Today:

• OpenAI's New PhD-Level AI

• Meta Updates AI Tag Visibility

• Mistral AI Launches Groundbreaking Pixtral 12B

• OpenAI Eyes $150B Valuation in New Funding

• AI Execs, White House Confer on Energy Demands

• Meta Trained AI with Public Posts Since 2007

• DataGemma Tackles AI Hallucinations with Real Data

• Chinese AI Chipmakers Eye IPOs to Rival Nvidia

OpenAI o1 CRUSHES PHD Level Experts! [HIDDEN THOUGHTS]

OpenAI's new model, O1, is a significant advancement in artificial intelligence, demonstrating near-human expertise in math, coding, and science. This model, trained with reinforcement learning, uses a "Chain of Thought" process to meticulously think through problems before providing solutions. It has excelled in solving complex issues that previous models could not, showing superior performance even compared to PhD-level experts. 

Unlike earlier models, O1's internal reasoning, or "Chain of Thought," remains hidden from users, enhancing safety and robustness. 

WATCH THE VIDEO ON YOUTUBE

Facebook and Instagram are making AI labels less prominent on edited content

Meta is updating its AI content labeling on Facebook, Instagram, and Threads. The "AI Info" tag, previously visible beneath the user's name, will now be hidden within a menu. This change, starting next week, aims to reflect the degree of AI usage in edited images and videos more accurately. 

Meta introduced this label following criticism of their earlier "Made with AI" tag, which incorrectly marked genuine photos as AI-manipulated. Despite enhancing user interface aesthetics, this shift may complicate efforts to discern AI-altered content, especially as editing tools improve and become more widespread on devices.

THE VERGE

Pixtral 12B is here: Mistral’s new multimodal AI can analyze images without any limits

Mistral AI, a French startup, has released Pixtral 12B, its first multimodal AI model, which can process both language and images. This model is not publicly available but can be downloaded from Hugging Face or GitHub for individual testing. Pixtral 12B stands out by supporting an unlimited number of images of any size, with a complex structure featuring 40 layers and 32 attention heads, indicating robust computational power. 

Mistral AI has positioned itself aggressively in the AI industry, collaborating with major companies and securing significant funding, indicating its intention to compete with leading AI labs. The full capabilities and performance of Pixtral 12B will become clearer once it is accessible via API.

VENTUREBEAT

OpenAI Fundraising Set to Vault Startup’s Valuation to $150 Billion

OpenAI is currently negotiating to raise $6.5 billion in equity financing, which would elevate its valuation to $150 billion. This new valuation marks a significant increase from its previous $86 billion earlier this year, positioning it among the most valuable startups globally. 

The discussions also involve securing a $5 billion credit line from banks, underscoring OpenAI's robust growth trajectory and expanding influence in the tech sector.

BLOOMBERG

Nvidia, OpenAI, Anthropic and Google execs meet with White House to talk AI energy and data centers

OpenAI CEO Sam Altman and other tech leaders are meeting at the White House to address the growing energy demands of AI technologies, which are expected to significantly increase U.S. power needs. This meeting, the first of its kind, involves discussions on how to support AI's expansion in a sustainable way, considering its potential to tackle major issues like climate change and health crises. 

The dialogue will include strategies for enhancing AI infrastructure and aligning it with environmental goals, as high energy use by AI could challenge America's power grid and transition to renewable energy.

CNBC

Meta fed its AI on almost everything you’ve posted publicly since 2007

Meta has been using publicly posted content from Facebook and Instagram users since 2007 to train its AI models. This includes all text and photos unless they were specifically set to private. The practice was confirmed by Meta's global privacy director, Melinda Claybaugh, during an inquiry into AI adoption. 

Despite privacy concerns, users outside the EU, where regulations allow opting out, do not have the option to prevent their public posts from being used for AI training. This has raised concerns about the lack of privacy protections and the potential exploitation of users' data, particularly the data of children. Claybaugh acknowledged that Meta does not scrape data from users under 18 but did not clarify the handling of accounts created by users when they were minors

THE VERGE

DataGemma: Using real-world data to address AI hallucinations

DataGemma introduces the world's first open models that integrate large language models (LLMs) with real-world data from Google's Data Commons to reduce inaccuracies known as "hallucinations." These advanced AI models can generate content, summarize texts, and even write code but sometimes create false information. By linking these models to a vast database of reliable data from global organizations, DataGemma aims to improve the accuracy of AI outputs. 

This new approach utilizes specific techniques like Retrieval-Interleaved Generation and Retrieval-Augmented Generation to enhance fact-checking and contextual understanding, thereby minimizing errors and increasing trust in AI-generated content.

GOOGLE

Two Chinese AI Chipmakers Seek IPOs to Mount Challenge to Nvidia

Two leading Chinese AI chip companies, Shanghai Enflame Technology Co. and Shanghai Biren Intelligent Technology Co., are preparing for initial public offerings (IPOs) on Shanghai's STAR board, potentially as early as 2024. This move is seen as part of China's broader effort to challenge Nvidia Corp. in the AI chip market. 

Enflame aims to raise up to 2 billion yuan ($280 million), leveraging the STAR board's receptiveness to startups that are growing quickly yet operating at a loss. Biren also plans to list on the same exchange, with both companies targeting to file their IPO documents within this year, planning a debut by early 2025.

