Welcome back!

OpenAI is reportedly developing Orion, its next flagship AI model designed to surpass GPT-4o. Orion will use synthetic training data generated by its advanced reasoning system, "Project Strawberry". Will Orion redefine the future of AI? Let's unpack...

Today’s Summary:

• 🌠 OpenAI preps Orion model

• 🌟 Google debuts 3 new Gemini AI models

• 🎥 CogVideoX: Open-source video generation AI

• 📝 Anthropic reveals Claude prompts

• 📜 OpenAI backs California AI watermark bill

• ⚡ Cerebras chips run AI 20 times faster

• 🛠️ 2 new tools

TOP STORY

OpenAI is developing Orion, its next flagship model

The Summary: OpenAI is working on Orion, its next-generation large language model (LLM) designed to surpass GPT-4o. Orion uses advanced reasoning capabilities from “project Strawberry” to enhance its training data.

Key details:

• Orion uses reasoning breakthroughs from 'Strawberry' project, recently presented to US national security, to generate high-quality training data

• The model is designed to improve accuracy in AI responses

• OpenAI’s MRR has tripled to $283M compared to last year

• Orion is part of OpenAI's strategy to outpace competitors in conversational AI

Why it matters: Orion represents a major effort by OpenAI to advance AI technology. With improved reasoning and reduced “hallucinations”, Orion will set new standards for reliability and versatility in AI applications.

GOOGLE

Google unveils trio of experimental Gemini AI models

The Summary: Google has introduced three new experimental AI models, including an even more compact 8B-parameter version of Gemini 1.5 Flash, and upgraded versions of Gemini 1.5 Pro and Flash. The new models increase performance in multimodal tasks, complex prompts, and coding.

Key details:

• Gemini 1.5 Pro exp-0827: Outperforms previous versions and ranks #2 on the LMSys leaderboard

• Gemini 1.5 Flash exp-0827: Mini model that performs comparably to GPT-4o-mini according to LMSys

• Gemini 1.5 Flash-8B: New 8-billion-parameter micro model designed for high-volume multimodal tasks

• Free tier available for testing via Google AI Studio and Gemini API

• Automatic redirection from old to new Pro model starts Sep 3

Why it matters: These experimental releases provide developers with early access to improved versions of Gemini AI. The introduction of a smaller, yet powerful model offers an efficient option for AI applications that don’t require the full capabilities of larger models.

Try it in Google AI Studio

VIDEO AI

CogVideoX open-source video generation AI

The Summary: Researchers from Tsinghua University and Zhipu AI have released CogVideoX, an open-source text-to-video model capable of creating coherent videos up to six seconds long. With 5 billion parameters, CogVideoX offers strong performance and runs on consumer-grade hardware.

Key details:

• Generates 720×480 videos at 8 fps

• Implements a 3D Variational Autoencoder for efficient compression

• Code and model weights are publicly available

• Runs on consumer-grade hardware like RTX 3060 GPUs

• Initial user feedback shows strong prompt adherence and quality, even with complex scenarios

Why it matters: CogVideoX delivers usable results where previous open-source video generation models struggled with quality. This advancement opens up new possibilities for small companies and developers to experiment with video AI generation on local hardware.

Try it on Hugging Face - Research paper

QUICK NEWS

Quick news

• Anthropic publishes Claude system prompts

• OpenAI supports California AI bill requiring watermarking of AI images and videos

• Cerebras AI chips make AI processing 20 times faster

TOOLS

🥇 New tools

• TheySaid - Conversational surveys

• LM Studio 0.3 - Run LLMs locally on your computer (new version)