Dauntless · Systems

Open-loop digest

September 13, 2026

33 items · 14.3 KB

Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron

Agentic Frameworks, Tooling, Skills

  • RoastMyHarness, https://www.reddit.com/r/LocalLLaMA/comments/1weufjc/ — Small benchmarking engine that runs DeepSWE coding tasks against bare Pi as a control and your custom harness (extension, skill, AGENTS.md, etc.) as the variant, then diffs what improved, what broke, and the token cost. Solves the "I added a tool and feel like it helped but can't prove it" problem that has no clean answer in vanilla eval harnesses. Runnable today via a Pi extension wizard.

  • smolbenchmark, https://www.reddit.com/r/LocalLLaMA/comments/1weekio/ — Ranks models that fit your hardware (8 GB tablets through Jetson Orin Super) by decode speed, tok/J, ITL latency, power draw, and thermals. Currently 13 families, ~1000 configs on a Jetson reference. Useful as a second opinion before pulling a 30B MoE onto a 32 GB card: it surfaces the thermals/bandwidth trade-offs that a single-pass tok/s number hides. Repo not linked in the thread; check the comment for the GitHub slug.

  • AgentsDock, https://agentsdock.net (HN #17, 66 pts) — IDE purpose-built for agentic AI research. Listed on HN without a detailed feature breakdown in the scrape; the pitch is a dedicated editor surface for building and debugging agent loops rather than repurposing a general IDE. Worth a 5-minute look if your current harness-editing workflow (vi + terminal + Ollama) is getting unwieldy.

  • Lorivo, https://www.reddit.com/r/LocalLLaMA/comments/1weum85/ — Serverless LoRA-adapter hosting on vLLM: shares the base model weights across adapters so you pay only for the rank-8–32 delta instead of a full GPU instance per fine-tune. Relevant if you want to A/B a CodeFinetuner output (covered 9/12) without keeping a 24/7 GPU. Cloud-hosted, so not a local replacement, but cuts the cost of iterating on small adapters.

Notable Research

  • SenseNova-U1.5: Towards Native Unified Visual Intelligence, source: https://huggingface.co/papers/date/2026-09-11 (Sep 11, 222 upvotes, 65 authors, submitted by taesiri) — Large multi-institution paper on a unified visual-intelligence architecture. No direct arXiv or code link visible in the scrape; the 65-author list and "native unified" framing suggest a foundation-model-level contribution rather than an incremental task paper. Flagging for awareness; no runnable artifact identified yet.

  • An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics, source: https://huggingface.co/papers/date/2026-09-11 (Sep 11, 32 upvotes, NVIDIA) — Open training recipe for pushing a Nemotron-class model to IMO-gold math. The "open recipe" framing means the data mixture, RL loop, and reward design are published, not just the final weights. If you're distilling reasoning into your 27B local primary (the per-token gating distillation paper from 9/12 gives the mechanism), this is a concrete task-level target and reward design to copy. No code link in the scrape; check NVIDIA's GitHub post-publication.

  • From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge, https://arxiv.org/abs/2609.11859 (53 pp., 13 figs) — Mechanistic study of how decoder-only LLMs surface stored facts versus hallucinate them. Directly relevant to the educational-content pipeline: understanding the retrieval-internal-knowledge path tells you where a confidently-wrong science fact for a 7-year-old is coming from, and whether a RAG override or a self-consistency check is the right fix. Pure paper, no code link in the abstract page.

  • SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem, source: https://huggingface.co/papers/date/2026-09-11 (Sep 11, 122 upvotes, KAIST AI) — Synthetic spatial-reasoning training task for vision-language models. If your game pipeline ever needs a local VLM to validate that a generated level's physics or spatial layout is coherent ("can this block stack actually be placed?"), this is the training-data pattern. No code link visible in scrape.

Frontier Lab Updates

  • OpenAI image-generation update (via Simon Willison, https://simonwillison.net/) — OpenAI's image models (GPT-Image family) report "more than 3 billion images across ChatGPT Images and the GPT-Image models in the API" as of this release. The update targets multi-turn instruction-following, faster response, and better subject preservation in reference-photo workflows. Vendor-reported; no independent benchmark. Relevant to your gpt-image-2 / Codex CLI 2D art pipeline only if the API surface changes; the scrape does not list specific API diffs.

  • Intern-S2-397B (via Reddit, https://www.reddit.com/r/LocalLLaMA/comments/1wf3wt2/) — InternLM's multimodal foundation model for "scientific intelligence and long-horizon agents." 397B parameters; not runnable locally. Notable for the training methodology: visual pre-training directly from raw scientific-literature pages plus multi-task RL and long-horizon agent RL. The pre-training paradigm (jointly modeling layout + text from pages) is a pattern worth stealing if you're building a document-grounded RAG layer for curriculum content. Field news, not actionable on 32 GB.

