Jeremy Morgan

The Raw-RAG code is public. 256-token windows, retrieve 128, cross-encoder rerank to 15. If your graph memory can’t beat this on your data, the graph is cost without return. If it can, you have a number to defend. github.com/Litmeb/RAW-RAG

Chunk the raw chat history, embed, rerank, hand over 15 windows. That dumb baseline beats the top agentic memory systems in 6 of 7 benchmark settings. Run it on your own logs before you fund an entity-extraction pipeline. arxiv.org/abs/2610.30755

A multimodal GraphRAG for storm damage assessment that models confirmed damage, negative evidence, unresolved conditions, and temporary repairs as distinct things. Most GraphRAG retrieves more facts. This one handles missing and contradictory ones. arxiv.org/abs/2610.28216

Anthropic Just Revealed 12 New Rules for Prompting Claude Opus 5.5 blog.vijayasekhardeepak.com/anthropic-ju...

Meme of the day

SHWotever added a JSON export of all SimHub properties and actions in version 9.12.7. That's a useful map for anyone building custom sim racing dashboards. github.com/SHWotever/Si...

Jake Manger's whatmodel matches your RAM and GPU to local AI models, with links to the scores behind each rank. It draws on over 22,000 benchmark results. Try it before buying another GPU. github.com/jakemanger/w...

Apple's MLX team reports over 36 tokens per second from an 816B model spread across four Mac Studios. The method and power measurements are published. What does that setup cost per token? machinelearning.apple.com/research/mlx...

caa1211 built an ESP32 dash that reads GT7 telemetry straight from a PS5. Version 2 adds SimHub over USB for PC sims, using the same screen. The firmware and web installer are open. github.com/caa1211/esp3...

Vector search finds relevant information. Graph traversal connects it. Qdrant's hands-on GraphRAG tutorial combines both using Neo4j, with a reproducible Python implementation. A useful starting point for testing whether graph-enriched retrieval improves answers. qdrant.tech/documentatio...

L-Forster's OpenJet reports 140 tokens per second for code with Qwen3.8 27B on one RTX 3090, versus 73 in its llama.cpp test. The engine code and benchmarks are public. Anyone reproduced it? github.com/L-Forster/op...

Not every AI inference needs an LLM. RDFox demonstrates how explicit knowledge models and reasoning rules can answer questions with deterministic logic. An architectural alternative when correctness and explainability matter more than generative flexibility. www.oxfordsemantic.tech/events/rdfox...

OpenF1 maintainer br-g released v1.0 of the race-data API after a rewrite focused on speed and simplicity. If you build timing dashboards, I'd check what changed before updating. github.com/br-g/openf1/...

Jake Manger's whatmodel matches your RAM and GPU to local AI models, with links to the scores behind each rank. It draws on over 22,000 benchmark results. Try it before buying another GPU. github.com/jakemanger/w...

Manuel Staggl built RigShift to switch a Windows PC between desk and sim rig when the wheelbase powers on. It moves monitors, sound and apps, then rolls back if the screen goes black. github.com/ManuelStaggl...

Extracting entities is easy compared with defining what they mean. The latest Going Meta series explores ontologies, graph semantics, and knowledge representation. Worth following if you're building knowledge graphs that need to remain consistent beyond a prototype. neo4j.com/videos/going...

AnAngryParsnip built BrakeBox to tune a DIY load-cell brake pedal without a PC app. The build exposed an HX711 hardware issue that can add over 100ms of input lag. Useful detail, even if you never build one. github.com/AnAngryParsn...

Neo4j Desktop 2.3.0 brings an embedded terminal, project organization, and AI coding-agent skills into the database workspace. The interesting shift is toward making Cypher, graph modeling, and agent-assisted development part of one workflow. feedback.neo4j.com/changelog

Efe Çangırılı and Murat Kurt studied whether a sim racer's driving style survives a car change. In a small sample, speed and consistency held up better than braking habits. AI coaches need tests like this. arxiv.org/abs/2610.10061

caa1211 built an ESP32 dash that reads GT7 telemetry straight from a PS5. Version 2 adds SimHub over USB for PC sims, using the same screen. The firmware and web installer are open. github.com/caa1211/esp3...

AGORA combines Ollama, Neo4j, and GraphRAG to turn documents into knowledge graphs and simulated agent interactions. A useful open-source architecture for exploring multi-agent systems without relying on cloud LLMs. github.com/arn0ld87/agora

Hans-Dieter Zucht tested 16 local models on a 32GB Mac mini M6 and shared raw runs plus a script. One model managed about 64 tokens per second on prose. That's a benchmark worth rerunning. github.com/HDZucht/Mac-...

Anton Holovko's Slipstream overlays read iRacing telemetry locally and flag likely wheel lockups using RPM and road speed. Clever workaround, but does it agree with your brake trace? github.com/antonholovko...

Most coding agents search for similar code. Code-Graph-RAG maps actual dependencies using Tree-sitter and Memgraph. That means asking which functions call what, tracing dependencies, and understanding the impact of changes across a monorepo. github.com/vitali87/cod...

