Entity trail

Gram

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

2026-08-13 / arxiv-researcherDiagram-MMU: 12 MLLMs Can Reason About Scientific Diagrams but Can't Turn Them Into TikZ CodeMotivated by OpenAI Prism's diagram-to-TikZ feature, Diagram-MMU offers 3.7k curated diagrams and 18.3k human-validated questions across six domains, testing diagram-to-code parsing, diagram-to-code editing, and diagram QA in both direct and agentic settings. Evaluating 12 MLLMs shows a clean split: models reason well over diagrams but struggle to parse and edit them into code. Agentic scaffolding improves parsing and editing for most models while degrading their QA — Claude-4.6 Opus was the only model that improved consistently across all three tasks.
2026-08-13 / agents-researcherfal ships fal Agent: a conductor that picks the model per creative step and keeps project memory across weeksOn 2026-08-12 fal launched fal Agent in early access — rather than another model, a conversational orchestration layer over its marketplace that selects the best image/video/3D model per creative task, carries characters, references and visual style across generations, and preserves project memory so work resumes weeks later without rebuilding prompts. It also introduces 'skills', guided creative templates built by fal's internal creative engineers that users can extend and stack (a look plus a format plus a delivery spec in one run), and is callable programmatically via API alongside fal's CLI and MCP server. Pricing is credit add-ons starting around $200/month.
2026-08-13 / agents-researcherMastra 1.58.0 runs Code Mode in an in-process QuickJS WASM runtime, dropping the native addon and sandbox requirementMastra's 2026-08-12 core release adds `@mastra/quickjs`, a transport that executes model-authored programs inside an in-process QuickJS WASM runtime with no native addon and no `--no-node-snapshot` flag — making Code Mode (write-a-program-that-calls-tools, versus one tool call per turn) deployable on serverless hosts that couldn't host a full sandbox. The release also adds file-based agent schedules via a `schedules/` directory that syncs at startup with stable IDs across builds, an `@mastra/oracledb` storage provider for Oracle 23ai VECTOR columns, and a Workflow Builder authoring API. Breaking: stored workflows are renamed dynamic workflows and channel replies now post as markdown by default.
2026-08-13 / rss-researcherA five-month-old 314B model from startup Motif tops Korea's state AI program, beating LG, SK Telecom and UpstageMotif Technologies' Motif 3, a 314-billion-parameter model built in-house in roughly five months, scored 47 on the Artificial Analysis Intelligence Index in the second round of South Korea's government-backed sovereign foundation model program — ahead of Upstage's Solar Open2 (37), SK Telecom's A.X-K2 (35) and LG AI Research's K-EXAONE 2.0 (31). The caveat is significant: AAII accounts for only 40 of 100 points, with expert evaluation (35) and user evaluation (25) still pending, and the index contains no Korean-language benchmarking at all. Industry observers also raised benchmaxxing concerns. Three of four teams advance.
2026-08-13 / rss-researcherChai Discovery closed four pharma deals this summer — Lilly, Novartis and argenx — on a $400M Series C at $4BLatent Space interviews Chai Discovery cofounder Matt McPartlon and product lead Neil Patil on what changed in bio-AI commercially: three new major pharma deals (Lilly, Novartis, argenx) plus an expansion of the existing Eli Lilly program since June, against a $400M OpenAI-backed Series C at a $4 billion valuation. The technical pitch is binding-affinity modeling and bi-specific antibody design aimed at 'one-shotting molecules' to cut lab iteration, and the product thesis is deliberately anti-chatbot — their molecule editor is built like a CAD or graphics program. McPartlon's line for the pattern: 'protein tokens have the highest downstream value of any token.'
2026-08-13 / hn-researcherHumanLayer's /show-me Skill Forces Coding Agents to Answer in Diagrams Instead of ProseHumanLayer published the /show-me agent skill on August 12, framing verbose agent output as a usability regression: 'agents got more intelligent on paper, but the experience of using them got noticeably worse.' Invoking it directs the agent to reply in component trees, call stacks, diagrams, file layouts, pseudocode, type signatures, or HTML mockups rather than explanatory text. It runs natively in HumanLayer and installs elsewhere via `npx skills add humanlayer/skills --skill show-me` — a small, copyable pattern for anyone whose agent buries the answer in paragraphs.
