Today:

• OpenAI, Broadcom Plans New AI Chip

• Mistral, NVIDIA Release NeMo 12B

• OpenAI Dropped from Lawsuit

• China Tests Socialist AI

• Tech Giants Form AI Coalition

• OpenAI Upgrades ChatGPT Enterprise

• Malaysia Develops AI Hub

• Google Gemini Joins Olympic Broadcast

• AI Boosts Cancer Detection Rates

OpenAI and Broadcom in talks about developing new AI chip

OpenAI is in talks with Broadcom and other semiconductor designers to create a new chip, aiming to reduce reliance on Nvidia and strengthen its supply chain. This effort, led by CEO Sam Altman, seeks to enhance chip, energy, and computing capacity for advanced AI models. 

OpenAI collaborates with partners like Microsoft, governments, and investors to stay at the forefront of AI technology. Despite Nvidia's dominance, OpenAI aims for self-reliance in developing AI that surpasses human capabilities. Significant financial backing, including Microsoft's $13 billion investment, is essential for these ambitions.

FINANCIAL TIMES

Mistral AI and NVIDIA Unveil Mistral NeMo 12B, a Cutting-Edge Enterprise AI Model

Mistral AI and NVIDIA have introduced Mistral NeMo 12B, a top-tier AI language model for enterprise use. This model, designed for tasks like chatbots and coding, combines Mistral AI’s data expertise with NVIDIA’s hardware and software. It boasts high accuracy and efficiency, handling complex information well with its 128K context length. 

Trained on NVIDIA's DGX Cloud AI platform, it uses advanced techniques for better performance. The model, which can be deployed quickly and easily, offers strong support and security, making it ideal for diverse enterprise applications.

NVIDIA

OpenAI Dropped From First Ever AI Programming Copyright Lawsuit

OpenAI has been dismissed from a copyright lawsuit filed by open-source programmers who claimed the AI tool Copilot used their code without proper credit. The case continues against GitHub and Microsoft, who collaborated with OpenAI on Copilot. Filed in 2022, this was the first major copyright lawsuit against OpenAI. 

The programmers argued Copilot violated the Digital Millennium Copyright Act by generating copyrighted code without proper licensing. Although the judge dismissed these claims, the case will proceed on the grounds that Copilot breached open-source licensing agreements. The case is Doe 1 v. GitHub Inc.

BLOOMBERG

China deploys censors to create socialist AI 

Chinese government officials are testing AI companies' large language models to ensure they align with socialist values, expanding the country's censorship efforts. The Cyberspace Administration of China has mandated tech giants like ByteDance and Alibaba to undergo government reviews of their AI models, focusing on political sensitivities and President Xi Jinping.

China's strict approval process has led AI groups to learn how to censor large language models effectively, including filtering out sensitive keywords and questions that violate socialist values. To comply with regulations, Chinese chatbots reject queries on topics like the Tiananmen Square massacre or President Xi Jinping, showcasing the stringent control over AI content in the country.

FINANCIAL TIMES

The biggest names in AI have teamed up to promote AI security

Major tech companies including Google, OpenAI, Microsoft, Amazon, Nvidia, and Intel have formed the Coalition for Secure AI (CoSAI) to tackle AI security issues. CoSAI, operating under the Organization for the Advancement of Structured Information Standards (OASIS), aims to develop best practices, address AI challenges, and secure AI applications. 

The coalition will provide open-source tools and frameworks to improve AI security and privacy, responding to concerns about confidential information leaks and automated discrimination. Other members include IBM, PayPal, Cisco, and Anthropic.

THE VERGE

OpenAI gives more control over ChatGPT Enterprise

OpenAI has introduced new control features for its ChatGPT Enterprise to enhance compliance, data security, and user management. The Enterprise Compliance API now records interactions, enabling data audits. Integration with providers like Microsoft and Palo Alto Networks supports compliance activities such as data retention and policy enforcement. 

OpenAI is launching a user identity management system to simplify access control. Admins now have broader control over custom GPTs, including setting approved domains and managing permissions. ChatGPT Enterprise, released in 2023, offers robust security and faster processing, and is crucial for OpenAI's expansion in the enterprise sector.

