gde

Gemma 4 E2B served three ways on one Cloud TPU v5e chip: a pure-JAX engine in Python, the same graph written in Rust and compiled by the libtpu that JAX loads, and vLLM, all timed by one client. bf16 beats 4-bit in the JAX engine, the Rust engine matches the JAX engine's tokens, and over HTTP it decodes at vLLM's rate with the fastest first token.

We now have AuthService for authentication and SignInModalComponent for users to enter their email...

How to enable fair, zero-cost, air-gapped AI assistance in technical classrooms and...

In this blog post, we build a Sign-In modal component (SignInModalComponent) using the stable Angular...

Howdy friends! If you spend as much time playing with Google’s AI ecosystem as I do, you’re probably...

A tour of the A2A CLI against Rust agents on A2A v1.0: the official Rust SDK, a plain stdin/stdout program wrapped by a2a server --exec, and an ADK coordinator on Gemini, each timed from the CLI and from a Rust client.

How my workflow evolved from solo prompting in AI Studio to managing a multi-model AI team, running...

A practical, step-by-step engineering guide to packaging team workflows, MCP tools, custom agents, and deterministic guardrails into a single versioned Antigravity plugin.

Background I have a primary LINE Bot that I use heavily every day:...

Recap In the previous post "Comprehensive Analysis of Managed Agents API", I created an...

Why maintaining six separate parsers for Make, Gherkin, EBNF, DOT, SysV, and prompt checklists collapsed into a single Smalltalk-80 keyword cascade—and how 40-year-old Xerox PARC technology gave AI coding agents mathematical zero-amnesia.

Gemma 4 E2B with the model in Rust, served from a TPU v5e-1 VM through PyO3 and a C++ shim over ATen, with no tch and no pinned PyTorch, an OpenAI endpoint, an MCP server and a local chat UI.

A Rust crate reads the checkpoint, runs all 35 layers of Gemma 4 E2B and drives the decode loop on one TPU v5e chip in Colab, with every operation a TorchTPU kernel and the same tokens as compiled Python.

Anyone with the public URL can use my tokens to perform image analysis and text-to-speech, which can...

A deep dive into everything new in RackDoc DCIM v1.1: a 4-state background Heartbeat engine (Grey/Green/Orange/Red), independent Frontend SSH/HTTP/HTTPS launcher, 48-port switch CAD geometry, NOC Kiosk Mode, Rack Trays, and Active Directory RBAC.

The incident that never shows up in your deploy log Here is a scenario that's easy to...

Foreword: After the AI Agent Goes Online, What is the Next Bottleneck? Welcome back to...

Introduction: Why Decoupling Prompts Isn't Enough? In the previous post [Agent Platform...

Introduction: From "Manual Prompt Tuning" to "Two-Layer Recursive Autonomous...

Every public profile you own says something slightly different about you. A step by step guide to treating your bio as source code, with one file, one tracker and an agent skill that checks the live sites.

Day 33: From main.js Leaks to Hidden Tailwind CSS v4 Bloat — How I Built...

Why do an eager intern with a Hacker News blog post, a C-level executive stepping off an airplane with an in-flight magazine, and a trillion-parameter LLM all walk straight into the exact same production bog? And why even Google DeepMind can only bake half of our 600 scars into model weights.

An agent skill goes stale in six separate ways while its checks stay green. A step by step guide to measuring that drift and closing it with a feedback loop, using one publishing skill's own audit.

Gemma 4 E2B from BF16 to Q2_K on a Tesla T4, an NVIDIA L4, a GTX 1650 laptop and an RTX 3050 desktop, with byte-identical GGUF files and the same llama.cpp build. The ladder keeps its shape on all four cards, 218 questions die on all four, and any two GPUs grade 2 to 4 questions differently at BF16.

How we built a full-stack visual DCIM platform with interactive 2D floor plans, Front/Back 42U rack elevations, Bezier patch cabling, environmental sensors, and live network monitoring using Argon.

When a production Kubernetes cluster starts throwing 5xx errors , the hardest part of incident...

Using static Service Account (SA) keys in CI/CD pipelines has long been a common method for...

Why 90% of AI coding frustration comes from the Quota vs. Prompt-Compliance Death Loop, why Rich Sutton's Bitter Lesson fails in open-world software, and how we compile mid-flight collisions into 4-Vector Synthetic Scars and a Six-Pillar Critic.

Gemma 4 E2B from BF16 to Q2_K, speed against GSM8K accuracy, on an AWS L4 with byte-identical GGUF files and the same llama.cpp build as the Tesla T4. The L4 decodes 1.25 to 1.89 times faster, Q5_K_M is still the best trade, and switching GPUs changes nearly as many grades as quantizing from BF16 to Q5_K_M.

Gemma 4 E2B quantized from BF16 to Q2_K on one Tesla T4, every rung asked the same 250 GSM8K questions, and each answer buried at the rung where it first goes wrong. Q5_K_M and above hold the score; Q3_K_M loses 38; Q2_K gets 3 right. With thinking on, Q3_K_M wins back 17.