WORK
BENCH
A build-it-yourself manual for the gap between using software and understanding it.
Every lesson makes a problem happen on your own machine, explains it in plain words, and then has you build the fix. You predict a number first, then measure it. Jargon gets defined in the margin, every time.
- How slow is slow?How long does it actually take to read from RAM, an SSD, or the network?The Machine
- Where do my variables live?What are the stack and the heap, and where does a Python object actually sit in memory?The Machine
- Processes, threads and the GIL.What is the difference between a process and a thread, and why does Python have a lock that ruins threads?The Machine
- Waiting vs working.What do sync and async actually mean, and why does my server stall under load?The Machine
- Why ~17 tokens per second?Why is LLM decoding limited by memory speed, and how fast should my Spark be before I run anything?Inference
- Baseline: my Spark vs Fireworks.How does a naive model server on my own hardware compare to a commercial inference API?Inference
- Lessons done
- 0
- Lessons started
- 0
- Tasks checked
- 0
- Written / on map
- 63 / 63
█ done ▓ started ▒ ready ░ not written yet
Table of Contents.
Status1.THE MACHINE
What your code turns into once it runs: memory, processes, threads, waiting.
2.THE WIRE
How bytes get from one machine to another, from IP packets to streaming tokens.
3.DATA
Storing things so they survive, stay correct, and come back fast.
4.SCALE
Answering "how does it scale?" with numbers instead of vibes.
5.INFERENCE
The anchor project: serve a model on your DGX Spark and chase Fireworks, one optimization at a time.
- Why ~17 tokens per second?Ready
- Baseline: my Spark vs Fireworks.Ready
- How much memory does a model need?Ready
- The KV cache, by hand.Ready
- Batching and continuous batching.Ready
- Paged attention and prefix caching.Ready
- Quantization.Ready
- Speculative decoding.Ready
- Inside vLLM and SGLang.Ready
- FlashAttention and kernels.Ready
6.TRAINING
From numbers and memory to training your own models end to end.
7.EVALUATION
Knowing whether something actually got better.
8.AGENTS & RL
Long tasks, environments, rewards: how models learn to act.