Nvidia Introduces AI Workbench Toolkit for Simplified Generative AI Model Tuning and Deployment

By

-

August 9, 2023

nvidia ai workbench for generative AI — Image Credits; wccftech

A unified, user-friendly toolkit called NVIDIA AI Workbench, which the company just unveiled, enables developers to swiftly build, test, and customize pre-trained generative AI models on a workstation or PC before scaling them to almost any data center, public cloud, or NVIDIA DGX Cloud.

With the aid of AI Workbench, starting an enterprise AI project is no longer difficult. Developers can use a streamlined interface running on a local system to access models from well-known sources like Hugging Face, GitHub, and NVIDIA NGC and modify them using unique data. The models can then be simply shared between various other platforms.

Manuvir Das, vice president of enterprise computing at NVIDIA said, “Enterprises around the world are racing to find the right infrastructure and build generative AI models and applications. NVIDIA AI Workbench offers a streamlined path for cross-organizational teams to develop the AI-based applications that are increasingly crucial in modern business.”

Although there are now hundreds of thousands of pretrained models accessible, customizing them using the many open-source tools may require searching through numerous internet repositories for the appropriate framework, tools, and containers as well as using the appropriate skills to customize a model for a particular use case.

Developers may quickly customize and execute generative AI with NVIDIA AI Workbench. As a result, they are able to compile into a single developer toolkit all essential enterprise-grade models, frameworks, software development kits, and libraries from open-source sources and the NVIDIA AI platform.

Leading providers of AI infrastructure, such as Dell Technologies, Hewlett Packard Enterprise, HP, Lambda, Lenovo, and Supermicro, are embracing AI Workbench for its capacity to enhance their most recent lineup of multi-GPU capable desktop workstations, high-end mobile workstations, and virtual workstations.

Nvidia, Hugging Face to Make Generative AI Supercomputing Available to Developers

By

Sahil Pawar

-

August 9, 2023

Nvidia Hugging Face generative AI supercomputing available developers — Image Credits; Nvidia

Hugging Face and Nvidia are working together to increase access to AI compute. Nvidia announced this week that it will offer a new Hugging Face service called Training Cluster as a Service to make the development of fresh and unique generative AI models for the workplace simpler. The announcement was timed to coincide with the annual SIGGRAPH conference.

The all-encompassing cloud-based AI “supercomputer” from Nvidia, DGX Cloud, will power Training Cluster as a Service when it launches in the upcoming months. The DGX Cloud offers access to a cloud instance with eight Nvidia H100 or A100 GPUs, 640GB of GPU memory, Nvidia’s AI Enterprise software for building big language models and AI applications, as well as consultations with Nvidia experts.

Companies can sign up for DGX Cloud on their own. The monthly price per instance starts at $36,999. However, Training Cluster as a Service combines the DGX Cloud infrastructure with the Hugging Face platform’s more than 250,000 models and over 50,000 datasets, making it a useful starting point for any AI project.

Hugging Face co-founder and CEO Clément Delangue said, “Our collaboration will bring Nvidia’s most advanced AI supercomputing to Hugging Face to enable companies to take their AI destiny into their own hands with open source to help the open source community easily access the software and speed they need to contribute to what’s coming next.”

The partnership between Hugging Face and Nvidia comes as the business is apparently seeking new funding at a $4 billion valuation. Hugging Face, which was founded in 2014 by Delangue, Julien Chaumond, and Thomas Wolf, has grown quickly over the previous nine years, transitioning from a consumer app to a hub for all things AI model-related.

Hugging Face is becoming the go-to place for AI developers to exchange ideas these days. Hugging Face has evolved into the GitHub equivalent for developers looking to learn more about the most recent models and APIs in order to avoid being rendered obsolete by the generative AI technology, as it gains popularity.

Nvidia Introduces New AI Chip GH200 Grace Hopper

By

Sahil Pawar

-

August 9, 2023

nvidia-grace-hopper-superchip — Image Credits: Nvidia

On Tuesday, Nvidia unveiled GH200 Grace Hopper superchip, which is made to run AI models, as it tries to fend off AMD, Google, and Amazon competitors in the budding AI market.

The 72-core Grace CPU and 141 GB of HBM3e memory, which is organized into six 24 GB stacks and has a 6,144-bit memory interface, are the foundation of the new GH200 Grace Hopper superchip. Although Nvidia installs 144 GB of memory physically, only 141 GB are usable for higher yields.

