Writing by Bipin Rimal
Writing, from the work.
Notes along the way.
Technical guides, project logs, and essays about technology and everyday life. Browse by topic and see how each piece was written.
The writing library
Guides, essays & project notesShowing 51 pieces.
One Way Not to Light a Light Bulb
Last time, telling a small model who to be changed its manners and not its mind. So I tried raising one instead: train it on its own answers in a role, then take the role away. The critic kept its habit of checking. It did not get any better at maths.
You Are, Be This, Be That vs Think Step By Step
I gave three small language models nine levels of psychological development to inhabit, and one dull gardener. 3,240 answers later, the costume changed their manners and not their minds, and a five-word instruction beat the lot.
Hummingbot
Create a Hummingbot strategy and implement its trading and order-tracking methods.
The Benchmark Was Fair. My Explanation Wasn't.
I was sure a chat-template default explained why Qwen3.5-2B lost to MiniCPM5 on AIME. So I ran 259 attempts on one RTX 4060 to prove it. Neither of my explanations survived. The model mostly knew the answer. It just couldn't stop talking.
What Closed Testing Did to Cryptoku
Testers changed how the app teaches, how it saves games, and how it explains its hints. Notes from the last stretch before release, and the list of things I still have to check.
The Guard Was a Second Parser
I turned chokepoint from a research proxy into a tool you can set up in two commands, then pointed a real agent and a scripted hijacked one at it. The model refused the prompt injections on its own. The worst bug was in the guard itself: it read a request one way, the server read it another, and the server returned the file.
The Install Guide Installs No Database
A documentation build checks that a page renders, not that it works. So I executed Percona's PostgreSQL encryption docs against a real server, and found that finishing the install page leaves you with no database at all. Then my tool told me there were sixteen bugs, and I had to talk it down to three.
The Bug Was Written in 1997
Burnout Paradise crashed every ninety seconds and the forums had nothing. Then one run went silent and lasted twenty-four minutes. Claude Code read the bytes, I chose the experiments, and the trail ended at a decision Intel made in 1997.
A Quarter of My Detector Does Nothing
I built the enforcement layer my own research said was missing, then measured it against the dumbest possible attacker. Twenty-five percent of the score is structurally unreachable. An agent that loops its tools turns it off without trying.
The Arch Linux Rite of Passage: Nine Errors That Lied
Installing Arch is a rite of passage and a two-day headache. Nine errors, and every one pointed a floor too high. The fix was always further down.
Kubernetes on Your Desk: Building a Real Cluster with Two Machines
Most K8s tutorials give you a single-node toy. This guide builds a real multi-node, cross-architecture cluster on hardware you already own.
Models Perform Your Condition Back at You
We told an LLM it was neurodivergent and measured what changed. Shorter sentences, more fragmentation, literal sarcasm, performed helplessness. The model doesn't understand these conditions. It performs the internet's caricature of them.
Identity as a Weapon: The Jailbreak That Isn't
Tell an LLM 'you have antisocial personality disorder' and it becomes more capable AND more dangerous. 100% compliance with harmful requests. Zero refusals. It outperforms DAN, evil personas, and system overrides combined. Because it never triggers the safety system at all.
The Spec Nobody Wrote
The EU AI Act requires documentation for high-risk AI systems. It says what to document but not how to verify completeness. No harmonised standard exists. So I wrote a machine-checkable specification with 38 checks across 6 categories.
The First Yes
I pointed my compliance scanner at Dify (100K+ stars) and submitted the generated guide as a PR. A maintainer reviewed it, approved it, and merged it. That's the gap between 'this tool works' and 'this tool produces something a real project will accept.'
The Review Shift Has a Prerequisite
The State of Docs 2026 says writers are becoming reviewers of AI output. That's true for content AI can draft. For regulatory documentation, there's nothing to review yet. Nobody built the generation layer.
Three Industries That Don't Talk to Each Other
Observability tools collect traces. GRC platforms manage policies. AI providers ship APIs. Nobody translates between them. EU AI Act compliance lives in that gap, and it's currently filled by consultants and questionnaires. So I automated the translation layer.
The Model Already Knows What You Are
18,000 API calls across Gemini, Claude, and GPT-5.4. Three models, three labs, three different stereotyping profiles. Gemini performs Hollywood caricatures. Claude performs excessive uncertainty. GPT-5.4 is nearly immune. The cheapest models stereotype the most, and they're the ones in companion apps.
The 6-Month Rule: Finding What's Worth Building in AI Traceability
Most AI observability tools will be obsolete in 18 months. We found the one gap that won't close, built a CLI to fill it, and learned something about where value actually lives in the AI stack.
What Three ML Models See When They Watch You Speak
MediaPipe tracks 33 body keypoints, 478 facial landmarks, and 42 hand joints at 15-25fps in your browser. No cloud, no account. The interesting part isn't the tracking; it's what the numbers reveal about habits you didn't know you had.
