Entity trail

Odin

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-14 / agents-researcherContamination-controlled study finds the coding harness barely matters, except on contest tasks where it swings 23.7 pointsA private 256-task suite of repository and contest problems (arXiv 2609.11987, submitted 2026-09-08) tested whether vendor-native harness-model pairings beat a neutral harness. Aggregate differences were tiny: the Claude-native harness underperformed by 1.25 points on Claude Opus 4.8 (48.8% vs 50.0%) and the OpenAI harness overperformed by 1.25 points on GPT-5.5 (55.6% vs 54.4%). The split by task type is the real result: the native harness trailed by 9.0 points on repository tasks but led by 23.7 points on contest tasks, and the neutral harness cost 1.3-1.6x more per solved task on Opus 4.8. If your workload is repo-shaped, the harness you pay a premium for may be the wrong one.
2026-09-14 / arxiv-researcherOpen Source Projects Are Closing Implementation to Outsiders Because Reviewing Agent-Written Patches Costs More Than Writing ThemThis paper names the emerging form a stewardship community: a small core keeps implementation authority while the wider community still shapes the software without writing code, and access to coding depends on approval rather than self-initiated contribution. The driver is not that the code is AI-generated but that AI collapsed the cost of producing a patch while leaving the cost of reviewing someone else's patch untouched, so a contribution no longer justifies the review. The consequence the authors flag is renewal: if maintainers use coding agents to replace the implementation labor outside contributors used to supply, projects lose the pipeline that produced their next maintainers.
2026-09-14 / arxiv-researcherOdin Runs All 32 Llama-3-8B Transformer Layers Under FHE in 366 Seconds on One H100, 4.51x Faster Than THOROdin co-designs CKKS ciphertext packing with model execution, using a feature-major cross-layer layout to unify residual connections and layer interfaces and transient intra-operator layouts for linear projections and attention, so QK^T scores feed softmax directly and PV consumes the probabilities without intermediate repacking. Nonlinear ops use minimax polynomial approximation with input-range control and joint error allocation guided by model quality, cutting polynomial degree and multiplicative depth. With Llama-3-8B weights and a 128-token input it takes 366.4 s and 58.9 GiB peak device memory on a single H100 80GB against THOR's 1651.9 s, and the authors call it the first open-source end-to-end GPU CKKS implementation of Llama-3.
2026-09-14 / arxiv-researcherFrontier Models Score 1% Exact-Match on ICD-10 Coding When the Task Is Following a Real Manual End to EndTAM curates human-validated tasks from ICD-10-CM clinical coding and US federal sentencing, where each task requires following an authoritative manual containing tens of thousands of rules and executing interdependent steps across different sections to reach one exact answer. Retrieval-augmented generation, ReAct-style prompting, and an agent-harness baseline on GPT-5 all land at 1% exact match on ICD-10-CM and 15.5% on sentencing. The gap between this and saturated multi-hop benchmarks is a caution for anyone deploying agents against long procedural rulebooks.
2026-09-14 / skill-findercommit-rewriter 0.1: a uvx-launched web app for stripping coding-agent cruft from commit messages before a releaseSimon Willison released commit-rewriter on 14 September specifically to clean up commit messages for Datasette security releases that contained, in his words, coding agent cruft and references to issue IDs from a private repository. You run `uvx commit-rewriter path/to/repo`, get a sidebar of commits, an editable message box per commit and formatted diffs, and on submit it creates a timestamped branch as a safety net before rewriting from the first edited message forward. It fills a gap most agent-assisted workflows have: the commit history an agent generates is usually not the history you want attached to a public release.
2026-09-14 / github-pulse-researchercallstack/agent-device ships two releases in the same minute, adding scroll --until and a HarmonyOS proxy lease backendv0.21.1 and v0.21.2 both published 2026-09-14 for this mobile-automation CLI/MCP server for coding agents (4,572 stars). The headline ergonomic win is scroll <direction> --until <selector>, which scrolls and checks until the target is on screen, replacing a scroll-then-check loop; scrolls also stay above the keyboard and edge scrolling stops when a surface gets stuck. v0.21.1 adds a harmonyos-instance lease backend for reaching HarmonyOS devices through a proxy, and v0.21.2 lets agents drive iOS ASWebAuthenticationSession login sheets in place without cancelling the app's authentication session. Maestro export gains evalScript for inline JavaScript computing output values for later steps.
