THE NOTEBOOK / 23 ESSAYS
AI gets interesting when you look past the announcement. These are notes on what works, what falls apart, and how to build something useful yourself.
New to AI? Start with the learning guideAI & culture · 5 min read
Why synthetic polish spreads on LinkedIn, how it changes what we reward, and a practical way to make AI-assisted writing sound like you again.
Read the essay : LinkedIn AI Slop Is Making Us Worse at ThinkingResearch, useful disagreements, and things to build.
Benchmarks · 3 min read
Passing coding tasks is not the same as safely shipping changes. A practical look at what agent benchmarks reveal, what they leave out, and how to test a coding agent yourself.
Read the essay : Coding Agent Benchmarks Miss the Hard PartBuilding · 3 min read
A clear engineering guide to retrieval-augmented generation and long context windows, with a decision framework based on freshness, corpus size, citations, and failure cost.
Read the essay : RAG vs Long Context: Choose by Failure ModeAI & culture · 3 min read
A practical workflow for writing original technical content that uses AI without losing your examples, judgement, evidence, or natural voice.
Read the essay : Write Technical Posts Without AI SlopBuilding · 3 min read
Measure AI project cost in retries, review, latency, maintenance and errors. A practical worksheet to decide whether a feature is earning its complexity.
Read the essay : The Real Cost of AI WorkBenchmarks · 4 min read
How to read AI benchmarks without being fooled by a leaderboard. Learn to inspect task design, scoring, baselines, contamination, variance, and the gap between a score and your job.
Read the essay : An AI Benchmark Score Is Not a VerdictBenchmarks · 4 min read
What training data leakage means for AI benchmark scores, what contamination can and cannot prove, and how to evaluate models on genuinely fresh work.
Read the essay : Benchmark Contamination: When Tests LeakLearning · 4 min read
A practical AI learning path from Python and machine learning basics to language models, evaluation and a first useful project, with a weekly study loop.
Read the essay : How to Start Learning AI: A Practical PathBuilding · 3 min read
A step-by-step framework to choose a personal AI project, define its success measure, build a safe prototype and learn from the results.
Read the essay : Build an AI Project That Solves a Real ProblemBuilding · 3 min read
A practical first step for starting a startup: speak to potential users, test whether the problem happens often, measure current workarounds and avoid mistaking compliments for demand.
Read the essay : How to Start a Startup: Test the Problem FirstBuilding · 3 min read
Decide whether to automate a repeated task using frequency, stability, error cost and human judgement. Includes a practical automation scorecard and safe rollout approach.
Read the essay : When to Automate a WorkflowBuilding · 4 min read
A practical setup for using Claude Code with the current BMAD-METHOD: install versioned workflows, give the agent a clear project brief, and keep agile planning proportional.
Read the essay : Setting Up Claude Code with the BMAD MethodBuilding · 3 min read
A practical guide to repository instructions, task context, source-of-truth documents, code discovery and verification for coding agents like Claude Code.
Read the essay : Context Engineering for Coding AgentsBenchmarks · 3 min read
Create a trustworthy first evaluation set for an AI project: collect representative examples, write success criteria before comparing models, score failures and retain regression tests.
Read the essay : Build a Practical AI Evaluation SetLearning · 3 min read
A beginner-friendly explanation of tokens, context windows, transformer attention and next-token prediction, with examples that help debug real language-model applications.
Read the essay : Tokens, Context and Transformers: An LLM GuideLearning · 3 min read
A structured reading workflow for AI research papers: identify the research question, inspect methods and baselines, read result tables carefully and find limitations before trusting the headline.
Read the essay : How to Read an AI Research PaperBuilding · 3 min read
A launch checklist for a small AI feature covering user confirmation, output validation, data handling, monitoring, fallback behaviour and regression tests.
Read the essay : Ship an AI Feature With a Safe FallbackResearch · 5 min read
A technical look at DeepSeek Harness, Cordis, its plugin architecture, traceable sessions and runtime modes, plus what builders should examine before calling it the future of agents.
Read the essay : DeepSeek Harness and Cordis, ExplainedResearch · 6 min read
Google reports strong Gemini 4 Argon results, but what do the benchmark suites measure, which gains transfer to real work, and what does a different model cost?
Read the essay : Gemini 4 Argon Benchmarks: What the Scores MissResearch · 4 min read
A close look at Phonon-2's compact weights, reported word error rate, Apple Neural Engine speed, licensing and the real evaluation questions for local transcription.
Read the essay : Phonon-2: Open Speech Recognition on DeviceResearch · 4 min read
How to assess OpenAI's October 2026 mathematics results: what the Lean formalizations establish, what the human-written manuscripts still need, and why discovery is not the same as verification.
Read the essay : What AI-Generated Math Proofs Actually VerifyBenchmarks · 4 min read
Understand how WER works, which transcript errors matter in practice, and how to build a speech recognition evaluation using your actual microphones, accents and vocabulary.
Read the essay : Why Word Error Rate Isn't EnoughBenchmarks · 3 min read
A repeatable method for comparing AI models on your own tasks, controlling prompts and tools while tracking success, retries, latency and total operating cost.
Read the essay : Compare AI Models on Cost, Latency and Quality
