Topic

Large language models, once they have to serve real traffic

Benchmarks measure the model. Production measures everything around it: latency, cost per token, evaluation you can trust, and the failure modes that only appear at volume. Written from inside systems that had to keep running.

132 pieces

Teardowns

34

China-Available LLM APIs From Mainland Networks: The Real Map Three gates decide what you can call from a mainland connection: the network, the account and the payment rail. Only one is a firewall, and it is the one that matters least.
Keycloak Single Sign-On Across Distributed Sites, and the Parts That Bite Realms, PKCE, token clocks, brokering and the database nobody budgeted for. What actually breaks when single sign-on spans a dozen sites, and what to decide before it does.
Do Chinese AI Models Work for African Languages? The Answer Has Never Been Published Qwen3 claims 119 languages. Open the actual list and Swahili is the only Niger-Congo entry. Yoruba appears in the technical report once, as the language they left out of the evaluation.
When a Knowledge Graph Beats a Vector Database Vector search retrieves what looks similar. It cannot tell you which shipped assemblies a bad plating bath touches two hops out. That is traversal, and traversal is a different operation.
Harness Engineering: The AI Agent Guardrails Nobody Demos The model is a component you swap. The harness around it, tool schemas, validation, sandboxes, budgets, permissions and audit logs, is where nearly every production AI incident actually starts.
When AI Agents for Non-Technical Teams Reach the Shared Drive Agentic tools have moved out of the private chat window and into the folders a team depends on. What that relocation actually changes, checked against vendor documentation as of August 2026.
Tool calling reliability: DeepSeek, Qwen, Kimi open weights and what to check before you build Four separate properties hide inside the phrase "tool calling works". Here is what each of the three Chinese open-weight families actually documents, where the calls break and how to test it yourself.
How AI Text Watermarking Works, and What It Cannot Prove The mark is not added to finished text. It is made of the model's word choices, and it can show a model was involved without ever showing how much of the writing was yours.
pass^k vs pass@k agent reliability: your user only gets one run pass@k counts whether a success exists in k tries. pass^k counts whether every try worked. At a 0.61 success rate those read 0.99 and 0.09. Your user gets one run, so publish the second number.
How to Test LLM Applications Without Calling the Model The model is the only part of an LLM application that cannot be tested deterministically. Everything around it can be, and that surrounding code is where most production incidents actually start.
What Kimi, DeepSeek, Qwen and Doubao actually cost from Shenzhen Benchmark parity is not workflow parity. The four Chinese frontier models are genuinely competitive on price and quality, and the reasons a Western team still cannot use them are not technical.
Signing Off on AI Agent Security Without Lying to Yourself A language model cannot tell your instructions from the text it retrieves, because both arrive as the same token stream. Every real control follows from accepting that rather than trying to fix it.
Designing APIs for AI Agents When the Caller Cannot Read Your Docs An API read by a model has different requirements from one read by a developer. Names become prompt, errors become instructions and idempotency stops being optional once the retry loop is autonomous.
AI Visual Inspection Fails at the Lighting Long Before the Model Classical vision still wins geometric checks. Learned models earn their place on variable and rare defects. Accuracy is set by lighting, fixturing and the false-reject threshold you can afford.
Eight Seconds at a Time: What AI Video Generation Is Actually Good For Generative video in August 2026 is a shot generator, not a film generator. The documented limits explain why, and they explain exactly where it earns its place.
The agent control loop is where production agents fail Prompt engineering was never the hard part. The loop is: what the agent sees, what it may touch, what stops it and what a step costs. Most agent failures are loop failures, not model failures.
AI Legacy Code Modernization Is Not a Translation Problem Legacy modernization was never mainly a translation problem. The hard parts are comprehension and behaviour preservation, and a model's real value sits almost entirely in the first one.
