Dauntless · Systems

Open-loop digest

September 9, 2026

17 items · 4.8 KB

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

Agentic Frameworks, Tooling, Skills

  • Procedural Graphs, https://huggingface.co/papers — Self-evolving execution structures for LLM agents that replace static tool-use routing with dynamic graph construction; enables long-horizon rover loops and agentic pipelines to adapt tool dependencies on-the-fly without external orchestrators. [Source: https://huggingface.co/papers]
  • Environments as Scaffold, https://huggingface.co/papers — Boots self-evolving agents in long-horizon tasks via enriched environmental feedback rather than synthetic rewards; stabilizes step-drift and improves agent reliability before local deployment, replacing brittle stress tests. [Source: https://huggingface.co/papers]
  • MOLE, https://huggingface.co/papers — Detects insider threats and prompt-injection vectors in deployed AI agents at the policy level; enables mandatory security gating for local skill deployments that static linting or basic MCP validation misses. [Source: https://huggingface.co/papers]

Notable Research

  • BeaconKV, https://huggingface.co/papers — KV cache compression guided by beacon queries specifically targets large reasoning models; directly reduces KV-cache pressure and inference latency for your 32GB-constrained local loops without accuracy collapse. [Source: https://huggingface.co/papers]
  • SQS (Sparse Quantized Sub-distributions), https://huggingface.co/papers — Bayesian compression framework for DNNs that quantizes sub-distributions sparsely; offers a tighter GGUF/nf4 alternative to standard PTQ, preserving mathematical/coding fidelity for models in your weight class. [Source: https://huggingface.co/papers]
  • Online Draft Co-Training for Speculative Decoding, https://huggingface.co/papers — Optimizes speculative decoding via co-training drafts with the target model; cuts step-latency and KV-cache overhead during local long-context RL post-training inference, applicable to your reasoning pipelines. [Source: https://huggingface.co/papers]

Frontier Lab Updates

  • OpenAI RSI Day & Defensive AI Pivot, https://simonwillison.net/ — Internal logs confirm OpenAI's research focus has shifted to Recursive Self-Improvement (RSI) cycles, explicitly framing powerful aligned AI as a primary deployment pillar for defensive infrastructure against rogue agents and automated vulnerability discovery. Signals competitive pressure is moving from raw benchmark chasing to agent stability and operational defense. [Source: https://simonwillison.net/]
  • Internal Reasoning Stack Acceleration, https://simonwillison.net/ — Documents a sharp inflection in OpenAI's internal researcher spend on coding agents starting late July, coinciding with access to the model later released as GPT-6 Astra; indicates rapid iteration on their next-generation reasoning architecture ahead of public rollout. [Source: https://simonwillison.net/]

Models to Download & Try

  • nex-agi/Nex-N2.5-mini, https://huggingface.co/nex-agi/Nex-N2.5-mini — 35B params | ~18GB weights (INT4) + ~6–8GB KV cache at num_ctx=32k–64k; fits within your 32GB budget with room for context overhead. Introduces a new architecture targeting agentic routing and spatial reasoning, designed to augment Qwen 3.6:35b in multi-step tool-use without expanding footprint. [Source: https://huggingface.co/models?sort=trending]
  • ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF, https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF — 27B params | ~14GB weights (INT4 GGUF) + ~5–7GB KV cache; comfortably fits 32GB with num_ctx=32k+. Applies Gated Scalar Quantization (GSQ) to the Qwen 3.8 base, preserving math/coding holdout performance where standard AWQ/IQ2 variants degrade. [Source: https://huggingface.co/models?sort=trending]
  • IFM/K2-Horizon-MoVA-36B-A4B, https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B — 37B params (A4B denotes 4-bit activations) | ~19GB weights + compressed KV cache; VRAM fit similar to Nex-N2.5-mini but optimized for multi-view spatial and VLA reasoning, offering a vision-augmented alternative to text-only routing for your aerospace/robotics scraper pipelines. [Source: https://huggingface.co/models?sort=trending]

Skipped as Already Covered

  • GPT-6 Astra rollout metrics, pricing ($10m/$50m), and ARC-AGI/ExploitBench benchmarks (covered 9/4, 9/7)
  • Claude Fable 5.1 science benchmark scores & enterprise thresholds (covered 9/3, 9/4, 9/7)
  • OpenAI internal agent spend acceleration and RSI operational focus (covered 9/7 adjacent context)
  • Qwen/Qwen3.8-27B & community GGUF quant variants (DavidAU, HauhauCS, orcarouter) (covered 9/1, 9/4, 9/5, 9/8)
  • MiniMax-H3 multimodal routing specs & weights (covered 9/1, 9/5, 9/7)
  • Experiential zero-markup BYOK gateway & PostHog self-driving observability platform (covered 9/7, 9/8)