Llama 3 Unveiled, Most Business Leaders Unprepared for GenAI Security, Mona Lisa Rapping …

The AI Weekly Breakthrough | Issue 7 | April 23, 2024

Welcome to The AI Weekly Breakthrough, a roundup of the news, technologies, and companies changing the way we work and live

Mona Lisa Rapping: Microsoft’s VASA-1 Animates Art

Researchers at Microsoft have developed VASA-1, an AI that produces realistic talking face videos from a single static image and audio clip. While impressively realistic, the technology, which can animate photos or artwork with life-like facial expressions synchronized with audio, still shows signs of being AI-generated on closer inspection. Demonstrations include a rapping Mona Lisa and other figures delivering speeches or songs. VASA-1, not yet released for public use, could be used to create lifelike avatars that emulate human conversational behaviors, but the team is cautious of misuse.

Llama 3: “The Most Capable Openly Available LLM to Date”/h2>

Meta has unveiled Llama 3, the latest upgrade to its open-source LLM. Featuring two models with 8B and 70B parameters, Llama 3 offers enhanced reasoning, code generation and instruction-following capabilities and, if you take Meta’s PR as fact, is “the most capable openly available LLM to date.” Llama 3 was pre-trained on over 15 trillion tokens collected from publicly available sources, which is 7 times larger than the one used to train Llama 2. Meta has also integrated its AI chatbot, powered by Llama 3, into the search bars of its major apps (Facebook, Messenger, Instagram, and WhatsApp) across multiple countries.

AI Examined in Stanford’s 2024 AI Index Report

Stanford just released its 2024 AI Index Report. Published by the Institute for Human-Centered Artificial Intelligence (HAI), the exhaustive 500-page report examines key trends related to AI, including technical advancements in AI, public perceptions of AI, and the geopolitical landscape surrounding the development of AI. This year’s report, the 7th edition, includes roundups of AI’s impact on the economy, education, policy and governance, as well as an expanded chapter on responsible AI, and new chapters focusing on AI in science and medicine.

GenAI Survey: 80% of Leaders Concerned about Data Privacy and Security

Most U.S. business leaders are concerned about data privacy and security when it comes to using Generative AI. A survey of 300 US GenAI strategy or data analytics decision makers, sponsored by SAS, shows that despite enthusiasm for GenAI’s potential to boost productivity, 80% of business leaders are concerned about data privacy and security issues, 10% feel prepared for regulatory compliance, and 93% lack a comprehensive governance framework for GenAI. The survey identified gaps in preparedness in four key areas of implementation:

1. Increasing trust in data usage and achieving compliance
2. Integrating GenAI into existing systems and processes
3. Finding enough GenAI talent
4. Predicting costs

The study underscores the need for identifying high-value use cases to deliver sustainable and scalable solutions.

Adobe Announces New AI Tools For Premiere Pro

Adobe has launched new generative AI tools in Premiere Pro to simplify video editing by enabling users to modify scenes and extend clips. These features are part of Adobe’s Firefly suite, which integrates across its Creative Cloud. The update also supports third-party AI models, enhancing content creation capabilities. Additionally, new audio workflows will be available in May to enhance sound quality. Adobe is committed to responsible AI use and provides transparency on the AI models used through the Content Authenticity Initiative.

GPT-4 Can Interpret, Exploit Security Advisories

New research indicates that GPT-4 can autonomously exploit real-world security vulnerabilities when presented with CVE advisories (publicly disclosed computer security flaws). This capability surpasses several other LLMs and open-source vulnerability scanners in testing. While the paper acknowledges a limited dataset of 15 CVE advisories, GPT-4 showed proficiency in exploiting 87% of them, a stark contrast to the 0% success rate of competing models and scanners evaluated.

Augment Yourself 🤖

• Explore this definitive guide to understanding and tackling the challenges of AI hallucinations.
• Maximize your RAG system’s effectiveness with these proven data enrichment strategies.
• Shield Your RAG system from these four unstructured data risks.
• Find out how to seamlessly link structured and unstructured data with a knowledge graph.
• Read new studies testing AI against data analysts and resident physicians.

🔥 For more AI News brought to you via email subscribe to our newsletter here.
👀 Want to know more about Shelf’s suite of AI solutions? Check out our website here.