# External references — articles worth reading Curated articles, papers, and projects relevant to posttrainllm. Each entry is one-sentence-what + one-sentence-why-for-us + link, per the docs preference for leaning on authoritative external sources rather than re-explaining them. Updated 2026-06-08; GPU systems references added 2026-09-29. --- ## LLM internals — pedagogical explainers - **[The Illustrated Transformer — Jay Alammar](https://jalammar.github.io/illustrated-transformer/)** Visual explainer of the original transformer; got a 2025 refresh with animations. _Why for us_: the bar for "explain a transformer with diagrams" — any posttrainllm learning doc should link here rather than redraw the diagrams. - **[Transformer Explainer (Poloclub)](https://poloclub.github.io/transformer-explainer/)** Interactive GPT-2 running live in the browser; click any layer to see values flow. _Why for us_: pair with our forward-pass walkthrough sessions — lets readers poke a real model after reading our annotated code. - **[Lil'Log (Lilian Weng)](https://lilianweng.github.io/)** Long-form deep technical essays on inference, training, evals, and agents. _Why for us_: the quality bar for technical writing we should aim for in posttrainllm's own learning docs. - **[Lil'Log — Why We Think (May 2025)](https://lilianweng.github.io/posts/2025-05-01-thinking/)** 40-min read on test-time compute and why small models + better inference can beat scaling up. _Why for us_: direct intellectual backing for the posttrainllm thesis (specialists + on-device > frontier APIs). - **[Karpathy — 2025 LLM Year in Review](https://karpathy.bearblog.dev/year-in-review-2025/)** Opinionated short post on the paradigm shifts of the year. _Why for us_: useful framing for the HN-launch positioning of posttrainllm relative to where the field is. - **[ByteByteGo — How Transformers Architecture Powers Modern LLMs](https://blog.bytebytego.com/p/how-transformers-architecture-powers)** Diagram-heavy walkthrough of the seven-step decode loop. _Why for us_: closest reference for what a posttrainllm explainer post should look like structurally. - **[Stephen Wolfram — What Is ChatGPT Doing](https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/)** First-principles essay on next-token prediction without assuming any ML background. _Why for us_: the bar for accessibility; useful model for our `learn.md` entry-point. - **[Maxleiter — They're Made Out of Weights](https://maxleiter.com/blog/weights)** Terry Bisson "Made Out of Meat" parody — short dialogue framing LLMs as numbers all the way down. _Why for us_: example of literary/vibe-shaped writing about LLMs; inspiration for a Sarthak voice piece on specialists. - **[Terry Bisson — They're Made Out of Meat (original)](https://terrybisson.com/theyre-made-out-of-meat-2/)** The 1991 sci-fi dialogue Leiter's piece riffs on. _Why for us_: read first to get the joke. - **[Janelle Shane — AI Weirdness](https://www.aiweirdness.com/)** Long-running humor/AI internals blog (e.g., what neural nets name things weirdly). _Why for us_: closest practicing tradition for whimsy + ML mechanics; good rhythm reference. --- ## GPU systems and performance - **[Learn Inference](https://learn-inference.com/)** Interactive companion to _Inference Engineering_ covering inference mechanics, latency metrics, hardware, software, and serving techniques. _Why for us_: use its prefill/decode and KV explanations in Week 2, then its metric simulations in Week 3 before collecting local measurements. - **[Wafer AI performance-engineering resources](https://github.com/wafer-ai/gpu-perf-engineering-resources)** Curated primary-source index from GPU fundamentals through kernels, inference engines, and distributed serving. _Why for us_: select specific sources for Weeks 4, 6–7, and 10; treat CUDA and cluster material as boundary mapping until run on that hardware. - **[Modal GPU Glossary](https://modal.com/gpu-glossary)** An interlinked reference spanning NVIDIA GPU hardware, CUDA's programming model and software stack, and performance vocabulary. _Why for us_: use its roofline and arithmetic-intensity explanations for bottleneck predictions in inference Weeks 3–5, and its thread/memory hierarchy and coalescing entries for the Week 4 kernel diagrams. Treat NVIDIA-specific hardware details as a Mac-to-CUDA comparison, not a Mac execution result. - **[GPU MODE Triton Puzzles](https://github.com/gpu-mode/Triton-Puzzles)** Interpreter-backed puzzles that teach Triton loads, stores, masks, and increasingly complex kernels without requiring a GPU. _Why for us_: optional Week 4 prediction-and-check practice; verify the local setup before treating it as a Mac lab, and do not claim GPU timings. - **[GPU MODE