AI highlights of the week: NVIDIA showed off the charm of AI at the Computer Show, and Lao Huang hosted a speech "Opening the Future Moment of AI"
In this week, the focus of the AI world has been contracted by NVIDIA, "super chip", AI NPC, AI composing and singing... This "old yellow show" at the Taipei International Computer Show once again tells everyone "the next AI The big moment has now come."
COMPUTEX 2023 - NVIDIA's AI Showcase
Lao Huang releases "super chip" and shows off AI applications
"Super Chip" GH200 is fully put into production
On May 29 (Monday), Nvidia CEO Huang Renxun dressed in an iconic leather jacket and appeared to host the COMPUTEX NVIDIA keynote speech.
The presentation had barely begun when he announced that the GH200 superchip is now in full production , combining the energy-efficient Grace CPU with the high-performance H100 Tensor Core GPU into a single superchip.
Also released at the same time is the DGX GH200 large-memory AI supercomputer that can provide enterprises with super AI performance. Through NVIDIA NVLink, the supercomputer combines up to 256 GH200 superchips into a single data center-sized GPU.
It can provide 1 exaflop performance and 144 TB shared memory, which is nearly 500 times higher than a single A100 320GB system, and significantly improves the efficiency of development tasks such as building large language models and complex algorithms.
Google Cloud, Meta, and Microsoft are among the first companies expected to plug into the DGX GH200 , exploring the supercomputer's capabilities for generative AI workloads.
AI affects the way people play games
In his speech, Huang Renxun also released the NVIDIA Avatar Cloud Engine (ACE) game development version. Developers can use this foundry service to build and deploy custom speech, dialogue and animation AI models.
At the scene, NVIDIA also showed a random dialogue scene with the store manager NPC in the bar of "Cyberpunk 2077" to show the power of this engine: it will give the game NPCs dialogue skills, so that they can use Evolved lifelike personalities answer questions.
NVIDIA ACE Game Development Edition includes AI-based models such as NVIDIA Riva that detect and transcribe player speech. NVIDIA NeMo generates customized responses based on text prompts and animates facial expressions with NVIDIA Omniverse Audio2Face technology.
AI plays "Minecraft" more slippery
New Advances in AI Agent Research
Over the weekend, two teams made new breakthroughs in AI playing Minecraft.
Why the game "My World"? Because **"My World" involves many activities such as survival, exploration and creation, this is a miniature model that simulates the real world. ** Therefore, many teams such as DeepMind and OpenAI are conducting related research on AI agents, hoping to seek answers to the real world through it.
Among them, domestic research teams from Shangtang, Tsinghua University and other institutions launched Ghost in the Minecraft (GITM), a generalist AI agent capable of autonomous learning and solving tasks. This "little devil" not only has a better performance in "Minecraft", but also can spend less on training costs.
In foreign countries, Voyager, an AI agent launched by NVIDIA chief scientist Jim Fan and others, also performed amazingly in "My World". The Voyager connected to GPT-4 not only masters basic survival skills such as digging, building houses, collecting, and hunting, but also continuously explores different places on its own, and even builds portals and uses fences to raise animals... These high-end gameplay, They will too.
How big is the potential of AI agents? We don't know yet, as their quest in Minecraft continues...
Alibaba Cloud Releases AI Assistant for Work and Study "Tongyi Hearing"
Based on the large model of Tongyi Qianwen, it is now in public beta
On June 1 (Thursday), Alibaba Cloud released a new product of the Tongyi family - "Tongyi Tingwu", an AI assistant for work and study.
This product adopts the understanding and summarization technology provided by the Tongyi Qianwen large model, which can provide assistance for users' work and study, and can be applied to scenarios such as conference discussions, teaching training, research interviews, and video viewing.
It can help users efficiently transcribe, retrieve, translate, summarize, and organize audio and video content anytime, anywhere, such as automatically taking notes, organizing interviews, and extracting PPTs.
"Tongyi Listening" has started public beta testing on the official website. During the public testing period, users can receive more than 100 hours of listening and comprehension transcription time for free.
Midjourney update plan: Prompts analyzer will be launched
Help users understand which words are effective
On May 31 (Wednesday), Midjourney held an "Office Hours" user exchange meeting on the official Discord channel, answered users' questions, and introduced their current situation.
In this exchange meeting, the Midjourney development team admitted that the development of v6 has encountered some problems and it will take some time , but we can still receive some new personalities and features before v6, such as a tool that can analyze our Prompts, so that we You can know which prompts are effective and which words are not used in AI drawing.
The new Outpainting (extended redrawing) function is expected to meet users earlier than the Inpainting (partial redrawing) function.
At the COMPUTEX keynote speech, Huang Renxun said: "We are now at a turning point in a new era of computing. Accelerated computing and AI have been adopted by almost all computing and cloud companies in the world." Experienced computer popularization, Internet revolution, smartphones From scratch, we are now ushering in new changes and opportunities.
What will happen in the future and how AI will affect people's daily life is worth imagining, trying, and looking forward to.