Meta AI Introduces Transfusion, a New Method | Tencent Debuts AI Teammates | Daily Brief

🌐 Meta AI Debuts Transfusion, a New Unified Approach

🤖 Amazon's AI Assistant, Amazon Q

🚀 Grok-2 Mini Inference Stack Rewritten

🔧 Meshy 3D Generation Tool New Release

📢 Xinchen AI's Lingo Voice Model Opens for Beta Testing

🤖️ SenseTime's Yuanluobo AI Chess Robot Coming to Market

📈 Boshixiangyuan Closes 130 Million RMB Series A

🌟 Qianxun SI Completes Strategic Funding Round

🎮 Tencent Unveils Voice-Command FPS AI Teammate

Meta AI Debuts Transfusion, a New Unified Approach

Meta AI's newly developed "Transfusion[1]" method integrates language models and image generation models into a single unified AI system capable of processing both text and image data. Experiments show that Transfusion achieves image generation results comparable to established systems like DALL-E 2, while also handling text efficiently. Compared to similar approaches, Transfusion demonstrates superior scaling efficiency with significantly reduced compute, and integrating image data actually improves its text processing capabilities.

The Transfusion method combines the strengths of language models in handling discrete data (such as text) with the ability of diffusion models to generate continuous data (such as images). It employs a single Transformer architecture applicable to all modalities, enabling end-to-end training with different loss functions for text and image data: next-token prediction for text, and diffusion for images.

Amazon's AI Assistant, Amazon Q

Amazon's CEO shared on his LinkedIn the internal results of Amazon Q, estimating that its integration across company systems has saved roughly the equivalent of 4,500 developer-years of work.

Amazon introduced Amazon Q at the AWS re:Invent conference in November 2023 as a new chatbot to help enterprises and employees better leverage AWS. As a GenAI software development assistant, Amazon Q is designed to simplify and accelerate repetitive tasks — analyzing existing code, proposing modifications, implementing changes, updating package dependencies, refactoring outdated and inefficient code, and integrating security best practices.

Grok-2 Mini Inference Stack Rewritten

Developers Igor Babuschkin, Lianmin Zheng, and Saeed Maleki spent three days rewriting Grok-2's inference stack using the SGLang language. Babuschkin committed to continuing to improve Grok-2-mini's processing speed and shared some news about the API, signaling xAI's ongoing push toward high-performance, low-overhead solutions.

The improved Grok-2 scored highly on the LMSYS Chatbot Arena leaderboard, tying with Google's Gemini-1.5 Pro for second place, just behind OpenAI's ChatGPT-4o. Grok-2 mini also rose to 5th place, demonstrating its strong performance.

Meshy 3D Generation Tool New Release

Meshy, the startup led by Ethan Yuanming Hu, has rolled out a new version of its 3D AIGC tool with cleaner, more detailed geometry and a more streamlined workflow, significantly boosting 3D generation capabilities. All users can try out this advanced 3D generation tool for free.

"Meshy-4[2]" represents a breakthrough in 3D AI generation technology. Whether text-to-3D or image-to-3D, it produces models with clean hard surfaces and intricate details. The user interface has been updated, particularly the text-to-3D workflow, which is now split into two discrete steps — modeling and texturing — improving flexibility and control over results. Meshy-4 also expands the model selector, letting users choose from different generation models to match various modeling styles and needs.

Xinchen AI's Lingo Voice Model Opens for Beta Testing

Xinchen AI, invested in by Talking Tom, launched its "Xinchen Lingo" voice large model this August — China's first end-to-end voice large model. Beta registration officially opened on August 24.

Compared to traditional text-to-speech (TTS) technology, Xinchen Lingo integrates speech recognition, natural language processing, intent recognition, dialogue management, and speech synthesis into a single end-to-end model, enabling complete voice-to-voice interaction. The Xinchen Lingo voice model matches GPT-4o's voice capabilities, making it the first domestic model to reach this level.

SenseTime's Yuanluobo AI Chess Robot Coming to Market

SenseTime's home robotics brand "Yuanluobo" is set to release an AI chess robot designed specifically for international chess, featuring 0.5mm precision and stability for reliable vertical pick-and-place operations. Previously, SenseTime had already launched Yuanluobo AI Chess Robot in Xiangqi and Go versions, both equipped with move confirmation, directional keys, and robotic arm functionality.

International chess pieces are three-dimensional with varied shapes, posing challenges for robotic grasping and recognition. The new robot addresses this with 4 degrees of freedom designed around the 3D pieces, simulating human arm movements for precise grasping. The Yuanluobo AI Chess Robot not only accompanies children in learning and playing but also helps develop thinking skills and protect eyesight. It can be used for official chess skill rating assessments, combining traditional chess culture with AI technology.

Boshixiangyuan Closes 130 Million RMB Series A

Boshixiangyuan recently completed a 130 million RMB Series A funding round led by China Fortune-Tech Capital, Beijing 5G Industry Fund, Jinfu, Beihang Investment, and returning investor Langmafeng. The funding will support Boshixiangyuan's market expansion in high-performance cameras and sensors and accelerate its global business deployment.

Founded in 2022, Boshixiangyuan is a supplier focused on core components for high-performance machine vision, primarily engaged in R&D, production, and sales of vision inspection and measurement systems. Its product portfolio includes 3D cameras, smart cameras, DLP projectors, and high-speed cameras, widely used in intelligent industries such as semiconductors, new energy, consumer electronics, and automotive.

Qianxun SI Completes Strategic Funding Round

Qianxun SI recently completed a new strategic funding round, bringing its total valuation to over 16 billion RMB. Investors include Beijing Information Industry Development Investment Fund, Shanghai Shuyu Dingyuan Private Equity Investment Fund Partnership, Beijing Dongcheng District Technology Innovation Industry Investment Fund, and Qingdao Zhonghe Xingyao Venture Capital Fund.

Founded in 2015, Qianxun SI is a spatiotemporal intelligence service provider focused on developing spatiotemporal intelligence infrastructure. Based on BeiDou satellite system positioning data (compatible with GPS, GLONASS, and Galileo), it leverages over 5,000 GNSS satellite-based and ground-based augmentation stations worldwide, self-developed positioning algorithms, and a large-scale internet service platform to provide users with centimeter-level positioning, millimeter-level perception, and nanosecond-level timing services.

Tencent Unveils Voice-Command FPS AI Teammate

At Gamescom 2024, Tencent's Morefun Studio introduced an innovative AI technology called "F.A.C.U.L." — the world's first voice-command FPS AI teammate, developed specifically for Arena Breakout: Infinite, the international PC version of Arena Breakout.

F.A.C.U.L. leverages generative AI to provide advanced voice input and real-time speech synthesis, with environmental recognition capabilities that allow it to understand player voice commands and infer player intent. Unlike traditional hotkeys or command wheels, F.A.C.U.L. lets players direct AI teammates through natural voice commands, supporting multiple simultaneous instructions for a more intuitive gaming experience. The AI teammate can recognize over 10,000 in-game objects, enhancing environmental understanding to enable real-time situational awareness and autonomous decision-making, allowing more flexible responses to player commands.

You ultimately become what you are becoming. Every choice you make comes from the questioning of your life.

— W. Somerset Maugham

Editorial Team

Editor: Yuki

Design: Ivan

For business inquiries, contact on WeChat: Rwkfbcianvd

References

[1] Transfusion: https://the-decoder.com/metas-transfusion-blends-language-models-and-image-generation-into-one-unified-model/

[2] Meshy-4: https://www.meshy.ai/zh/blog/meshy-4-break-grounds