Hello Tuners,

This week, we turned our eyes again on OpenAI, with more co-founders leaving the company. The direction in which the Closed Source Giant seems to be going has seemed to continue to disappoint the original founders.

However, barring employee retention and shipping models, OpenAI continues to build products around its existing line-ups and invest in groundbreaking ventures

• OpenAI Shipping New Teams for Startups Instead of Models

• LLMs Powering The Next Leap in Humanoid Robotics

• Hardware Remains Top Choice for Investors

• OpenAI JSON Generation Gets a Glow-Up

OpenAI Shipping New Teams for Startups Instead of Models

John Schulman, a co-founder of OpenAI, has made a significant move by joining rival AI startup Anthropic. Schulman, a pivotal figure in the development of OpenAI's ChatGPT through his leadership of the reinforcement training team, left to deepen his focus on AI alignment and engage in more hands-on technical work. He expressed confidence that OpenAI would continue to thrive in his absence and excitement about his new opportunities at Anthropic.

In related developments, OpenAI president and co-founder Greg Brockman is taking an extended leave to rest and recharge, and product manager Peter Deng has also exited the company. These departures mark a transition period for OpenAI, which has been scrutinized for its approach to AI safety research. Despite the changes, OpenAI CEO Sam Altman praised Schulman for his contributions and friendship, underscoring the high regard his former colleagues hold him. With Schulman’s rule, three of OpenAI's 11 founders remain active.

LLMs Powering The Next Leap in Humanoid Robotics

Figure has officially introduced its latest humanoid robot, Figure 02, a significant upgrade from its predecessor, Figure 01. The unveiling video, reminiscent of high-end consumer electronics promotions, highlights the robot's deliberate movements and showcases a demo area filled with similar robots. These robots are seen performing practical tasks like carting totes, a typical application for humanoids in various industries.

Figure 02 benefits from an enhanced collaboration with OpenAI, following Figure's $675 million Series B funding round in February, which valued the company at $2.6 billion. This new model is designed to work seamlessly alongside human colleagues on factory floors and is equipped with speakers and microphones for effective communication. As the robotics industry increasingly integrates advanced neural network technology, Figure 02 represents a crucial step toward more versatile and interactive humanoid robots.

Hardware Remains Top Choice for Investors

Groq Inc., an artificial intelligence startup, has raised $640 million in new funding, highlighting significant investor interest in AI chip innovation. BlackRock Inc. funds led this substantial investment round. They saw participation from the investment divisions of tech giants such as Cisco Systems Inc. and Samsung Electronics Co. With this new funding, Groq’s valuation has soared to $2.8 billion.

Groq specializes in designing advanced semiconductors and software specifically tailored to enhance the performance of AI tasks. Their technology aims to address the critical bottleneck in AI computing power demand, offering solutions that could significantly improve efficiency in AI system operations. This investment underscores the growing importance and demand for optimized AI hardware solutions in the rapidly evolving tech landscape.

OpenAI JSON Generation Gets a Glow-Up

While OpenAI has improved the reliability of generating valid JSON, it hasn't adhered to specific schemas. Addressing this gap, OpenAI has now unveiled Structured Outputs in the API, a groundbreaking feature that guarantees model-generated outputs will conform precisely to JSON Schemas defined by developers.

Structured Outputs are available in two forms: function calling and response format. This feature is supported by models such as gpt-4-0613 and gpt-3.5-turbo-0613 and later versions. The latest model, gpt-4o-2024-08-06, has demonstrated perfect reliability in adhering to complex JSON schemas, marking a significant improvement over its predecessors. This feature presents a boom for data scientists, making the compilation of sophisticated synthetic data a work of minutes.

LLM Of The Week

ShieldGemma: Generative AI Content Moderation Based on Gemma

Google has unveiled ShieldGemma, a comprehensive suite of LLM-based safety content moderation models built on the advanced Gemma 2 framework. ShieldGemma includes classifiers tailored to detect and manage key harm types such as dangerous content, toxicity, hate speech, and more, making it a robust tool for maintaining safer online environments. In addition to these classifiers, ShieldGemma introduces an LLM-based data curation pipeline designed to support various safety-related applications and tasks.

Performance-wise, ShieldGemma demonstrates significant improvements over existing models, achieving a +10.8% increase in Area Under Precision-Recall Curve (AU-PRC) on public benchmarks compared to Llama Guard and a +4.3% improvement over WildCard. These metrics underscore ShieldGemShieldGemma's capability to moderate harmful content and ensure a safer digital experience for users.

Weekly Research Spotlight 🔍

Improving Retrieval Augmented Language Model with Self-Reasoning

Baidu has introduced an innovative end-to-end self-reasoning framework designed to enhance the reliability and traceability of Retrieval-Augmented Generation (RAG) systems. This new approach leverages the reasoning trajectories generated by Large Language Models (LLMs) to carry out three key processes, significantly improving the accuracy of these systems.

The first process is relevance-aware, where the LLM judges the relevance between retrieved documents and the question posed. The second process is evidence-aware selection, which involves the model choosing and citing relevant documents and automatically selecting snippets of key sentences as evidence. The final process, trajectory analysis, generates a concise analysis based on the self-reasoning trajectories from the previous two processes and provides the final inferred answer.

This method enables the model to be more selective, better reason, and effectively distinguish between relevant and irrelevant documents. Impressively, Baidu achieves performance comparable to GPT-4 with only 2,000 training samples generated by GPT-4, showcasing its efficiency and potential in improving RAG systems.

Best Prompt of the Week 🎨

Today's Goal: Try new things 🧪

Acting as a Career Development Planner

Prompt: I want you to act as a career development planner for UI/UX designers. You will create a structured daily plan designed to help individuals learn and master Figma as part of their journey into the UI/UX industry. You will identify a target client profile, develop key strategies and action plans, select the tools and resources for effective Figma training, and outline any additional activities needed to support their career transition. My first suggestion request is: "I need h” lp creating a daily activity plan for a UI/UX designer who wants to learn Figma to advance their career in the UI/UX industry."

This Week's Must Read Gem 💎