AI Daily Report: Unveiling the Generative Models Behind "Apple Intelligence"

06/14 2024 484

The underlying generative models behind Apple Intelligence have been released, with the 3B model outperforming Gemma-7B and the server model rivaling GPT-3.5-Turbo.

Microsoft Copilot GPTs will be discontinued on July 10, and any GPTs created by users will be cleared. Microsoft's explanation is "strategic adjustment," likely due to the inability to turn a profit.

What other hot topics in the AI industry, both domestic and international, are worth paying attention to over the past day? Let Raven Jun take you through them.

/ 01 / Large Models

1) The models behind "Apple Intelligence" are revealed: The 3B model outperforms Gemma-7B, and the server model rivals GPT-3.5-Turbo

Apple mentioned in an updated blog post that Apple Intelligence is composed of various highly intelligent generative models. The blog details two models: a device-side language model with approximately 3 billion parameters and a larger server-based language model that runs on Apple servers through private cloud computing.

2) Microsoft Copilot GPTs will be discontinued next month! Launched only 3 months ago, but cut due to inability to turn a profit

Microsoft announced that Copilot GPTs will be discontinued on July 10, and any GPTs created by users will be cleared. Microsoft's official explanation is a strategic adjustment - shifting the focus of GPT towards commercial and enterprise scenarios. The reason behind this may be a lack of profitability or inability to compete with OpenAI.

3) Luma AI launches a heavyweight text-to-video model that can generate 120 frames in 120 seconds

Luma AI's newly released text-to-video model, Dream Machine, is now freely available and can generate high-quality videos comparable to OpenAI's Sora. The model supports physical simulations and can generate 120 frames of video in 120 seconds, equivalent to 5 seconds of smooth animation, ensuring video authenticity and consistency.

The official also pointed out some current issues with the model. For example, cars may appear distorted during perspective switches, and dogs' movements may not use their paws correctly. Luma AI promises to continuously optimize the model.

4) Stable diffusion 3 officially opens source: The powerful text-to-image model SD3-M debuts

SD3-M is a powerful text-to-image model with 2 billion parameters, efficient inference speed, and excellent generation results. Stability AI has open-sourced the SD3-M weights, providing users with a free trial opportunity. The model uses the MMDiT architecture, achieving significant improvements in image quality, layout, and text prompt understanding. Users can experience the SD3-M generation effect through an online demo, but it is currently only for academic research, and commercial needs should contact Stability AI. Open-sourcing SD3-M provides users with opportunities to explore the application potential of text-to-image models.

5) The Chinese version of LMSYS is here! ByteDance's coze.cn builds a big model arena: Anonymous PK effect, users as judges

ByteDance's coze.cn has launched a "Model Square" to organize a competition among domestic large models. Through user participation, two models are anonymously matched, and scores are given based on the performance of the generated content. This feature is similar to the authoritative foreign large model arena Chatbot Arena.

6) Suno officially launches an audio input feature, allowing users to create songs with any sound

Suno's new feature allows users to create songs from any sound, available to Pro and Premium users. Users can capture inspiration anytime, anywhere, and transform sounds from daily life into beautiful musical works.

/ 02 / AI Applications

1) Apple's market capitalization regains global first place, striving to introduce Apple Intelligence to China

On June 12, when US stocks opened, Apple's share price continued to rise, increasing by 2.86%, with a market capitalization exceeding $3.2 trillion, surpassing Microsoft and NVIDIA to regain the top global market capitalization.

In an interview, Apple's Software Engineering Chief Federighi talked about the Chinese market, stating that Apple is working to bring "Apple Intelligence" to all customers and is advancing this process. Regarding the future development of Apple Intelligence, he emphasized that online models remain important as they can provide users with more information.

Apple Intelligence will begin testing in the US this fall, and according to MacRumors, users interested in testing Apple Intelligence will need to enter a waiting list.

2) ByteDance responds to developing an AI phone: It's just a proposal for manufacturers to refer to

Recently, there were rumors that ByteDance had secretly launched a project to develop an AI phone two months ago. Regarding this information, ByteDance responded to a First Financial journalist, stating that the information is inaccurate. In fact, they are exploring large model software solutions based on mobile phones, providing them to mobile phone manufacturers for reference. Currently, they have no plans to develop and sell their own phones.

/ 03 / Investment and Financing Intelligence

1) OpenAI's latest financial performance revealed: Annual revenue of $3.4 billion

It is reported that OpenAI CEO Sam Altman told employees that the company's annual revenue reached $3.4 billion in the past six months or so. In comparison, OpenAI's annual revenue was only $1.6 billion at the end of 2023 and approximately $1 billion last summer.

2) Hong Kong's $62 billion investment fund supports AI unicorn SmartMore and plans to build the first Hong Kong AI research institute

Hong Kong Investment Management Limited, a fund organization established by the Hong Kong Special Administrative Region government with a capital of HK$62 billion (US$8 billion), announced the signing of a strategic cooperation agreement with AI unicorn SmartMore Group. In his speech, Chief Executive of the Hong Kong Special Administrative Region Carrie Lam revealed plans to establish Hong Kong's first artificial intelligence research institute.

3) Astrocade AI announces the completion of a $12 million seed round of financing

Astrocade AI announced the completion of a $12 million seed round of financing. Founded in 2022, the company's main product is an UGC AI gaming platform called Astrocade, where users can create their own casual games through natural language and share them with other players on the platform.

/ 04 / AI Infrastructure

1) MimicBrush: Upload a reference image to achieve partial style redrawing of the original image

MimicBrush is a zero-reference image editing technology proposed by a research team at the University of Hong Kong. It achieves image editing through self-supervised learning without requiring users to accurately describe the editing effect. Its innovation lies in automatically understanding reference images, improving editing accuracy and efficiency.

2) Guangdong introduces 45 major AI policies, aiming to produce 100 million AI phones

Guangdong issued the "Several Measures of Guangdong Province on Empowering Thousands of Industries with Artificial Intelligence." By 2027, it aims to create over 100 large-scale intelligent terminal products in eight categories, including mobile phones, computers, home appliances, and robots. The production of AI phones is expected to reach over 100 million units, and the scale of the AI core industry will exceed 440 billion yuan.

3) Andrew Ng opens source an AI agent machine translation project

Andrew Ng's latest open-source AI agent machine translation project, Translation Agent, utilizes a reflective agent workflow and LLM technology to provide highly customized translation services, allowing users to flexibly set tone, regional characteristics, and a glossary of professional terms.

4) Uizard releases Autodesigner 2.0 AI design engine

Uizard has released the Autodesigner 2.0 AI design engine, which combines proprietary models, Anthropic AI and OpenAI technologies, as well as Stability AI's image generation technology. This simplifies the UI design process, improving design efficiency and innovation.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.