  • Dario Amodei, "We must pace the frontier", https://darioamodei.com (HN #25, 669 pts) — Anthropic CEO's open letter on AI safety and pacing. Pure policy commentary with no engineering artifact. Reporting for completeness; no action item.

Memory, Grounding & Factual Correctness

  • Memory as Plans: World-Action Modeling with Memory-Grounded Planning, source: https://huggingface.co/papers/date/2026-09-11 (Sep 11, 35 upvotes, 8 authors) — Couples persistent memory with action-planning so the agent's "memory" is not a retrieval store but a structured plan that constrains future actions. The key idea: memory is the plan, not a lookup table the plan references. This is a different architecture than the MAPLE approach (covered 9/11) where memory and planning are separate components; here they are unified. No code link in scrape.

  • RAG-Safety-Bench: Reliable Evaluation of Retrieval-Augmented LLM Safety, https://arxiv.org/abs/2609.11758 (EMNLP 2026 main conference) — Benchmark specifically for evaluating whether RAG-augmented LLMs produce safe, grounded output. Relevant as an evaluation gate: before shipping a generated lesson for a 6-year-old, you need a structured way to check that the RAG layer actually constrained the model rather than the model ignoring the retrieved context and fabricating. EMNLP main-conference acceptance is a quality signal; check the paper for the rubric structure.

  • When Agents Disagree: Bayesian Backward Reasoning as a Label-Free Anchor for Multi-Agent Collective Decision-Making, https://arxiv.org/abs/2609.11709 — Uses Bayesian backward inference to resolve disagreement among multiple agent outputs without requiring ground-truth labels. Applicable as a lightweight consensus mechanism when your frontier-model "judge" (Astra) and your local workhorse (Qwen3.8) disagree on whether a generated puzzle is solvable: instead of picking one, you get a probability-weighted resolution. No code link in the abstract.

Games, Engines & 3D

  • Three.js Native Gaussian Splats, https://ben3d.ca (HN #7, 17 pts) — Tutorial / API guide for using Gaussian splatting natively in Three.js (no separate splat renderer). If you're building a 3D level viewer or a game asset preview in Godot's web export, this is the fastest path to rendering a Gaussian-splat scene without pulling in a heavy dependency. Directly usable in a web-based level inspector.

  • Gaussian Light Transport, https://arxiv.org/abs/2609.11430 (SIGGRAPH Asia 2026 Conference Track, 11 pp., 8 figs, project page: https://patrick-attimont.com/projects/gaussian-light-transport/) — New lighting model that operates in Gaussian-splat space rather than mesh/radiance-map space. If your game assets are moving toward Gaussian representations (as the Three.js splat path suggests), this is the physically-based lighting layer that would make them look correct under dynamic lights rather than baked. SIGGRAPH Asia conference track = peer-reviewed.

  • SceneHI: High-Resolution 3D-Consistent Scene Texturing with Controllable Illumination, https://arxiv.org/abs/2609.10363 (ECCV 2026, 7 authors) — Generates high-resolution textures for 3D scenes while keeping illumination controllable (you can relight the textured mesh without re-baking). Directly relevant to your texture/material synthesis pipeline: if you're generating game materials and need to swap from "outdoor afternoon" to "indoor warm" lighting without re-generating the texture, this is the method. ECCV 2026 acceptance.

  • Hologram Representation via Quadratic Phase Gaussian Splatting, https://arxiv.org/abs/2609.11434 (SIGGRAPH Asia 2026 Technical Communications) — Extends Gaussian splatting to encode holographic (light-field) information via quadratic phase functions. Not immediately actionable for game content, but if you're following the Gaussian-splatting rendering pipeline (Three.js splats, Gaussian Light Transport), this is the next step in fidelity. SIGGRAPH Asia technical track.

  • Learning Realistic Athletic Sprinting Without Demonstrations, https://arxiv.org/abs/2609.11083 (17 pp., 23 figs, 7 authors incl. C. Karen Liu, Kayvon Fatahalian) — Learns physically plausible sprinting motion without any reference motion-capture data. If your game features a running character (kids' games very much do), this is a path to procedural animation that doesn't require a mocap pipeline. No project page or code link in the abstract.

Models to Download & Try

  • Edge0-35B-A3B-preview, https://huggingface.co/Edge0/Edge0-35B-A3B-preview — 35B total / ~3B active MoE, text generation. Updated 3 days ago, 3.55k downloads, 663 likes. At ~3B active parameters with Q4 quantization, this should fit comfortably in 32 GB with room for a large KV cache. "Preview" tag means not final; benchmark the actual output quality before trusting it. The MoE architecture means inference speed should be close to a 3B dense model while having 35B total capacity. No Ollama tag visible in the scrape; check for a GGUF upload or use ollama run with a direct HF reference.

