Nuntia digest

Technical AI news, updated 13 November at 11:06 am IST. This page has no script. The full page is at https://news.naklitechie.com; feeds are at feed.xml.

Must read this week

  1. OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience.… — On Continual-ARC, our new benchmark at @letsunifyai , none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both. (Daniel Lenton (X); @DanielLenton1 on X; 1399 likes)
  2. InstantHMR : Real-Time Human Mesh Recovery on Your Phone — Real-time 3D human mesh recovery from a single image on mobile devices is a challenging problem with many applications, including robotics, biomechanics, AR/VR, and exergames. Current state-of-the-art models are either too slow to run efficiently on mobile devices or sacrifice accuracy for speed. (arXiv; Mohamed Adjel + 2 authors; 2 upvotes; code released)
  3. tencent/Qwen3.8-27B-DSpine-b16 — DSpine drafts a whole block of tokens in one pass, passing each position's predicted features to the next position at every layer, so each draft token is conditioned on its predecessors without a second step. (Hugging Face; text-generation; licence apache-2.0; 212 likes)
  4. Today, we're releasing the highest accuracy Qwen3.8-27B GGUFs.… — Our new Dynamic V3.1 GGUFs lead by 5–10% on average across Terminal Bench, DeepSWE & more. We’re working on even better version for Qwen4 Run on 12GB RAM. GGUF: https://huggingface.co/unsloth/Qwen3. (Unsloth AI (X); @UnslothAI on X; 3259 likes)
  5. Persistent Agents are the next unlock of AI usage Our spend on AI is now 45% persistent agents Unfortunately for Anthropic, Claude Tags… — sucks OpenAI's Slack agent is even worse We primarily use Perplexity Computer, which uses a lot of different models Perplexity's Harness mogs (Dylan Patel (X); @dylan522p on X; 1150 likes)
  6. Introducing 𝙻𝚒𝚚𝚞𝚒𝚍𝙲𝚘𝚍𝚎𝚌, our ~74M-parameter streaming audio codec powering LFM3-Audio-2B and LFM3-Omni-2B.… — An audio codec turns speech into tokens a language model can predict, then converts those tokens back into audio. We built ours around two requirements: tokens (Liquid AI (X); @liquidai on X; 1101 likes)
  7. Kinetic — an exact-output local inference engine (Velocity) (Picked; Velocity (Jakub Kozica, CEO and founder, author of the technical report))
  8. AtomicBot-ai/atomic-agent, Fusion mode (MIT, 3,361 stars, TypeScript): a cloud orchestrator delegates bounded task waves to local workers, with models by role (plan, build, review, explore, utility), contracts, a judge, escalation, replanning, branch isolation and turn budgets (Picked; AtomicBot-ai (Atomic Agent))
  9. LibrePaper (LibrePaper/librepaper, MIT, Rust, 102 stars, created 2 Sep 2026) — a free, self-hostable Overleaf alternative: live multi-person editing and comments for LaTeX, Typst, Markdown, HTML and Quarto, compiled in the browser by WebAssembly engines; Git sync by a CLI and a GitHub Action, agent tokens (Picked; Vincent Arel-Bundock (@VincentAB), a professor in Montreal who writes R packages (marginaleffects, modelsummary), per his profile line)
  10. dphnAI/nf12 (AlpinDale, Apache-2.0) — lossless fixed-length 12-bit BF16 weights for NVIDIA GPUs, with fused decode-in-GEMM kernels for dense, MoE and embedding layers; 1.25× Qwen3-8B serving and −25% weight memory on an RTX PRO 6000 (Picked; Alpin Dale (AlpinDale, dphnAI))

Models

  1. VideoRebirthAI/hyperflow-h3-flash (Hugging Face; quantised from MiniMaxAI/MiniMax-H3; image-text-to-video; licence other)
  2. tencent/Qwen3.8-27B-DSpine-b16 — DSpine drafts a whole block of tokens in one pass, passing each position's predicted features to the next position at every layer, so each draft token is conditioned on its predecessors without a second step. (Hugging Face; text-generation; licence apache-2.0; 212 likes)
  3. MiniMaxAI/MiniMax-H3-Image — November 11, 2026 — Google enters the decision model market. At AI Engineer, Cloudflare showed which parts of a build failure a classifier could handle. TypeSafe AI showed off their production customers. (Hugging Face; text-to-image; licence other; minimax-h3)
  4. PrunaAI/P-H3-Image-Comfy (Hugging Face; quantised from PrunaAI/P-H3-Image; text-to-image; licence other)
  5. Comfy-Org/MiniMax-H3-Image (Hugging Face; text-to-image; licence other; diffusion-single-file)
  6. ccore/Qwen3.5-4B-3D-Coder (Hugging Face; licence apache-2.0; GGUF; 1.1k likes)
  7. Zyphra/ZONOS3-Mini (Hugging Face; text-to-speech; licence apache-2.0; zonos3)
  8. mradermacher/gemma-4-31B-it-ARA-heretic-GGUF (Hugging Face; quantised from richardyoung/gemma-4-31B-it-ARA-heretic; licence apache-2.0; transformers)
  9. ConwayResearch/Underdog-Woof (Hugging Face; text-to-speech; licence other; numpy)
  10. Scantist/GLM-5.3-SLIM-GGUF (Hugging Face; quantised from zai-org/GLM-5.3; text-generation; licence other)

