What is ELYZA? Exploring the Japanese Large Language Model: Features, Capabilities, and How to Use It

The world is currently experiencing a surge in generative AI, spearheaded by OpenAI's ChatGPT. Japan is no exception, with many companies investing heavily in this technology.

In this article, we'll introduce "ELYZA," a Japanese Large Language Model (LLM) gaining attention, covering its overview and key features.

What is ELYZA?

ELYZA is the name of a generative AI model and tool, and also the name of a startup company originating from the University of Tokyo.

ELYZA (the company) is a noteworthy startup in the AI field, stemming from Professor Yutaka Matsuo's laboratory, known as a pioneer in AI research.

Recently, news of its acquisition by KDDI Group, a major telecommunications company, has been reported, promising further growth with enhanced resources.

As an LLM, ELYZA is based on the Japanese language, making it particularly user-friendly and capable of natural conversations for Japanese speakers.

What is an LLM (Large Language Model)?

LLM stands for Large Language Model, referring to the underlying mechanism of generative AI itself.

It's a system that builds the "brain" of generative AI based on massive amounts of text data, as seen in the popular ChatGPT.

ELYZA, a pioneer in Japanese LLMs, has undergone several version updates. Let's explore some of its previous models.

ELYZA's Model Lineup

While the latest model is the "ELYZA-japanese-Llama-2-70b," boasting 70 billion parameters comparable to global models, ELYZA has previously released models such as:

• ELYZA-japanese-Llama-2-13b

• ELYZA-japanese-Llama-2-13b-instruct

• ELYZA-japanese-Llama-2-13b-fast / ELYZA-japanese-Llama-2-13b-fast-instruct

Now, let's compare the performance of the latest model, "ELYZA-japanese-Llama-2-70b," with ChatGPT.

ELYZA vs. ChatGPT: A Performance Comparison

Let's compare the performance of ELYZA with ChatGPT, one of the world's leading generative AIs (as of March 2024).

Basic User Interface (UI)

First, let's look at the basic UI.

Reflecting their origins, ELYZA (Japanese company) features a Japanese UI, while OpenAI (foreign company) uses an English UI.

ELYZA

ChatGPT

However, GPT also supports text input in Japanese (prompts), so once you get used to the basic operations, the difference isn't significant.

Parameter Count

Next is the parameter count.

The parameter count is an indicator of the AI model's size. According to OpenAI's research, it directly impacts AI performance.

• ELYZA: 70 billion

• GPT: Up to 200 billion

Currently, there's a gap, but ELYZA is expected to gradually increase its parameter count, potentially narrowing the difference in the medium to long term.

Multimodality

Finally, let's consider multimodality.

Multimodality refers to the breadth of data types an AI model can handle.

• ELYZA: Text only

• GPT: Text + Image Data + Audio Data (including paid plans)

Based on these comparisons, GPT currently appears to be a more powerful AI model.

However, considering the fundamental difference in base languages (Japanese vs. English), ELYZA may excel in "ease of use for Japanese speakers."

What Can You Do with ELYZA?

For those familiar with ChatGPT, here are three examples of what you can do with ELYZA.

Everyday Conversations via Text

First, you can engage in everyday conversations via text.

As shown in the image, you can have in-depth conversations beyond simple Q&A.

However, the data used to train AI is typically from the past, so it cannot answer real-time questions (e.g., "What's the weather in Tokyo today?").

Translation

Second, ELYZA can translate between different languages.

It can handle common English ↔ Japanese translations, as well as translations between other languages.

Consultation Based on Specific Conditions

Third, you can use ELYZA for consultations by inputting specific conditions.

In this example, we pretended to be a café owner near Shibuya Station and asked for advice.

ELYZA responded with the following:

The image doesn't show the full response, but ELYZA provided a total of five strategies and specific measures.

By asking follow-up questions, you can have a discussion with ELYZA, demonstrating how AI can provide answers for simple consultations.

ELYZA has wide-ranging applications from simple conversations to complex topics, but let's also examine its weaknesses.

ELYZA's Weaknesses

This article will highlight three limitations.

Communicating Subtle Emotional Nuances

First, ELYZA struggles with subtle emotional nuances.

This isn't unique to ELYZA, as generative AI generally infers answers based on input data and outputs them as "answers."

Therefore, it's highly unlikely to understand implied meanings or unspoken expectations.

Note: This is the current situation (as of March 2024), and it may change in the future.

If you want to consult about something involving your feelings, be sure to verbalize and express them in the text you provide to the AI.

This may lead to some helpful insights.

Processing Based on the Latest Information

Second, ELYZA is not good at processing based on the latest information.

As mentioned earlier, the data used for AI training is from the past, so asking "What's the weather today?" won't yield the desired response.

Processing Requiring Specialized Knowledge and Expertise

Third, ELYZA struggles with processing requiring specialized knowledge and expertise.

A good example is medical-related information.

In the image, the user describes a situation and asks ELYZA for the best course of action, but the response is fairly generic.

While this provides some information, the key is to decide which of these options is most appropriate and take action.

This decision-making process requires experience and expertise, which current AI cannot provide.

However, generative AI excels at listing options and thinking logically. ELYZA, with its user-friendly Japanese language base, offers a wide range of potential applications.

Now that we've covered the basics, let's briefly explain how to use ELYZA.

How to Use ELYZA

It only takes three steps:

Step 1: Search for "ELYZA for JP"

First, search for "ELYZA for JP" using your preferred search engine.

Step 2: Click "ELYZA for JP (Demo Version)"

The search results will likely include a URL like the image below. Click it.

Step 3: Enter Text Information

After clicking, the following screen will appear. Simply enter your text in the "Ask the AI Assistant" section at the bottom.

There are currently no usage limits, so feel free to experiment.

Important Considerations When Using ELYZA

Finally, here are three things to keep in mind when using ELYZA.

The Information May Not Always Be Accurate

First, the information provided may not always be accurate.

As mentioned previously, ELYZA's answers are based on the data it has been trained on. This means it may struggle with information that requires up-to-date knowledge or highly specialized questions, and may even provide incorrect answers.

Therefore, it's important not to blindly accept everything ELYZA says. It's recommended to use it in conjunction with search engines when you need to know unfamiliar information.

Effective Use Requires Strong Questioning and Communication Skills

Second, effective use requires strong questioning and communication skills.

This also echoes a previous point, but AI (including ELYZA) still struggles with subtle nuances and emotions.

To get the answers you want, you need to carefully consider what information you should provide to ELYZA.

While you can initially experiment with it as a game, it's crucial to keep this in mind when using it for professional purposes.

It's Still Difficult to Completely Replace Humans

Third, it's still difficult to completely replace humans.

Due to its struggles with providing correct context and understanding emotional nuances, completely replacing humans is currently difficult with existing technology.

Conversely, this highlights the value of human intervention. Emotional intelligence will become increasingly important in future careers and work styles.

Conclusion

This article provided an overview of the Japanese LLM "ELYZA" and a brief explanation of how to use it.

Hopefully, this article has been helpful in promoting the use of generative AI, including ELYZA.