Today:

• OpenAI SearchGPT

• AI Triumphs at Math Olympiad

• Groq Unveils Largest AI Model

• Udio v1.5 Enhances Audio Quality

• Google Play Adds AI Summaries

Are we ready for self-improving AI?

OpenAI announces SearchGPT, its AI-powered search engine

OpenAI has launched SearchGPT, an AI-powered search engine prototype. Unlike traditional search engines, it provides organized and summarized search results with clear attributions. SearchGPT aims to offer a more interactive experience by allowing users to ask follow-up questions and view visual answers. 

Initially, it will be available to 10,000 test users. OpenAI developed SearchGPT with input from news organizations and plans to integrate it into ChatGPT. The goal is to challenge Google and Perplexity by offering more advanced AI search capabilities. The prototype is free for now, but monetization strategies are needed due to high operational costs.

THE VERGE

AI achieves silver-medal standard solving International Mathematical Olympiad problems

DeepMind's new AI models, AlphaProof and AlphaGeometry 2, achieved a silver-medal level in the International Mathematical Olympiad (IMO). AlphaProof uses reinforcement learning to solve algebra and number theory problems, while AlphaGeometry 2 tackles complex geometry challenges. The models solved four out of six IMO problems, scoring 28 points. 

These systems mark significant progress in AI’s mathematical reasoning, enhancing the ability to solve intricate problems and assist mathematicians. This achievement showcases AI's potential in advancing mathematical research and applications.

GOOGLE DEEPMIND

Now Available on Groq: The Largest and Most Capable Openly Available Foundation Model to Date, Llama 3.1 405B

Groq has launched Llama 3.1 405B, the largest openly available AI model, via GroqCloud and GroqChat. This partnership with Meta enables developers to access advanced AI models without proprietary restrictions. Llama 3.1 offers increased context length, custom tool calling, and robust safety features. Groq’s LPU AI inference technology provides unparalleled speed, supporting real-time applications in healthcare, dynamic pricing, predictive maintenance, and personalized learning. 

This launch represents a significant step in open AI innovation, allowing over 300,000 developers to build sophisticated applications quickly and efficiently. Early API access is available to select customers, with broader access coming soon.

GROQ

Introducing v1.5

Udio v1.5 is the latest music model, offering better audio quality, key control, and enhanced global language support. New features include a dedicated creation page, stem downloads, audio-to-audio remixing, and shareable lyric videos. The platform now produces 48kHz-stereo tracks with improved clarity, instrument separation, and musicality. 

Users can split tracks into vocals, bass, drums, and other elements for advanced remixing. Key control allows for specific musical key generation, and the audio-to-audio feature lets users remix their own tracks. Enhanced language support broadens accessibility, and the shareable lyric videos boost social media engagement.

UDIO

Google is updating the Play Store with AI-powered app reviews and curated spaces

Google Play Store is updating to create a more engaging experience. Key updates include AI-generated review summaries, FAQs, and app highlights to help users make informed decisions quickly. Shared spaces will provide curated content hubs, starting with cricket and Japanese manga. 

The gaming section will feature enhanced details, YouTube videos, and developer notes, akin to Steam. The personalized Collections feature will offer custom app categories based on past purchases. Some updates are available now, while others are in early access or still in development.

ENGADGET

🧠RESEARCH

MovieDreamer: Hierarchical Generation for Coherent Long Visual Sequence

MovieDreamer is a new method for creating long videos with complex stories and consistent characters. It combines autoregressive models for narrative coherence and diffusion rendering for high-quality visuals. This approach improves on current techniques, producing superior visual and narrative quality over extended durations, akin to traditional movie production.

CoD, Towards an Interpretable Medical Agent using Chain of Diagnosis

Chain-of-Diagnosis (CoD) improves the transparency of medical diagnoses made by large language models. By mimicking a physician's thought process, CoD creates a clear diagnostic chain and outputs disease confidence levels. This method enhances interpretability and control, leading to better diagnostic accuracy. DiagnosisGPT, using CoD, outperforms other models in diagnosing 9604 diseases.

