DeepMind AGI Safety 🦺, OpenAI Nonprofit Guidance Commission 👥, Hate Speech Benchmark 📚

Google DeepMind has outlined its perspective on building safe and secure artificial general intelligence, emphasizing oversight and safeguards ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2025-04-04 Agentic AI is Changing Everything We Know About Process Orchestration – Learn How (Sponsor) Traditional automation is no match for today's complexity. Agentic workflows—autonomous, goal-driven agents collaborating to achieve business goals—require more than prompt engineering. They need a powerful workflow engine to orchestrate tasks, decisions, and outcomes at scale—with humans in the loop when it matters most. The latest webinar by Orkes explored how combining agentic AI and process orchestration can enable businesses to integrate AI-driven decision-making, streamline workflows, and scale with ease. Tune in to learn: Why Traditional BPMN Tools Fff rations to financial services Watch the recording 🚀 Headlines & Launches DeepMind's Approach to AGI Safety (11 minute read) Google DeepMind outlines its perspective on building safe and secure artificial general intelligence, emphasizing oversight and technical safeguards as AGI capabilities progress. OpenAI just made its first cybersecurity investment (3 minute read) OpenAI invested in Adaptive Security, a startup using AI to simulate and train employees against social engineering hacks. Adaptive Security secured $43M in Series A funding, with plans to bolster its platform amid growing AI threats. Co-founded by veteran entrepreneur Brian Long, the startup has quickly amassed over 100 clients since launching in 2023. OpenAI Nonprofit Guidance Commission (3 minute read) OpenAI is forming a new expert commission to shape how its philanthropic arm supports communities using AI, aiming to align AI innovation with real-world nonprofit needs. 🧠 Research & Innovation Articulated Kinematics Distillation from Video Diffusion Models (18 minute read) This work introduces Articulated Kinematics Distillation (AKD), a framework that leverages skeleton-based animation and generative diffusion models to produce high-fidelity, physically plausible character motions with reduced complexity. It ensures structural consistency and outperforms existing methods in 3D coherence and expressive motion quality by using Score Distillation Sampling to guide joint-level control. Enhanced LoRA-based Fine Tuning (18 minute read) MetaLoRA introduces dynamic parameter generation using meta-learning principles, enhancing the flexibility and task-awareness of LoRA-based fine-tuning strategies. Backdoor Attacks in CLIP (19 minute read) CLIP models are highly vulnerable to poisoning backdoor attacks, achieving nearly 100% attack success with minimal poisoned data. An efficient detection method is to use local outlier detection to uncover unintentional backdoors in existing datasets. 🧑‍💻 Engineering & Resources Flash Attention reimagined with Kvax (Sponsor) Nebius open-sourced Kvax: a JAX-based Flash Attention implementation optimized for long-context training. Achieve faster, denser packing and state-of-the-art performance with context parallelism and efficient doc mask computation—built by Nebius' AI R&D team. Read more here and explore the code Large Small Net (GitHub Repo) A new family of lightweight vision models inspired by the dynamic heteroscale capability of the human visual system, i.e., "See Large, Focus Small". LSNet achieves state-of-the-art performance and efficiency trade-offs across various vision tasks. It introduces a new type of convolution kernel. Pplx Cuda Kernels (GitHub Repo) Perplexity has released some of its MoE kernels, which outperform DeepSeek at scale while being slightly more flexible and less opinionated about the MoE architecture. HateBench for Evaluating Hate Speech (6 minute read) HateBench provides a framework for evaluating hate speech detection models on LLM-generated content along with a manually annotated dataset and code for analyzing adversarial and stealthy hate campaigns. 🎁 Miscellaneous Zonos TTS (8 minute read) A compelling Apache 2.0 model for speech generation and voice cloning. It has multi language support and expressive real time generation. Hugging Face's AI Agents Course (14 minute read) Hugging Face released an AI agents course today. This free course will take you on a journey, from beginner to expert, in understanding, using, and building AI agents. Google is shipping Gemini models faster than its AI safety reports (4 minute read) Google has launched an AI reasoning model, Gemini 2.5 Pro. The model leads in coding and math capabilities, but Google hasn't released safety reports yet. Google plans to publish these reports after collecting feedback from experimental releases, though this approach raises transparency concerns. Despite pledges for transparency, Google seems to prioritize speed in model deployments, diverging from industry norms for responsible AI practices. ⚡ Quick Links Code with Claude Developer Conference (1 minute read) Anthropic has announced its first developer event focused on hands-on sessions and best practices for building with Claude - it is set to take place in May in San Francisco. The LLM Course from Hugging Face (2 minute read) Hugging Face has revamped its popular NLP course into a broader LLM curriculum with expanded chapters on fine-tuning, reasoning models, and modern AI agent workflows. Devin, the viral coding AI agent, gets a new pay-as-you-go plan (2 minute read) Cognition launched a $20 entry-level plan for its AI programming tool, Devin, which has been improved to handle complex coding tasks better. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://refer.tldr.tech/0b6a6dc1/2 Track your referrals here. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian & Andrew Carr Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.