Genome Language Models: What an Engineer Should Actually Take From Evo 2 DNA is a sequence and transformers are good at sequences. That analogy carries genome language models further than it should. What Evo 2 actually does, what it is used for and where the pitch breaks.
Do LLMs Work in Yoruba, Hausa and Igbo? What the Published Numbers Actually Say One of Nigeria's three major languages appears in the vendors' own published scores. The other two do not. Here is what the model cards, the African benchmarks and the tokenisers actually document.
Why a digital twin in manufacturing stalls at level two A digital twin is only useful when live telemetry feeds it and a decision comes back out. A 3D model that never updates is a rendering, and the data feed was always the expensive part.
Where to Rent GPUs in Africa: What Is Actually Plugged In For anything larger than a fine-tune, one country on the continent has rentable modern GPUs. What is actually plugged in, what is only announced and what either costs to use.
The AI Data Privacy Checklist I Run Before Shipping a Feature Retention windows, zero-retention gaps, personal data in vector stores and the erasure problem, written for the engineer who has to configure it before Friday.
LLM Evals Are the Regression Tests for Probabilistic Software Teams that cannot measure their AI system cannot improve it, so every change becomes an argument. Evals are the regression tests of probabilistic software and the first thing to build.
Five Prompt Engineering Best Practices That Hold Up Most prompting advice is folklore repeated without a control. Five practices survive testing, four popular ones do not, and the research reversed direction as the models improved.
Context Engineering Is Deciding What Earns a Place in the Window Windows grew to a million tokens and the industry assumed retrieval was solved. Attention still degrades, cost still scales, and everything you paste competes for the model's attention.
How Many Defect Images Do You Need to Train a Defect Detection Model: 25 or 12,568 Vendors quote 100 images per class. Published work says 25 can be enough, or 12,568 may not be. The number depends on how many ways your defect can look, and here is how to measure yours.
AI Agent vs Agentic AI: Four Rungs, and Where Your Vendor Actually Sits An agent is a program: a model called in a loop with tools. Agentic is a property of that loop, meaning how much of the goal, the plan and the action the model chooses. A dial, not a category.
OT Security Breaks the IT Playbook, and AI Agents Make It Worse Availability outranks confidentiality, you cannot patch a machine mid-run, and Modbus has no password because in 1979 the wall was the password. Then we connected all of it to AI.
How DeepSeek Builds World-Class AI on a Shoestring Budget The real story behind the $5.6 million model that panicked Silicon Valley, explained so you actually understand it.
Self-Balancing Robot Simulation: What Transfers and What Does Not The inverted pendulum is the standard teaching model because it fails informatively. Simulating one is easy. Transferring the controller is not, and the reasons are always the same handful.
What Actually Moves Through the Africa China Technology Corridor Trade numbers are the worst lens. From Shenzhen, six flows explain more: handsets and their preinstall slate, vendor certifications, open model weights, payment rails, components and people.
Generative AI for engineering design is cheap. Verification is where the money goes. The generative step in engineering costs minutes of compute. Proving the output manufacturable, simulation-clean and safe costs weeks, and that ratio decides which applications are worth doing.
Context Engineering vs Prompt Engineering: One Task, Two Fixes, Two Bills Prompt engineering optimises the wording. Context engineering decides what is in the window at all. As windows grew the bottleneck moved from phrasing to selection, and the bill moved with it.
The break-even on self-hosted open weights is 72 percent utilisation Renting two H100s to serve gpt-oss-120b costs $4,803 a month and beats a serverless endpoint only above 72 percent sustained utilisation, which almost nobody's traffic reaches.

Essays

16

Signals

1

Taalas Never Sold a Chip. AMD Bought the Memory Trick Underneath It The HC1 burns Llama 3.1 8B into silicon and quotes 17,000 tokens a second. Taalas' own product page calls it a technology demonstrator. That phrase is the whole story.

Projects

1

Presenting technical work in a second language changes the preparation Second place at Tencent's OpenClaw hackathon in Shenzhen, pitched and defended in Mandarin. What changes about preparation, and what reading Chinese sources first-hand actually buys.

Tools

12 of 80

All 80 matching tools