Namazu AI Adapts Global Models for Japan

Sakana AI has unveiled its Namazu AI series, a suite of prototype models designed to adapt massive open-weight foundation models for specific national requirements, starting with Japan. This development tackles the challenge of integrating global AI advancements into local contexts, ensuring cultural relevance and safety.

The company's research focuses on post-training technology, a crucial step in refining pre-trained models. This process aims to meet the unique cultural, ethical, and security demands of different regions, a necessity as the cost of pre-training models restricts development to a few major players.

Adapting Global Models for Local Needs

The Namazu series, initially released in an alpha version, demonstrates Sakana AI's ability to fine-tune existing frontier models. The primary goal is to preserve the core capabilities of these powerful models while rectifying issues like bias and censorship that can arise from their original training data and development environments.

Sakana AI highlights that overseas models can inadvertently reflect the ideologies or information control tendencies of their origin countries. Their post-training technology seeks to mitigate these effects, making the AI more suitable for Japanese users.

Namazu AI: Performance and Bias Correction

Performance benchmarks indicate that the Namazu models maintain parity with their base models across key areas like reasoning, knowledge, and coding. This suggests that the fine-tuning process does not significantly degrade the model's fundamental abilities.

Crucially, Namazu shows marked improvements in neutrality and factual accuracy, particularly on politically sensitive topics relevant to Japan and other nations. In one test, a base model refused to answer 72% of related questions, while the Namazu-adapted version reduced this refusal rate to near zero, providing objective, multi-faceted responses.

Japanese language capabilities were also evaluated, with the top-tier Namazu model achieving performance comparable to other models of similar scale.

Sakana Chat: Public Access and Web Integration

To bring Namazu AI to users, Sakana AI has launched Sakana Chat, a new conversational service. This platform integrates a web search function, allowing the AI to gather and synthesize real-time information for its responses.

Sakana Chat's beta testing involved around 1,000 users, whose feedback was instrumental in refining the Namazu models. The public release aims to further accelerate improvements based on broader user interaction.

Examples showcase Namazu's ability to compare international AI trends from recent news, explain sensitive topics like government internet censorship objectively, and answer philosophical questions within specified constraints, even in English.

Sakana AI acknowledges the contributions of the broader AI community and the computational resources provided by GMO Internet. The company views this project as proof that large models can be safely adapted for specific markets through appropriate post-training, paving the way for more sophisticated AI solutions beyond chatbots.