Systems engineering monograph / Compute workbench
Work through the Whybit stack: local AI hardware, systems and audio tools, foundational texts, then the episode breakdown that ties them together.
#ad As an Amazon Associate I earn from qualifying purchases.
Open the Systems Desk
Episode breakdown index
No records match that search.
Product tier 1
The parts that carry the model: the graphics cards, the memory, the power and the machines they live in.
The component behind Why AI Confidently Makes Things Up; How Does A Model Route One Token To One Expert?.
ServesWhy AI Confidently Makes Things Up, How Does A Model Route One Token To One Expert?
The component behind Why Does a Language Model Sound Sure When It Is Wrong?; Why Can A Model Think Longer Before It Answers?.
ServesWhy Does a Language Model Sound Sure When It Is Wrong?, Why Can A Model Think Longer Before It Answers?
The component behind What Does Fine-Tuning Change?.
The component behind Why Do Experts Disagree on When AI Matches Humans?; How Does A Model Guess Five Tokens And Then Check Itself?.
ServesWhy Do Experts Disagree on When AI Matches Humans?, How Does A Model Guess Five Tokens And Then Check Itself?
The component behind Where Does a Language Model's Context Stop?.
The component behind Why Attention Gets Expensive When The Context Gets Long?.
ServesWhy Attention Gets Expensive When The Context Gets Long?
The component behind What Architecture Is a Language Model Usually Built On?.
ServesWhat Architecture Is a Language Model Usually Built On?
The component behind How Do Words Become Vectors You Can Measure?; Why Does A GPU Move A Model Faster Than A CPU?.
ServesHow Do Words Become Vectors You Can Measure?, Why Does A GPU Move A Model Faster Than A CPU?
The component behind How Does a Language Model Pick the Next Word?.
Product tier 2
The desk around the machine: the storage, the network, the switch and the meter that proves what it drew.
The component behind How Can a Model Answer From Documents It Was Not Trained On?.
ServesHow Can a Model Answer From Documents It Was Not Trained On?
The component behind What Does Bad Training Data Do to the Answer?.
The component behind What Are The Layers Of Memory A Model Has To Move Through?.
ServesWhat Are The Layers Of Memory A Model Has To Move Through?
The component behind What Changes When a Language Model Can Use a Tool?.
The component behind Why Does Rounding Numbers Save Memory And Cost Accuracy?; Why Does A Model That Explains Its Reasoning Still Get It Wrong?.
ServesWhy Does Rounding Numbers Save Memory And Cost Accuracy?, Why Does A Model That Explains Its Reasoning Still Get It Wrong?
Product tier 3
The texts that explain the model and the system around it.
Publication: 1st edition, 2025, O'Reilly
A practical guide to application-layer reliability, retrieval, evaluation and model-system behaviour.
ServesWhy AI Confidently Makes Things Up, How Can a Model Answer From Documents It Was Not Trained On?, Why Does A Model That Explains Its Reasoning Still Get It Wrong?
Publication: 1st edition, 2022, O'Reilly
Maps the full machine learning system from data and model pipelines to deployment and monitoring.
ServesWhy Does a Language Model Sound Sure When It Is Wrong?, What Changes When a Language Model Can Use a Tool?, What Does Fine-Tuning Change?
Publication: 2nd edition, 2017, O'Reilly
Explains storage, retrieval, replication and the systems foundations behind reliable data applications.
ServesWhere Does a Language Model's Context Stop?, What Does Bad Training Data Do to the Answer?
Publication: 1st edition, 2023, MIT Press
Provides a clear mathematical foundation for transformers, attention, optimisation and generalisation.
ServesWhat Architecture Is a Language Model Usually Built On?, What Are The Layers Of Memory A Model Has To Move Through?
Publication: 1st edition, 2016, MIT Press
A canonical reference for representation learning, optimisation and the foundations of deep networks.
ServesWhy Attention Gets Expensive When The Context Gets Long?, How Does A Model Route One Token To One Expert?, How Does A Model Guess Five Tokens And Then Check Itself?, Why Does Rounding Numbers Save Memory And Cost Accuracy?
Publication: 2nd edition, 2018, MIT Press
Explains reward, value, policy optimisation and learning through feedback.
Publication: 3rd edition, 2022, O'Reilly
A practical foundation for loading, cleaning, transforming and analysing data with pandas.
Publication: 1st edition, 2018, O'Reilly
Frames the production concerns of training pipelines, packaging and deployment.
Publication: 2nd edition, MIT Press, 2018
A full treatment of agents, planning, search and the decision machinery behind a model that has to act.