Papers

  1. BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model — Driving world models aim to simulate how a scene unfolds under different ego actions. This requires plausible motion of surrounding vehicles as well as precise ego-trajectory control. (arXiv; Junyoung Seo + 5 authors; 21 upvotes)
  2. On the Design of Hyper-Connections: Simplification, Stability, and Scalability — A systematic study of Hyper-Connections architectures identifies the simpler efficient design SimpleHC, which removes residual mixing and uses shared memory access to match mHC performance while adding minimal training overhead and improving stability. (arXiv; Xiangdong Zhang + 10 authors; 13 upvotes; code released)
  3. Duo: A Reciprocal World Model for Full-Duplex Interaction — A reciprocal world model enables real-time full-duplex audiovisual interaction by jointly modeling intent and realization using autoregressive reasoning, synchronized attention, and dual rollout distillation. (arXiv; Hengyuan Cao + 19 authors; 8 upvotes)
  4. G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation — Discrete autoregressive models offer a scalable framework for image generation. However, they face a fundamental trade-off: long token sequences preserve representation quality but slow generation, while aggressive spatial compression reduces representation capacity. (arXiv; Xu Huang + 10 authors; 6 upvotes; code released)
  5. WorldDraft: If Code Builds the World, Can Video Models Just Render It? — Multimodal coding agents decouple 4D world construction from visual generation, repurposing video models from world generators toward generative renderers. This division of labor raises a fundamental question: if code builds the world, can video models just render it? (arXiv; Zhiyu Hou + 9 authors; 3 upvotes; code released)
  6. Scaling Mathematical Proof Supervision from Structured Textbooks — A large-scale corpus of structured proof-supervision records derived from university mathematics textbooks improves language model mathematical reasoning by 16.7 percentage points over a base model when used for fine-tuning. (arXiv; Youqing Fang + 9 authors; 42 upvotes)
  7. SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering — Low-bit quantized LLM agents can recover much of their task success through closed-loop guidance from a full-precision teacher, followed by unassisted student training on corrected and reference turns. (arXiv; Genming Cui + 5 authors; 33 upvotes; code released)
  8. InstantHMR : Real-Time Human Mesh Recovery on Your Phone — Real-time 3D human mesh recovery from a single image on mobile devices is a challenging problem with many applications, including robotics, biomechanics, AR/VR, and exergames. Current state-of-the-art models are either too slow to run efficiently on mobile devices or sacrifice accuracy for speed. (arXiv; Mohamed Adjel + 2 authors; 2 upvotes; code released)
  9. Incentivizing Self-Learning in Large Language Models via Reinforcement Learning — Large language models (LLMs) draw on parametric knowledge and retrieval-augmented generation (RAG) to solve knowledge-intensive tasks. (arXiv; Lingkai Kong + 8 authors; 28 upvotes)
  10. Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents — Web-facing AI agents repeatedly access webpages, documents, repositories, and tool responses, yet reading interfaces often expose entire objects when only sparse evidence is needed. This over-reading increases token and latency costs and can dilute relevant evidence. (arXiv; Zedong Liu + 7 authors; 19 upvotes)

