NVIDIA Nemotron 3 Embed Ranks #1 on RTEB

ginobefun (@hongming731) 转发了 BestBlogs (@BestBlogsDev) 的帖子： BestBlogs Daily · 07-17 # Nemotron 3 Embed / Kimi K3 / Inkling / Fable / Bun [1] ★ Deep Dive · NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB， Advancing Agentic Retrieval NVIDIA open-sources Nemotron 3 Embed, a retrieval family whose 8B model ranks #1 on RTEB (78.5%) with a 32k context and multilingual plus code retrieval, plus a 1B and a Blackwell NVFP4 build that doubles throughput at 99%+ of BF16 accuracy. The payoff is agentic: better retrieval surfaces evidence earlier, so agents loop and reason less, burning fewer downstream tokens — the foundation under RAG and agent memory. Source: Hugging Face - Blog https://t.co/yRfqaOVunf [2] ★ Deep Dive · How to Make Your AI Agent's Actions Reliable (No Code) Calling an API is the easy half of an agent; the hard part is firing it only when it should, and carrying the right value forward. This no-code guide argues a prompt is not a boundary — the model's choice to act is a probabilistic judgment that can be coaxed or prompt-injected. The fix splits each action into model judgment versus server guarantees: a 'Run only when' gate checked before any request, and 'Save to memory' to carry one value forward, so the chat cannot talk past it. Source: Hacker News - Newest: "AI Agent" https://t.co/mnQ79S9s1L [3] ★ Deep Dive · The Pulse: What can we learn from Bun's rapid Rust rewrite with AI? Jarred Sumner used Anthropic's Fable to rewrite Bun from Zig to Rust — 535K lines, 11 days, $165K — turning a year-long migration into a sprint. The method wasn't 'rewrite this, zero mistakes': 3 hours of prep yielded a 600-line porting guide, a trial run was adversarially reviewed, then work split across 64 parallel agents. The takeaway: a thoroughly-tested project plus disciplined orchestration makes an unthinkable rewrite feasible — not AI doing it solo. Source: The Pragmatic Engineer https://t.co/xxetGahYEA [4] Welcome Inkling by Thinking Machines Inkling is a large open multimodal model (~1T params, 1M context) that natively accepts image, text, and audio inputs, with agentic capabilities and day-0 support in major inference engines. Source: Hugging Face - Blog https://t.co/Pv9R5nTLQE [5] Computer-Use 2.0: Agents Just Got Multi-Cursor — Francesco Bonacci， Cua [Video] A conference talk that maps computer-use agents from foreground screenshot loops to background execution, then connects reliable evaluation and sandbox infrastructure to scalable agent training. Source: AI Engineer https://t.co/jW1vUMgN8M [6] Forward Deployed Engineering at Cursor — Pauline Brunet [Video] Cursor's forward deployment leader offers a practical framework for deciding where FDE belongs, how to scope it around measurable change, and how to build a team that improves both customer adoption and the product roadmap. Source: AI Engineer https://t.co/d6FCOvsSXf [7] Kimi K3， and what we can still learn from the pelican benchmark Simon Willison reviews the newly released Kimi K3 model from Moonshot AI, runs his signature 'pelican riding a bicycle' test, and reflects on the test's evolving utility as a quick model evaluation tool. Source: Simon Willison's Weblog https://t.co/1RmvxJymCQ [8] The Archaeologist’s Copilot This article presents a systematic approach to using AI as an 'archaeologist' rather than a 'tourist' when dealing with legacy codebases, demonstrating through a real case study how to analyze, contain, and gradually modernize a 2005-era Java project without breaking its fragile functionality. Source: Martin Fowler https://t.co/AubO6Lcoye [9] Danau5tin/ai-trains-ai: RL-training an AI agent to RL-train AI agents An RL agent learns to design AI training jobs, improving itself and generalizing to new tasks. Source: Hacker News https://t.co/EVrxRUpQqk [10] WTF Is the Context Layer? The Missing Infrastructure for Production Agents — Prukalpa Sankar [Video] Atlan founder Prukalpa Sankar explains why production AI agents need a shared, governed context layer that turns a company’s facts, expertise, norms, and feedback into reusable machine-usable knowledge. Source: AI Engineer https://t.co/hPQUvlZLRz --- https://t.co/Kyws10KILE · Discover high-quality content that truly fits you BestBlogs is an AI-powered personal reading assistant that helps you discover high-quality content that truly fits you. Follow the sources and topics you care about, and get a daily brief that fits you better every day. Try it and follow us. Read online: https://t.co/OdKrfzvn52 > **引用原帖 BestBlogs (@BestBlogsDev):** > https://t.co/zdGLIWi5TO > https://x.com/BestBlogsDev/status/2077908819792040345