VENTUREBEAT

Malaysia builds AI hub on Singapore’s doorstep

Southern Malaysia's Johor state is transforming from palm oil plantations to a digital economy hub, attracting billions in investments for data centers from companies like TikTok, Nvidia, and Microsoft. This shift leverages cheaper land and abundant energy, while remaining close to Singapore's financial center. Malaysia aims to emulate China's Shenzhen by integrating Johor with Singapore through a proposed special economic zone, enhancing economic ties. 

The initiative is supported by Malaysia's pro-business king. Despite concerns about resource pressure and job creation, the integration with Singapore and rapid data center growth position Johor for significant economic transformation.

FINANCIAL TIMES

Google’s Gemini AI will be all over the Paris Olympics broadcast

Google, the official AI sponsor for Team USA, will feature its AI tools prominently in the Paris 2024 Olympics broadcast. NBCUniversal’s coverage will include Google Maps’ 3D views of venues, showcasing events with photorealistic models. Announcers will use Google Search AI Overviews to answer Olympic-related questions. 

Comedian Leslie Jones will interact with Google’s AI, Gemini, for learning new sports. Five Olympic and Paralympic athletes will appear in social videos using Google Lens and other tools to explore Paris. This initiative highlights Google’s AI capabilities and integrates them into the Olympics' viewing experience.

THE VERGE

GPs use AI to boost cancer detection rates in England by 8%

The “C the Signs” AI tool, used by about 15% of GP practices in England, has boosted cancer detection rates from 58.7% to 66.0%. This AI analyzes patient records, including medical history and demographics, to spot cancer risks and suggest tests. Tested in 2021, the tool’s success led to detecting over 50 types of cancer. It also aids in faster diagnoses and patient tracking. 

Developed by GP Bea Bakshi, the tool addresses late cancer diagnoses. Despite challenges, the tool is part of the NHS's strategy to improve early cancer detection and care through advanced technology and awareness.

THE GUARDIAN

🧠RESEARCH

Scaling Laws with Vocabulary: Larger Models Deserve Larger Vocabularies

The study highlights the importance of larger vocabularies for large language models (LLMs). Researchers found that optimal vocabulary size depends on the computing power available and that most LLMs use too small vocabularies. Increasing vocabulary size improves performance, demonstrating the need to balance model parameters with vocabulary size for efficient scaling.

Scaling Retrieval-Based Language Models with a Trillion-Token Datastore

The paper explores how increasing the datastore size during inference improves language model (LM) performance. Researchers created a 1.4 trillion-token datastore and found that even smaller models with large datastores outperform larger models on knowledge-intensive tasks. They emphasize considering datastore size for efficient LM scaling and have open-sourced their datastore and code.

Shape of Motion: 4D Reconstruction from a Single Video

The paper presents a new method for reconstructing 3D motion from single videos. Unlike previous approaches, this method captures full-sequence 3D motion in dynamic scenes using a compact motion representation and data-driven priors. It successfully decomposes scenes into moving parts and achieves top performance in motion estimation and view synthesis, even in complex scenes.

AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases

The study introduces AgentPoison, a novel attack method targeting language model (LLM) agents by poisoning their memory or knowledge bases. This backdoor attack manipulates the retrieval process to trigger malicious responses while keeping normal performance intact. Tested on various agents, including autonomous driving and healthcare, AgentPoison shows over 80% success with minimal impact on regular tasks, highlighting significant security vulnerabilities.

GoldFinch: High Performance RWKV/Transformer Hybrid with Linear Pre-Fill and Extreme KV-Cache Compression

The paper presents GoldFinch, a hybrid model combining Linear Attention and Transformer techniques. It features a novel KV-Cache compression method, achieving significant cache size reduction and improved modeling performance. GoldFinch enables efficient inference of large context lengths, even on limited hardware. The authors provide their trained weights and code for community use.

🛠️TOP TOOLS

Flow Studio - Produced film-quality videos with stories, characters, and sound in one click.

CharacterGen - generate 3D characters from single images using advanced multi-view pose calibration techniques.

Waveform AI - Collect Data using AI voice forms

Tripo AI - Generates high-quality and ready-to-use 3D models with text or images in seconds

Blaze - Scan millions of online signals to help modern companies target the right potential customers as soon as they show interest.

📲SOCIAL MEDIA