT also has a context window of 128K tokens, it performs better in maintaining emotional tone and stylistic consistency. Multiple studies have shown that GPT's output is emotionally closer to that of human expert translations and resonates more. It can better maintain a consistent narrative voice, being "the most consistent and reliable model" in terms of meaning, sentence structure, and contextual coherence. This ability to produce stable output that is faithful to the original text is an excellent embodiment of "sense of proportion." It understands that good translation is not about piling up information but about selective and focused presentation.
Another aspect can also confirm this difference. We noticed that some users reported that Gemini's safety filters are sometimes overly "sensitive," interrupting translations due to individual vocabulary when handling completely normal academic or historical texts. This also reflects the model's slight shortcomings in understanding real contexts and grasping "sense of proportion"—it sees the "trees" (sensitive words) but fails to understand the whole "forest" (academic context).
In summary, true contextual understanding is not just about processing how long a text can be but how deeply it can comprehend the text's intent, tone, and style, and reproduce it in an appropriate manner. For us explorers navigating the ocean of knowledge, an AI partner with a "sense of proportion" is far more valuable than a "repository" that merely dumps information.
Outlook and Commitment: A New Beginning, A Better Experience
After explaining our difficult yet firm choice in detail, I am now excited to officially announce to everyone: the free translation service integrated with the new GPT engine is currently in the final stages of internal testing and will be fully launched to all users within this week!
This means that the long wait times and unstable translation quality during peak hours that you have long complained about will be greatly alleviated. We understand that every minute of waiting consumes your patience, and every unsatisfactory translation result betrays your trust. This upgrade is aimed at putting an end to all of that.
Making this decision was not easy. Choosing a more expensive option means significant pressure for a team that is still growing. But we repeatedly ask ourselves: What is the purpose of Doclingo's existence? The answer remains the same: to create value for users. We firmly believe that an excellent and reliable user experience is the core and soul of the product and should never be compromised by cost. Therefore, this upgrade is not just a technical iteration but a solemn fulfillment of our commitment to "user first." We are willing to invest more just to gain your undistracted focus and smoothness while reading literature.
Of course, a new beginning requires us to open it together with you. A more powerful engine is just the starting point, and your real experience is the only standard for measuring the value of our work. Therefore, we sincerely invite every user to fully experience, use, and evaluate the new engine once it goes live.
- Are long and complex paragraphs more natural and fluent, maintaining the "sense of proportion" faithful to the original text?
- Have those annoying issues with names and organizations been resolved?
- Is your paper translation more precise and professional?
Please share your real experience with us through the feedback channels within the product. Every like you give us is the greatest encouragement; every criticism is the most valuable driving force for our optimization and iteration. We promise to carefully read and analyze every piece of feedback and incorporate it into our future product roadmap, forming a transparent and efficient feedback loop.
This is not just the end of an upgrade but the beginning of us working together with you to refine a top-notch translation tool. We are confident about the future and look forward to witnessing every progress of Doclingo with you.