Ring Attention experiments](https://github.com/gpu-mode/ring-attention)** Notebooks, experiments, and references for blockwise attention across GPUs. _Why for us_: a concrete Week 10 source to trace communication at the single-device-to-cluster boundary without requiring a cluster run. --- ## Apple Neural Engine / Core ML / Mac-native LLMs (Critical reading for the ANE arc — see `docs/learn/ane-research/dossier.md` for our synthesized dossier of these and adjacent sources.) - **[Apple ML Research — Deploying Transformers on the Apple Neural Engine (2022)](https://machinelearning.apple.com/research/neural-engine-transformers)** The canonical paper introducing the `(B, C, 1, S)` layout and ane_gelu / ane_silu / LayerNormANE patterns. _Why for us_: the reference our ANE M7 layout port works from. - **[Apple ML Research — On-Device Llama 3.1 with Core ML](https://machinelearning.apple.com/research/core-ml-on-device-llama)** Apple's own walkthrough showing Llama-3.1-8B at ~33 tok/s on M1 Max via Core ML. _Why for us_: direct precedent for the Qwen3 ANE conversion we're attempting; their KV-cache pattern matters. - **[HuggingFace — Running Mistral 7B with Core ML (WWDC '24)](https://huggingface.co/blog/mistral-coreml)** Step-by-step conversion of Mistral 7B to .mlpackage via coremltools. _Why for us_: the cleanest community walkthrough; pairs with our `scripts/ane/qwen3_to_coreml.py`. - **[HuggingFace — Releasing Swift Transformers](https://huggingface.co/blog/swift-coreml-llm)** Swift-side LLM inference plumbing. _Why for us_: parallel work to TinyGPTServe — worth checking for things they've already solved that we shouldn't re-implement. - **[CoreML-LLM (john-rocky, GitHub)](https://github.com/john-rocky/CoreML-LLM)** Community project hitting 52 tok/s on ANE for Gemma 4 / Qwen3.5 / Qwen3-VL with zero GPU contention. _Why for us_: closest direct competitor to posttrainllm's ANE work; understand their patterns before shipping our own. - **[Orion — Characterizing and Programming Apple's Neural Engine for LLM Training and Inference (arXiv)](https://arxiv.org/html/2603.06728v1)** Recent reverse-engineering paper on programming the ANE directly. _Why for us_: cited in our M8 research arcs; useful for going beyond what coremltools exposes. - **[BrightCoding — Stop Wasting GPU Cycles: CoreML-LLM Unlocks ANE](https://www.blog.brightcoding.dev/2026/05/23/stop-wasting-gpu-cycles-coreml-llm-unlocks-ane-for-insane-on-device-speed)** Community post benchmarking the CoreML-LLM project on ANE. _Why for us_: real-world tok/s numbers on the same silicon we target; sanity check for our benchmarks. --- ## Distillation + specialist models - **Predibase — 12 Best Practices for Distilling Small LMs from GPT** Practitioner checklist for teacher→student distillation at production scale. _Why for us_: closest analog to the Pace specialist arc; cross-check our methodology against their twelve points. - **[Labelbox — End-to-end distillation with Gemini](https://labelbox.com/guides/end-to-end-workflow-for-knowledge-distillation-with-nlp/)** Concrete walkthrough: label with teacher, train student, compare. _Why for us_: structure to borrow for the posttrainllm "factory" docs. - **[Nebius — The Concept Behind Distilling an LLM](https://nebius.com/blog/posts/concept-behind-distilling-llm)** Accessible intro for non-practitioners. _Why for us_: useful as a hook section in a future "specialists beat frontier" article. - **["Need a Small Specialized Language Model? Plan Early!" (arXiv)](https://arxiv.org/pdf/2402.01093)** Paper on lifecycle considerations for small specialist LMs. _Why for us_: backs the architectural choices we're making for Pace planner v6 / v7. --- ## Philosophical / literary takes on LLMs - **[A Philosophical Introduction to Language Models, Part II (arXiv)](https://arxiv.org/html/2405.03207v1)** Academic essay on what's encoded in model weights and what LLMs can be said to "know." _Why for us_: vocabulary for the "what is the model actually doing" framing we keep returning to in mech-interp work. - **[Reviving the Philosophical Dialogue with LLMs (paper)](https://philpapers.org/archive/SMIRTP-8.pdf)** Argument for treating LLM conversations as philosophical exercises. _Why for us_: orthogonal reading for thinking about Pace's voice-companion shape. --- ## Where this fits - See `curriculum.md` for the structured 7-session learning track. - See `docs/learn/ane-research/dossier.md` for our synthesized ANE research notes that draw on the Mac/Apple section above. - See `docs/learn/app-intents-comparison.md` for the App Intents study feeding v7 verb taxonomy. - New entries: add tight one-line `what` + `why-for-us` + link. Don't re-explain content that already has an authoritative source.