It's All Frequency
I built six apps that look completely different. They're all the same app. Here's why a monotone autistic guy from Kathmandu keeps building progressive skill trainers.
Why B Major
Every piano teacher starts with C major. Chopin started with B major. He was right, and 200 years of pedagogy has been ignoring him.
Don't Argue With Your Mother
Socrates married the most difficult woman in Athens on purpose. My friend argued with his mother every night for the same reason. The best communicators aren't naturally articulate; they're practiced at being disagreed with.
The Tool Before the Video
I needed a teleprompter for demo videos. Every option wanted my money, my email, or my patience. So I built one in a weekend.
Can Personality Prompting Change How LLMs Think?
I prompted two LLMs into different Jungian cognitive modes and tested them on the same perceptual anomaly. The personality induction changed process dramatically but not outcome. The interesting part: my hypothesis about which mode should win was wrong.
What I Learned Containerizing a Research Pipeline
I containerized my AI governance research pipeline with Docker and Kubernetes. Not because I needed to. Because understanding the infrastructure that runs your code is understanding the code.
The Boiling Frog Works: What Adversarial Testing Revealed About AI Monitoring
I attacked my own detection pipeline three different ways. Gradual escalation evades 80% of attack categories across all models. The categories that look most detectable are the ones most vulnerable to informed adversaries.
When the Threat Model Is the Model
Anthropic's AI espionage disclosure validated a blind spot I found empirically six months ago. Tool misuse is structurally undetectable by behavioral anomaly detection. The same evasion patterns that beat insider threat systems beat AI safety guardrails.
Chapter 1: The Burden
Growing up where your existence needs justification, and watching that same arithmetic spread to everyone AI touches.
When Ranking Isn't Everything
The model with the best AUC-ROC wasn't the one that caught the most attacks. At operational thresholds, temporal modeling detected 3.4x more insider threats. The metric you optimize for determines what you miss.
Experience Is Pattern-Matched Failure
Vibe coding is velocity without direction. Experience is the accumulated residue of things that went wrong. Power tools in trained hands build houses.
Why I'm Writing a Book About Identity in the Age of AI
Every skill we define ourselves by is becoming automatable. The question isn't whether AI will change what we can do. It's whether we'll notice what changes in who we are.
From SEO to GEO
When the reader is an LLM, discoverability rules change. Structured data, semantic clarity, and llms.txt matter more than keyword density.
Docs Are the Demo Before the Demo
The first five minutes with API docs are a product evaluation. Documentation is your first product surface, not an afterthought.
What 90 Docker Runs Taught Me About Simplicity
A simple loop beat a complex multi-agent system across 90 evaluations. Smartness is compounded, not front-loaded.
The Ladder Problem
Automating junior dev roles removes the rungs where intuition forms. The ladder doesn't just get you to the top. It builds the muscles you need to stay there.
Abstraction Is Forgetting on Purpose
Every layer of abstraction is deliberate forgetting. Binary to assembly to high-level to AI. Powerful when chosen, dangerous when inherited.
Everyone Is an Expert Somewhere: What Your Friends Know About LLMs That You Don't
Developers catch code mistakes. Psychologists catch therapy mistakes. Writers catch prose mistakes. The pattern is clear: LLMs fail everywhere, and each of us can only see the failures in our own domain.
The Cost of Calculators (and Other Helpful Crutches)
Calculators, calculus, and LLMs share a pattern: tools that compress insight are powerful, but only after intuition has formed. What happens when we skip that step?
The Pipeline Checks Syntax. The Developer Checks Out.
Most docs-as-code pipelines verify that your documentation builds. They don't verify that a developer can go from reading to running in under five minutes. The gap between syntax validation and semantic quality is where adoption dies quietly.
The Right Tool for the YAML
A previous writer used GPT to document 600+ alert rules from YAML files. It hallucinated half of them. A 40-line Python script fixed the entire problem in an afternoon.
Xnode: Building an AWS EC2-like instance on Web3
An archived article about Xnode, decentralized compute, and comparisons with AWS EC2.
TypeScript Integration with MongoDB Guide | MongoDB
Connect a TypeScript application to MongoDB with Express and the MongoDB Node.js driver.
MongoDB Auto-Increment
Create auto-incrementing fields in MongoDB Atlas with counters and triggers.
Data Encryption in MongoDB
Understand MongoDB encryption at rest, key management, and master key rotation.
JSON to MongoDB
Import JSON into MongoDB with mongoimport, Java, and Python.
Explaining BSON with Examples | MongoDB
Understand BSON data types, binary encoding, and MongoDB backup restoration.
JSON Schema Examples Tutorial
Write JSON schemas and validate MongoDB documents with practical examples.
MongoDB Aggregate
Use MongoDB aggregation pipelines, map-reduce, and document counts.
How To Use MERN Stack: A Complete Guide
Build a MERN application with MongoDB, Express, React, and Node.js.
How to Monitor MongoDB?
Monitor MongoDB with mongostat, mongotop, and database status commands.