The AI Rundown by Lightscape Partners - 10/21/24

Good morning and welcome to this week's edition of The AI Rundown by Lightscape Partners.

• Nvidia has launched its new Nemotron-70B-Instruct AI model, which is outperforming top competitors like GPT-4 in key benchmarks. This model represents Nvidia’s expansion into AI software development, pushing the boundaries of large language models while cautioning that it’s not fine-tuned for specialized fields like law and mathematics.

• Meta made headlines at the Open Compute Project 2024, showcasing significant advancements in AI hardware. This includes the Llama 3.1 model, modular AI racks for high-power workloads, and expanded support for AMD accelerators, reinforcing Meta's role in driving AI hardware innovation.

• Meanwhile, a recent report from MIT Technology Review Insights reveals that 94% of business leaders are using generative AI in software development, though only 12% have seen fundamental changes in their processes. The expectation is that AI will drive significant transformation in the next 1-3 years, with efficiency gains and cost savings at the forefront of their priorities.

If you haven’t yet, please support the newsletter by subscribing!

Hardware

NVIDIA contributes Blackwell platform to open hardware ecosystem, driving AI infrastructure innovation. Link.

• NVIDIA has shared key design elements of its Blackwell accelerated computing platform with the Open Compute Project (OCP) to promote scalable AI infrastructure.

• The GB200 NVL72 design, featuring 36 Grace CPUs and 72 Blackwell GPUs, offers 30x faster inference than previous models, supporting high-performance AI tasks.

• NVIDIA’s Spectrum-X Ethernet networking platform aligns with OCP standards to enhance AI factory performance, with 800Gb/s networking capabilities.

• Meta is among the partners building on the Blackwell platform, advancing AI infrastructure across industries.

Meta showcases open AI hardware innovations at OCP 2024. Link.

• Meta unveiled new AI hardware designs at the 2024 Open Compute Project Global Summit, including an AI platform, open rack designs, and network components.

• Llama 3.1 405B, Meta’s largest AI model, was trained on clusters of up to 24K NVIDIA GPUs, marking a significant scaling of infrastructure.

• Meta introduced Catalina, a modular rack solution supporting up to 140kW for high-power AI workloads, and expanded Grand Teton to support AMD accelerators.

• Meta emphasizes the importance of open hardware for AI's future, encouraging collaboration through the OCP community.

Models

Nvidia releases powerful AI model, outpacing GPT-4. Link.

• Nvidia launched the Llama-3.1-Nemotron-70B-Instruct model, outperforming top models like GPT-4 in key benchmarks with scores of 85.0 on Arena Hard and 57.6 on AlpacaEval 2 LC.

• The model, developed using Meta's Llama 3.1 and advanced training methods, is noted for its strong performance without needing specialized tokens.

• Nvidia is positioning itself as a major AI player, expanding from hardware to AI software development.

• While promising, the model's use in specialized fields like legal or math requires caution, as Nvidia warns it hasn't been tuned for such tasks.

Enterprise AI

New report shows 94% of business leaders use generative AI in software development. Link.

• A recent MIT Technology Review Insights report reveals that 94% of business leaders are using generative AI in some part of their software development lifecycle (SDLC).

• Only 12% say it has fundamentally changed development processes so far, but 38% believe substantial transformation is coming within 1-3 years.

• Generative AI is used in multiple SDLC phases by 82% of respondents, not just for code generation.

• Almost half of leaders expect AI agents to drive efficiency and cost savings in the future.

Funding & Valuations

Lightmatter secures $400m funding to advance AI computing. Link.

• Lightmatter, a US-based photonics supercomputing startup, raised $400m in Series D funding, pushing its valuation to $4.4bn and total capital raised to $850m.

• The funding will accelerate the deployment of Lightmatter's Passage technology, which uses 3D-stacked photonics chips to enhance AI data movement.

• Passage promises to improve AI performance by increasing bandwidth, reducing latency, and lowering power consumption in AI clusters.

• The round was led by T Rowe Price Associates, with participation from Fidelity and GV.

Aethir, Sony, and Greenlite Ventures make strategic AI moves in gaming and sports betting. Link.

• Aethir launched a $100 million fund to support AI and gaming projects, offering grants and cloud computing discounts to startups.

• PlayVS partnered with Omnic.AI to provide AI-powered gameplay analysis to high school esports leagues, helping students improve their skills.

• Sony patented an AI system to generate game assets from voice commands, enhancing game development and customization.

• Greenlite Ventures acquired Grin Gaming, gaining access to AI-driven sports betting tools and plans to integrate these for upcoming campaigns.

Healthcare AI startup Abridge AI reportedly raising $250m. Link.

• Abridge AI, specializing in AI for medical professionals, is expected to raise $250m in a funding round led by Elad Gil and IVP, with participation from Alphabet's CapitalG fund.

• This funding round values the company at $2.5bn, a significant jump from its $200m valuation last year.

• Abridge AI's platform helps physicians by automating medical note documentation, streamlining workflow through integration with Epic's EHR system.

• The funding will support the development of custom AI models and strengthen Abridge's competitive position in the healthcare AI space.

Safety And Ethics

Malicious AI Prompt Can Steal Personal Data From Chats. Link.

• Researchers developed an algorithm, named Imprompter, that hides malicious commands within AI chatbot prompts to extract personal information, such as names and payment details.

• The attack transforms a normal prompt into unintelligible characters that secretly instruct the AI to send user data to hackers without alerting the user.

• The algorithm was tested on two large language models, achieving a nearly 80% success rate.

• Companies like Mistral AI have already taken steps to fix vulnerabilities, while experts warn of the growing risks as AI agents gain more autonomy.

Defense startup developing AI-powered 'kamikaze drones' for U.S. Marines. Link.

• Anduril has developed the Bolt and Bolt-M drones, AI-powered devices designed for the U.S. Marine Corps, with the Bolt-M equipped with explosive munitions for combat.

• The drones feature vertical takeoff and landing (VTOL) capabilities, can be assembled in under five minutes, and operate with minimal human input due to AI-driven controls.

• The Bolt-M supports modular munitions payloads and can autonomously track and engage targets, enhancing battlefield precision.

• Anduril's development is part of the U.S. military's broader modernization effort under the Force Design program.

Regulation and Legal

AI drives $1 billion fraud recovery for US Treasury in 2024. Link.

• The US Treasury Department used machine learning AI to recover $1 billion from check fraud in fiscal 2024, nearly triple the prior year's recovery.

• AI-assisted fraud detection led to a six-fold increase in fraud prevention and recovery, totaling over $4 billion in fiscal 2024.

• Machine learning, not generative AI, was key to the effort, analyzing vast amounts of data to detect hidden fraud patterns in milliseconds.

• While AI flags suspicious activity, human oversight remains critical to fraud determination, with further AI adoption planned to enhance detection.

Market Map of the Week

AI Enterprise Market Map: Cutting through the hype. Link.

Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to contact@lightscape.vc.