Parse notes
Sponsored
You know that moment when a company goes from “never heard of” to “suddenly everywhere? That's RAD Intel's [4] right now. They're helping global brands like Hasbro and Omnicom truly understand the audiences they want to attract in real time — not just guess. It's AI that actually works. Pre-IPO, and legit.
OpenAI has raised $40B at a $300B valuation to support AI research, infrastructure scaling, and its growing user base and partnered with SoftBank to expand AGI development.
Google DeepMind has revealed its Gemini Robotics models, which extend Gemini 2.0 with fine-tuning for physical actions.
OpenAI has launched OpenAI Academy, a free platform that offers a range of AI courses, from beginner lessons to advanced topics like AI safety and governance. Catering to various audiences, the platform aims to broaden access to AI education and encourage critical thinking about AI's societal impacts. Initial reactions highlight the platform's accessibility and comprehensive approach to demystifying AI for a global audience.
Following the recent trend in tracking systems, this work adds dense pixel tracking to improve long term segmentation with Dino and SAM2.
A strong 32B model trained with RL on top of Qwen that surpasses many substantially larger models on agentic coding tasks.
There is a huge open question in RL reasoning efforts whether or not a strong enough base model is required and necessary for emergent reasoning. This work explored many cases of scaling RL on a base mode. It found that strong base models are very helpful for reasoning convergence.
Easi3R is a 3D vision system that can better estimate 3D scenes with high motion. It improves substantially over previous methods in full scene reconstruction. This was achieved by masking moving objects and learning those and the background separately.
SEED-Bench-R1 is a benchmark that evaluates post-training methods like RL and SFT for multimodal LLMs in complex video-based tasks, showing RL's strengths in perception and data efficiency but challenges in logical coherence.
UniOcc is a unified framework for forecasting and flow prediction in driving scenarios built for multi-dataset training and cross-domain evaluation in both real and synthetic environments.
Nexthop AI secured $110 million led by Lightspeed Venture Partners to enhance networking solutions for hyperscalers, focusing on cost and power efficiency. The round also saw participation from Kleiner Perkins, WestBridge Capital, Battery Ventures, and Emergent Ventures. CEO Anshul Sadana emphasized Nexthop AI's role in innovating alongside cloud companies.
Alibaba chairman Joe Tsai warns of a potential AI bubble due to massive investments in data centers without clear demand. Despite $52 billion committed to AI development, concerns are rising about overinvestment. Recent disruptions like Chinese startup DeepSeek highlight investor fears of overpaying in the AI sector.
Anthropic has introduced the Economic Index, which uses Claude Sonnet 3.7 to assess AI's influence on different job sectors. The model evaluates how AI affects productivity, automation risks, and labor market trends. This initiative aims to provide data-driven insights for policymakers and businesses navigating AI-driven economic shifts.
Amazon's Alexa Fund is expanding its investment focus to include AI startups, investing in companies like NinjaTech AI, Hedra, Ario, and HeyBoss.
A Dartmouth study found that AI therapy chatbot Therabot significantly reduced symptoms in participants with depression, anxiety, and eating disorders.
Databricks has developed Test-time Adaptive Optimization (TAO), a method that leverages reinforcement learning and synthetic data to improve AI models without the need for clean labeled data.