The California-based company said that the HBM3e processor, which is 50% faster than current HBM3 technology, will power its next-generation GH200 Grace Hopper platform.

Developers will be able to run expanded Large Language Models (LLMs) as a result of its dual configuration, which will provide up to 3.5x more memory capacity and 3x more bandwidth than the currently available chips on the market.

CEO Jensen Huang stated at a presentation on Tuesday that the new technology would help “scale-out of the world’s data centers.” In addition, he predicted that “the inference cost of large language models will drop significantly,” alluding to the generative phase of AI computing that comes after LLM training.

The latest product launch by Nvidia comes after the hype surrounding AI technology propelled the company’s value above $1 trillion in May. Nvidia became one of the market’s brightest stars in 2023 due to soaring demand for its GPU chips and a forecasted shift in data center infrastructure.

According to estimations, Nvidia currently holds a market share of over 80% for AI chips. Graphics processing units, or GPUs, are the company’s area of expertise. These processors are now the one of choices for the large AI models that support generative AI applications, including Google’s Bard and OpenAI’s ChatGPT.

Despite all this, Nvidia’s chips are hard to come by as tech behemoths, cloud service providers, and startups compete for GPU power to create their own AI models.

OpenAI Introduces New Web Crawler GPTBot to Consume more Open Web

By

Sahil Pawar

-

August 8, 2023

To increase its dataset for training its upcoming generation of AI systems, OpenAI has introduced a new web crawling bot called GPTBot. According to OpenAI, the web crawler will gather information from websites that are freely accessible to the public while avoiding content that is paywalled, sensitive, or illegal.

However, the system is opt-out. GPTBot will presume available information is open for use by default, similar to other search engines like Google, Bing, and Yandex. The owner of a website must include a “disallow” rule in a common server file in order to stop the OpenAI web crawler from digesting that webpage.

Additionally, according to OpenAI, GPTBot will check scrapped material in advance to weed out personally identifiable information (PII) and anything that contravenes its rules. However, some technological ethicists believe that the opt-out strategy still poses consent-related concerns.

Some commenters on Hacker News defended OpenAI’s action by arguing that it needs to amass as much information as possible if people want to have a powerful generative AI tool in the future. Another person who was more concerned with privacy complained that “OpenAI isn’t even quoting in moderation. It obscures the original by creating a derivative work without citing it.”

The launch of GPTBot comes in response to recent criticism of OpenAI for previously illegally collecting data to train Large Language Models (LLMs) like ChatGPT. The business changed its privacy policy in April to address these issues.

Meanwhile, a recent GPT-5 trademark filing appears to hint that OpenAI might be working on its next version of the GPT AI model. Large-scale web scraping would probably be used by the new system to refresh and increase its training data. However, there is no official announcement concerning GPT-5 as of yet.

IIM Lucknow Introduces Executive Programme in AI for Business

By

Sahil Pawar

-

August 8, 2023

An executive programme in AI for Business has been launched in partnership between the Indian Institute of Management (IIM) Lucknow and Imarticus Learning. It seeks to provide graduates with at least five years of relevant work experience with the knowledge and skill sets required for AI and machine learning.

Classes for this executive curriculum will be offered on weekends, either Sunday or Saturday, and will last for six months. It will be entirely online. The training will conclude with a three-day campus immersion.

The start date for the course is October 1. The course fee is Rs. 2.35 lakh + GST (Including the registration price of Rs. 47,000 + GST). There are 50 seats for students in total.

The goal of this programme is to give ambitious professionals a solid foundation in AI. The programme offers a pedagogy that mixes project-based learning with the case-based methodology used by the IIM and focuses on real-world business outcomes.

This will aid in developing abilities including teamwork, critical thinking, and problem solving. The curriculum also provides a chance to network with influential businesspeople and subject matter experts. There are eight modules in the curriculum.

Candidates must have earned a bachelor’s or master’s degree in computer science, engineering, mathematics, statistics, economics, or another related field with a minimum of 50% on their final exam to be eligible for the programme.

An offer letter will be given to shortlisted candidates. Candidates will obtain a certificate from IIM Lucknow once the course is finished. Candidates who meet the requirements can apply here.

