MOSS, a ChatGPT-like program, goes open source
Shanghai based Fudan University’s MOSS, a ChatGPT-like AI bot, has become open source in an effort to establish a ‘China path’ for the use of creative AI tools. According to Qiu Xipeng, the director of MOSS’s developers and a researcher at the School of Computer Science at Fudan University, organisations and programmers can now access MOSS’s codes on GitHub for free, starting mid-April.
With significant backers like Microsoft, ChatGPT and its creator OpenAI have access to extensive financial and data resources. Chinese developers, including the tech giant Baidu, have scarce means in comparison. The majority of data is separate and not shared, leading businesses to perform a lot of redundant work.
Compared to ChatGPT, MOSS has fewer parameters and interactive statistics. However, when it has access to external knowledge bases or learns to look for or use tools in particular areas, its capabilities will be improved, according to experts.
On February 20, just after the pilot test for MOSS went live, the website was inundated with fervent visitors. According to Qiu, MOSS will be “in line” with moral principles. “We hope to promote this ecology with an open source (generative language) model to avoid duplicated works,” Qiu told a Beyond ChatGPT Forum organised by the School of Management, Fudan University. “On the model, you can focus on respective segment development, such as e-commerce, education and medication.”
Painting tool Tux Paint adds 15 new magic tools
A new version of the cross-platform, open source, free digital painting and drawing app Tux Paint for kids is now accessible. About ten months after Tux Paint 0.9.28, Tux Paint 0.9.29 is here with fifteen new Magic Tools, such as Maze for creating random mazes, Googly Eyes for adding eyeballs to your creations, Fur for adding fur to a drawing, and Circles and Rays for adding brush stroke effects.
The following new Magic Tools are part of this release: Saturate and Desaturate, a tool to increase and decrease colour saturation; Double Vision, a tool to simulate diplopia; and 3D Glasses, a tool to create 3D images that can be viewed with those red/cyan anaglyph glasses. Remove Color and Keep Color let you fully desaturate sections of an image that match a chosen colour.
Transformers are revolutionising machine learning (ML) capabilities and ushering in a new age of generative AI. However, how do data scientists create models that completely utilise the power of transformers? The open source Kubeflow initiative aims to help with that.
Since the release of Kubeflow
1.6 in September 2022, the widely used open source MLOps platform hasn’t received an update until now, when Kubeflow 1.7 became broadly available. At its core, Kubeflow is an open source machine learning toolkit that enables businesses to install and manage ML workflows on Kubernetes infrastructure built for the cloud. The Kubeflow 1.7 update includes, among its themes, an emphasis on improving support for transformer based models.
Model makers must learn how to efficiently use resources as they transition to transformer based models. Workload placement and auto scaling can be aided by Kubeflow 1.7, which can help to optimise processes and use fewer resources.
The addition of ‘Parallelfor’ statements, which allows developers to more effectively use parallel processes across AI accelerator hardware, benefits the Kubeflow Pipelines component in particular in the 1.7 update. Another advantage of Kubeflow 1.7 is its ability to be integrated with an increasing number of cloud native technologies that can facilitate the deployment of MLOps workflows.