開発スタッフがアプリ関連の最新技術やニュースをご紹介します。

Meta社の新型ヘッドセット、Meta Quest 3 が昨日(2023/10/10)発売開始されました！

Quest 1やQuest 2 はVR(Virtual Reality 仮想現実)ヘッドセットでしたが
Quest 3ではMR(Mixed Reality 複合現実)ヘッドセットとして登場です。

Apple社から来年(2024年)発売予定の Vision Pro もMRヘッドセットとしての使用がアピールされていたので、今後はVRをさらに発展させたMRが主流となっていきそうです。

VRの仮想空間では現実世界と切り離された仮想空間の中でゲームをプレイしている人は
その他の人(ゲームをプレイしている人を横から見ている人)と現実では同じ空間にいても
隔離されて孤立してしまうという問題がありました。

VR体験を同じ部屋にいる他の人といっしょに共有できないという点がVRの弱点かなと思っていましたので、その弱点が今回のMR対応によって解消されたように思います。

MRではパススルーカメラで見えている現実世界の上に仮想のオブジェクトを配置して遊ぶことができ、
それをリアルタイムで複数のヘッドセットと共有することができます。同じ部屋に集まって複数人で
遊ぶ際などQuest 3が人数分用意できれば、みんなで同じ空間を共有しながら遊ぶことができます。

以下のYouTubeで、複数人で同じ部屋でMRのボードゲームやシューティングゲームをする様子が紹介されています。
https://www.youtube.com/shorts/r9txAdERckY

これは今までにない楽しい体験ができそうです。

木曜日担当：nishida

Tweet

Edgeブラウザ上などで利用できるBing ChatにDALL-E 3が搭載されたようです。

もともDALL-E 2を利用した画像生成は可能でした。
DALL-E 3に変わったことでプロンプトを細かく設定することができ、
より写実的な画像を出力可能になったそうです。

Bing Chatの入力欄に「〜の絵を描いてください」と入力することで
イラストを出力してくれました。

出力された画像の例を見る限りではありますが、
画像生成AIでよく発生する指の破綻が少ないように思えます。

水曜担当：Tanaka

Tweet

Today, I would like to share about 3 free alternatives to GPT-4. Let’s take a look.

LlaMA 2

LlaMA 2 is a cutting-edge open-source large language model developed by Meta AI. It’s available for commercial use and comes with pre-trained models, fine-tuned models, and code resources. You can find all these assets on HuggingFace. You can also get a feel for how well the model performs by trying it out on HuggingChat. By making LlaMA 2 openly accessible, Meta AI is empowering researchers and developers to create innovative applications that leverage advanced language capabilities.

PaLM 2

Google AI’s PaLM 2 is Google’s latest large language model, excelling in advanced reasoning tasks like coding, mathematics, classification, question answering, translation, multilingual proficiency, and natural language generation. It surpasses previous state-of-the-art large language models like the original PaLM due to its optimized compute-scaling approach, improved dataset mixture, and architectural enhancements. You can access PaLM 2 for free using Bard. While there’s still room for improvement compared to GPT-4 in terms of quality and performance, it offers impressive capabilities.

Claude 2

Claude 2 represents the latest version of Anthropic’s conversational AI assistant. It offers improved performance, longer responses, and can be accessed through an API as well as a new public beta website, claude.ai. Developers at Anthropic have focused on enhancing its capabilities in areas such as coding, mathematics, and logical reasoning compared to earlier Claude versions. For instance, Claude 2 recently achieved a score of 76.5% on the multiple-choice section of the Bar exam, a significant improvement from Claude 1.3’s 73.0%. You can access various Claude models on Poe and experience their performance firsthand.

Conclusion

Despite GPT-4 being unavailable to the public, there are other promising open-source large language models emerging as alternatives that are accessible to everyone. While these models may not be as massive as GPT-4, they show that cutting-edge language AI is evolving rapidly and becoming more widely available. These freely accessible models excel in fields such as mathematics, coding, and logical reasoning, making them suitable replacements for various applications.

Asahi

Tweet

Amazon has officially launched Bedrock, a generative AI service that provides access to various AI models, both from Amazon and third-party partners, via an API.

Bedrock empowers AWS customers to build applications using these generative AI models and customize them with their own data. Brands and developers can create AI agents for tasks like travel booking, inventory management, and insurance claims processing. In the near future, Bedrock will incorporate Llama 2, an open-source large language model from Meta, along with models from AI21 Labs, Anthropic, Cohere, and Stability AI.

Amazon claims that Bedrock will be the first fully managed generative AI service to offer Llama 2, including its 13-billion- and 70-billion-parameter versions. While similar to Google’s Vertex AI, Amazon argues that Bedrock’s advantage lies in its seamless integration with existing AWS services, like AWS PrivateLink, ensuring secure connections to a company’s virtual private cloud. This perceived advantage may vary depending on the specific customer and their cloud infrastructure.

Swami Sivasubramanian, VP of data and AI at AWS, highlighted the increasing interest in generative AI, driven by data availability, scalable computing, and machine learning advancements. He emphasized that Bedrock’s launch democratizes generative AI, making it accessible to businesses of all sizes and their employees, from developers to data analysts.

Additionally, Amazon introduced its Titan Embeddings model, a first-party model that converts text into numerical representations known as embeddings, supporting around 25 languages and text segments of up to 8,192 tokens, aligning with OpenAI’s latest embeddings model. These announcements reflect Amazon’s commitment to the rapidly growing generative AI market following its initial challenges with Bedrock’s availability.

Yuuma

Tweet

ChatGPTに音声による会話、画像の認識の機能が追加されると、OpenAI公式ブログで発表がありました。
https://openai.com/blog/chatgpt-can-now-see-hear-and-speak

Tweet