Zoom’s Updated Terms Say It can now use Customer Data to Train AI

By

Sahil Pawar

-

August 8, 2023

Zoom’s updated terms can now use customer data train AI — Image Credits: Zoom

The most recent revisions to the terms of service state that Zoom intends to use some of customer data to train its artificial intelligence models. If you read through the conditions on software licensing, beta services, and compliance in the most recent revision to the video platform’s terms of service, the small print appears to indicate a significant choice Zoom made regarding its AI strategy.

The modification, which became effective on July 27, gives Zoom the ability to use specific consumer data for developing and fine-tuning its AI or machine learning models. Customer information on product usage, telemetry and diagnostic data, and other comparable material or data gathered by the company are all examples of the “service-generated data” Zoom may now employ to train its AI.

In accordance with the terms of Zoom, “You consent to Zoom’s access, use, collection, creation, modification, distribution, processing, sharing, maintenance, and storage of Service Generated Data for any purpose, to the extent and in the manner permitted by applicable Law, including for the purpose of machine learning or artificial intelligence (including for the purpose of training and tuning of algorithms and models).”

Messages, files, and documents from customers do not appear to fall under this category. The company Zoom stated in a subsequent blog post that “for AI, we do not use audio, video, or chat content for training our models without customer consent.”
The upgrade comes amid a heated public discussion over how much personal data, no matter how aggregated or anonymized, should be used to train AI systems. A large portion of online text or photos are used to train chatbots like OpenAI’s ChatGPT, Google’s Bard, and Microsoft’s Bing as well as image-generation programmes like Midjourney and Stable Diffusion. Recent months have seen a rise in legal actions brought by authors or creatives who claim to see their own work reflected in the results of generative AI tools.

Alibaba Open Sources Two LLM Models

By

Suchitra Shenoy

-

August 8, 2023

Alibaba Cloud open sources its LLM Models

Alibaba Cloud, the digital technology backbone of the Chinese tech giant, Alibaba Group Holding, has open-sourced two of its large language models (LLMs). With this move, Alibaba intends to expand its influence in the generative AI field.

The two open-source models, Qwen-7B and Qwen-7B-Chat, are smaller versions of Tongyi Qianwen, which is Alibaba’s largest language model. Roughly translated to “seeking truth by asking a thousand questions,” Tongyi Qianwen is the LLM launched by Alibaba’s cloud computing service unit in April.

Both open-source models have each been trained on 7 billion parameters. Qwen-7B-Chat is a fine-tuned version of Qwen-7B and can conduct human-like conversations.

As per the company’s statement, the models’ internal mechanisms, including the codes and documentation, will be made freely accessible to scholars, researchers, and commercial institutions worldwide. They can access it through Alibaba Cloud’s AI model repository ModelScope, and the US collaborative AI platform Hugging Face.

This development comes after Meta released its open-source LLM—Llama 2—with Microsoft on July 16.

While companies with fewer than 100 million monthly active users can deploy the open-source models for commercial use free of charge, those with more users will have to request a license from Alibaba Cloud. This is similar to Meta’s Llama 2, which requires a license from companies with more than 700 million users.

Set to be spun off from its parent company next year to become a publicly listed company, Alibaba Cloud has been doubling down on generative AI development and commercialization amid the global frenzy around ChatGPT.

Zhou Jingren, chief technology officer of Alibaba Cloud Intelligence, said, “We aim to promote inclusive technologies and enable more developers and small and medium-sized enterprises to reap the benefits of generative AI.”

IIT Kharagpur Introduces a 3-Month Certification Course on AI in Partnership with TCS iON through AI4ICPS

By

Tejaswini Kasture

-

August 8, 2023

IIT Kharagpur Introduced AI Course in Partnership with TCS iON — Image Source: anmnewsenglish

In celebration of the third anniversary of the National Education Policy 2020 (NEP 2020), the AI for Interdisciplinary Cyber Physical Systems (AI4ICPS), a national AI Hub, established by the Indian Institute of Technology (IIT) Kharagpur under the guidance of the Indian Government’s Department of Science and Technology (DST), has unveiled the initiation of the “Hands-on AI for the real-world applications (HAAI)” program.

IIT Kharagpur’s director VK Tewari inaugurated the course in the presence of dignitaries, including Venguswamy Ramaswamy, global head of TCS iON, as well as Amit Patra, deputy director, and Pabitra Mitra, project director of AI4ICPS.

