Workstation 部落格
全部文章
關於 AI 基礎設施、MLOps、模型以及上線生產的實操指南。
What an AI Factory Actually Delivers: One Rack, 32 Blackwell Ultra GPUs and the Work It Does for You
An AI factory turns your own data into answers, agents, forecasts and trained models on hardware you control. Here is what goes in, what comes out, how the work flows every day, and what a rack of four Compal SGX30-2 servers with 32 NVIDIA Blackwell Ultra GPUs needs to run.
Balinder Walia閱讀 
One Sign-In for Every App, Run on Our Own Cluster
Five demo apps in five languages never see a password and never talk to Google: they send users to a self-hosted OpsAPI, which signs them in and hands back a token the app checks with one API call. What it costs to add an app, why admin dashboards belong behind a VPN or zero-trust broker, and the honest limits.
Balinder Walia閱讀 
Keep Your Test Sites Private: VPN at the Edge, One Sign-In for Every App
Staging and test sites hold real-looking data, yet they're often protected by an IP allowlist alone. Here's how two layers keep them private: a VPN checked at the edge proves where a request came from, and one OpsAPI sign-in with MFA proves who sent it, across every app.
Balinder Walia閱讀 
Put Internal Apps Behind Your VPN, Without Rebuilding Them
Staging sites and admin panels are often 'private' because of an IP allowlist. Put them behind the wslvpn Zero Trust VPN, enforced at the WSLProxy edge: check the tunnel and the person, fail closed, and see why someone was blocked. One field per domain, no app changes, no reload.
Balinder Walia閱讀 
Balancing Pruning and Diversity
Test-time reasoning is increasingly the place where quality is won or lost. Generating many candidate reasoning paths and then selecting the best ones lets…
LLM Drift Detection in Regulated Environments
Regulated industries do not have the luxury of treating a Large Language Model deployment as an experiment. In banking, health tech, fintech and tax returns, a…
Hybrid Execution for Coding Agents: Macro Skills, Micro Recovery
Bundled macro skills make coding agents cheap but brittle; primitive tool calls make them robust but costly. Here is how to run macros first, verify every result, and fall back to small steps only when the workspace shows it is needed.
Generative AI's Hidden Impact on Developers: What Surveys Miss
Developers say AI makes them faster, but the evidence shows hidden costs: almost-right answers, a gap between felt and measured speed, tool admin work and accessibility barriers. Here is what the research says and how to measure it in your team.
Pixel-Space Diffusion Is Back: What Changed and How to Evaluate It
Diffusion directly on pixels now competes with latent diffusion, without a separately trained autoencoder. What HDiT, SiD2, JiT and 2026 research changed, where pixel models still lose detail, and a practical way to test the trade-off on your own images.
Cutting LLM Inference Costs: Lessons from Cloudflare's 68% Cut
Cloudflare cut its AI code reviewer's average cost per review by 68%. Here is what that number does and does not mean, and a measured playbook for your own workload: prompt cache breakpoints, fewer agent calls, right-sized models and batch.