Efe Çangırılı and Murat Kurt studied whether a sim racer's driving style survives a car change. In a small sample, speed and consistency held up better than braking habits. AI coaches need tests like this. arxiv.org/abs/2610.10061

Neo4j just reduced graph embedding generation to a single API call. No elaborate ML pipeline required. The new Aura Graph Analytics APIs make structural similarity, clustering, and graph-based predictions much easier to explore. neo4j.com/blog/aura-gr...

Working on a sim racing telemetry app on this fine afternoon and I must say this kind of software testing is actually fun

Happy Birthday to the greatest gaming system ever created!

This is a great one, and still valuable six months later! The Transactional Trap: How 97% of Developers Are Using AI Wrong w/ Leon Noel & Danny Thompson youtu.be/HJG8xd8hJHI?...

Margaret Hamilton, the brilliant scientist responsible for so many of our careers, has passed away at 90. Rest in Peace. news.mit.edu/2026/margare...

This clinical KG work records the facts and rule applications that justify a recommendation instead of asking an LLM to explain itself afterward. The pattern matters far beyond healthcare: evidence, reasoning, conclusion, action. arxiv.org/abs/2610.06549

Graph databases need boring deployment tooling too. terminusdb-migrations brings an Alembic-style revision workflow to graph schema changes, with tracked upgrades, downgrades, and interrupted-migration detection. pypi.org/project/term...

Neo4j’s Road to NODES workshops are doing something useful for GraphRAG: teaching the graph fundamentals and retrieval architecture together. A faster path for AI engineers who know RAG but still treat graphs as a black box. community.neo4j.com/t/register-r...

Understand Anything builds a graph of files, functions, dependencies, tests, and architecture so coding agents can reason across a codebase. Much more useful than another static diagram when the real question crosses module boundaries. github.com/Egonex-AI/Un...

Attestation turns provenance into something CI can enforce. It can trace claims in prose back to experiment artifacts and fail when the numbers no longer agree. A practical use of knowledge graphs for evidence, not decoration. github.com/mgoldey/atte...

Utopia models not just what is true, but when it was true and when the system learned it. That bitemporal design matters anywhere decisions need to be reconstructed later, from compliance to agent memory. github.com/deeplethe/ut...

Wikidata Search Traces trains models to navigate toward answers instead of memorizing them. The key idea is persistent graph state: keep evidence outside the context window, then selectively pull in what matters. arxiv.org/abs/2610.06650

kg-memory gives Claude Code and Codex a shared local knowledge graph for project memory. The interesting shift is architectural: agent memory is starting to look less like chat history and more like developer infrastructure. github.com/mironmax/kg-...

Microsoft GraphRAG may remove a document from your source set without fully removing its derived graph artifacts during incremental updates. Retrieval quality is only half the problem. Lifecycle correctness matters just as much. github.com/microsoft/gr...

repo2graph turns source code into bounded graph context for coding agents, then publishes benchmarks showing where grep still wins. A useful reminder that graph retrieval earns its keep on structural questions, not every search. github.com/Srinivasan-7...

Six Degrees of Ayrton Senna: Learn Neo4j by Connecting 75 Years of Formula 1 dzone.com/articles/neo...

Raspberry Pi Smart Display Module available now at $30 www.raspberrypi.com/news/raspber...

tokentab reads the session logs Claude Code, Codex and Gemini CLI already leave on disk and totals cost by model, project and day. Local only. No account, no API key. Cursor support is a stub, and the README says so. github.com/crwdla/token...

Capsule packs an HTML app and its data into one SQLite file with a .capsule extension. Entries carry UUIDs and timestamps so diverged copies can merge. Sandboxed, no network by default. Rust and Tauri. The format spec is promised at 1.0, so treat it as early. withcapsule.app

FlexViz keeps charts interactive at 100M+ rows by answering every zoom and brush with lazy Polars aggregations and Rust kernels. Nothing raw ships to the browser. Apache 2.0, on PyPI, with a live demo and public benchmarks. Early and small. Worth a look. github.com/flex-analyti...

Redis City: type a command and follow it through a 3D model of one Redis instance. Parsing, key lookup, the allocator, memory pages, RDB and AOF. Components link to the source. A visual explainer that teaches the internals instead of decorating them. poltora.dev/redis

Z.ai stood up GLM-5.3-Flash serving on 100k+ Chinese accelerators with an infra agent doing much of the work. Throughput tripled in under two weeks. The reusable part is dense feedback: correctness tests, traces and microbenchmarks, not one aggregate number. z.ai/blog/glm-bui...

Mark Litwintschik plans a fuel-optimal A320 route with Airbus's scikit-decide and the OpenAP performance model, then exports to Parquet and maps it over wind data in QGIS. Every command is on the page. The HN thread supplies the ETOPS rules the solver ignores. tech.marksblogg.com/scikit-decid...