BLOOMBERG

🧠RESEARCH

PingPong: A Benchmark for Role-Playing Language Models with User Emulation and Multi-Model Evaluation

The paper introduces a new benchmark for testing how well language models can act out roles in conversations. It uses language models to mimic users and evaluate dialogues through three parts: a player model, an interrogator model, and a judge model. The tests show good agreement with human opinions, offering a way to measure conversation quality in interactive settings.

MEDIC: Towards a Comprehensive Framework for Evaluating LLMs in Clinical Applications

MEDIC is a framework for evaluating Large Language Models (LLMs) in healthcare. It assesses models across five areas: medical reasoning, ethics, data understanding, learning, and safety. MEDIC uses a new method to measure model performance in tasks like answering medical questions and note generation. Results highlight differences in model types and sizes, helping select the right model for specific healthcare uses.

LLaMA-Omni: Seamless Speech Interaction with Large Language Models

LLaMA-Omni is a model for seamless speech interaction with large language models. It offers low-latency, high-quality responses without needing speech transcription, generating text and speech directly from spoken instructions. Built on the Llama-3.1-8B-Instruct model and a new dataset, LLaMA-Omni shows improved response quality and speed, with a latency of only 226ms. 

GroUSE: A Benchmark to Evaluate Evaluators in Grounded Question Answering

GroUSE is a benchmark designed to evaluate how well models can judge grounded answers generated by Retrieval-Augmented Generation (RAG) systems. It assesses judges on their ability to spot 7 types of errors, using 144 tests. Findings show that even advanced judges like GPT-4 overlook some errors. The paper proposes improvements for evaluation frameworks, noting that training on GPT-4’s methods enhances the judges’ accuracy and reliability in evaluating answers.

Towards a Unified View of Preference Learning for Large Language Models: A Survey

The paper surveys how to align Large Language Models (LLMs) with human preferences, crucial for enhancing model performance with minimal data. It breaks down alignment strategies into four components—model, data, feedback, and algorithm—and provides a unified framework to connect these strategies. This comprehensive approach clarifies existing methods and suggests potential synergies, exploring challenges and future research directions in preference learning for LLMs.

🛠️TOP TOOLS

Adobe Firefly - Firefly models and services power generative AI features in Adobe creative apps.

SaleStack - AI-powered sales tools at your fingertips.

Aicado - The easiest way to integrate AI into your business

Hoop - AI Task Management For Busy Professionals

Verse - Turn your inspiration into creation

📲SOCIAL MEDIA

🗞️MORE NEWS

Hacker tricks ChatGPT into giving out detailed instructions for making homemade bombs

A hacker manipulated ChatGPT into providing bomb-making instructions by framing questions within a fictional context, bypassing safety protocols. This exploit, termed "jailbreaking," raises concerns about AI security, as it was able to generate detailed, actionable explosive creation methods despite ethical safeguards.

NotebookLM now lets you listen to a conversation about your sources

Google's NotebookLM now offers an Audio Overview feature, converting uploaded documents into engaging audio discussions. This tool, using AI hosts, facilitates in-depth conversations on provided materials, making complex information easily digestible. While still experimental and only available in English, this feature represents a step forward in personalized learning and information processing.

Meta Will Soon Get a 100,000 GPU Cluster Too; What’s Life At Character Like Now?

Meta Platforms is finalizing a supercomputing cluster surpassing 100,000 Nvidia H100 server chips in the U.S., aimed at training its advanced Llama 4 model. This step, costing over $2 billion for the chips alone, highlights the intense competition among tech giants to enhance AI capabilities, with significant developments expected by late fall.

Cybever unveils AI-based 3D world creation platform

Cybever is launching an AI-driven 3D world creation platform to simplify and democratize the development process for creators and developers, making it more efficient and flexible. The beta, set to release by month-end, will feature tools for asset retrieval, placement, and generation, supporting collaborative workflows and reducing development times. With ethical AI practices, Cybever partners with asset marketplaces to ensure proper licensing, aiming to empower creators across all industries.

Bringing generative AI to video with Adobe Firefly Video Model

Adobe's Firefly Video Model, launching in beta later this year, introduces advanced AI tools for video editors. Developed with professional input and designed for commercial safety, this model aims to enhance creativity and efficiency in video editing by simplifying complex tasks like filling gaps and removing unwanted elements, all while ensuring the creative process is streamlined and collaboration-friendly.

Gemini’s chatty voice mode is out now for free on Android

Google has launched its Gemini Live voice chat mode for all Android users, offering it for free. Previously exclusive to Gemini Advanced subscribers, this feature allows users to engage in voice conversations with the AI, ask questions aloud, and choose from different voices. Currently available only in English, Google plans to expand Gemini Live to iOS and other languages in the future.

Amazon starts testing ads in its Rufus chatbot

Amazon is beginning to test advertisements within its Rufus chatbot, a shopping-focused AI. Ads will appear based on search contexts and conversations, enhancing brand and product discovery for users. This move aligns with Microsoft's previous ad integrations in Copilot and underscores the broader industry trend of monetizing AI functionalities to offset operational costs.

Gusto Opens Waitlist for AI Assistant for Small Business Owners

Gusto is introducing an AI assistant named Gus to its suite of services for small business owners, which includes HR, payroll, and compliance. This new assistant aims to simplify tasks by providing quick answers and personalized insights based on user data on the platform. Gus will handle queries in plain English and perform actions like approving paid time off and modifying employee salaries, pending owner approval. The beta version of Gus is set to become more capable in the coming months as it develops further.