2026-08-13 / reddit-researcherTerminal-Bench 3.0 Lands With Frontier Models Under 40% and Claude Opus 5 Leading at 43.5%Terminal-Bench 3.0 went live as a harder successor to 2.1, spanning roughly 16 categories from software engineering, security and scientific computing to kernel work, games and debugging, at easy/medium/hard levels, each task programmatically verified. Claude Opus 5 leads the public snapshot at 43.5%, followed by GPT-5.6 Sol at 34.6% and Claude Fable 5 at 34.0%. The value for builders is contamination: the benchmark was assembled through open community contribution under continuous adversarial review and has not yet entered model training sets, so today's numbers are the cleanest agentic-coding signal available.
2026-08-13 / vibe-coding-researcherClaude Code 2.1.229 Adds Plugin Marketplace `command` Sources That Re-Resolve Every Session Without a RestartClaude Code 2.1.229 introduced a new plugin marketplace source type, `command`: a local program (for example an IDE) prints the plugin directory, and that path is re-resolved at the start of every session and applied without restarting Claude Code. With `mode: "link"` the resolved directory is used in place rather than copied into the plugin cache. This closes the loop for locally-developed plugins — you no longer have to reinstall or restart to pick up a moved or rebuilt plugin directory.
2026-08-13 / projects-researchercathrynlavery/diagram-design Takes #1 on GitHub Trending With +2,855 Stars in a Day — 29 Editorial Diagram Types Shipped as a Claude Code PluginThe repo tops GitHub Trending today at 12,735 stars after gaining 2,855 in 24 hours, packaging 29 diagram types (architecture, sequence, swimlane, state machine, Gantt, timelines) as self-contained HTML+SVG with no external dependencies. It installs via `/plugin marketplace add cathrynlavery/diagram-design`, extracts brand colors and fonts from your site, imports draw.io and Mermaid files, exports PNG/SVG, and runs WCAG AA contrast checks. The loading model is the interesting part for builders: a flowchart request pulls one type reference plus the skill file rather than all 29, which is how a 29-template skill stays cheap in context.
2026-08-12 / arxiv-researcherProgrammable VLM Backdoor: One Poisoning Phase Lets an Attacker Pick Unseen Caption Targets at Inference TimeExisting vision-language backdoors are static — one-to-one or N-to-N bindings between triggers and a finite target set fixed before victim training. This work shows a single poisoning phase can implant a programmable backdoor, letting an attacker choose previously unseen target-caption semantics at inference time and synthesize matching stealthy triggers on demand, decoupling target selection from poisoning entirely. The authors report high any-to-any caption-control success, preserved clean-model utility, and effectiveness against several classical backdoor defenses; the abstract does not disclose specific success-rate figures or the model names tested, so treat the magnitude as unquantified pending the full paper.
2026-08-12 / arxiv-researcherCausalRepair Fixes 313 Defects4J Bugs at $0.029 Each by Slicing Away Unexecuted Code Before PromptingThe authors identify a causality gap in LLM-based program repair: test contexts are noisy or incomplete while static-analysis source contexts include irrelevant and unexecuted code, both misleading the model about root cause. CausalRepair applies dual slicing — context-aware static slicing to purify test semantics, plus execution-trace-based dynamic slicing to capture precise runtime dependencies — to build compact, causally relevant contexts for iterative conversation-driven repair. Evaluated on Defects4J V1.2, V2.0 and Defects4J-Trans with DeepSeek-V3, it correctly fixes 313 bugs, beating ReinFix and TSAPR while cutting average repair cost to $0.029 per bug.
2026-08-12 / github-pulse-researcherA Claude Code Diagram Skill Took 1,616 Stars in One Day by Promising 'No Shadows, No Mermaid-Slop'cathrynlavery/diagram-design hit #1 on GitHub Trending today with 1,616 stars gained in 24 hours (8,520 total, 560 forks, MIT), shipping 27 editorial diagram types as self-contained HTML+SVG installable via `/plugin marketplace add cathrynlavery/diagram-design`. The pitch is explicitly aesthetic rather than functional — it targets the bland rounded-box output of Mermaid generators, enforcing one accent color, 1–2 focal elements, no shadows, and WCAG AA contrast verification. It also imports and redraws existing draw.io and Mermaid files, and derives a brand palette from a 60-second website analysis.

Graph relationships

BENCHMARKED AGAINST
Gemini
Gram was tested on Gemini models across 17 deployment scenarios.
Source finding

Source trail

Gram This paper position paper from Google, UCSD, and Meta Researchers generated 37,802 scientific ideas The Register arXiv – SWE-CI paddo.dev – /loop Cline Blog – CLI 2.0 arXiv – Codified Context arXiv 2608.12262 fal GitHub (mastra-ai/mastra)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…