Today:

• Meta AI Joins Apple

• OpenAI Acquires Rockset

• ByteDance Partners with Broadcom

• Stability AI Welcomes Akkaraju

• Synthesia's Multilingual Video Creation

• HeyGen Raises $60M, Valued at $500M

• Character.AI's Advances in AI Inference

Arch-rivals Apple and Meta reportedly discuss AI partnership for iOS 18

Apple and Meta, traditionally rivals, are reportedly discussing a partnership to integrate Meta's AI models into iOS 18's Apple Intelligence feature. This collaboration would be similar to Apple's existing deal with ChatGPT, which allows users to access third-party AI for complex queries. Under such agreements, external AI models can't collect user data and must delete information after processing requests. 

Apple aims to offer multiple AI model options, with Meta potentially joining alongside ChatGPT. Apple Intelligence, set to launch with iOS 18, iPadOS 18, and macOS Sequoia updates, will initially be available in US English on newer devices, with gradual expansion to other languages and regions.

9TO5MAC

OpenAI’s first acquisition is an enterprise data startup

OpenAI has acquired Rockset, an enterprise data startup, marking its first acquisition. The deal aims to enhance OpenAI's retrieval infrastructure across its products. Rockset's technology helps businesses turn data into actionable insights, which OpenAI plans to offer its customers. While financial details weren't disclosed, Rockset had previously raised $105 million in funding. 

OpenAI's COO Brad Lightcap emphasized the benefits of integrating Rockset's foundation into their products. Rockset's CEO Venkat Venaramani assured current customers of a gradual transition off the platform. Some Rockset team members will join OpenAI as part of the acquisition.

THE VERGE

China's ByteDance working with Broadcom to develop advanced AI chip

ByteDance, the Chinese company behind TikTok, is working with US chip designer Broadcom to create a cutting-edge AI chip. This 5-nanometer chip will comply with US export rules and be made by Taiwan's TSMC. The partnership aims to secure a steady supply of high-end chips for ByteDance amid US-China tensions. 

This collaboration is rare, as there haven't been public US-China chip partnerships at this advanced level since 2022 due to US restrictions. ByteDance needs these chips to power its AI-driven apps like TikTok and Doubao. The company has also been stockpiling chips from Nvidia and Huawei to support its AI efforts.

REUTERS

Stability AI gets new leadership as gen AI innovations continue to roll out

Stability AI, the company behind Stable Diffusion, is undergoing significant changes. CEO Emad Mostaque resigned to pursue decentralized AI, with COO Shan Shan Wong and CTO Christian Laforte stepping in as interim co-CEOs. This shift comes amid financial struggles, with the company reportedly facing low revenue and high debt. In June 2024, Prem Akkaraju was appointed as the new CEO, accompanied by a funding round led by tech entrepreneur Sean Parker. 

The leadership change reflects the volatile nature of the AI startup landscape and intensifying competition in the field. Stability AI's future now depends on its ability to adapt and find a sustainable business model in the rapidly evolving AI industry.

VENTUREBEAT

Nvidia-backed AI startup Synthesia now lets you make multilingual video presentations using just your phone or webcam

Synthesia, a British AI startup backed by Nvidia, has introduced new features that allow users to create AI avatars using just a webcam or phone. The company now offers full-body avatars with hand movements and a screen recording tool with AI avatar guidance. Users can make multilingual video presentations with voice cloning capabilities. 

Synthesia aims to make video creation accessible to a wider audience, moving beyond its initial focus on high-budget productions. The company's latest innovations include an AI video assistant for summarizing content and plans to customize avatars for specific regions. Valued at $1 billion, Synthesia is positioning itself as a comprehensive video production suite for large corporations.

CNBC

AI Video Startup HeyGen Valued at $500 Million in Funding Round

HeyGen, an AI startup that creates realistic digital avatars, has raised $60 million in funding, valuing the company at $500 million. Led by Benchmark, the investment will help HeyGen expand its AI-powered video creation tools. Founded in 2020, HeyGen allows users to generate lifelike avatars that speak in their voice and translate content into various languages. The technology has been adopted by businesses for training videos, ads, and chatbots. 

While concerns about deepfakes exist, HeyGen's CEO Joshua Xu believes the tool opens new possibilities for visual storytelling. The company has been profitable since 2023 and plans to enhance security and AI ethics with the new funding.

BLOOMBERG

Optimizing AI Inference at Character.AI

Character.AI, a company aiming to develop advanced artificial intelligence, has shared insights into their efforts to make AI language models more efficient. They currently handle an impressive 20,000 queries per second, which is about 20% of Google Search's volume. The company has managed to reduce the cost of an hour-long conversation to just one cent through various optimization techniques. They're focusing on improving "inference," the process by which AI generates responses, to make their technology more cost-effective and scalable.

While they claim to be working towards artificial general intelligence (AGI), some experts question whether they mean true AGI or simply more human-like AI conversations.

CHARACTER AI

🧠RESEARCH

Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges

This paper examines the performance and biases of large language models (LLMs) used as judges to evaluate other models. Using TriviaQA as a benchmark, the study assesses various LLMs and finds that while some align well with human evaluations, others do not. Key findings highlight alignment issues and suggest improvements for future use.

LongRAG: Enhancing Retrieval-Augmented Generation with Long-context LLMs

The paper introduces LongRAG, a new framework that improves retrieval-augmented generation (RAG) by using longer text units for both retrieval and reading. By processing larger sections of text, LongRAG reduces the number of units needed and enhances performance, achieving high recall and accuracy in answering questions from Wikipedia.

nabla^2DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials

This paper introduces nabla²DFT, a comprehensive quantum chemistry dataset designed for training neural network potentials (NNPs). The dataset includes a wide range of molecular structures and properties, offering extensive data for accurate drug discovery simulations. It also provides a new benchmark for evaluating NNPs in molecular property prediction and conformational optimization tasks.

HARE: HumAn pRiors, a key to small language model Efficiency

This paper introduces HARE-1.1B, a small language model (SLM) optimized using human priors to enhance data efficiency. By focusing on a concise, high-quality dataset with semantic diversity, HARE-1.1B outperforms existing SLMs in resource-constrained settings, demonstrating the value of incorporating human insights into language model training.

Instruction Pre-Training: Language Models are Supervised Multitask Learners

This paper presents Instruction Pre-Training, a framework that enhances language models by using supervised multitask learning. By generating 200 million instruction-response pairs across 40+ tasks, this method significantly improves model performance. The approach enables smaller models, like Llama3-8B, to match or surpass larger ones, such as Llama3-70B.

🛠️TOP TOOLS

Claude - AI assistant built for work and trained to be safe, accurate, and secure.

Captions - The next generation of storytelling.

Lens Studio - Custom creation of ML models, 2D and 3D assets, and more for anyone to use.

Mojo - Create amazing logo animations with AI reveals

Amazon Q - Generative AI–powered assistant for accelerating software development and leveraging companies' internal data

📲SOCIAL MEDIA