the archive
26 long-form articles — researched, source-verified, and written to be read rather than skimmed.
Six frontier models refuse ≥98% of real CVE reproduce-and-patch tasks. Mistral Large 4 completes them — and gets 82% right, the highest score ever measured. Charts, independent failure data, and the honest anatomy of what a defender’s model can and cannot do.
ai · 18 min · 2026-10-07
Anthropic’s own postings say the quiet part: formal certifications are explicitly optional, and the screening question asks whether you have shipped an LLM system to production for real users. The full path: what the market actually hires for, the seven-layer builder’s stack, the learning science on why watching fails, the honest failure numbers — and a build ladder to start this month.
ai · 21 min · 2026-10-05
Three weeks after Jev launched as the only System One model, the category has five new open models on Ollama (Nimble 9B, Tev1 4B/0.8B, Clef 27B, Clef-Flash 9B), a native /v1/systemone endpoint in Ollama 0.35, llama.cpp support, and a three-way benchmark war. The full map of the wave, with the honest reading of every number.
ai · 15 min · 2026-10-03
Two models defined the decision-model category in September 2026. Jev is a hosted frontier API; Laya is Apache 2.0 weights you can run on your own hardware. For a public body that difference decides whether the technology is admissible at all — measured against the EU four-level sovereignty framework.
ai · 18 min · 2026-09-27
Two camps rule agent building — hand everything to the model, or route every path by hand. n8n owns the middle: a fair-code canvas with 2,273 integrations, LangChain-powered agents, per-execution pricing, and a production ops story most coverage skips. Everything numeric here traces to a primary source.
tech · 15 min · 2026-09-27
Prompt engineering asked what words; context engineering asks what configuration of context is most likely to produce the desired behavior — curated every turn, because the window is a budget. The full map: the 2025 coining, the measured evidence, the four levers, a 14-pattern playbook with code, and the honest counter-theses.
ai · 31 min · 2026-09-18
Most of the data that decides whether a business works is a table. On tables, gradient-boosted trees still win — not from inertia, but from inductive bias. The published evidence, the mechanism, the operational reasons, and the exact line where deep learning takes over.
ai · 12 min · 2026-09-17
Frontier models only run in the cloud — but does that mean handing your business secrets to a provider? Zero Data Retention is the answer. This article breaks down the technology, compares the policies of Ollama Cloud, OpenAI, Gemini, and Anthropic against their own documentation, shows the risks that ZDR does not solve, and delivers a verifiable due-diligence plan for companies.
ai · 18 min · 2026-09-16
A model nobody can use is just a liability. Gradio is the interface layer that made machine learning usable: wrap any Python function in a browser UI with one line, from a single gr.Interface to full multi-page Blocks apps and ChatInterface LLM frontends — then share it with a public URL or deploy it to Hugging Face Spaces, the home of ML demos. Created by Abubakar Abid at Stanford and stewarded by Hugging Face, Gradio has become the default way the world touches AI. This is the complete tour.
tech · 17 min · 2026-09-13
Your teams are already using AI — a dozen SaaS subscriptions, data leaking at every boundary, and zero governance. Open WebUI is the antidote: one self-hosted, provider-agnostic command center that connects local Ollama models and cloud APIs, wraps them in SSO, RBAC, audit logging and RAG, and runs entirely inside your own perimeter. Here is why organizations from Samsung Semiconductor to Astellas Pharma are deploying it — and how you can be running it tonight.
ai · 15 min · 2026-09-13
Streamlit turns plain Python scripts into interactive web apps without writing a single line of HTML, CSS, or JavaScript. Here is an exhaustive tour: how its rerun architecture works, every widget type, caching and session state, multi-page apps, custom components, theming, deployment, security, and how it compares to Dash, Gradio, and Panel.
tech · 18 min · 2026-09-13
Most knowledge graphs are a snapshot frozen at index time. Graphiti is built for the opposite: facts that change, get superseded, and need to be queried as they were true. Here is how temporal context graphs work, why they beat static GraphRAG for agents, and a working Docker Compose stack to run one tonight.
