Blog
Writing on GPUs, LLMs, MLOps, Kubernetes — and mindset · 3517 posts
#2026-03 765#english 592#culture 264#deep-dive 254#kubernetes 247#career 229#ai 216#llm 208#devops 193#2026-04 146#security 141#database 114#observability 113#communication 109#history 107#architecture 100#productivity 96#finance 88#economy 84#mindset 81#psychology 80#ai-papers 79#food 78#it 78#travel 78#deep-learning 77#japanese 77#networking 77#performance 72#business-travel 70#linux 70#gpu 69#ai-agent 66#cs-fundamentals 63#postgresql 60#rag 58#self-improvement 55#learning 53#mlops 53#ai-platform 51
The Claim of 100x Cheaper Is True Only When the Task Was Narrowed — Verification and Break-Even
A case study published in August 2026 reports that a 4-billion-parameter-class open model, post-trained with reinforcement learning, matched frontier models on a retrieval task while cutting per-request cost by an order
2026-08-09 · 8 min read #llm#cost#fine-tuning#retrieval#open-modelsIt Was Not the CPU That Was Slow but the Syscall Path — Lessons From Turning a Phone Into a Server
Through the case of turning a CMF Phone 1 into a personal infrastructure server, this post lays out where the real cost of a Linux compatibility layer actually lands. After failing to flash postmarketOS, the author kept
2026-08-09 · 9 min read #linux#termux#chroot#proot#self-hostingA Screen Is Written With Ten Letters — The One Question to Ask Before Building a Custom Widget
Jakob Nielsen has laid out the claim that almost every user interface is assembled from about ten elements, the way every English word is written with twenty-six letters. The list is useful not because it tells you what
2026-08-09 · 7 min read #ui-design#ux#design-system#frontend#usabilityWhen Two Services That Passed the Type Checker Halt Waiting for Each Other — Choreography as a Different Approach
Using a language that guarantees memory safety does not stop two services from halting while each waits for a message from the other, because the field of view of a type checker ends at one process. Choreographic program
2026-08-09 · 8 min read #programming-languages#compiler#distributed-systems#type-systems#concurrencyWhen the Artifact Gets Cheap, Where Does Assessment Move — Why Denmark Chose the Oral Defense
The Danish Ministry of Education has issued an immediate package that makes an oral defense mandatory for exam assignments written at home. When the cost of writing approaches zero, the submission alone tells you nothing
2026-08-09 · 8 min read #assessment#hiring#code-review#education#engineering-cultureWhen Friction Disappears, Taste Does Not Remain — the Path to Growing Taste Disappears
Taste Is All That Is Left, the essay that drew attention in August 2026, says that as making things got cheap, the only ability left scarce is judging what is worth making. This post agrees with the diagnosis and then go
2026-08-09 · 8 min read #career#craft#ai#code-review#mentoringYour Deploy Is the Load Test — What Happens When Nobody Designs the Cost of Seeding the Cache
A walk through how Canva moved the in-memory session revocation cache in its gateway from MySQL to S3. The problem was not the steady-state lookup cost but the startup cost of hundreds of pods seeding their caches simult
2026-08-09 · 8 min read #architecture#caching#s3#scalability#deploymentWhy the Claim That Code Was Never the Hard Part Makes People So Angry
An essay that reached the top of Hacker News in August 2026 argues that saying code was never the hard part is an insult to every programmer. This post agrees with the rebuttal but locates the cause somewhere else. That
2026-08-09 · 8 min read #career#craft#ai#engineering-culture#skillsData Residency Is a Replication Topology, Not a Dropdown
Using the document Fastmail published when it opened an EU data region on August 3, 2026 as a textbook, this post lays out why the location of your data is not settled by picking a region once. The primary copy, the repl
2026-08-09 · 9 min read #architecture#data-residency#gdpr#replication#complianceIn Eval-Driven Development, the First Thing to Calibrate Is the Judge
The eval-driven development retrospective Airbnb Engineering published in July 2026 is less a plea to write the eval set first than a plea to earn the right to treat the grading model as an instrument. This post lays out
2026-08-09 · 9 min read #ai#llm#eval-driven-development#llm-as-judge#evaluationWhat Does a QR Code with a Photo Inside Pay For It — Error Correction Is a Budget
Take apart the technique for putting a photograph inside a QR code and it turns out to be a budget allocation decision rather than a design decision, because the slack spent on making it pretty was set aside for crumpled
2026-08-09 · 8 min read #algorithm#qr-code#error-correction#dithering#image-processingWhat LLMs Cannot Do Is Not the Proof, It Is Setting Up the Premise
The ICML 2026 position paper Position: LLMs can not jump argues that generative AI has mastered induction and is rapidly conquering deduction, yet remains structurally unable to reach abduction, the act of producing a ne
2026-08-09 · 8 min read #ai#llm#reasoning#abduction#researchExcel Shortcuts That Actually Save Time — Understand the Rule, Don't Memorize the List
Instead of listing 200 Excel shortcuts, this post groups them by the task you're actually doing. Once you understand why the arrow-key family always knows exactly where the data ends, or what order the absolute-reference
2026-08-02 · 16 min read #excel#shortcuts#productivity#office#keyboardHow to Read a Difficult Book — Five Judgment Calls That Matter More Than Finishing
Most of the time, what makes a difficult book difficult isn't unfamiliar words — it's the shape of the argument. Grounded in Adler's method for reading and in what reading research actually supports, this piece lays out
2026-08-02 · 14 min read #storytelling#reading#books#learning#culturePutting Observability Data Into ClickHouse — Schema, Rollups, TTL, and Splitting the Work
When traces and logs grow to multiple terabytes a day, a single search engine or time-series database starts to buckle. This post lays out why ClickHouse fits observability data so well, in terms of columnar storage, com
2026-08-02 · 16 min read #observability#clickhouse#opentelemetry#data-modeling#costWhat LLM Ops Actually Does — Reproducibility, Contamination, Checkpoints, Promotion, and Rollback
This post organizes LLM Ops as a list of responsibilities, not a list of tools. It covers what belongs in a run manifest that lets you reconstruct a training run, how to prevent and audit eval-set contamination, the form
2026-08-02 · 14 min read #mlops#llmops#reproducibility#evaluation#model-registryThe Four Kinds of Multi-GPU Parallelism — What You Split and What You Communicate
A numbers-first breakdown of what data parallelism, tensor parallelism, pipeline parallelism, and context parallelism each split, and what they pay in communication for it. It starts by building the per-parameter 16-byte
2026-08-02 · 13 min read #mlops#distributed-training#multi-gpu#fsdp#deepspeedLessons from Published Training Runs — What Was Tried, and What Failed
Seven published large-scale training technical reports and logbooks, stripped down to just the failures and the responses to them, not the scoreboard. The statistics of 419 interruptions over 54 days from Llama 3 405B le
2026-08-02 · 13 min read #mlops#llm-training#case-study#training-stability#scalingBuilding an AI Blog-Writing Pipeline — Verification, Not Drafting, Is the Bottleneck
This post breaks down the process of writing a blog post with AI into seven stages, from topic collection to checking performance, and separates out exactly where the model genuinely helps and where you must never hand o
2026-08-02 · 17 min read #ai-writing#content-pipeline#fact-checking#automation#bloggingHow to Actually Read a Model Card — Pulling Out What You Need in 5 Minutes
A model card is written in a way that, read top to bottom, keeps you from finding what you need. The benchmark table takes up half the screen while the license and chat template pass by in a single line. This post flips
2026-08-02 · 19 min read #llm#huggingface#model-card#license#tokenizer