Engineering Notes.
Notes on large-scale training, GPU infrastructure, and what we're learning from engineering communities.
The Hidden Backdoor Risk in "Uncensored" AI Models
A proof-of-concept shows how a model downloaded from Hugging Face can carry a backdoor planted directly in its weights — and the part that should worry you most is how little it costs to pull off, at any model size.
DFlash-2: Benchmarking Z-Lab's Successor to DFlash for Accuracy and Throughput Gains
Z-Lab's inco.ai adds a Lightweight Path Selector and a Local Convolution layer to DFlash's diffusion-based draft model. We benchmark DFlash-2 against the original on Muse Glimmer 30B to see exactly where the gains show up.
Qwen3.8-27B: An Architecture and Hands-On Look at Alibaba's Compact Dense Model
Alibaba released Qwen3.8-27B shortly after its massive 2.4T parameter sibling. We examine its architecture and verify its hands-on performance against Muse Glimmer 30B.
Muse Glimmer 30B: An Architecture and Hands-On Look at Meta's New Mid-Range Model
We read through Meta's Muse Glimmer 30B architecture, then run it hands-on on a single GPU — standard inference, DFlash speculative decoding, and its Vision encoder.
Trying “DFlash,” a Diffusion-Model Approach to Parallel Draft-Token Generation, on Gemma
In this third part, we verify DFlash, an unusual MTP draft model that uses diffusion to predict multiple tokens all at once, comparing its real-world performance against Gemma's Assistant model.
The Hidden Frictions of Renting GPUs
Clicking "Rent GPU" is the easy part. Reading 77 Reddit accounts of what actually happens next — setup, storage, availability, and recovery — tells a different story.
MTP in Practice: Benchmarking Gemma's Speculative Decoding on a Real GPU
We ran Gemma-4-12B's MTP drafter on a consumer RTX 3060 and measured, turn by turn, exactly how much faster it gets. Part 2 of 2: the implementation/benchmark edition.
MTP: The Low-Risk, High-Reward Bet Behind Faster Generation
Qwen and Gemma both predict several tokens ahead to speed up generation — but they've built completely different machinery to do it. Part 1 of 2: the concept edition.
Getting Started with AI Agent Development: A Cloud Architecture Guide
From GPU selection to infrastructure design — the fundamentals every developer needs to know to run autonomous AI agents on their own infrastructure.
Stop comparing GPU clouds only by $/hour
The cheapest GPU instance is not always the cheapest way to finish a workload.
From Vibe Coding to GPU Spec-Driven Development.
The development style AI-era engineers need to know: why Vibe Coding hits a wall, and how Spec-Driven Development (SDD) picks up where it leaves off.
What AI engineers actually care about when choosing a GPU cloud.
We reviewed 100 Reddit threads on GPU cloud decisions. Price was the primary pain point in 23—but in 74, cost was not the primary decision criterion.
Get early access to Compute Cluster
Priority allocation and founding-customer pricing for early signups.