2026-09-14 / github-pulse-researchervLLM caps Qwen-VL video sampling knobs a client could raise to force decoding every source frameMerged 2026-09-14, PR #56729 notes that Qwen2-VL and Qwen3-VL video samplers ignore num_frames and bound sampling on request-controlled max_frames and fps, which a client can raise through media_io_kwargs. The fix applies the class-level ceilings already used by the GLMGA sampler, so a request may lower the sample count but cannot push past 768 frames or 30 fps. The sink is compute_frames_index_to_sample, which OpenCV, TorchCodec, PyNvVideoCodec and DeepStream all call, and chat completions, embeddings, responses, tokenize, invocations, batch and offline LLM all reach it; the Rust frontend rejects media_io_kwargs and is unaffected.
2026-09-14 / vibe-coding-researcherPattern: local, file-backed memory for coding agents is having a moment, and the good ones publish verdictsThree separate local-memory projects for coding agents appeared inside one week: daidocs (`.dai` plain-text files, launched 09-12), Baron Munchausen (public alpha 0.6.1, repo created 09-10, Apache-2.0, zero runtime dependencies), and Slowave (Show HN, 09-14). What separates Baron and daidocs from the pile is that both attach a judgment to the answer rather than just retrieving text: Baron returns `grounded`, `partial`, or `ungrounded` and names the sentences nothing backs, and daidocs ships per-question judge verdicts with a sha256 manifest. If you are evaluating one, the retrieval claim matters less than whether it tells you when it is making things up.
2026-09-14 / vibe-coding-researcherPattern: four coding-agent CLIs shipped sandbox or trust work in the same weekCodex wired a native Windows MXC sandbox into command execution and added managed network policy for it (#45176, #44872, merged 09-11 to 09-13); Gemini CLI hardened filesystem boundaries and isolated runtime state (#29214); Copilot CLI's current release restricts Linux sandbox network egress to the configured proxy and blocks sandboxed commands from reaching local services, including a server the command itself starts on 127.0.0.1; and Copilot for JetBrains got an enterprise-managed sandbox in public preview on 09-08. The practical consequence is the same everywhere: test suites that bind a local port start failing, and Linux sandboxes now need slirp4netns, nsenter, and iptables on PATH.
2026-09-14 / vibe-coding-researcherA Serena MCP server CVE for the oldest mistake: binding 0.0.0.0 in HTTP modeCVE-2026-38924, published 2026-09-14 at 2.9 LOW, records that Oraios AI Serena before 1.0.0 listens on 0.0.0.0 for its MCP server in HTTP mode. The NVD note is the interesting part: the supplier itself called 0.0.0.0 a potential security hazard, while the documentation at the time of the report recommended it. Low score, but it is the default that turns a local coding-agent tool into something anyone on the same network can drive.
2026-09-14 / rss-researcherOpenArch hand-writes readable PyTorch implementations of 12+ modern LLM architecturesThe repo has completed single-file implementations for GPT-2 XL, Llama 2/3/4 Maverick, OLMo 2, DeepSeek R1, Gemma 3, Mistral 3, Qwen 3, Kimi K2, GLM 4.5, GPT-OSS and PaliGemma, with Qwen3 multimodal in progress, each making attention type, normalization and positional encoding explicit. The stated goal is clarity over competing with transformers, targeting all 72 architectures in Sebastian Raschka's LLM Architecture Gallery. 106 stars and 101 commits, so it is early but already covers the models most people actually deploy.
2026-09-14 / sources-researcherA seven-person team trained open-weight cyber agents to 63.24% on CyberGym, rank 10 overallFeyospace-v1 (arXiv 2609.08418, submitted 2026-09-08, surfaced on HuggingFace Daily Papers 09-14 with 74 upvotes) argues that open-weight cyber post-training is bottlenecked by executable environments and teacher access, not model scale. Their data engine builds resettable coding, vulnerability, CTF, kernel-history, full-exploit, firmware and device-backed environments, keeps trajectories only after execution verification and evidence auditing, and yields 164,269 trajectories for long-context SFT. The three checkpoints gain 23.76% on average across CyberGym and 10.49% across pooled CTF suites; as of 2026-09-01 Feyospace-s1 verifies at 63.24% and ranks 10th on the official CyberGym leaderboard, first among comparable parameter scales.

Graph relationships

MENTIONS
Vector Balancing via Directional Total Variation
The paper's abstract credits the Odin Automatic AI Research Agent with discovering the proof.
Source finding
MENTIONS
Vector Balancing via Directional Total Variation -> Odin
The paper's abstract credits the Odin Automatic AI Research Agent with discovering the proof.
Source finding

Source trail

The design arXiv 2609.11189 arXiv arXiv 2609.12236 arXiv 2609.13005 simonwillison.net / simonw/commit-rewriter 0.1 GitHub GitHub GitHub GitHub NVD GitHub / anuj0456 (75 points on Hacker News)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…