Engines and tools

  1. llama.cpp: 10 builds in 7 days, latest b12343 (GitHub; ggml-org/llama.cpp; 10 releases in 7 days)
  2. I'm making LLMs work on small and old devices (Hacker News; 314 points and 104 comments on Hacker News)
  3. vllm v0.34.0rc2 released (GitHub; vllm-project/vllm; 3 releases in 7 days)
  4. unsloth Dynamic v3.1 GGUFs + Performance Fixes released (GitHub; unslothai/unsloth; 3 releases in 7 days)
  5. Kinetic — an exact-output local inference engine (Velocity) (Picked; Velocity (Jakub Kozica, CEO and founder, author of the technical report))
  6. Diffusers 0.43.0: MiniMax-H3-Image, Pika H3 Flex, P-H3-Image, Kandinsky S2V, and PAN 2 released (GitHub; huggingface/diffusers; 1 releases in 7 days)
  7. AtomicBot-ai/atomic-agent, Fusion mode (MIT, 3,361 stars, TypeScript): a cloud orchestrator delegates bounded task waves to local workers, with models by role (plan, build, review, explore, utility), contracts, a judge, escalation, replanning, branch isolation and turn budgets (Picked; AtomicBot-ai (Atomic Agent))
  8. LibrePaper (LibrePaper/librepaper, MIT, Rust, 102 stars, created 2 Sep 2026) — a free, self-hostable Overleaf alternative: live multi-person editing and comments for LaTeX, Typst, Markdown, HTML and Quarto, compiled in the browser by WebAssembly engines; Git sync by a CLI and a GitHub Action, agent tokens (Picked; Vincent Arel-Bundock (@VincentAB), a professor in Montreal who writes R packages (marginaleffects, modelsummary), per his profile line)
  9. litellm: 10 builds in 7 days, latest v1.111.0-dev.1 (GitHub; BerriAI/litellm; 10 releases in 7 days)
  10. pytorch: 10 builds in 7 days, latest trunk/97730ab49ff79c5a8d60497192a5d226152e1b19 (GitHub; pytorch/pytorch; 10 releases in 7 days)

Business of AI

  1. The FBI says North Korean hackers stole $400M+ in crypto in the last 18 months using "FaceStealer", a scheme involving fake video calls and AI deepfakes (A.J. Vicens/Reuters) (Techmeme)
  2. Nectar launches an AI agent to automate creator discovery and outreach for brands — Nectar's new AI agent automates influencer marketing, helping brands discover creators, manage outreach and partnerships, and track campaign results. (TechCrunch AI)
  3. Anthropic IPO (Hacker News; 873 points and 839 comments on Hacker News)
  4. Former Nvidia CEO Huang had a relationship with a second employee, sources say (Hacker News; 182 points and 216 comments on Hacker News)
  5. United Airlines blocks AI agents from booking flights, citing cybersecurity risks — The airline said it will not allow customers to use AI agents to book flights and manage reservations. (TechCrunch AI)
  6. Meta launches an outbound calling feature for Muse, such as to businesses, available for paid US subscribers; Gemini and startup Instinct offer a similar option (Sarah Perez/TechCrunch) (Techmeme)
  7. Mirendil, founded by former Anthropic researchers to build self-improving AI, raised a $1B Series A led by Kleiner Perkins and Clearlake Capital (Sarah Klearman/Wall Street Journal) (Techmeme)
  8. Sources: Nvidia faced pressure to boost its 25% residual value support backstop on a planned $500B AI financing vehicle with asset managers before Huang's exit (Bloomberg) (Techmeme)
  9. New York-based Hatz AI, which makes tools for managed services providers to help small businesses adopt AI, raised a $24M Series A (CJ Fairfield/CRN) (Techmeme)
  10. Analysis of open weights: six models from DeepSeek, Moonshot, MiniMax, and Ant Group released since April carry fingerprints of export-controlled Blackwell GPUs (Will DePue/Unemployment Labs) (Techmeme)

