Recommended Learning Sequence

This sequence is for building systems judgment around local AI inference, MLX, open model architecture, speculative decoding, and agent workflows.

Primary Path

1. Local AI Model Systems

Purpose: build the first mental model for model size, memory, KV cache, MoE active vs total parameters, MLX execution, and DFlash as a concrete case study.

Do this first if the question is: "What can fit on my machine, and why?"

Expected artifact:

• a local inference design memo for one target model and one machine

2. NanoTransformer in MLX

Purpose: implement before optimizing. Build transformer primitives directly enough that config files, cache shapes, and benchmark numbers stop feeling abstract.

Do this before the advanced systems material. It is the course that turns intuition into implementation evidence.

Expected artifacts:

• RMSNorm, RoPE, attention, KV cache, and decoder-block implementations
• logit matching report
• memory trace
• systems judgment report

3. Local AI Model Systems Advanced Lab

Purpose: return to architecture and runtime questions after you have implemented the core path. Study attention variants, position scaling, MoE routing, cache layout, scheduler design, profiling, speculative correctness, and DFlash internals.

Expected artifact:

• advanced local inference architecture review

4. Speculative and MoE Systems Lab

Purpose: implement and measure the techniques that change inference economics: quantization, Medusa-style speculation, MoE routing, paged allocation, runtime comparison, and evidence reports.

Expected artifacts:

• quantization error report
• acceptance histogram
• MoE load-forensics report
• page allocator report
• final evidence report

Agent Workflow Side Path

Take the RLM sequence when you want to connect local model systems to agent workflows:

1. Recursive Language Models Foundations
2. Recursive Language Models Applied Patterns
3. Recursive Language Models Systems Lab

The RLM track is useful, but it is a side path for the local inference goal. It becomes more valuable after the systems path, because you will understand what local execution can and cannot support.

Four-Week Sprint

Week 1:

• Local AI Model Systems
• memory estimator
• first local inference design memo

Week 2:

• NanoTransformer in MLX chapters 1 and 2
• RMSNorm, RoPE, attention, KV cache

Week 3:

• NanoTransformer in MLX chapters 3 and 4
• logit matching, memory trace, benchmark harness

Week 4:

• Local AI Model Systems Advanced Lab chapters 1 and 2
• config archaeology, attention variants, MoE routing, KV cache profiling

Eight-Week Plan

Weeks 1-2:

• Local AI Model Systems
• NanoTransformer in MLX chapters 1 and 2

Weeks 3-4:

• NanoTransformer in MLX chapters 3 and 4
• complete logit matching and systems judgment report

Weeks 5-6:

• Local AI Model Systems Advanced Lab
• DFlash deep dive and architecture review

Weeks 7-8:

• Speculative and MoE Systems Lab
• choose one capstone: speculation, MoE, quantization, or paged cache

What To Defer

Defer these until the primary path is solid:

• broad RLHF or DPO
• multimodal model systems
• full vLLM clone
• CUDA kernel writing as a main path
• agent framework surveys

They are interesting, but they do not remove the current root confusion as directly as building and measuring the inference path.