TLDR AI 2024-12-24

Genesis, an open-source simulation platform, accelerates robot training 430,000 times faster than real-world conditions using text-generated 3D worlds. Developed by a team led by Carnegie Mellon University, it processes physics calculations 80 times faster than existing simulators using standard GPUs, facilitating rapid neural network training for robotics. Genesis uses Python and offers an accessible, non-proprietary solution for creating dynamic, physics-based environments without manual programming effort.

Waymo's autonomous vehicles significantly outperform human drivers in safety, showing an 88% reduction in property damage and a 92% reduction in bodily injury claims according to a study that analyzed 25.3 million miles of driverless operation. Compared to new cars with advanced safety tech, Waymo still achieves an 86% reduction in property damage and 90% in bodily injury claims. The data supports the potential of autonomous vehicles to reduce traffic fatalities, though more extensive data is needed for comprehensive comparisons.

Google DeepMind and OpenAI are intensifying their competition as OpenAI shifts towards a more commercial focus. This rivalry is driving both companies to innovate and push AI technology boundaries.

Meta has released a number of multimodal video understanding models with strong long context video performance.

Most texture generation systems rely on pretrained 2D images diffusion models. This work tried to approach the problem directly in UV texture space, which resulted in a set of inductive biases that strongly improved the underlying results of the system.

Stag-1 is a 4D driving simulation that reconstructs real-world scenes and generates lifelike videos from any perspective.

One of the NanoGPT speed run record holders has written an in depth blog post about the newest Muon optimizer that is used in many of the winner runs.

The new Driv3R system revolutionizes 4D reconstruction for autonomous vehicles by eliminating slow global alignment processes.

Zyphra has written performant backwards kernels for AMD chips.

AI is rapidly transforming drug discovery, with biologics developers securing $1.6B in 2024 funding, indicating an expanding focus beyond small molecules. M&A activity in AI drug discovery is surging, with major big pharma players acquiring startups and bolstering in-house AI capabilities. This reflects a growing urgency to integrate AI in drug R&D to reduce costs and accelerate development.

Google's Quantum group's "Willow" is a 105-qubit superconducting chip that showcases advancements in error correction and a new quantum supremacy experiment. The chip demonstrates increased coherence times and gate fidelity, marking a significant milestone in quantum fault-tolerance. Despite its achievements, challenges remain in reaching fully fault-tolerant quantum operations, and verification of results continues to be complex.

OpenAI has made its digital editing space, Canvas, available to all ChatGPT users, integrating it with GPT-4o. Canvas enhances the chat interface by allowing real-time editing and running Python code directly. It's now a default feature in custom GPTs, providing advanced functionalities for users.

Google's Files app now supports Gemini, allowing Advanced subscribers to ask questions about PDFs via an "Ask about this PDF" button.

Perplexity has closed a $500 million funding round led by Institutional Venture Partners that values the company at $9 billion.

Apple is collaborating with Broadcom to develop its first server chip for AI applications.