ai · 14 min · 2026-09-09
Running AI locally sounds like it needs a PhD and a $10,000 computer. It does not. Local AI is a model file on your disk and a program that runs it — no cloud, no API keys, no subscription. Here are the real building blocks and the four ways to run a model on your own machine, from a clickable app to pure code.
tech · 14 min · 2026-09-08
LangChain 1.0 collapsed a sprawling ecosystem into one import path and made create_agent the standard way to build agents. This is the full tour — from a weather tool to structured output, memory, RAG, and the middleware system that lets you swap prompts and models at runtime — updated through the latest release, v1.4.0, which brings MCP support into the core package, and set in the modern three-layer stack of Deep Agents, LangChain, and LangGraph.
ai · 18 min · 2026-09-08
A diffusion model does not "draw" an image — it learns to reverse a process that turns an image into noise, then walks that reverse path from pure noise back to a picture. This is the full engineering tour: the DDPM objective, score matching and the SDE/ODE view, the U-Net-to-DiT architecture shift, the samplers that make it fast, and the conditioning stack that makes it obey a prompt.
ai · 22 min · 2026-09-02
Construction and cartography independently invented the same idea: a model as a database of objects with properties, relationships, and measurements, where geometry is one view, not the definition. Biology has a name for that pattern — convergent evolution. This article traces the two lineages, verifies the convergence in the standards themselves, and shows why AI is the selection pressure that makes the commitment urgent.
architecture · 18 min · 2026-08-31
AI agents are not a faster pencil — they are a new agent of production. From site analysis and generative ideation to drafting, fully functional BIM, and facility management, this is how the data generated at every stage compounds into value, and why the architect who does not integrate AI is forfeiting the data dividend.
architecture · 45 min · 2026-08-31
Every AI conversation starts from a blank slate — unless you give the agent a memory layer. That single decision separates a demo that repeats itself from a system that compounds. Here is how vector-store memory, a derivation worker, and a background embedder actually fit together, and why a "latest version" check is the smallest real example of agentic memory you will use all year.
knowledge · 12 min · 2026-08-29
Maps are becoming reasoning engines, not just pictures. From conversational querying and geospatial foundation models to autonomous monitoring, AI is collapsing the distance between a human question and a spatial answer. The numbers, the systems, and the outlook.
ai · 14 min · 2026-08-29
Your data lives in Postgres, in Parquet on S3, in an old SQLite file, in a pandas DataFrame, in a GeoPackage and in a folder of JSON dumps. DuckDB attaches to all of it with plain SQL — no servers, no pipelines. An exhaustive, source-backed tour from the vectorized engine to R-tree joins and RAG.
tech · 34 min · 2026-08-28
The model is the engine; the harness is the car. Why the real value — and the real risk — of AI now lives in the agentic layer that wraps a model in eyes, hands, memory, and schedules — with the market numbers, the DeepSeek cost revolution, and the EU AI Act timeline to back it.
ai · 24 min · 2026-08-24
Plain RAG is brilliant at finding a needle, but it fails when an answer means connecting the dots across a whole corpus. GraphRAG builds a graph first, then summarizes it. This is how, and when each retrieval approach wins.
ai · 13 min · 2026-08-24
A voice agent is never one thing — it is a microphone, an STT, an LLM, a TTS and a speaker wired together in real time. Pipecat models that wiring as an explicit pipeline of composable processors. Here is how, and how it compares.
tech · 13 min · 2026-08-23
What separates a coding tool from an AI employee is memory, preferences and a self-improvement loop — plus guardrails that hold when the context window compacts. A research-grounded look at autonomy, memory, and the failures that teach us.
ai · 12 min · 2026-08-22
From Polanyi and Nonaka to RAG and knowledge graphs — why most knowledge management is still an archive, and what it actually takes to make an org’s knowledge queryable instead of just stored.
knowledge · 12 min · 2026-08-21
The self-improvement loop, curated memory, external providers like Honcho, MCP connections, cron autopilot, skills — and how a SOUL.md file turns an agent into a colleague. The full recipe for an AI employee that remembers and improves.
ai · 11 min · 2026-08-20
