Writing · 7 essays

Notes in public

Essays on AI architecture, agentic systems, neurosymbolic reasoning, and where current models stop working.

AI Architecture
2
AI Research
1
Agentic AI
3
Neurosymbolic AI
1

Featured · AI Architecture

26 min read

How to Train Your LLM

The whole training pipeline in one place, because when I went looking for that I found a dozen posts each covering one stage. Pre-training through preference optimisation and tool use, with what DeepSeek-V3 and Kimi K2 actually did rather than what the papers claim in the abstract.

August 10, 2025Read
  1. AI Architecture

    July 26, 2025

    The Anatomy of a Modern LLM

    Attention, the feed-forward block, where normalisation goes, what the tokeniser decides for you. Which of those choices are load-bearing and which are inherited. I wrote it as the reference I wanted while re-deriving the same diagram for the fourth time.

    10 min read
    • LLM
    • Transformers
    • Neural Networks
    • AI Architecture
  2. Neurosymbolic AI

    July 17, 2025

    Why AI Cannot Invent: The Abstraction Problem

    A model can describe a mechanism it cannot operate. My argument is that the gap is between interface-level and implementation-level understanding, that scale does not close it, and that what is missing is a symbolic layer with something to be wrong about.

    11 min read
    • Neurosymbolic AI
    • AI Limitations
    • Symbolic Reasoning
    • Philosophy
    • Research
  3. AI Research

    June 21, 2025

    Where are the LLM-Driven Scientific Breakthroughs?

    We were promised discoveries and got better literature reviews. My read: the bottleneck is the experimental loop, not the reasoning, and language models are strongest exactly where science is already cheap.

    7 min read
    • Scientific AI
    • Research
    • LLM Applications
    • Innovation
  4. Agentic AI

    April 4, 2025

    Local Operator: AI Agents That Solve Problems Creatively

    The agent environment I built to run on your own machine. It plans, writes code, executes it and reads what came back, with nothing leaving the device. Written when every framework on offer assumed a cloud sandbox and an API bill.

    5 min read
    • AI Agents
    • Local Operator
    • Open Source
    • Privacy
    • Automation
  5. Agentic AI

    February 23, 2025

    Path to Generic User Interfaces: Personal Budgeting with Local Operator

    Interfaces generated for the question you asked, instead of designed in advance for every question you might ask. My own budget is the worked example, because it is where a fixed dashboard always missed the thing I wanted to know.

    14 min read
    • UI/UX
    • AI Agents
    • Local Operator
    • Personal Finance
  6. Agentic AI

    February 3, 2025

    Local Operator: Self-Validating AI Agent for Fast Bootstrapping

    How the agent checks its own work: write code, run it, read the error, revise — so a task survives being wrong on the first attempt. The architecture behind that loop, and the point where self-validation stops helping.

    17 min read
    • AI Agents
    • Local Operator
    • Automation
    • Self-Validation