Blog
Research, product launches, engineering deep-dives, and company news.
Jev and Laya bring structured decisions to AI applications
Jev and Laya answer bounded questions with choices, scores and probabilities. Their practical value depends on the workload, the evidence behind their confidence estimates and the software that decides when to escalate.
Generating tool-use training data from executable workflows
ToolGrad constructs executable tool-call sequences before generating the requests used to train language models. The study examines how this ordering affects data generation, performance on unfamiliar tools, and the limits of scaling synthetic examples.
Synthetic neuron shapes improve automated brain mapping
MoGen uses generated 3D neuron fragments to improve the reconstruction of real brain tissue. The research connects advances in generative modeling to one of connectomics' central challenges: reducing errors that still require human correction.
Every Vector Database Ranking Is an Artifact of Its Workload
Three peer-reviewed and preprint benchmarks each report a vector database ranking that inverts under a different one of pgvector, Milvus, and Qdrant when a single measurement condition changes, which means none of the rankings, including our own from June, describe a system rather than a workload.
Vector Databases in 2026: A Systematic Performance Analysis of Indexing Architectures, Quantization Methods, and Production-Scale Retrieval
A comprehensive empirical and theoretical survey of vector database systems as of June 2026. We analyze indexing architectures (HNSW, IVF, DiskANN, CAGRA), quantization strategies (scalar, product, binary, RaBitQ), and filtered search performance across Qdrant, Milvus, Pinecone, Weaviate, pgvector, LanceDB, and Turbopuffer — with production benchmarks at 1M, 10M, 100M, and 1B+ vector scales.
Thinking From the Beginning and Other Ways of Working Things Out
Some questions you should work out from the beginning, and most you should not. A plain guide to the ways people work things out, what each one costs, and how each one fails.
A Survey of Open Weights Image Generation Models Running Locally in 2026
Every open-weights model that competes with Midjourney, GPT Image, Nano Banana, and Grok Imagine, organized the way a team actually has to evaluate one: base model lineage, the fine-tune ecosystem that grew on top of it, editing capability, hardware requirements, and the license governing the weights. The best open model by preference score is non-commercial. The most-downloaded fine-tune base is not the newest architecture. Renting a GPU beats buying one for almost every workload measured here.
Running an AI Assistant on Company Data Without the Data Leaving Vietnam
Vietnam's Decree 356/2025 counts cloud processing as a cross-border data transfer, which turns an architecture preference into a legal question. This is the cost model for all three routes, the break-even volume for self-hosting, and the arithmetic showing a hundred-person company sits 155 times below it.
How Letters Became Software
A history of type from cast metal to variable fonts, told as six changes in what a letter is: an object, a photograph, a grid of pixels, an outline, a program, a parameter space. Each stage froze an assumption the next one had to break, and the bill for the frozen assumptions was usually paid by the world's non-Latin scripts.