Former General Joins OpenAI Board, Pope Weighs in on Techno-Human Condition, NVIDIA's Synthetic Data Pipeline ... and more

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

Former NSA Head Joins OpenAI Board and Safety Committee

Retired General Paul Nakasone , formerly head of the NSA, is joining OpenAI's board and safety committee. This appointment likely aims to address concerns about OpenAI's rapid development of AI models and potential safety risks. Nakasone's expertise and experience in cybersecurity are seen as valuable assets, but also raise questions about OpenAI's data practices and its safety committee's effectiveness.

Pope to G7: AI Is ‘Neither Objective nor Neutral’

In an address to the G7 summit, Pope Francis discussed the threat and promise of AI, the ‘techno-human condition’, human vs algorithmic decision-making, AI-written essays, and the necessity of political collaboration on technology.

NVIDIA Releases Open Synthetic Data Generation Pipeline for Training LLMs

NVIDIA announced Nemotron-4 340B , a family of open models that developers can use to generate synthetic data for training LLMs for commercial applications. High-quality training data plays a critical role in the performance, accuracy and quality of responses from a custom LLM, but robust datasets can be prohibitively expensive and difficult to access.

Samsung's Integrated AI Chip Production

Samsung Electronics announced a one-stop shop for AI chip production , integrating their memory chip, foundry, and chip packaging services. This integrated approach reduces AI chip production time by 20%. Samsung, uniquely positioned with its combined services, foresees global chip industry revenue hitting $778 billion by 2028. The company stressed its advanced chip technologies aiming for improved chip performance and reduced power consumption.

Picsart and Getty Partner for Responsible AI Imagery

Picsart and Getty Images have partnered to launch a model for responsibly-crafted, commercially-safe AI-generated imagery. This custom model, trained exclusively on Getty Images' licensed content, simplifies content creation for marketers and small businesses and enables users to generate unique, commercially licensed images directly within the platform.

Tesla Shareholders Sue Musk for Starting Competing AI Company

Tesla shareholders are suing Elon Musk and Tesla's board, accusing them of breaching fiduciary duties by allowing Musk to create and run a competing AI company, xAI, and diverting resources from Tesla to the new startup. The lawsuit comes just before Tesla's annual meeting where shareholders will vote on Musk's $56 billion compensation package.

Paris-based AI Startup Mistral AI Raises $640M

Parisian AI startup Mistral AI closed a $640 million Series B funding round led by General Catalyst, valuing the company at $6 billion. The company develops foundational AI models to compete with OpenAI and Anthropic, offering both open-source and proprietary models with various licensing restrictions. Mistral AI also offers a free chat assistant and looks to convert its engineering achievements into revenue by attracting corporate customers.

Databricks Expands Mosaic AI to Help Enterprises Build with LLMs

Databricks is expanding its Mosaic AI platform with a variety of new features to help enterprises build applications using LLMs. These new features include tools for building and evaluating LLM-based applications, fine-tuning models with private data, and querying and managing any open source or proprietary model in a governed way. Databricks is making these changes in response to how its customers are now working with LLMs, as enterprises are moving away from using solely open source models and are now using a variety of models from different vendors.

Runway Introduces Gen-3 Alpha

Runway's Gen-3 Alpha is a new model for large-scale multimodal training, significantly improving consistency, fidelity, and motion compared to its predecessor, Gen-2. This model, trained on videos and images, supports Runway's tools like Text to Video, Image to Video, and Advanced Camera Controls. It introduces new safeguards, including visual moderation and C2PA standards.

Anthropic Opens Beta Steering API for Developers

Anthropic is inviting developers to sign up for access to their Beta Steering API , now in research preview. This beta offering allows developers to adjust internal features of their language models for experimentation. While access is limited and prioritized based on the information provided by applicants, successful participants will receive documentation, example code, and possibly a Slack channel for discussions with the Anthropic team.

Luma Labs Introduces New AI Video Generator

Luma Labs has just released Dream Machine , an innovative AI model that generates high-quality, realistic videos from text and images in record time. This scalable and efficient transformer model produces 120 frames in 2 minutes, offering creators rapid iteration and exploration of new ideas. Capable of lifelike motion and consistent character interactions, Dream Machine is now available to everyone.

