This hands-on WorkBuddy report covers coding, visual reasoning, planning, and design, and records several task inputs, durations, and specific errors. However, it omits model and client configurations, complete execution traces, and a consistent scoring method. It offers insight into the author's experience, but is not enough to reproduce the tests or show that Hy4 generally outperforms other models.
Tasks it can help assess: Refer to its approach to testing experiences with multi-turn development tasks, visual spatial reasoning problems, cross-checking records, business planning, and poster generation; understand performance and errors on the specific tasks in the report.
Tasks it cannot support generalization to: Inferring Hy4's success rate, reliability, or cost advantage across all coding, office, visual, or long-horizon agent tasks, or treating the author's subjective aesthetic judgments as blind-test results.
Model version covered: The author says this was the official Hy4 release; no API model ID, snapshot, or version number is provided.
Test environment or client: WorkBuddy; some steps mention a real Chrome browser, automated simulation, and online deployment. The WorkBuddy version, operating system, model entry point, tool permissions, and deployment platform are not specified.
Reasoning level and parameters: Not specified.
The author says they did not run a standard leaderboard test, but instead used practical work tasks as test cases and observed task decomposition, tool use, problem discovery, and iteration. The article includes these cases:
Tower defense game: Starting with the broad request “make a game, from scratch to ready for release,” development took place over three rounds and about 55 minutes in total. The report describes simulated matches, browser checks, balance adjustments, an architecture refactor, map expansion, and deployment; the detailed requirements for the last two rounds are not fully listed.
Two-robot path planning: Given a map image and explicit rules, the model was asked to provide routes and a timestep-by-timestep schedule, check for vertex and edge-swap conflicts, and prove that a passing space was necessary and that the completion time was earliest possible.
Cross-checking orders and receiving records: The model was asked to compare codes, names, and quantities; distinguish record discrepancies from facts about the physical goods; and determine whether a label roll could be confirmed and whether its quality inspection was complete. The complete original answer links to Tencent Docs, which was not opened during this review.
Business plan presentation: The model was asked to assess the opportunity to start a technology media/integrated marketing studio in 2027, analyze supply and demand, differentiation, strategy, resources, and monetization, deliver a 12-slide presentation, and compare it with Hy3.
Two poster tasks: One asked for a concept poster for “When Machines Start Thinking”; the other asked for the key visual for the “BUILD DIFFERENT 2027” developer conference. Both were compared subjectively with Hy3.
Tower defense: The author reports that the three rounds took 14 minutes 10 seconds, 18 minutes, and 23 minutes 18 seconds. Simulations found that a small number of towers made the game too easy to complete, that stacking technology and upgrades let 38 towers clear hard mode with full health, and that one map was too difficult. The model then changed the difficulty, tower limit, and map parameters. Browser regression checks also found a clipped mobile dialog and map cards crowding the start button. The author says the issues were fixed, regression checks passed, and the game went live. A friend played for about five minutes and said it was “pretty good”; the author also noted that later gameplay devolved into idle play and that automated simulation cannot prove a game is fun.
Robot scheduling: The article presents a set of conflict-free routes and says both robots reached their targets at t=8, matching the theoretical lower bound derived in the article. This result is based on the grid rules specified there; the author explicitly excludes real robot speed, size, turning radius, and safety distance.
Record cross-check: The author reports two exact matches and three items requiring review. The model did not assert that record discrepancies meant the wrong goods had been shipped, nor did it treat pending as having passed quality inspection. However, the original answer referred to three discrepancies as “the other two items” and inferred from “Counted by: J. Lee” that a person had filled out the record; the author says the latter inference lacked image evidence. The count was corrected during editing in Tencent Docs, but no version history that could be checked is provided here.
Business plan presentation: In the order presented in the article, Hy3 took 18 minutes 57 seconds and Hy4 took 20 minutes 31 seconds. The author found Hy4 more specific about competition, profit, resource constraints, and what not to do, but also denser, with smaller text. The financial assumptions and industry data in both versions require manual verification.
Posters: For the first task, the author thought Hy4's concept of “an object misaligned with its shadow” was imaginative, but difficult to understand without the explanation. For the second, the author says Hy4 was nearly two minutes faster and had better detail, while noting that the two outputs were highly similar, their design language was still common, and the divider near the date felt crowded. Both evaluations are the author's subjective visual judgments; no blind-review panel or scoring rubric is reported.
The following figures are reported directly in the article. Without original logs or an independent retest, they should not be treated as third-party-verified results.
| Item | Reported in the article | Nature of evidence |
|---|---|---|
| Three tower defense iterations | 14:10, 18:00, and 23:18; about 55 minutes total | Task durations reported by the author; no execution logs attached |
| Business presentation | Hy3: 18:57; Hy4: 20:31 | Single-generation durations reported by the author |
| Two-robot scheduling | Both robots reached their targets at t=8; the author says this matches the theoretical lower bound | Routes and reasoning summary in the article; the map image and complete proof were not included in this review |
| Record verification | Two matches, three items requiring review; the original answer contained a counting error and an unsupported inference | Author's summary, including disclosed model errors; the complete original answer is in an external document |
The article also repeats scores and percentages from evaluations conducted by Tencent or its partners. These are institutional figures included as background and are not raw results from the author's independent tasks in this article; they should not be conflated as results from the same test.
The article's main value is that it preserves task processes and failure cases instead of giving only aggregate scores: simulation exposed balance problems, browser checks found interaction layout issues, and the record task showed the model making errors in counting and evidence attribution. It also acknowledges that automated completion does not mean a game is fun, business assumptions need verification, and visual creativity requires human judgment.
The evidence has clear limits: the article gives no precise Hy4 or WorkBuddy version, model parameters, device environment, full prompts for each round, number of human interventions, or reviewable step-by-step logs. Requirements for the last two tower defense rounds are incomplete; the Hy3 comparison does not say whether it was run under the same conditions; and the subjective presentation and poster evaluations had no blind review or predefined rubric. The map and design drafts appear as images, while the complete reasoning and original answer to the record task link to Tencent Docs; these attachments were not checked during this review. Therefore, “Hy4 is better at making judgments” should be understood as the author's experience-based conclusion from these examples, not as an overall model ranking.
The author's profile confirms that the publishing account is Hogwarts Testing and Development Society and that its profile associates it with CeTest (Beijing) Technology Co., Ltd.; the original article says it was published with the author's authorization. The page does not identify the individual tester, and no original version elsewhere could be confirmed, so the Tencent Cloud page is used as the sole primary source.
This article organizes evidence from the original report; it is not an independent retest. Reproducing similar conclusions would require the complete images and task inputs, exact Hy4 and Hy3 model and client configurations, logs from runs under the same conditions, records of human intervention, viewable output artifacts, and predefined pass criteria. Reusing only the prompts listed in the article would not reproduce the complete results of the three tower defense rounds, image tasks, or presentation comparison.
Hy4