  • Laguna XS 2.1 / Laguna XS.2, https://ollama.com/library/laguna-xs-2.1 / https://ollama.com/library/laguna-xs.2 — 33B total / 3B active MoE, explicitly designed for "agentic coding and long-horizon work on a local machine." Pullable from Ollama directly. At 3B active, Q4 weights are ~2 GB of active compute; total weights ~17–20 GB at Q4, leaving 12+ GB for KV cache at 131K context. The XS.2 tag suggests a minor revision over 2.1; check the model card for the diff. Claim: "designed for agentic coding" — unverified in the scrape, no benchmark numbers listed.

  • muse-glimmer (Meta), https://ollama.com/library/muse-glimmer — 30B dense, Apache 2.0, described as "Meta's latest open model built for always-on local agents... tuned for tool use, long tasks, and failure recovery." Runs on a single GPU. At 30B dense Q4 that's ~17–18 GB, fitting in 32 GB with ~14 GB for context. The "failure recovery" framing is interesting: it implies the model was specifically trained on agent-loop edge cases (tool errors, partial completions) rather than just chat. No benchmark numbers in the scrape; check the HF model card for SWE-bench or tool-use scores.

  • MiniCPM5-2B / MiniCPM5-2B-GGUF (openbmb), https://huggingface.co/openbmb/MiniCPM5-2B / https://huggingface.co/openbmb/MiniCPM5-2B-GGUF — 3B text generation, updated 1 day ago. Small enough to run as a fast router/classifier alongside your 27B primary (e.g., routing which agent skill to invoke, classifying a child's reading level before generation). The GGUF variant is directly pullable. 150k downloads on the base model in ~1 day suggests active adoption. No benchmark comparison to Qwen3.8 in the scrape.

  • DFlash2 speculative draft for Qwen3.8-27B, https://huggingface.co/HermiHg/Qwen3.8-27B-DFlash2-Q2_K_S-MIX-GGUF — Community-reported: paired with ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF (IQ3_XXS) at 128K context, ~60 tok/s on a 16 GB RX 9070 XT. The commenter notes this outperforms the built-in MTP path because MTP "seems to multiply the required VRAM," and that speculative decoding increases sensitivity to GTT (GPU texture/temp) overflow — turning off speculative decoding is faster when you're already over VRAM. On your 32 GB V620 the headroom is better, so the 60 tok/s figure should be a floor, not a ceiling. Community-reported, single user, single GPU; not independently reproduced.

  • bartowski/Qwen3.8-27B-GGUF (per-tensor layout update), https://huggingface.co/bartowski/Qwen3.8-27B-GGUF (via https://www.reddit.com/r/LocalLLaMA/comments/1webfsq/) — Updated GGUF quantization using per-tensor layout maps (new tensor-type layouts in bartowski's uploads). If you're running Qwen3.8-27B on llama.cpp, this is a drop-in weight swap that may shift the quantization error distribution. The Reddit thread links a blog post on the per-tensor layout method. No benchmark delta reported; check the model card for the specific quant levels affected.

  • Agnes-3.0-Flash (via https://www.reddit.com/r/LocalLLaMA/comments/1wers2v/) — 33B model claiming to beat Qwen3.8-27B on the ArtificialAnalysis intelligence index. Caveat is in the scrape itself: the HF model is tagged "Preview," the AA listing says "proprietary," and the HF page and website reference an "openness" narrative that doesn't match the proprietary AA tag. The commenter expresses confusion about whether the evaluated model and the uploaded model are the same. Do not act on the AA score until the model is unambiguously open-weight and the AA listing is corrected. No direct model URL in the scrape.

Skipped as Already Covered

  • NCP-ArchPreview (275 upvotes, 28 authors) — covered 9/11
  • EvoSafeHarness (Johns Hopkins, 48 upvotes) — covered 9/11
  • Recursive Code World Models / World in World — covered 9/11
  • Negative Self-Distillation, HyQuant, COBRA-Skills, MAPLE — covered 9/11
  • Domain-Specific Hallucination Detection, OmniHallu, FlexComp, K/V-Cache Interventions, Per-Token Gating, EXYGEN — covered 9/12
  • Qwen3.8-27B GGUF variants (ISTA-DASLab, unsloth, DavidAU), GLM-5.3 / MiniMax-H3 / DeepSeek-V4.1-Flash — covered across 9/1–9/12
  • OpenAI RSI day, Dario Amodei "pace the frontier," GPT-6 Astra internal spend — covered 9/7, 9/9, 9/10