Skip to stories

Vol. I · No. 86 · Independent daily intelligence

Signal

Papers and systems worth your time

Frontier AI is converging on one theme: inference-time control and efficiency matter as much as training. On the 3D side, world models and Gaussian splatting are moving from demos to streamable, editable, and trainable systems.

Topic:
Source:
Signal:

Today’s edition

The front page

23 stories to scan

  1. ● Top story

    Long-Context Generation Is a Sampling Problem

    This paper argues that long-generation degradation in open-weight LMs often comes from sampling dynamics under higher temperature, not just model weakness. It reframes the fix as inference-time decoding control rather than more training.

    Useful if you care about decoding behavior, temperature tradeoffs, and how generation policies shape long-horizon quality.

  2. ● Top story

    Streaming 3D Gaussian Splatting Worlds on the Web

    World Labs details a streamable level-of-detail system for 3D Gaussian splatting in Spark 2.0. The emphasis is on serving large splat worlds interactively over the web without giving up responsiveness.

    Good implementation read on LOD, streaming, and bandwidth-aware delivery for large neural 3D scenes.

  3. ● Top story

    Nimony: A New Compiler for Nim

    A Hacker News-discovered compiler project written for Nim and discussed heavily by readers. The thread centers on compiler design choices and implementation details rather than a language announcement.

    Strong systems read on compiler architecture, self-hosting, and build complexity.

  4. ● Top story

    Test-Time Compute for Generative Visual Reasoning

    The paper studies how multimodal models spend inference budget when they reason by producing sketches, pixel rollouts, or video-like intermediate states. It focuses on allocating diffusion steps more selectively instead of running a full image-generation schedule for every visual thought.

    Shows a concrete compute-allocation strategy for multimodal reasoning systems, with direct implications for latency and cost.

  5. ● Top story

    Atlas: A World Model for Spatial Intelligence

    World Labs introduces Atlas as a new omni world model for spatial intelligence. It sits in the same technical cluster as their world-generation work, with a focus on learned spatial representation.

    Relevant for world-model representation design and where spatial intelligence is heading.

  6. ● Top story

    Marble: A Multimodal World Model

    World Labs says Marble is a frontier multimodal world model now available broadly. The post is a core reference point for their text/image/video-to-world direction.

    Good context on the productization of world-model generation and multimodal scene understanding.

  7. ● Top story

    RTFM: A Real-Time Frame Model

    This research preview generates video in real time as the user interacts with it. The interesting part is the interactive frame-generation loop, not just the model output quality.

    Important if you care about low-latency generative video systems and interactive world models.

  8. ● Top story

    TrnRT: Open-Source LLM Inference on AWS Trainium

    An open-source inference engine targets AWS Trainium for LLM serving. The post is a practical signal on hardware-specific runtime work outside the CUDA default path.

    Worth reading for Trainium execution details, kernel/runtime constraints, and deployment tradeoffs.

  9. ● Top story

    NEBench R1: Nuclear-Domain Reasoning Benchmark

    This benchmark measures LLM competence in nuclear engineering with an emphasis on reasoning quality, not just final-answer accuracy. It targets a knowledge-heavy, safety-critical domain where shallow benchmarks are inadequate.

    Useful if you care about domain-specific eval design and reasoning-centric assessment.

  10. ● Top story

    Language Models Can Rewrite Their Own Weights

    The paper explores direct weight editing by language models rather than relying on prompts or retraining. The central claim is that models may be able to modify a smaller model’s parameters in-place to preserve changes beyond a single inference call.

    Interesting for weight-space adaptation, persistent edits, and the boundary between inference and training.

  11. ● Top story

    Removing Screen-Space Artifacts from 3D Gaussian Splatting

    The paper targets artifacts caused by 2D-image-plane occlusions that break the clean multi-view assumptions behind 3DGS. It proposes an unsupervised method to clean those inconsistencies before splat training or rendering.

    Useful for real-capture 3DGS pipelines where input noise and occlusion handling matter.