Posts

  1. I'll be giving a talk tomorrow at the SGLang Summit at 10am! Oftentimes, I find explanations in ML infra to be unsatisfying.… — A common failure mode is that they'll explain the *literal* tradeoffs but fail to provide a mental model for how to reason about the choice. This talk (Horace He (X); @cHHillee on X; 217 likes)
  2. Inkling 2 Small from Thinking Machines ( @thinkymachines ) is now on Ollama with zero data retention! The 594B mixture-of-experts model… — with 1M context window demonstrates strong visual interpretation and creation, agentic knowledge work, and finance and quantitative reasoning. (ollama (X); @ollama on X; 130 likes)
  3. We optimized DeepSeek-V4.1-Flash on vLLM for Vera Rubin NVL72 and reached 58× the throughput of GB300 NVL72 at 1.7 million tokens/s/GPU and… — 387 tokens/s/user on SemiAnalysis AgentX. Across the curve: 2.52 million tokens/s/GPU at peak throughput (2.7× GB300) 1.7 million (vLLM (X); @vllm_project on X; 160 likes)
  4. Introducing 𝙻𝚒𝚚𝚞𝚒𝚍𝙲𝚘𝚍𝚎𝚌, our ~74M-parameter streaming audio codec powering LFM3-Audio-2B and LFM3-Omni-2B.… — An audio codec turns speech into tokens a language model can predict, then converts those tokens back into audio. We built ours around two requirements: tokens (Liquid AI (X); @liquidai on X; 1101 likes)
  5. 🚨 Our Generative AI Lab is releasing Primer: an open source agent-driven system where a teacher talks through lessons & gets back a… — playable simulation for their students, grounded in pedagogy. We hope educators & researchers build on it. Paper/prompts/code are open: gail.wharton.upenn.edu/primer/ (Ethan Mollick (Bluesky); Ethan Mollick on Bluesky; 89 likes)
  6. Meet Gemini 4 Hydrogen: built for rapid coding, multimodal synthesis, and everyday workflows.… — It finishes tasks 2.5x faster than 3.8 Flash and surpasses Argon on the @ArtificialAnlys Intelligence Index. (Google DeepMind (X); @GoogleDeepMind on X; 6322 likes)
  7. I'm not sure if people realize this, but what has been done to Jensen by the US is significantly WORSE than what happened to Jack Ma, for a… — comparable defiance of the state. Jack got "disciplined", but kept a position in Alibaba, and there's been no lurid character assassination. (Teortaxes (DeepSeek 推特铁粉 2023 – (X); @teortaxesTex on X; 2071 likes)
  8. This whole thing sucks, and I wish none of it had happened.… — This whole thing sucks, and I wish none of it had happened. It’s a really sad day! But I want to make one thing clear: we’ll keep fighting for open-source AI, harder than ever. This is too important to slow down now. Let’s go! (clem (X); @ClementDelangue on X; 8492 likes)
  9. Sad but still bullish on @nvidia ! I’m seeing a lot of people criticizing its structure and the lack of a clear CEO successor.… — Having seen it from up close, I think a lot of this criticism is ill-informed and overly simplistic (as often happens when a company is run differently (clem (X); @ClementDelangue on X; 1009 likes)
  10. Our initial public offering: https://ipo.anthropic.com Legal disclaimer: Anthropic has filed a registration statement (including a… — prospectus) with the SEC for the offering to which this communication relates. Before you invest, you should read the prospectus in that registration (Anthropic (X); @AnthropicAI on X; 9636 likes)

Blogs and discussion

  1. Brave Search API now accepts agentic payments via the x402 protocol (Hacker News; 18 points and 4 comments on Hacker News)
  2. Nvidia employees petition to bring back Jensen Huang (Hacker News; 292 points and 226 comments on Hacker News)
  3. Lazarus: a closed-form weight edit that turns dormant neurons into trigger-activated payload writers — Microsoft researchers have discovered “Lazarus,” a backdoor technique that can reprogram dormant neurons in an AI model, causing an otherwise normal AI agent to execute malicious commands when a hidden trigger appears. The threat is especially serious for AI agents with tool or terminal access. (Picked; Yuchen Zeng, Giorgio Severi, Blake Bullwinkel, Eric Horvitz, Dimitris Papailiopoulos (Microsoft))
  4. Gemini 4 Hydrogen (Hacker News; 495 points and 317 comments on Hacker News)
  5. Chinese frontier AI models carry the fingerprint of export-controlled GPUs (Hacker News; 233 points and 62 comments on Hacker News)
  6. Gemini 4 Hydrogen and Orchestrator (Hacker News; 357 points and 96 comments on Hacker News)
  7. Edge0-145B: Frontier AI Capabilities on Your Own Computer (Hacker News; 158 points and 39 comments on Hacker News)
  8. Deploy frontier AI models on AWS Trainium with SGLang-Neuron — SGLang-Neuron is now publicly available through an AWS Deep Learning Container. With SGLang-Neuron, you can serve models and generate reinforcement learning rollouts with SGLang on AWS Trainium using the same SGLang interface you already know, all running on native PyTorch. (AWS Machine Learning)
  9. How Natera designed an agentic AI system for lab ops with Amazon Bedrock AgentCore — Natera built SWARM, an agentic AI system on Amazon Bedrock AgentCore that captures institutional lab-operations knowledge and deploys it at scale. (AWS Machine Learning)
  10. Microsoft Research Focus – Edition #98 — What's new in research at Microsoft? Host Anson Ho explores the latest, including AI-powered universal translation, a breakthrough in a 31-year-old mathematics challenge, and advances making AI coding agents faster and more efficient. (Microsoft Research)