Good morning. It’s Monday, December 23rd.

Did you know: AI generated videos of UAP are starting to be released online?

In today’s email:

• OpenAI’s o3 Model

• Meta’s Llama 4

• xAI’s Grok App

• Quantum Teleportation Over the Web

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

OpenAI’s o3 Model Shatters Programming Benchmarks

OpenAI’s o3 model significantly raised the bar in both reasoning and safety, stepping up from the o1. Built on a reinforcement learning framework, o3 is designed for goal-oriented tasks like programming and mathematics, constructing intricate reasoning chains similar to AlphaGo. Enhanced by scaled "train-time compute" and bolstered "test-time compute," it easily outpaces o1 in structured problem-solving.

On the safety front, o3 introduces "deliberative alignment," a technique that integrates OpenAI’s safety protocols during inference. This involves teaching models to consider ethical guidelines through synthetic data, ensuring responses remain aligned with safety standards while minimizing latency. The new approach promises safer, more reliable AI systems, reducing harmful outputs and enhancing benign query handling, as demonstrated by benchmarks like Pareto.

Though o3 is slated for full deployment in 2025, and is currently open to applicants in the research field.

And if you're curious why OpenAI jumped from o1 to o3, apparently they ran into a little copyright hiccup with 'o2'. Read more.

Meta to Launch Multiple Llama 4 Releases in 2025, Focused on Reasoning and Speech

Meta is gearing up for multiple releases of Llama 4 throughout 2025, sharpening its focus on advanced reasoning and voice interaction. With over 650 million downloads and 85,000 variations, Llama has rapidly gained traction, outpacing proprietary models.

The Llama 3.x updates laid the groundwork with multimodal capabilities and resource efficiency, but Llama 4 will demand significantly more computational power, amplifying its reasoning and voice capabilities for applications in business automation and customer service.

Meta expects AI interactions to shift from text to voice, a transition already underway with 600 million monthly users of Meta AI. However, Llama 4 will initially be unavailable in Europe due to regulatory concerns. With a growing web of hardware and software partners, Meta continues to fortify Llama’s position in the open-source AI ecosystem. Read more.

xAI Trials Standalone iOS App for Grok Chatbot

xAI, Elon Musk’s AI initiative, is testing a standalone iOS app for its Grok chatbot, once exclusive to 𝕏. Now in beta, the app pulls live data from X and the web, offering an array of generative AI tools — from text rewriting and summarising to Q&A and image generation. With a focus on accuracy, curiosity, and utility, Grok promises users a highly interactive experience.

The company is also prepping Grok.com, a site that will further broaden access to the platform. Grok’s image model, praised for its photorealistic renderings, offers near-total freedom, even for content featuring public figures or copyrighted materials. After being gated for 𝕏’s premium users, Grok is now free to all, marking a significant push for more widespread adoption. Read more.

Quantum Teleportation Meets Internet Traffic in First Demo

Northwestern engineers have made history by performing quantum teleportation over a fiber optic cable already carrying standard Internet traffic. The feat blends quantum communication with existing infrastructure, cutting out the need for costly new networks. Harnessing quantum entanglement, they transferred quantum states between distant photons, sidestepping the physical travel of data.

The team tackled the challenge of photon interference by strategically placing their quantum particles in less congested wavelengths, ensuring clean transmission despite the high-speed traffic around them. The test, which spanned 30 km, confirms that quantum teleportation can coexist with traditional Internet traffic, opening the door for secure, scalable quantum networks without the need for dedicated fiber. Next up: longer distances and real-world deployment. Read more.

• ElevenLabs launches Flash, its fastest text-to-speech AI yet

• tldraw’s 'Computer' Turns Natural Language into AI Workflows

• Google is expanding Gemini's in-depth research mode to 40 languages

• OpenAI’s search lead leaves after seven months

• Researchers open-source Genesis simulation platform for training robots

• Trump AI and crypto 'czar' role is being scaled back as David Sacks clings to Silicon Valley commitments

• Arizona's getting an online charter school taught entirely by AI

• Sriram Krishnan named Trump's senior policy advisor for AI

• GitHub welcomes o1 model as OpenAI upgrades ChatGPT memory ​

• Hugging Face Shows How Test-Time Scaling Lets Small Models Outthink the Giants

• Google's ‘AI Mode’ to Merge Search and Chatbot

📄 Parallelized Autoregressive Visual Generation

📄 Affordance-Aware Object Insertion via Mask-Aware Dual Diffusion ​​

📄 SCOPE: Optimizing Key-Value Cache Compression in Long-context Generation

📄 Offline Reinforcement Learning for LLM Multi-Step Reasoning

📄 How to Synthesize Text Data without Model Collapse?

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!