Today:

• AI Safety Leak Sparks Internal War

• OpenAI Launches Tokyo Hub

• Adobe Premiere Pro AI Video Tools

• Cohere's Private Beta for New Compass Model

• Altman Unveils ChatGPT Enterprise to Top Execs

• ChatGPT Marks 13% Traffic Growth in March 2024

SECRET WAR to Control AGI | AI Doomer $755M War Chest | Vitalik Buterin, X-risk & Techno Optimism

In a recent event, researchers from OpenAI, including those identified as Ally of Ilia, were terminated due to leaking confidential information. This follows the controversial dismissal of CEO Sam Altman and other leaks that were later verified. The focus of the leaks remains unclear, though they are connected to AI safety. 

However, the true intentions and actions of those within this movement are questioned, as they seem to diverge from their altruistic beginnings, becoming secretive and possibly pursuing a more controlling agenda under the guise of global safety and governance. 

WATCH THE VIDEO ON YOUTUBE

OpenAI expands into Asia with Tokyo hub, Japanese variant of GPT-4

OpenAI is pushing into Asia with a new Tokyo office led by Tadao Nagasaki, ex-President of AWS Japan. They're customizing GPT-4 for Japanese, joining a trend of region-specific AI models. OpenAI aims to collaborate closely with locals, catering to cultural nuances. Nagasaki will lead market engagement and team building. The move targets partnerships with big Japanese brands like Rakuten and Toyota. OpenAI's Japanese GPT-4 promises better performance and speed, appealing to diverse local needs. 

Despite competition, including from Sakana, OpenAI plans to work with the Japanese government on AI policy alignment. Microsoft's $2.9 billion investment in Japan adds momentum to the AI scene.

VENTUREBEAT

Adobe Premiere Pro is getting generative AI video tools — and hopefully OpenAI’s Sora

Adobe is enhancing Premiere Pro with generative AI video tools, including a Firefly model. Users can extend shots and manipulate objects using text prompts, similar to Photoshop's Generative Fill. Third-party integrations with Runway, Pika Labs, and potentially OpenAI's Sora models are in the works, offering more options for users. 

Adobe's move to collaborate with third-party AI companies aims to provide greater flexibility to Premiere Pro users. Despite no set release date, the integration of these AI models signals Adobe's commitment to advancing its video editing capabilities.

THE VERGE

Cohere Compass Private Beta: A New Multi-Aspect Embedding Model

Cohere has launched its private beta for Compass, a new embedding model tailored for multi-aspect data. Traditional approaches struggle with data containing multiple concepts and relationships, leading to semantic confusion. Cohere Compass addresses this by transforming data into JSON documents and storing representations in vector databases. 

This enables accurate indexing and searching, facilitating tasks like retrieving specific information from emails or GitHub issues. The private beta invites partners to test the system and provide feedback, highlighting Cohere's commitment to advancing enterprise data management.

COHERE

OpenAI's Altman pitches ChatGPT Enterprise to large firms, including some Microsoft customers

OpenAI's CEO Sam Altman recently pitched ChatGPT Enterprise to Fortune 500 execs in SF, NY, and London. The move signals OpenAI's push into corporate AI services, potentially competing with backer Microsoft. Altman showcased ChatGPT Enterprise, APIs, and text-to-video models, emphasizing data privacy. 

Despite Microsoft's offering, Altman highlighted direct collaboration and access to latest models as perks. OpenAI aims to diversify revenue beyond consumer products like ChatGPT. The enterprise push includes Hollywood interest in Sora video tool, despite concerns. OpenAI, valued at $86 billion, eyes $1 billion revenue in 2024. Attendees included execs from finance, healthcare, and energy sectors, signaling broad interest.

REUTERS

ChatGPT Leads with 13% Month-on-Month Traffic Growth, Hits 1.77 Billion Visits in March 2024

In March 2024, ChatGPT experienced a notable 13% month-on-month traffic growth, reaching 1.77 billion visits. This surge indicates a reversal from previous declines since summer 2023, with a robust 33% year-on-year increase. OpenAI's introduction of custom GPTs in fall 2023 also contributed to this momentum, with approximately 56.5 million visits recorded. Third-party GPTs witnessed a 3% uptick from February, totaling 50.5 million visits in March, marking a substantial 68% increase from November 2023. 

Despite this growth, ChatGPT's monthly traffic increase trails behind platforms like Claude and Gemini. Character AI and Perplexity also experienced significant traffic boosts, highlighting the broader trend of AI platform expansion and innovation.

DIGITAL INFORMATION WORLD

🧠RESEARCH

COCONut: Modernizing COCO Segmentation

The COCO Segmentation benchmark has seen slow progress, prompting a reevaluation. COCONut, a new dataset, enhances annotation quality and expands to 383K images with 5.18M masks, unifying semantic, instance, and panoptic segmentation. It aims to improve assessment of neural network progress in the vision community.

