AI research, dev logs, and thoughts.https://boxwrench.dev/https://boxwrench.dev/posts/interesting-finds-2026-09-16/https://boxwrench.dev/posts/interesting-finds-2026-09-16/Five notes: Tailcat's accountless tunnels, Claude Money's trust question, System One models that cannot hallucinate, recursive self-improvement through dreaming, and an inbox for long-running agents.Mon, 14 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-09-11/https://boxwrench.dev/posts/interesting-finds-2026-09-11/Five notes: RTX Spark AI PCs, the Dime headset mystery, building too much, Runway's realtime worlds, and an anti-hallucination startup.Fri, 11 Sep 2026 00:00:00 GMTnewsmodelshttps://boxwrench.dev/posts/interesting-finds-2026-09-10/https://boxwrench.dev/posts/interesting-finds-2026-09-10/Five notes: Nvidia's Hydra-0 action-flow world model, a random human life, Google's Putty vibe-coding experiment, Qwen-RobotWorld, and a visual atlas of agent systems.Thu, 10 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-09-11-b/https://boxwrench.dev/posts/interesting-finds-2026-09-11-b/Two more: Tencent's WeKnora knowledge platform and a paper showing sliding-window attention beats linear retrofits.Thu, 10 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-09-09/https://boxwrench.dev/posts/interesting-finds-2026-09-09/Six notes: Cognition at 47 billion, a Tailwind resource index, goal loops that deliver, AI and the direction problem, crisis as building permission, and the second-brain implementation gap.Wed, 09 Sep 2026 00:00:00 GMTnewsopinionhttps://boxwrench.dev/posts/interesting-finds-2026-09-08/https://boxwrench.dev/posts/interesting-finds-2026-09-08/Five notes: the Astra recurrent-depth safety fight, agentic video in Gemini, the abliteration business question, DLSS 5 perception splits, and the Fed watching token prices.Tue, 08 Sep 2026 00:00:00 GMTnewsmodelshttps://boxwrench.dev/posts/interesting-finds-2026-09-07/https://boxwrench.dev/posts/interesting-finds-2026-09-07/Five notes: a one-shot web-demo bench, Cloudflare cache transcoding, Copilot cost efficiency, new knobs for the brain, and Fable 5.1 worlds from code.Mon, 07 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/droppie-astra-build/https://boxwrench.dev/posts/droppie-astra-build/A tiny living water droplet built with GPT-6 Astra — soft-body physics, WebGPU refraction, and procedural audio.Mon, 07 Sep 2026 00:00:00 GMTmodelsnewshttps://boxwrench.dev/posts/interesting-finds-2026-09-06/https://boxwrench.dev/posts/interesting-finds-2026-09-06/Four notes: Raschka on looped transformers, TxBench antibody evals, Google Pics gated to paid tiers, and the Gemini Flash Cyber overlap.Sun, 06 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/kitten-rescue-manager-loop/https://boxwrench.dev/posts/kitten-rescue-manager-loop/A manager loop running Muse Spark built a complete small game in a single 24-minute run. The structure did the work, not the spend.Sun, 06 Sep 2026 00:00:00 GMTmodelsnewshttps://boxwrench.dev/posts/interesting-finds-2026-09-08-b/https://boxwrench.dev/posts/interesting-finds-2026-09-08-b/Five more notes: Shumer's 3D-world loop, a benchmark that scores agent construction, Qwen-Drive-1.0, loop engineering patterns, and AlphaGenome Atlas.Sun, 06 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-09-08-c/https://boxwrench.dev/posts/interesting-finds-2026-09-08-c/Eleven more: Images 2.5, Mercury 2.5 diffusion, ID-V2V restyling, cross-model KV transfer, Muse, Gaussian Splat Lite, the KV-caching explainer, tgrep, DIY weather forecasting, a 1-meter LiDAR viewer, and Switzerland's open-source trial.Sun, 06 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-09-05/https://boxwrench.dev/posts/interesting-finds-2026-09-05/Five notes: an SSH fighting game with a bot league, Muse Spark 1.3, the Muse superapp leak, test-time training as a scaling axis, and Meta's organizational second brain.Sat, 05 Sep 2026 00:00:00 GMTnewsmodelsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-09-04/https://boxwrench.dev/posts/interesting-finds-2026-09-04/Three preprints on harness as infrastructure: HarnessDev measures whether LLMs can build their own scaffolding, Harness-of-Harness makes multi-day autonomy composable, and Fast Weight Attention fixes the temporal alignment of recurrent memory.Fri, 04 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-09-03/https://boxwrench.dev/posts/interesting-finds-2026-09-03/Seven notes: California youth-safety regulation, industrial overcapacity as science catalyst, robotics hardware takeoff, in-context learning's GPT moment, AI as wet-lab co-pilot, a productivity-miracle claim, and Gemini 3.8 Flash / Flash Cyber.Thu, 03 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/gpt-6-astra/https://boxwrench.dev/posts/gpt-6-astra/OpenAI's largest training run ships with critical-rated cyber capability, looped computation that weakens chain-of-thought monitoring, and task-based pricing pressure.Thu, 03 Sep 2026 00:00:00 GMTmodelsnewshttps://boxwrench.dev/posts/normalized-lora/https://boxwrench.dev/posts/normalized-lora/A one-line normalization of LoRA's A matrix restores balanced early gradients, faster convergence, and mergeability without