The certification course is scheduled to begin on September 2 and conclude on December 2, with a course fee of ₹ 4,998. The program will comprise 80% practical learning, emphasizing hands-on experiential learning, along with 20% theoretical learning through live lectures.

This three months duration program consists of weekend sessions. It is conducted online and led by faculty members from IIT Kharagpur’s department of electronics, mathematics, computer science, and Electrical, as well as scientist from TCS.

The program also features real-world applications presented by industry experts and IIT Kharagpur’s researchers, offering hands-on practical learning, as mentioned by the Institute.

With the aim of educating 1,00,000 individuals and contributing to national development, this certification program seeks to democratize AI education for the skill development of both youth and professionals.

Amazon CEO Andy Jassy Says Every Single Team Working on Generative AI

By

Sahil Pawar

-

August 7, 2023

Amazon CEO Andy Jassy says every single team working generative AI — Image Credits: Bloomberg

Amazon CEO Andy Jassy underscored the company’s large investment in artificial intelligence during the company’s Q2 2023 earnings call on Thursday. Jassy disclosed that many generative AI projects are now being worked on by “every single one” of Amazon’s divisions.

He emphasized the crucial significance of AI throughout the entire organization while talking about the AWS infrastructure and services that can enable AI applications.

Jassy addressed generative artificial intelligence applications on the call, and he went on to explain that the firm provides infrastructure and services via AWS that can support many of these applications. However, he also emphasized how crucial AI is to the company as a whole.

Jassy also made a suggestion on possible AI-based upgrades for Alexa that might be shown during the forthcoming Amazon products event on September 20. Jassy discussed the creation of a better large language model (LLM) expressly for Alexa on a recent earnings call.

AWS notably highlighted in its official blog how it has been working to increase everyone’s access to and understanding of artificial intelligence and machine learning. Amazon is paying millions of dollars to build the technology because it believes that it can have a large positive impact on both people and businesses.

AWS has made investments over the years to democratize access to and understanding of generative AI, machine learning, and AI in general. Amazon Web Services (AWS) has been at the forefront of advancing knowledge about the role that these innovations can play in the lives of both people and organizations through announcement about the establishment of a $100 million Generative AI Innovation Centre programme.

Research Discovers Acoustic Attack that Steals Data from Keystrokes with 95% Accuracy

By

Sahil Pawar

-

August 7, 2023

acoustic attack steals data from keystrokes 95% accuracy — Image Credits: depositphotos

By listening to what you type on the keyboard, a deep learning model can collect private information including usernames, passwords, and messages. According to the paper, The sound-recognition system can catch and analyze keystrokes captured from a microphone with 95% accuracy after being trained by a group of academics from British universities.

When the model was evaluated with the well-known video conferencing services Zoom and Skype, the accuracy fell to 93% and 91.7%. The technique clarifies how deep learning might be used to create fresh forms of malware that can listen to keyboard input and steal data such as credit card numbers, messages, conversations, and other private information.

Sound-based attacks are more viable than other strategies, thanks to recent advances in machine learning and the availability of inexpensive, high-quality microphones on the market. Other strategies are frequently constrained by variables like data transfer speed and distance.

The researchers recorded the sound made by 36 keys on a MacBook Pro being pressed 25 times apiece to gather data for the sound-recognition system. Using an iPhone 13 mini, the audio was recorded when it was 17 centimeters from the PC.

Waveforms and spectrograms that identified each key were generated from the recordings. An image classifier dubbed “CoAtNet” was then trained using the distinctive sounds of each button to determine which key on the keyboard was pushed.

The method does not absolutely need access to the device microphone, though. Threat actors can also join a Zoom session as a participant to hear what users are typing by listening to the keystrokes. Users can protect themselves against such attacks, according to the research article, by altering their typing habits or employing complicated random passwords. The model can potentially be rendered less precise by using white noise or software that simulates keyboard sounds.

It is quite unlikely that upgrading to silent switches on a mechanical keyboard or fully moving to membrane keyboards will assist since the model was extremely accurate on keyboards used by Apple on laptops in the last two years, which are typically silent. Currently, implementing biometric identification methods like a fingerprint scanner, facial recognition, or an iris scanner is the best approach to counter such sound-based attackers.