Pre-training Small Base LMs with Fewer Tokens

The study introduces Inheritune, a method to create small base language models (LMs) from larger ones. By inheriting a few transformer blocks and training on a tiny subset of data, a 1.5B parameter LM is built from 1B tokens. Inheritune models perform competitively against larger ones on various benchmarks, using significantly fewer tokens and resources.

Scaling (Down) CLIP: A Comprehensive Analysis of Data, Architecture, and Training Strategies

This study examines scaling down Contrastive Language-Image Pre-training (CLIP) across data, architecture, and training methods. It highlights the importance of high-quality data over quantity, suggests optimal model sizes for different datasets, and compares CNN-based versus ViT-based architectures. Various training strategies are evaluated, with CLIP+Data Augmentation showing promise in achieving comparable performance with half the data, enhancing CLIP's practical applicability.

On the Robustness of Language Guidance for Low-Level Vision Tasks: Findings from Depth Estimation

This paper investigates the robustness of language guidance in monocular depth estimation, revealing insights and limitations. While incorporating natural language improves results, it's found that current methods struggle with low-level descriptions, are vulnerable to adversarial attacks, and suffer performance decline with distribution shift. The study underscores the need for careful consideration in deploying language-guided depth estimation in real-world scenarios.

Probing the 3D Awareness of Visual Foundation Models

This study examines the 3D awareness of visual foundation models, exploring whether they encode and represent the 3D structure of scenes consistently across views. Through task-specific probes and zero-shot inference, limitations in current models are revealed. The research offers insights into the capabilities and constraints of these models, providing avenues for further exploration.

🛠️TOP TOOLS

E2B - Creates a secure cloud sandbox

Wingman - Lightweight open-source chatbot that's functional and not ugly

Concepto - Design & code websites as simple as drawing what you want

Ocean - A smarter prospecting data platform that brings you 5x more high quality leads 

Meshy - Create stunning 3D Models with AI

📲SOCIAL MEDIA

🗞️MORE NEWS

Zoom gets its first major overhaul in 10 years, powered by generative AI

Zoom, a popular video-calling tool, undergoes its first major revamp in a decade with Zoom Workplace, embracing AI to enhance user experience. Changes include customizable themes, improved speaker views, document collaboration, and AI features like portrait lighting and noise suppression. The update aims to boost productivity and collaboration. ZDNET

Poe introduces multi-bot chat and plans enterprise tier to dominate AI chatbot market

Poe, the AI chatbot platform from Quora, expands with multi-bot chat, allowing users to engage with multiple AI models in one conversation. With $75 million funding, Poe aims to be a hub for diverse AI models. It offers monetization for developers and plans an enterprise tier, foreseeing AI's role in customer service. VENTUREBEAT

Limitless wants its AI wearable to be your memory’s assistant

Limitless introduces a broach-like AI wearable, the Pendant, focusing on recording and processing conversations to enhance memory. With a sleek design by Ammunition, it aims to blend seamlessly into attire while providing noticeable recording cues. Priced at $99, it promises efficient hardware akin to software development. FAST COMPANY

SAG-AFTRA union secures AI protections for artists in deal with major record labels

The Screen Actors Guild‐American Federation of Television and Radio Artists (SAG-AFTRA) secures a tentative agreement with major record labels like Warner Music Group and Sony Music Entertainment, ensuring increased minimum salaries and safeguards against AI usage in music creation. The deal reflects growing concerns about AI's impact on artistic expression and aims to protect artists' rights. REUTERS

Generative AI can turn your most precious memories into photos that never existed

Synthetic Memories, a project by Domestic Data Streamers, recreates lost moments from people's lives using generative AI like OpenAI’s DALL-E. Memories, often undocumented, from various cultures, including immigrant and refugee communities, are brought to life through blurred, dream-like images. The project also explores therapeutic applications, like dementia treatment. MIT

Gaps emerge in how doctors and patients see AI

A survey reveals 4 in 10 US doctors are open to using generative AI with patients, a shift in attitude over the past year. While doctors see benefits like time-saving, patients remain wary. Transparency on AI training data is crucial for both parties, as organizations lag in developing clear guidelines. AXIOS

Generative AI’s Open Secret: Everyone Is Copying Everyone Else

In the realm of artificial intelligence, a surprising trend emerges: startups offering AI chatbots at a fraction of the cost of top-tier models like GPT-4 may actually be utilizing data from those very models, including those from OpenAI. This creates a competitive landscape where low-cost services mimic high-end AI capabilities, despite undercutting their prices. THE INFORMATION