KAN or MLP: A Fairer Comparison

This paper compares KAN and MLP models across various tasks, maintaining equal parameters and FLOPs for fairness. MLP generally outperforms KAN, except in symbolic formula representation, where KAN's B-spline activation gives it an edge. Applying B-spline to MLP improves its symbolic performance. MLP also exhibits less forgetting in continual learning than KAN.

ChatQA 2: Bridging the Gap to Proprietary LLMs in Long Context and RAG Capabilities

ChatQA 2, a model based on Llama3, aims to match proprietary models like GPT-4-Turbo in long-context understanding and retrieval-augmented generation (RAG). It extends Llama3's context window from 8K to 128K tokens and improves performance through a three-stage tuning process. ChatQA 2 rivals GPT-4-Turbo in long-context tasks and excels in RAG benchmarks.

Stable Audio Open

Stable Audio Open introduces a new open-access text-to-audio model, addressing the lack of accessible generative models for artists and researchers. Trained with Creative Commons data, this model performs competitively with state-of-the-art alternatives. It excels in generating high-quality stereo sound at 44.1kHz, as demonstrated by its strong FDopenl3 results, which measure audio realism.

🛠️TOP TOOLS

HeyGen Labs Interactive Avatar - Create an avatar optimized for continuous streaming. 

Krea - The easiest way to generate high-quality visuals with AI.

Move - Bring realistic human motion to animated characters by turning 2D video into 3D motion data with proprietary technology that uses advanced AI, computer vision, biomechanics and physics.

Llama Tutor - Enter a topic you want to learn about along with the education level you want to be taught at and generate a personalized tutor tailored to you.

PixVerse V2 - AI-powered video creation platform with full potential of video creation.

📲SOCIAL MEDIA

🗞️MORE NEWS

AI models collapse when trained on recursively generated data

Generative AI models, like GPT, risk "model collapse" when trained on data produced by other AI models. This process leads to the gradual loss of original content diversity, especially rare events. Over generations, models trained on AI-generated data increasingly misrepresent reality, focusing on common patterns and missing the tails of the data distribution. This degradation is evident in various AI models, including language models and image generators. To avoid collapse and preserve AI performance, training must prioritize real human-generated data. NATURE

AI Video Generator Runway Trained on Thousands of YouTube Videos Without Permission

A leaked document obtained by 404 Media reveals that Runway's AI video generation tool, Gen-3, was trained on thousands of YouTube videos and pirated films without permission. This tool, praised in the AI community, was initially codenamed Jupiter and launched in June. Despite its acclaim and significant funding from Google and Nvidia, Runway’s co-founder, Anastasis Germanidis, did not disclose the specific sources of training data when asked by Techcrunch, only mentioning the use of curated internal datasets. 404 MEDIA

Why OpenAI Could Lose $5 Billion This Year

OpenAI, valued at $80 billion, could face losses up to $5 billion this year, based on internal data and insider insights analyzed by The Information. Despite being a rapidly growing business, the high operational costs are a significant challenge. If these projections hold, OpenAI will need to secure additional funding within the next year to sustain its operations.  THE INFORMATION

US lawmakers send a letter to OpenAI requesting government access

U.S. Senate Democrats and an independent lawmaker have sent a letter to OpenAI CEO Sam Altman, raising concerns about the company's safety standards and practices towards whistleblowers. The letter includes a request for OpenAI to allow U.S. government agencies access to its next foundation model for pre-deployment testing and evaluation. The lawmakers also seek a commitment from OpenAI to dedicate 20% of its computing resources to AI safety research. The scrutiny follows whistleblower allegations of inadequate safety measures for GPT-4 Omni and reports of retaliation against those raising concerns. COINTELEGRAPH

Who will control the future of AI?

Sam Altman, co-founder and CEO of OpenAI, argues that the future of artificial intelligence (AI) is at a crossroads between democratic and authoritarian control. He emphasizes the urgency for the U.S. and its allies to lead in AI development to ensure the technology benefits as many people as possible. Altman outlines four key areas: robust security measures, substantial investment in AI infrastructure and human capital, coherent commercial diplomacy policy, and creative models for global AI norms. He stresses that democratic nations must act now to prevent authoritarian regimes from dominating AI and using it to consolidate power. THE WASHINGTON POST