Stability AI Releases Stable Diffusion 3 Medium

Stability AI has released Stable Diffusion 3 Medium , a smaller yet powerful version of its flagship text-to-image generation model. Unlike its predecessor, which requires 8 billion parameters, SD3 Medium operates efficiently with just 2 billion parameters and 5GB of GPU VRAM, making it suitable for consumer-grade hardware. Despite its reduced size, SD3 Medium retains high-quality features such as photorealism, prompt adherence, and improved typography.

Zeta Labs Launches Jace AI Agent

Zeta Labs, founded by ex-Meta engineers, announced Jace, an LLM-powered AI agent capable of executing in-browser actions on command. Jace can automate complex tasks, from booking hotels to managing recruitment pipelines. Currently available for experimentation, Zeta Labs plans to refine Jace for broader release, aiming to offer a subscription model to meet high demand for automation in sectors like ecommerce, marketing, and sales.

AliCloud's Qwen2 : Advanced Open-Source LLM Models

AliCloud released a new open-source LLM model series, Qwen2 . Available in five sizes, these models, including Qwen2-0.5B to Qwen2-72B, offer state-of-the-art performance across multiple languages and capabilities. Notably, Qwen2 significantly enhances coding and mathematical proficiency and supports context lengths up to 128K tokens.

Apple's AI Strategy in a Nutshell

John Hwang, a seasoned tech analyst, delves into Apple's newly unveiled AI strategy at WWDC 2024 in his latest article. Hwang highlights Apple's vertically integrated approach, featuring the market-leading on-device AI model running on Apple silicon, and the company's shift to using its own data centers for AI inference. By integrating OpenAI’s ChatGPT and enhancing Siri's capabilities, Apple aims to dominate the AI landscape while reducing third-party dependency. Hwang's breakdown explains why this strategic move could significantly boost Apple's market position and revenue, making it a must-read for anyone interested in the future of AI.

MIT Researchers Enhance AI with Consensus Game Theory

Researchers from MIT have developed a "consensus game " where LLMs improve their accuracy and consistency by agreeing on answers through self-play. This novel method contrasts traditional AI milestones, like IBM's Deep Blue in chess, by using game theory not just for problem-solving but to refine the internal workings of AI. The implications suggest more reliable AI that holds potential for new applications.

Generative AI Propels Robotics Toward General Purpose

Integrating generative AI is enabling robots to perform a variety of tasks with greater efficiency. MIT’s “policy composition” (PoCo) method leverages generative AI to combine task-specific strategies, resulting in a 20% improvement in task performance. This approach allows robots to seamlessly switch between different tools and adapt to new tasks, bringing the dream of general-purpose robotics closer to reality.

Nadella Highlights AI and Digital Infrastructure at G7

Satya Nadella , CEO of Microsoft, attended the 2024 G7 Summit to discuss the Partnership for Global Infrastructure and Investment. He emphasized the importance of AI and digital infrastructure for global economic development, especially in low-income countries. Microsoft has recently invested in cloud and AI infrastructure in several developing countries. Nadella stressed that sustainable practices and public-private partnerships are crucial for building the future's infrastructure. He concluded by expressing optimism about global collaboration on using technology for growth.

Autonomous Driving with Musculoskeletal Humanoids

This paper details an autonomous driving project featuring the musculoskeletal humanoid, Musashi. Designed to mimic the human body precisely, Musashi boasts redundant sensors and a flexible structure, making it ideal for complex environmental interactions. The robot can sit in a car seat, operate acceleration and brake pedals, and steer using both arms.

Structuring Generative AI Prompting Techniques

This paper aims to clarify the concept of prompting in GenAI by creating a taxonomy of prompting techniques and analyzing their use. The authors introduce a comprehensive vocabulary of 33 terms, a taxonomy of 58 text-only prompting techniques, and 40 techniques for other modalities. They also conduct a meta-analysis of the literature on natural language prefix-prompting.

Linq raises $6.6M to use AI to make research easier for financial analysts.

GPTZero’s founders , still in their 20s, have a profitable AI detection startup and $10 million in the bank.

AWS Announces $230 million commitment for GenAI startups.

AI Engineer Summit - San Francisco - June 25 - 27

AI4 - Las Vegas - August 12 - 14

The AI Conference 2024 - San Francisco - September 10 - 11

World Summit AI - Amsterdam - October 9 - 10

Gitex Global - Dubai - October 14 - 18

Big Data Conference Europe - Vilnius - November 19 - 22