[Collection of large model daily reports on October 27] Zhipu AI launches the third generation of large base models; OpenAI forms a team to study “catastrophic” artificial intelligence risks, including nuclear threats; DeepMind: Who said convolutional networks are inferior to ViT? ; Boston Dynamics’ robot dog is equipped with a ChatGPT brain and acts as a tour guide, speaking with an old London accent.

Zhipu AI launches third-generation base model

Link: https://news.miracleplus.com/share_link/11223

On October 27, 2023, Zhipu AI launched the fully self-developed third-generation base model ChatGLM3 and related series of products at the 2023 China Computer Conference (CNCC).

OpenAI forms team to study ‘catastrophic’ AI risks, including nuclear threats

Link: https://news.miracleplus.com/share_link/11224

OpenAI announced today that it has formed a new team to evaluate, evaluate, and explore artificial intelligence models to protect against so-called “catastrophic risks.” The team, called Preparedness, will be led by Aleksander Madry, director of MIT’s Center for Deployable Machine Learning. (Madrid joined OpenAI in May as “Director of Preparedness,” according to LinkedIn.) A primary responsibility of Preparedness will be to track, predict, and guard against the dangers of future AI systems, from their ability to persuade and deceive humans. capabilities (such as phishing attacks) to malicious code generation capabilities.

DeepMind: Who said convolutional networks are inferior to ViT?

Link: https://news.miracleplus.com/share_link/11225

The early success of deep learning can be attributed to the development of convolutional neural networks (ConvNets). ConvNets have dominated computer vision benchmarks for nearly a decade. In recent years, however, they have been increasingly replaced by ViTs (Vision Transformers). Many people believe that ConvNets perform well on small or medium-sized data sets, but cannot compete with ViTs on larger network-sized data sets. At the same time, the CV community has moved from evaluating the performance of randomly initialized networks on specific datasets such as ImageNet to evaluating the performance of pretrained networks on large general datasets collected from the network. This raises an important question: Do Vision Transformers outperform pre-trained ConvNets architectures under similar computational budgets? In this article, researchers from Google DeepMind explored this issue by pre-training multiple NFNet models on the JFT-4B data set (a large labeled image data set used to train the basic model) of different scales. This results in a performance similar to ViTs on ImageNet.

Boston Dynamics’ robot dog is equipped with a ChatGPT brain and acts as a tour guide, speaking with an old London accent

Link: https://news.miracleplus.com/share_link/11226

In a newly released video, Boston Dynamics shows the results of integrating a robot dog with LLM: “Mr. Spot” wears a top hat, a mustache, big eyes and a British accent, and is showing people around the company’s facilities. In order to enable Spot to “speak,” Boston Dynamics used OpenAI’s ChatGPT API and some open source LLMs for training, then equipped the robot with speakers and added text-to-speech conversion capabilities. So you can see that while making sounds, Spot keeps opening its “mouth” and looks like it’s really talking. LLM’s so-called “emergent behavior” enables it to perform tasks outside of its intended training. Because of this, they can be adapted to a variety of applications. The Boston Dynamics team’s exploration of this began this summer with some proof-of-concept demonstrations of using LLM in robotics applications, and the ideas were expanded upon during an internal hackathon. In particular, they were interested in Spot’s demonstration of using LLM as an autonomous tool, with the team inspired by LLM’s apparent ability to role-play, replicate culture and nuance, shape plans and maintain coherence over time, as well as the recently released VQA Models (these models can add captions to images and answer simple questions about the images).

Microsoft has been betting on technology for more than 40 years, and this time it’s AI’s turn!

Link: https://news.miracleplus.com/share_link/11227

Matthew Ball is the former head of strategy for Amazon Prime Video and is currently a senior consultant at McKinsey and KKR. As the major Silicon Valley technology giants released their financial reports this week, Matthew published a paper called “Parallel Bets, Microsoft, and AI Strategies” In the article, Matthew reviewed in detail the competition between software and hardware in three different eras of PC, Mobile and AI, and elaborated on today’s AI market structure and strategic layout from Microsoft’s perspective.

The powerful combination of RLHF and AlphaGo core technologies, UW/Meta brings text generation capabilities to a new level

Link: https://news.miracleplus.com/share_link/11228

In a latest study, researchers from UW and Meta proposed a new decoding algorithm that applies the Monte-Carlo Tree Search (MCTS) algorithm used by AlphaGo to the proximal strategy optimized (Proximal Policy Optimization, PPO) trained RLHF language model, greatly improving the quality of text generated by the model.

The United Nations establishes an AI advisory body to provide support for the international community to strengthen governance

Link: https://news.miracleplus.com/share_link/11229

IT House reported on October 27 that on October 26, local time, United Nations Secretary-General Antonio Guterres announced the establishment of a new artificial intelligence (AI) advisory body to provide support for the international community to strengthen the governance of AI. Guterres said that current AI has transformative potential that is difficult to grasp. If AI is used maliciously, it may lead to the destruction of people’s trust in various institutions and the weakening of social cohesion. “For these reasons, I issue a call for a multi-stakeholder, multi-disciplinary global dialogue on the governance of artificial intelligence to maximize its benefits to humanity (all of humanity), and contain and reduce risks.”

Luminar Neo brings generative AI tools to amateur photographers

Link: https://news.miracleplus.com/share_link/11230

Adobe is making generative AI models available to professional photographers and designers, while another company is working to bring easy-to-use generative AI tools to a wider range of photographers, including hobbyists. Starting today, Luminar Neo, Skylum’s (formerly Macphun) photo editing software, is rolling out a suite of generative artificial intelligence features to its Mac and Windows desktop apps, allowing users to remove unwanted objects from images, expand the canvas, or replace and Add specific elements to your photos.

After helping sift through 400 million photos, GoodOnes renames itself Ollie

Link: https://news.miracleplus.com/share_link/11231

GoodOnes, an artificial intelligence photo classification application, received US$3.6 million in seed funding earlier this year and launched on the Apple App Store in April this year. After sorting through 400 million photos over the course of six months, it will be renamed Ollie and undergo a relaunch.