extra parameters.Wed, 02 Sep 2026 00:00:00 GMTresearchmodelshttps://boxwrench.dev/posts/interesting-finds-2026-09-02/https://boxwrench.dev/posts/interesting-finds-2026-09-02/Seven notes: Qwen Max 0902's 2.4T post-train, Quasar 438B tops Europe, H3-World turns H3 into a world model with 0.2% params, Fable 5.1 decodes a 373-year cipher, Humain M3 in Arabic, a cancer-vaccine primer, plus Dyson's $499 camera-toothbrush — noted with editorial.Wed, 02 Sep 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/golden-gate-park-web-demo-bench/https://boxwrench.dev/posts/golden-gate-park-web-demo-bench/The same Golden Gate Park Three.js prompt across several models, with the artifact, one-shot result, wall time, and run harness kept together.Wed, 02 Sep 2026 00:00:00 GMTmodelsresearchhttps://boxwrench.dev/posts/claude-fable-5-1-mythos-5-1/https://boxwrench.dev/posts/claude-fable-5-1-mythos-5-1/25% cheaper cache reads, 60% fewer cyber false positives, and early lab-validated science from a model that ships as two permissions, not two capabilities.Tue, 01 Sep 2026 00:00:00 GMTmodelsnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-08-27/https://boxwrench.dev/posts/interesting-finds-2026-08-27/Five notes: JIT-Agent synthesizes harnesses on the fly, Michigan Robotics open-sources its curriculum, DualView publishes exact phase-specialized weights, plus an ESP32 dashboard and the long-lived free-APIs directory.Thu, 27 Aug 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/qwen38-flash/https://boxwrench.dev/posts/qwen38-flash/Multimodal MoE with GDN+QSA attention, gated residual, n-gram embedding, and Muon — an open-weight preview of the Qwen4 line.Wed, 26 Aug 2026 00:00:00 GMTmodelsnewshttps://boxwrench.dev/posts/glm53-flash/https://boxwrench.dev/posts/glm53-flash/First natively multimodal in the GLM-5 line. 45 layers, hybrid linear-sparse attention with IndexPool, and Manifold-Constrained Hyper-Connections — MIT weights that stay local-feasible.Wed, 26 Aug 2026 00:00:00 GMTmodelsnewshttps://boxwrench.dev/posts/interesting-finds-2026-08-25/https://boxwrench.dev/posts/interesting-finds-2026-08-25/Four notes from the feed: an experimental vision model, a free agentic model window, CUDA on RISC-V, and a Linux utility for Logitech hardware.Tue, 25 Aug 2026 00:00:00 GMTnewsmodelshttps://boxwrench.dev/posts/interesting-finds-2026-08-25-b/https://boxwrench.dev/posts/interesting-finds-2026-08-25-b/Five more notes: hardware-aware kernels, a tighter matrix multiplication bound, AMD credits, and two SemiAnalysis signals on the CUDA moat.Tue, 25 Aug 2026 00:00:00 GMTnewsresearchhttps://boxwrench.dev/posts/interesting-finds-2026-08-25-c/https://boxwrench.dev/posts/interesting-finds-2026-08-25-c/Diffusion for language, skill-aware RL, explorative modeling, looped MoEs, and a map of 733k papers.Tue, 25 Aug 2026 00:00:00 GMTresearchmodelshttps://boxwrench.dev/posts/interesting-finds-2026-08-25-d/https://boxwrench.dev/posts/interesting-finds-2026-08-25-d/Sleep for hybrids, aggressive diffusion decoding, AI-found test-time controllers, a virtual cell, a DeepSeek J-Space report, and PagedAttention explained.Tue, 25 Aug 2026 00:00:00 GMTresearchmodelshttps://boxwrench.dev/posts/interesting-finds-2026-08-25-e/https://boxwrench.dev/posts/interesting-finds-2026-08-25-e/Meta's Hatch and Watermelon, Perplexity's local-first Portable Computer, Figure's Index dataset, Amazon's automated last mile, Jetson Orin Nano 2, and Keenable's knowledge index.Tue, 25 Aug 2026 00:00:00 GMTnewsresearchmodelshttps://boxwrench.dev/posts/diffusion-gemma/https://boxwrench.dev/posts/diffusion-gemma/Google DeepMind finetunes Gemma 4 into a discrete diffusion model that generates around 20 tokens per forward pass.Fri, 14 Aug 2026 00:00:00 GMTmodelsresearchhttps://boxwrench.dev/posts/dspark-adaptive-verification/https://boxwrench.dev/posts/dspark-adaptive-verification/One setting keeps speculative decoding fast from 1 to 256 concurrent users by skipping low-confidence drafts when the GPU is busy.Fri, 14 Aug 2026 00:00:00 GMTmodelsresearchhttps://boxwrench.dev/posts/qwen38-27b/https://boxwrench.dev/posts/qwen38-27b/27 billion dense, native vision-language, 262K to 1M context. Hybrid linear attention in a deployable package.Fri, 14 Aug 2026 00:00:00 GMTmodelsnewshttps://boxwrench.dev/posts/muse-glimmer-30b/https://boxwrench.dev/posts/muse-glimmer-30b/Meta distilled Spark for the workshop — a 30B model that runs locally and keeps the job on the bench.Mon, 10 Aug 2026 00:00:00 GMTnewsresearchmodelshttps://boxwrench.dev/posts/copilot-agentic-traces/https://boxwrench.dev/posts/copilot-agentic-traces/13.5 million GitHub Copilot sessions show what happens after you press enter.Sat, 08 Aug 2026 00:00:00 GMTresearchmodelshttps://boxwrench.dev/posts/political-superintelligence/https://boxwrench.dev/posts/political-superintelligence/Hall proposes AI that makes democracy cheaper to run. The question is who keeps it running.Sat, 08 Aug 2026 00:00:00 GMTresearchopinionhttps://boxwrench.dev/posts/weathernext/https://boxwrench.dev/posts/weathernext/Google DeepMind open sourced its most accurate weather model. The mechanism is simpler than it looks.Sat, 08 Aug 2026 00:00:00 GMTresearchmodels