#168 - OpenAI vs Scar Jo + safety researchers, MS AI updates, cool Anthropic research
May 28, 2024
auto_awesome
Host Gavin Purcell discusses OpenAI pausing Sky voice in ChatGPT, Microsoft's Copilot upgrade, Adobe Lightroom magic eraser, and free AI assistant for educators. Controversies at OpenAI, including safety researcher departures and a $250 million deal with WSJ owner. Microsoft's new image analyzing language model and Anthropic's interpretability research are also covered.
OpenAI addresses concerns on ChatGPT voice similarity to Scarlett Johansson.
Microsoft upgrades Copilot assistant with GPT-4o and focuses on PC future.
OpenAI founders defend after safety researchers leave; Strike major deal with WSJ.
Google introduces Trillium, enhancing speed with sixth-gen Tensor processors.
Deep dives
New OpenAI Research on AI Interpretability Reveals Insights into Black Box Models
OpenAI has made strides in interpretability research to shed light on understanding how large language models function internally. By examining neural net activations and outputs, they have built a dictionary of features to represent different neuron combinations. These features can be labeled, such as the 'Golden Gate' feature that activates for parts related to the Golden Gate Bridge. The research also uncovers features tied to coding, names, bias, sycophancy, and deception, providing key insights into model behavior.
Chabot Arena's New Category 'Heart Prompts' Challenged AI Models with Rigorous Criteria
Chabot Arena has introduced 'Heart Prompts', a new evaluation category designed to rigorously test AI models with complex and challenging user-submitted prompts. Models like Cloud Opus and GPT-4 compete in these more difficult and comprehensive prompts, providing a deeper assessment of their capabilities beyond typical benchmarks. These heart prompts assess specificity, real-world knowledge, and more, offering a tougher evaluation metric for cutting-edge AI models.
Google Unveils Trillium, Sixth Generation Tensor Processor with Impressive Speed Enhancements
Google has introduced Trillium, the latest iteration of their tensor processors, boasting significant speed improvements over the previous versions. The Trillium TPUs are 4.7 times faster than their predecessors, with plans to scale up to 256 units in cloud offerings. This advancement showcases Google's commitment to enhancing hardware infrastructure to support the demanding requirements of AI computing.
Microsoft Ventures into Visual AI with Five-Prime Vision Model for Image Analysis
Microsoft has expanded its language model offerings with Five-Prime Vision, a variant that focuses on image analysis and description. This multi-modal model integrates language and vision capabilities, allowing it to analyze and describe image content effectively. With a parameter size of 4.2 billion, Five-Prime Vision enhances Microsoft's AI offerings in the visual domain, catering to a broader range of applications requiring image understanding.
Scale AI Raises $1 Billion in Series F Round as Valuation Soars to $13.8 Billion
Scale AI, a data labeling startup specializing in data curation for machine learning models, has secured a significant $1 billion funding in their latest series F round. The company's valuation has doubled to $13.8 billion, reflecting the growing demand for data labeling services in the AI industry. Scale AI's infrastructure plays a crucial role in training and improving machine learning algorithms, contributing to the advancement of AI technology.
Implications of AI in Brain Science and Creativity
AI's black box insights could lead to broader understandings in brain science and creativity. Agencies like Anthropic are developing cutting-edge models to understand AI better and prevent harmful human-driven actions with AI. Despite claims of not knowing what's going on, advancements in interoperability techniques and deep learning contribute to an evolving understanding of AI's behaviors and applications.
Early Fusion Approach in Multimodal Training Models
Meta's chameleon mixed model emphasizes early fusion for understanding and generating images and text seamlessly. By treating images and text as a single stream for training models, early fusion allows for more comprehensive data processing and output capabilities. The approach differs from traditional methods by integrating image and text data in a cohesive manner, showing promise in multimodal training applications.
AI Regulation and Safety Efforts by Tech Giants
Tech giants commit to AI safety frameworks and potential kill switches to mitigate risks in developing advanced AI models. This initiative seeks to enforce safety measures and define intolerable risks associated with new AI models, emphasizing transparency and accountability. By collaborating with other entities, companies like Microsoft, Amazon, and OpenAI aim to establish robust safety guidelines and protections in AI development.
