The Daily AI Briefing - 18/04/2025

The Daily AI Briefing

Welcome to The Daily AI Briefing, here are today's headlines! In the rapidly evolving world of artificial intelligence, today we're covering Google's new Gemini model with an innovative "thinking budget," a breakthrough in protein-design AI scaling laws, practical AI applications in Google Sheets, Meta's latest perception research, and other notable developments in the AI landscape. These stories highlight the continued acceleration of AI capabilities across multiple domains. Our first headline features Google's launch of Gemini 2.5 Flash, a hybrid reasoning AI that introduces a novel "thinking budget" feature. This new model matches OpenAI's o4-mini while outperforming Claude 3.5 Sonnet on reasoning and STEM benchmarks. The standout innovation is its "thinking budget" system that allows developers to optimize the balance between response quality, cost, and speed by allocating up to 24,000 tokens for complex reasoning. This controllable reasoning gives users the flexibility to activate enhanced thinking capabilities only when needed, making it more cost-effective for high-volume use cases. The model shows significant improvements over Gemini 2.0 Flash and is available through Google AI Studio and Vertex AI, with experimental integration in the Gemini app already underway. In biotech news, Profluent has announced ProGen3, a groundbreaking family of AI models for protein design that demonstrates the first evidence of AI scaling laws in biology. Their 46-billion parameter model, trained on an unprecedented 3.4 billion protein sequences, successfully designed new antibodies that match approved therapeutics in performance while being distinct enough to avoid patent conflicts. Perhaps more remarkably, the platform created gene editing proteins less than half the size of CRISPR-Cas9, potentially revolutionizing gene therapy delivery methods. Profluent is making 20 "OpenAntibodies" available through royalty-free or upfront licensing, targeting diseases affecting 7 million patients. If these scaling trends continue, Profluent's approach could transform drug and gene-editor design from years-long laboratory work into a faster, more predictable engineering problem. For productivity enthusiasts, Google Sheets is rolling out an exciting new AI formula feature that allows users to generate content, analyze data, and create custom outputs directly within spreadsheets. The implementation is remarkably straightforward – simply type =AI("your prompt") in any cell with specific instructions like summarizing customer feedback or analyzing data patterns. The formula can be applied to multiple cells by dragging the corner handle down a column, enabling batch processing. For more sophisticated workflows, it can be combined with standard functions like IF() and CONCATENATE(). This practical application of AI in everyday tools demonstrates how artificial intelligence is becoming increasingly accessible and useful for non-technical users. Meanwhile, Meta's FAIR research team has published five new open-source AI research projects focused on perception and reasoning. Their Perception Encoder achieves state-of-the-art performance in visual understanding tasks like identifying camouflaged animals. The team also introduced the Meta Perception Language Model and PLM-VideoBench benchmark for improved video understanding. Another notable project, Locate 3D, enables precise object understanding with a dataset of 130,000 spatial language annotations. Finally, their Collaborative Reasoner framework demonstrates that AI systems working together can achieve nearly 30% better performance compared to working alone. These research projects represent crucial building blocks for developing more capable embodied AI agents. In brief updates, OpenAI's new o3 model scored an impressive 136 on the Mensa Norway IQ test (116 in offline testing), surpassing Gemini 2.5 Pro for the highest recorded score. Additionally, UC Berkeley's Chatbot Arena AI model testing platform