Publication: O'Reilly, Learning Python series
A working reference for the Python that the analysis and tooling in these episodes are written in.
Question and episode breakdown matrix
| No. | Question | Layer | Stack components | Watch |
|---|---|---|---|---|
| Q01 | Why AI Confidently Makes Things Up | Local AI & GPU Compute | NVIDIA GeForce RTX 5090 32GB graphics card, AI Engineering: Building Applications with Foundation Models | Watch the channel |
| Q02 | Why Does a Language Model Sound Sure When It Is Wrong? | Local AI & GPU Compute | NVIDIA GeForce RTX 4090 24GB graphics card, Designing Machine Learning Systems | Watch the channel |
| Q03 | Where Does a Language Model's Context Stop? | Local AI & GPU Compute | Crucial 128GB DDR5 5600MHz desktop memory kit, Designing Data-Intensive Applications | Watch the channel |
| Q04 | How Can a Model Answer From Documents It Was Not Trained On? | Systems Desk, Storage & Power | Synology DS923+ 4-bay NAS, AI Engineering: Building Applications with Foundation Models | Watch the channel |
| Q05 | What Changes When a Language Model Can Use a Tool? | Systems Desk, Storage & Power | TP-Link Omada 8-port PoE managed switch, Designing Machine Learning Systems | Watch the channel |
| Q06 | What Does Fine-Tuning Change? | Local AI & GPU Compute | NVIDIA RTX 6000 Ada 48GB professional graphics card, Designing Machine Learning Systems | Watch the channel |
| Q07 | What Does Bad Training Data Do to the Answer? | Systems Desk, Storage & Power | WD Elements 20TB portable external hard drive, Designing Data-Intensive Applications | Watch the channel |
| Q08 | What Architecture Is a Language Model Usually Built On? | Local AI & GPU Compute | Corsair RM1200x 1200W ATX 3.1 power supply, Understanding Deep Learning | Watch the channel |
| Q09 | Why Do Experts Disagree on When AI Matches Humans? | Local AI & GPU Compute | NVIDIA Jetson Orin Nano 8GB developer kit, Artificial Intelligence: A Modern Computational Agent Approach | Watch the channel |
| Q10 | Why Attention Gets Expensive When The Context Gets Long? | Local AI & GPU Compute | Noctua NH-D15 chromax.black CPU cooler, Deep Learning | Watch the channel |
| Q11 | How Does A Model Route One Token To One Expert? | Local AI & GPU Compute | NVIDIA GeForce RTX 5090 32GB graphics card, Deep Learning | Watch the channel |
| Q12 | How Does A Model Guess Five Tokens And Then Check Itself? | Local AI & GPU Compute | NVIDIA Jetson Orin Nano 8GB developer kit, Deep Learning | Watch the channel |
| Q13 | Why Can A Model Think Longer Before It Answers? | Local AI & GPU Compute | NVIDIA GeForce RTX 4090 24GB graphics card, Reinforcement Learning: An Introduction | Watch the channel |
| Q14 | How Do Words Become Vectors You Can Measure? | Local AI & GPU Compute | Intel NUC 13 Pro Core i5 mini PC, Python for Data Analysis | Watch the channel |
| Q15 | Why Does Rounding Numbers Save Memory And Cost Accuracy? | Systems Desk, Storage & Power | Owl 120V 240V power meter plug energy monitor, Deep Learning | Watch the channel |
| Q16 | Why Does A GPU Move A Model Faster Than A CPU? | Local AI & GPU Compute | Intel NUC 13 Pro Core i5 mini PC, Machine Learning Engineering | Watch the channel |
| Q17 | What Are The Layers Of Memory A Model Has To Move Through? | Systems Desk, Storage & Power | Samsung 990 PRO 4TB NVMe PCIe 4.0 SSD, Understanding Deep Learning | Watch the channel |
| Q18 | Why Does A Model That Explains Its Reasoning Still Get It Wrong? | Systems Desk, Storage & Power | Owl 120V 240V power meter plug energy monitor, AI Engineering: Building Applications with Foundation Models | Watch the channel |
| Q19 | How Does a Language Model Pick the Next Word? | Local AI & GPU Compute | Raspberry Pi 5 8GB developer board, Learning Python | Watch the channel |
This page sets no cookies and runs no analytics. The outbound Amazon links may set their own cookies, which Amazon controls. Nothing is tracked on this page.
Original editorial artwork only. Product links are affiliate search links, not scraped product pages. Prices can change on the store site.