  12. ● Top story

    Cerebras Disaggregated Inference

    Cerebras describes a disaggregated inference approach for serving large models. The discussion surfaced on Hacker News with attention on system architecture and serving tradeoffs.

    Useful for understanding memory/compute separation and inference-system design at scale.

  13. ● Top story

    ONNX Runtime Disappears from GitHub

    A GitHub-hosted infrastructure incident prompted broad discussion on HN. The story is notable mainly as a software supply-chain and availability event around a major runtime project.

    Worth a quick look if you track platform reliability, repository governance, or build-chain fragility.

  14. ● Top story

    Profiling and Fixing the Sampler in MoE Diffusion Decoding

    The paper profiles masked diffusion LLM decoding and finds sampler overhead can dominate even after weight quantization. It then focuses on the per-token sampling and stopping pipeline rather than only model weights.

    Good deep-dive on where diffusion-model inference time actually goes.

  15. ● Top story

    Turbo Pascal Compiler Written in Turbo Pascal

    A Hacker News favorite about a compiler implemented in the language it compiles. The discussion focuses on portability, bootstrapping, and old-school compiler engineering.

    Classic systems engineering signal: self-hosting, code size, and toolchain simplicity.

  16. ● Top story

    Gemini Robotics 2.5 Targets Industrial Dexterity

    Google DeepMind announces an updated VLA model for industrial and everyday robot tasks. The pitch is improved dexterity, reliability, and safety in physical manipulation.

    Worth skimming for the industrial robotics capability bar, though the technical detail is lighter than the papers.

  17. ● Top story

    DOOM-GS: Dynamic Out-of-Core Optimization for Massive Gaussian Splatting

    This work addresses Gaussian splatting when the primitives no longer fit in GPU memory. The core issue is making training and optimization work out-of-core instead of assuming the scene fits in-core.

    Relevant to scaling 3DGS beyond workstation-sized memory limits.

  18. ● Top story

    Cross3DGS: Vulkan-Based 3D Gaussian Splatting from Workstations to Mobile

    The paper proposes a portable 3DGS training stack built on Vulkan rather than CUDA alone. It is explicitly about making capture-device or laptop/mobile training feasible across heterogeneous GPUs and drivers.

    Good systems lesson on portability constraints in real 3D training pipelines.

  19. ● Top story

    Where Human-to-Robot Transfer Gains Actually Come From

    This paper rechecks which part of a human-to-robot transfer recipe drives the gains. It finds that subset size and training variability can explain much of the reported improvement, not only task-aware data selection.

    A good reminder to separate recipe effects from data volume and optimization noise in robotics ML.

  20. ● Top story

    Reducing Control Stalls in Vision-Language-Action Policies

    The method uses predicted-input lookahead plus selective KV-cache repair to hide inference latency in VLA execution. It aims to keep control loops moving while reconciling stale predictions with newly arrived inputs.

    Concrete latency-hiding design for robotics serving and embodied agents.

  21. ● Top story

    Readie: Low-Latency Code Sandboxes for AI Agents

    This work looks at code execution sandboxes designed to decouple agent harnesses from runtime execution. The focus is on reducing latency while keeping iterative code-generation loops usable.

    Practical systems paper for agentic workflows that depend on fast, isolated execution.

  22. ● Top story

    Astra for Law Tops OpenAI's Launch Evaluation

    OpenAI reports Astra for Law as leading a 24-area legal evaluation and opening access to ChatGPT Enterprise customers. The post is primarily a product/benchmark announcement rather than a technical deep dive.

    Only useful as a benchmark signal; there is little implementation detail here.

  23. ● Top story

    Hugging Face and OpenRouter Make Open-Weight Models Easier to Run

    Hugging Face and OpenRouter announce a partnership to simplify running open-weight models. The post is mainly ecosystem news with limited technical substance.

    Mostly platform positioning; skip if you want implementation depth.

End of today’s edition.

How Signal is made33 monitored sources · 6 on the roadmap