Babbitt | Metaverse daily must-read: Adobe releases three Firefly models; OPPO lightweight large model products open early adopters

**Abstract:**OPPO New Xiaobu 1.0 Beta experience based on AndesGPT large model officially opened; Software giant Adobe has released several generative AI innovations, including three new Firefly models designed to produce content that is safe for commercial use. TikTok introduces new features that now support publishing content directly from AI-powered Adobe apps, CapCut, Twitch, and more.

Image source: Generated by Unbounded AI

Hot Information:

OPPO New Xiaobu 1.0 Beta is available and is based on the AndesGPT large model

News on October 11, following the announcement of MediaTek and OPPO to cooperate to build a lightweight large-model end-side deployment solution, the OPPO new small cloth 1.0 beta version based on the AndesGPT large model was officially opened. OPPO officially uses the six words "smarter, more useful" to describe the new Xiaobu assistant, OPPO users wake up Xiaobu and say "Xiaobu big model", and after signing up, you can unlock a variety of scene experiences.

Adobe announces several generative AI innovations, including three new Firefly models and more

On October 10, software giant Adobe unveiled several generative AI innovations at Adobe MAX 2023, including three new Firefly models (Adobe Firefly Image 2 model, Adobe Firefly Vector model, and Adobe Firefly Design model) designed to generate content that is safe for commercial use. In addition, Adobe has released more than 100 important new AI features and updates in its Creative Cloud flagship app, including Adobe Illustrator, Adobe Photoshop, Adobe Lightroom, Adobe Premiere Pro, Adobe After Effects and Adobe Stock. Adobe Express also offers more artificial intelligence capabilities, with new capabilities for generating fills, text-to-templates, and translation.

Google AI releases SAMPO for a video dataset for large model training

It was reported on October 11 that Google AI announced the official launch of the multi-attribute video dataset SANPO. This is a dataset suitable for large model training, consisting of real-world data and synthetic data. The dataset supports a variety of intensive prediction tasks that are challenging for current models, including real and synthetic data with depth maps and video panorama masks, where each pixel is assigned a semantic category label.

Obi Zhongguang set up a new electronic technology subsidiary of 100 million yuan, including a number of AI businesses

Recently, Obi Zhongguang (Guangdong Shunde) Electronic Technology Co., Ltd. was established, the legal representative is Xiao Zhenzhong, the registered capital is 100 million yuan, the industry is computer, communication and other electronic equipment manufacturing, the business scope includes: artificial intelligence basic software development; AI application software development; Artificial intelligence theory and algorithm software development; software development; AI hardware sales, etc. The equity penetration map shows that the company is wholly owned by Obi Zhongguang Technology Group Co., Ltd.

Research: Fine-tuning ChatGPT with synthetic data outperforms GPT-4 on news summary tasks

It was reported on October 11 that Charlie George, a machine learning engineer at Elicit, recently published an article on the LangChain blog to introduce their research results using synthetic data to fine-tune ChatGPT to surpass GPT-4 on news summary tasks.

The results showed that the fine-tuned ChatGPT performed better on the news summary task than GPT-4 zero-sample, almost reaching the level of GPT-4 chain prompts, but its speed was 11 times faster and the cost was reduced by 63%. Compared to zero-shot GPT-4, fine-tuned ChatGPT has a 96% win rate in two-sample evaluations.

Opera Browser introduces an upgraded version of the AI tool Aria

As reported on October 11, Opera Browser launched an upgraded version of the AI tool Aria, which makes it easier for users to interact with AI, optimize queries and create content. It is reported that Aria's new tool "Refiner" has "reuse" and "paraphrasing" functions, and users can improve the AI's response based on previous answer elements. In addition, users can use Aria's Highlight feature to mark elements for post-processing, as well as train AI to mimic their writing style with My Style.

AI startup ElevenLabs has launched a new product, AI Dubbing, that translates any speech into more than 20 languages

On October 10, ElevenLabs, a speech cloning and synthesis startup founded by former Google and Palantir employees, announced AI Dubbing, which can translate any speech, including long-form content, into more than 20 different languages. ElevenLabs claims that the feature can deliver high-quality translated audio (depending on the length of the content) in minutes, while preserving the speaker's original voice, including their mood and intonation.

MediaTek announced that it will cooperate with OPPO to build a lightweight large-model end-side deployment solution

On October 11, MediaTek announced that it will cooperate with OPPO to jointly build a lightweight large-model end-side deployment solution to jointly promote the gradual landing of large-model capabilities on the end-side.

According to reports, MediaTek AI processor APU and AI development platform NeuroPilot have built a complete terminal-side AI and generative AI computing ecosystem, which can accelerate the application development and landing of edge AI computing and strengthen the performance of large language models and generative AI applications. MediaTek said that OPPO large-model AndesGPT uses 4-bit quantization technology to gradually land the large model on the end side and achieve better performance under the effect of accuracy without dropping points.

TikTok now supports publishing directly from AI-powered Adobe apps, CapCut, Twitch, and more

On October 11, TikTok launched new features that allow users to publish videos directly to the platform from a range of popular editing apps, including Adobe's AI video editing software Premiere Pro, AI creative app Adobe Express, and other apps like Twitch, SocialPilot, and CapCut.

Investment and financing:

AI startup TabbyML raised $3.2 million in funding, with participation from Yunqi Capital

TabbyML, founded by two former Google employees, was reported on October 10 to receive $3.2 million in seed funding, with participation from Yunqi Partners and ZooCap. The new funds will be used to develop its open-source code generator. Meng Zhang, the startup's co-founder, says self-hosted coding assistants like TabbyML have the advantage of being highly customizable compared to GitHub's Copilot.

AI startup Cleanlab raised $25 million, led by Menlo Ventures and TQ Ventures

It was reported on October 10 that Cleanlab, a startup co-founded by three PhDs from MIT, announced the completion of a $25 million Series A financing, co-led by Menlo Ventures and TQ Ventures, and Cleanlab was valued at $100 million. Also joining and partnering with Databricks is the round. Earlier this year, Databricks fine-tuned the OpenAI Davinci model provided by the API using Cleanlab and found that the process reduced error rates by 37% and test accuracy from 65% to 78% without adding any resources.

Metaverse Recommendation:

《PS+AI raw picture completed in one step, the effect is amazing! Adobe Firefly 2 Blockbuster Update: Fully upgraded models, perfect vector support"

Adobe's AIGC biograph platform, Firefly, recently received a major update. Adobe named it Firefly 2 directly. Adobe claims that Firefly 2 has greatly improved the quality of image output, especially in areas such as leaves, skin texture, and facial features, with more realistic and fine details than the previous generation model.

"When the world's largest image library, decided to "embrace AI"

A wave of the wave of the AI era, falling on the head of any company, is it either a tsunami or a "splash of wealth"?

As the world's largest photo agency, Getty Images, which focuses on "real", once banned AI-generated content from entering the image library. But recently, Getty launched its own AI generation tool, Generative AI by Getty Images.

What caused Getty to have a 180-degree change in attitude? As a content trading platform, is it possible for Getty to become another giant in the AI era?

All of these questions, Craig Peters, CEO of Getty Images, explained at a recent code conference how Getty's AIGC tools help creators in copyright and revenue, and the thinking behind it.