Handy AI is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
what to know for now
🧠 OpenAI launches o3 and o4-mini models with breakthrough reasoning. OpenAI's new o3 model achieves remarkable performance on reasoning benchmarks, scoring 88% on the ARC-AGI test and 96.7% on the 2024 AIME math competition. Alongside o3, OpenAI released o4-mini, a more efficient model with a 128K context window that scores 14.3% on Humanity's Last Exam, outperforming both Claude 3.7 Sonnet and DeepSeek R1. These models combine advanced reasoning with comprehensive tool usage capabilities, including web search, Python code execution, and visual analysis, marking a significant step toward more agentic AI systems.
🔍 Anthropic launches Claude Research and Google Workspace integration. Anthropic put out a new "agentic" Research feature for Claude that conducts multiple interconnected searches and provides cited results, alongside deep integration with Google Workspace services like Gmail and Drive. Read more
💡 Google's Gemini 2.5 Flash introduces cost-saving 'thinking budgets'. Google has launched Gemini 2.5 Flash with an innovative "thinking budget" feature that allows developers to control how much computational power is used for reasoning, reducing costs by up to 600% when turned down. The model achieves strong performance on key benchmarks, scoring 12.1% on Humanity's Last Exam and 78.3% on GPQA diamond. Read more
🔬 OpenAI announces GPT-4.1 with enhanced capabilities. OpenAI has unveiled GPT-4.1, featuring significant improvements in coding, instruction following, and long-context understanding. The model achieves state-of-the-art performance on SWE-bench, solving 55% of problems, and demonstrates enhanced agentic capabilities through improved tool usage and planning. Read more
Jake's Take: This study from Tsinghua University challenges the conventional wisdom around reinforcement learning for LLMs. The researchers show that while RLVR (Reinforcement Learning with Verifiable Rewards) improves sampling efficiency, it doesn't actually expand a model's reasoning capabilities beyond what's already present in the base model.
This finding suggests that current RL approaches may be optimizing for efficiency rather than truly expanding model capabilities. The researchers also found that distillation, unlike RL, is a more promising approach that can genuinely introduce new knowledge and expand reasoning boundaries.
what to know for later
🤖 Hugging Face acquires Pollen Robotics to democratize AI-powered robots. Hugging Face has acquired Pollen Robotics, the maker of the open-source humanoid robot Reachy 2, marking the company's fifth acquisition and expansion into physical AI. The $70,000 Reachy 2 robot, already in use at institutions like Cornell and Carnegie Mellon, combines advanced hardware with LeRobot, Hugging Face's open-source robotics library that has gained over 12,000 GitHub stars in its first year. Read more
🔄 OpenAI explores social network to rival X. OpenAI is developing a social network prototype that integrates ChatGPT's image generation capabilities into a social feed, with CEO Sam Altman actively seeking feedback on the project. The move comes after Musk's unsuccessful $97.4 billion bid to acquire OpenAI, and positions the company to compete directly with both X and Meta in the social media space. Read more
💻 OpenAI in talks to acquire Windsurf for $3 billion. OpenAI is reportedly in advanced negotiations to acquire Windsurf (formerly Codeium), a popular AI coding assistant that competes with Cursor and other coding tools. The potential deal, which would be OpenAI's largest acquisition to date, comes shortly after the company's $40 billion funding round at a $300 billion valuation. Read more
🎬 Fal.ai's Kling 2.0 turns static images into cinematic videos. Fal.ai has released Kling 2.0, their latest image-to-video AI model that can transform still images into dynamic 5-second video clips with sophisticated motion and cinematic quality. Read more
