The future of presentations, powered by AI

Gamma’s AI creates beautiful presentations, websites, and more. No design or coding skills required. Try it free today.

Today:

• OpenAI Introduces 'Use More Intelligence' Button

• Copyright Office Rules on AI Art

• OpenAI Eyes $340 Billion Valuation

• Ai2 Launches Tülu 3 Model

• One Million Records Leaked by DeepSeek

• Mistral Small 3 Model Released

OpenAI Introduces 'Use More Intelligence' Button

OpenAI has released an updated version of GPT-4o for ChatGPT, expanding its knowledge base through June 2024. The update improves its visual capabilities, enhancing its ability to understand spatial relationships in images and connect graphics with text. The model also performs better in math, science, programming, and general language tasks. A new "Use more intelligence" button allows for deeper processing before responses, though the updated model is not yet available in the API documentation.

Copyright Office Rules on AI Art

The U.S. Copyright Office has ruled that purely AI-generated art, created solely from text prompts, cannot be copyrighted, as it lacks sufficient human authorship. However, human-created works incorporating AI elements can still be copyrighted if the human's creative contribution is substantial. The decision clarifies that AI-generated outputs are not eligible for copyright protection, but human-modified works using AI tools can be, depending on the level of creative input.

OpenAI Eyes $340 Billion Valuation

OpenAI is in early talks to raise up to $40 billion in a funding round that would value the company at $340 billion. SoftBank is leading the round and is negotiating to invest between $15 billion and $25 billion. The investment would also support Stargate AI, an infrastructure venture. This funding round marks a significant step in OpenAI’s continued growth and innovation in artificial intelligence technology.

Ai2 Launches Tülu 3 Model

Ai2 has released Tülu 3 405B, an open-source large language model that outperforms DeepSeek v3 and matches GPT-4o in some areas, particularly in safety benchmarks. The model uses advanced post-training techniques, including reinforcement learning from verifiable rewards (RLVR), which enhances accuracy in complex tasks. Ai2 distinguishes itself by providing full access to training data, infrastructure code, and models, ensuring greater customization for users. Tülu 3 is available on Ai2's platform for public use.

One Million Records Leaked by DeepSeek

DeepSeek, a Chinese AI startup, suffered a major data breach exposing over a million sensitive records, including user chat history and internal operational details. The breach was caused by an unsecured ClickHouse database, which was discovered by cybersecurity firm Wiz. Although DeepSeek quickly secured the database, the incident highlights significant security risks in AI platforms. The breach has drawn regulatory attention from several countries, raising concerns about AI data security, especially as adoption grows.

Mistral Small 3 Model Released

Mistral AI has launched Mistral Small 3, a 24B-parameter model optimized for low latency and high performance. Released under the Apache 2.0 license, it competes with larger models like Llama 3.3 and Qwen 32B, delivering exceptional speed and efficiency. Mistral Small 3 excels in real-time applications, including virtual assistants, automated workflows, and specialized fine-tuning for domains like healthcare and finance. The model is available on platforms like Hugging Face, Ollama, and Kaggle, and is designed for both local inference and cloud deployment.

🧠RESEARCH

Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate

The paper challenges traditional supervised fine-tuning by introducing Critique Fine-Tuning (CFT), where models learn to critique noisy responses instead of imitating correct ones. CFT, tested on datasets like WebInstruct and MetaMath, improves performance by 4-10% on math benchmarks, offering a more effective way to enhance reasoning in language models.

Atla Selene Mini: A General Purpose Evaluation Model

Atla Selene Mini is a cutting-edge small language model that outperforms previous evaluators, including GPT-4o-mini, across 11 benchmarks. Trained with a unique data curation strategy and a combination of preference optimization and fine-tuning, it excels in scoring, classification, and preference tasks, showing strong zero-shot agreement with human evaluations.

Exploring the sustainable scaling of AI dilemma: A projective study of corporations' AI environmental impacts

This paper explores the environmental impact of AI, particularly large language models, highlighting energy consumption, hardware fabrication, and end-of-life processes. It proposes a methodology to estimate AI-related environmental impacts, forecasting a significant rise in AI electricity use by 2030. The authors call for coordinated efforts and standardized frameworks to mitigate these effects.

🛠️TOP TOOLS

Stylar - AI-powered image generation and editing tool designed to give users exceptional control over image composition and style. 

Sizzle AI - AI-powered learning companion designed to revolutionize education across various subjects.

Easy Peasy AI - AI-powered content creation platform designed to streamline various creative tasks.

Pixelcut - AI-powered photo editing platform designed to simplify and enhance image editing for content creators, e-commerce businesses, and social media marketers.

Accounting AI Solver - AI-powered tool designed to simplify tax-related tasks for individuals and small businesses.

📲SOCIAL MEDIA

🗞️MORE NEWS

• SoftBank is negotiating a $25 billion investment in OpenAI, aiming to become its largest backer, in addition to a $15 billion commitment for the Stargate project. This move would reduce OpenAI's reliance on Microsoft.

• Riffusion, an AI startup, launched a free platform allowing anyone to create original music using its Fuzz AI model. It generates personalized tracks based on text, audio, or visual prompts, gaining attention from industry professionals.

• Microsoft is making OpenAI’s o1 reasoning model, Think Deeper, available for free to all Copilot users. The tool helps answer complex questions by considering them from various perspectives, offering step-by-step solutions for tasks like coding and planning.

• Luma AI has launched Upscale to 4K, allowing users to enhance any DreamMachine video up to 4K resolution. Videos generated in 720p or 1080p can now be upscaled for enhanced clarity and stunning detail. Available now.

• Martin AI, a Y Combinator-backed startup founded by college dropouts Dawson Chen and Ethan Hou, raised $2 million to develop a personalized AI assistant. The assistant handles tasks like email, scheduling, and calls, aiming to rival offerings from OpenAI and Google.

• OpenAI will provide its AI models to U.S. National Laboratories, including Los Alamos, Lawrence Livermore, and Sandia, for nuclear weapons security and other scientific research. The collaboration aims to advance science and strengthen national security

• MIT researchers discovered that training AI agents in a less uncertain environment, known as the "indoor training effect," can enhance their performance in real-world scenarios. This finding challenges the traditional belief that training should match real-world conditions.