[Experiment Report] Completing Everything on Your Home PC! Running "Hayaoshi Encyclopedia" to Generate Quizzes from Wikipedia
While "having AI create quizzes" is easy, "building a system for real-time competition" is actually a task with quite high difficulty. In this experiment, to overcome this wall, I built "Hayaoshi Encyclopedia," which completes the AI execution entirely on my own PC.
- All inference and correctness judgments are completed on your own PC via a browser app running a local LLM (Ollama).
- Question text is displayed one character at a time, supporting "buzzer" responses via the Space key or buttons.
- When the correct answer is announced, the AI automatically displays the cited Wikipedia article and link used as evidence.
- Difficulty levels (Easy to Super Hard) and range specifications by category or era are possible.
- Implemented a mechanism to play online with distant friends just by sending a participation URL.
![[Experiment Report] Completing Everything on Your Home PC! Running "Hayaoshi Encyclopedia" to Generate Quizzes from Wikipedia](/img/hayaoshi-part1.png)
What kind of tool is "Hayaoshi Encyclopedia"?
"Hayaoshi Encyclopedia" is a browser app where a local AI (artificial intelligence running on your own PC) creates and presents quizzes using Japanese Wikipedia articles as sources. This is a tool I developed for myself with the help of AI (Claude).
Up to 4 people can participate simultaneously, and answers support free text input. A notable point is that everything from quiz generation to correctness judgment is completed on the host's PC. The major point of this experiment is that it runs using only the resources at hand, rather than calling external APIs.
- Tool built by the operator using AI (Claude) as an assistant
- Utilizes Wikipedia's vast knowledge as a quiz database
- AI judges the correctness of answers and also presents the cited sentences used as evidence
Deliberate Operability and Customizability
I didn't just want to present quizzes; I pursued the experience of a "buzzer quiz." The question text is displayed one character at a time, allowing users to react with the Space key or a button the instant they know the answer. This sets up an environment where viewers can compete in real-time.
Additionally, to improve playability, I included the following adjustment functions. Difficulty can be selected from "Easy" to "Super Hard," which fluctuates the Wikipedia article size, view counts, and the level of detail in hints. Furthermore, it is possible to freely narrow down the question range by specifying Wikipedia category names or eras such as "1990s."
- Buzzer Support: Displays question text character by character to compete for reaction speed
- Difficulty Adjustment: Fluctuates the detail of hints based on article scale and view counts
- Range Specification: Narrow down by category or era (e.g., Showa, 1990s, etc.)
- Remote Play: Play online with distant friends just by sending a participation URL
Measurement Environment and Operation Status
In this experiment, operation was confirmed in the following environment. The model used was "gemma4:12b-it-q8_0," and a PC equipped with an "RTX 4080" GPU was used.
When running local AI, PC specs determine the comfort of operation. With this configuration, I confirmed that the process of the AI generating quizzes and judging correctness operated smoothly without any stress.
- Model used: gemma4:12b-it-q8_0
- GPU used: RTX 4080
- Operation Point: Realizing real-time competition while managing the host's PC load
Summary and Next Preview
I discovered that by utilizing local LLMs (large language models running on your own PC), it is possible to build your own quiz app without worrying about privacy or costs. It is a very interesting discovery that such a sophisticated interactive experience can be created by running AI on your own PC.
In the next part (Part 2), I will explain in detail the tricks used to prevent "mistakes (hallucinations)" that inevitably get included in AI-generated quizzes using a three-step approach.
- With local AI, you can freely build unique mechanisms
- Possible to have free ways of play by leveraging your own PC's specs
- Next time, we will dive deep into how to prevent "AI lies"
- Those who want to try running AI on their own PC
- Those interested in building interesting tools using AI
- Those who want to know local AI specs and measurement results
- Local LLM
- Refers to a large language model that is downloaded and run inside your own PC without using cloud services (such as ChatGPT).
- Ollama
- A tool for easily running LLMs (large language models) on your own PC.
- Hallucination
- A phenomenon where AI tells plausible lies.
FAQ
Can I run it even if my PC specs are low?
According to official information, the required specs vary depending on the model size and the degree of quantization. While I used an RTX 4080 in this experiment, it is possible to operate on lower-spec PCs by selecting a lighter model.
I built "Hayaoshi Encyclopedia," which generates quizzes from Wikipedia on a home PC, and achieved real-time buzzer competition for up to 4 people by combining local AI (Ollama) and a GPU (RTX 4080). It supports difficulty and range customization as well as remote URL sharing, demonstrating the possibilities of free tool construction unique to local environments.