News

The Week in AI: NVIDIA's Supercomputer, Google's Cyber Defense Initiative, More

• By John K. Waters
• 02/16/2024

This edition of our weekly roundup of AI product-and-services news includes Google's new Cyber Defense Initiative, OpenAI's Sora launch, Bret Taylor's new AI startup, Anthropic's defense against election misinformation, and the first public preview of NVIDIA's Eos supercomputer.

NVIDIA Unveils "Eos" Supercomputer for GenAI
On Thursday, GPU maker NVIDIA released a video providing the first public look at Eos, its latest data-center-scale supercomputer. Already ranked as the ninth-fastest supercomputer in the world, Eos is an extremely large-scale NVIDIA DGX SuperPOD (the company's accelerated AI data center infrastructure platform). It comprises a cluster of 576 DGX H100 servers, each one featuring eight H100 GPUs for a total of 4,608 H100s linked with the company's Quantum-2 InfiniBand architecture. And it's tuned specifically for generative AI.

OpenAI Launches Sora Text-to-Video Model
This week, OpenAI introduced Sora, its new text-to-video and image-to-video generative AI model, and made it available to a select group of red teamers for assessment of critical areas of potential harms and risks, the company said. The prompt-based video-creation tool was designed to generate videos of up to one minute in length. It can also animate still images and turn them into videos, make existing videos longer, and/or fill in the gaps when a video skips. (See our earlier reporting on this announcement here.)

Google AI Takes on the "Defender's Dilemma"
Google launched a new initiative this week aimed at reversing what is known in cyber security circles as the "Defender's Dilemma." The dilemma, which states that breaches are inevitable because defenders must be right 100% of the time, while attackers have to be right just once, has been plaguing even the adept cyber security pros for decades. The newly announced Google Cyber Defense Initiative will be releasing new tools for defenders, and launching new research and AI security training. "These commitments are designed to help AI secure, empower and advance our collective digital future," the company said in a blog post.

The initiative includes a new "AI for Cybersecurity" cohort of 17 startups from the UK, US and EU under the Google for Startups Growth Academy’s AI for Cybersecurity Program.

The search engine giant also published a new report, "Secure, Empower, Advance: How AI Can Reverse the Defender's Dilemma," in which it explores AI security use cases and provides a "Roadmap to Digital Security."

Former Salesforce and Google Execs Launch AI Startup
Bret Taylor, Chairman of the OpenAI board and former co-CEO of Salesforce, and Clay Bavor, who spent 19 years at Google and led Google Labs, have launched a new AI startup: Sierra. The company's namesake offering is a conversational AI platform for businesses. The platform is designed to enable an organization to build an AI agent that is personalized to an individual business and its customers. "Enable your customers to get answers, solve problems, and take action through a natural, conversational experience," the company's website states.

The startup actually came out of stealth mode last month. At the time, Bloomberg reported that Sequoia Capital was expected to lead an $85 million investment in a deal that would value the company at nearly $1 billion. The company has already begun listing some marquee clients, including Weight Watchers, Sonos, and Sirius XM Holdings.

Anthropic Buttresses Claude with New Election Policies
Leading AI startup Anthropic announced that it is testing  a new set of policies for its AI-powered chatbot, Claude, which the company hopes will prevent the technology's use to spread misinformation during this election season. Dubbed Prompt Shield, the technology relies on a combination of AI detection models and rules to show a pop-up if a U.S.-based Claude user asks for voting information. The pop-up offers to redirect the user to TurboVote, a resource from the nonpartisan organization Democracy Works, where they can find up-to-date, accurate voting information. The company already has an Acceptable Use Policy (AUP) that aims to prohibit the use of its tools in political campaigning and lobbying—in other words, the company does not allow candidates to use Claude to build chatbots that can pretend to be them, and it doesn't allow anyone to use Claude for targeted political campaigns.

In company says it will be "trialing" an approach in which it uses its classifier and rules engine to identify election-related queries and redirect users to accurate, up-to-date authoritative voting information.

"While generative AI systems have a broad range of positive uses, our own research has shown that they can still be prone to hallucinations, where they produce incorrect information in response to some prompts," the company said in a blog post. "Our model is not trained frequently enough to provide real-time information about specific elections. For this reason, we proactively guide users away from our systems when they ask questions on topics where hallucinations would be unacceptable, such as election-related queries."

Cohere for AI Debuts Massively Multilingual GenAI LLM
Cohere for AI, a non-profit research lab focused on machine learning, launched a large language model (LLM) covering more than 100 languages this week. The new state-of-the-art, open-source, massively multilingual, generative LLM covers 101 different languages, which the organization says is more than double the number of languages covered by existing open-source models. Dubbed Aya, the model was created to help researchers "unlock the powerful potential of LLMs for dozens of languages and cultures largely ignored by most advanced models on the market today," the lab said in a blog post.

The non-profit group announced that it is open sourcing both the Aya model and its large,  multilingual instruction fine-tuned dataset. This data collection includes rare annotations from native and fluent speakers from around the world, ensuring that "AI technology can effectively serve a broad global audience that have had limited access to-date."

"Aya is part of a paradigm shift in how the ML community approaches massively multilingual AI research, representing not just technical progress, but also a change in how, where, and by whom research is done," the blog post stated.

Cohere for AI is supported by Cohere, whose Retrieval Augmented Generation (RAG) toolkit allows LLMs to accurately answer questions and solve tasks using enterprise data as the source of truth.