Welcome back!

OpenAI has released its new o1 “Full” model, showcasing major improvements in math and coding. Alongside the model, OpenAI introduced a new ChatGPT Pro subscription, aimed at researchers, with unlimited o1 and Voice access for $200/month. Let’s unpack...

Today’s Summary:

• 🚀 OpenAI launches o1 Full model

• 🦙 Meta Llama 3.3 matches 405B power

• 🕹️ DeepMind’s Genie 2 makes 3D worlds

• 📢 Google Veo and Imagen 3 available on API

• 🔥 xAI expands AI supercomputer in Memphis

• 🛡️ Anduril and OpenAI join forces for drone defense

• 🛠️ 2 new tools

TOP STORY

OpenAI launches Full o1 model with Pro Mode

The Summary: OpenAI has released the “Full” o1 model for Plus subscribers, introducing upgrades like image analysis and a 34% drop in errors on complex tasks. Alongside this, OpenAI launched a $200/month ChatGPT Pro tier aimed at scientists and researchers, offering enhanced computing power for demanding use cases. Medical researchers at Harvard and Berkeley Lab are among the first to receive Pro grants. The model shows remarkable improvements in math and coding benchmarks, with the Pro version solving 80% of Olympic math problems.

Key details:

• o1 Full scored 83% on Math Olympiad tests, up from GPT-4o's 13%

• o1 now supports image uploads, enabling detailed visual reasoning

• o1 Pro achieves 75% accuracy on competition coding tasks

• Hallucinations reduced from 61% to 44% on factual quizzes

• Available for $20/month with ChatGPT Plus, while the new “Pro” subscription offers unlimited use and additional computing power for $200/month

• Upcoming o1 features include web browsing, file uploads, and richer API integrations

Why it matters: OpenAI o1 model sets a new standard for reliability and versatility, equipping professionals with the cutting-edge reasoning tools. The introduction of ChatGPT Pro is a move towards specialized, high-performance AI services tailored to scientists and professionals.

OpenAI o1 System Card

META

Meta Llama 3.3 packs 405B power into 70B

The Summary: Meta just released Llama 3.3, a 70B parameter AI model that matches the performance of their previous 405B behemoth. Through advanced post-training methods and online preference tuning, Llama 3.3 delivers equal or better results on key benchmarks. The open-source model is available now on platforms like Hugging Face.

Key details:

• Outperforms GPT-4o, Gemini 1.5 Pro, and Amazon Nova Pro on MMLU language understanding tests

• Math reasoning jumps 9% to reach 77% accuracy through chain-of-thought processing

• Handles 128K context window while being 6x smaller than comparable models

• Meta’s next model, Llama 4, will require 10x more compute, driving plans for a $10B Louisiana data center

Why it matters: This leap in efficiency opens doors for AI development on standard hardware, not just tech giants data centers. This breakthrough levels the playing field, empowering smaller teams and researchers to work with cutting-edge AI locally.

Try it Now on HuggingFace

RESEARCH

DeepMind Genie 2 creates interactive worlds

The Summary: Just days after World Labs unveiled its groundbreaking 3D AI project, Google DeepMind introduced Genie 2, an AI model that generates playable 3D worlds from single images or text prompts. Users can explore these worlds using keyboard and mouse for up to a minute, with the AI maintaining memory of previously visited areas. The system handles character movement, physics, lighting, and NPC behaviors.

Key details:

• Generates 720p resolution worlds that stay consistent for up to 60 seconds

• Remembers and recreates off-screen areas when players return to them

• Trained on video data to learn physics, lighting, reflections, and smoke effects

• Works with DeepMind's SIMA agent to follow natural language commands like "open the blue door"

Why it matters: This rapid succession of innovations highlights fierce competition in 3D AI research. World Labs' announcement barely had time to resonate before Genie 2 entered the spotlight. These breakthroughs open exciting possibilities for AI training and game prototyping.

QUICK NEWS

Quick news

• Google API customers can generate media with Veo and Imagen 3

• xAI plans massive expansion of AI supercomputer in Memphis

• Anduril partners with OpenAI for drone defense

TOOLS

🥇 New tools

• Pointer - AI editing copilot for Google Docs

• Athina - Build, test and monitor AI apps and agents