Why the heck am I reading this? You have a thing you're putting off because it's boring and manual. You want the outcome, but not the toil. 10x velocity on shiny-new-thing is the wrong framing, nobody would have dared invest in your black hole of a backlog. Everyone's raving about …
Findings DeepSeek-V4-Flash-0731 runs fast on a single B300. I spent $40 to learn, so you don't have to! There is a reproducible issue with "Breakable CUDA Graph" enabled that causes a FlashMLA/TMA assertion and crash (triggered by my usual 8 harbor agents on regex-log) Workaround was to disable the …
Findings A single A100 with 80GB VRAM running Gemma4 31B with vLLM sustained 138 output tokens per second (serving 8 concurrent client agents). (extrapolates to a ceiling of ~500k tokens per hour) The bottleneck to more parallelism is the amount of VRAM available for KV cache; the model took roughly …
Reproducibility is under-appreciated When you say a bridge is engineered to carry a specified load - you must have reproducible evidence of that claim. In the era of AI and frenetic change, how to discern "good" among so many new Large Language Models? New LLM capabilities like "reasoning" or "tool use …
Untether your coding from your Laptop My "coding with AI" progression is likely similar to your own: (2022) "legacy coding" =] (2023) ChatGPT assisted "copy-paste" coding (2024) GitHub Copilot tab-completion coding (2025) Coding agents in the terminal: Claude Code and OpenAI Codex (2026) YOU ARE HERE We've all walked around holding …
Ambiguity in the Era of AI "Record who fed her cat food" Is it a human eating cat food, or is it a human feeding a cat? Record with video, or write it down? English is infamously imprecise - which is a challenge when writing software requirements. Throughout my software career …
AI Natives don't question the gifts from the gods At a recent SF hackathon we used ChatGPT to understand Clinical Trials, defaulted to Claude Code to code up everything, and used LangChain plus Minimax LLMs as the core solution. My teammates are AI natives: just graduating from university and they …
All you need is Agents and Human intent Took 3rd place (according to the AI judge) and won the "Scariest" award at the NEBULA:FOG, an AI × Security hackathon in San Francisco on March 14, 2026. My teammate's genius was giving human intent while 5+ (Claude Opus 4.6) agents …
Nobody Has the Answers I got my copy of "Refactoring" signed by both Martin Fowler and Kent Beck. That alone would have been worth the conference. What they shared in person at the Pragmatic Summit in SF was pure gold, Kent Beck put it bluntly: "At this moment, nobody knows …
Having the whole environment in the cloud can simplify setup and the hardware you need to experiment and learn. There are many ways to utilize AI/LLMs for different kinds of tasks, and accordingly a variety of types of AI models. Pre-Requisite Setup with Google Colab Working with models using …
To a hammer, everything looks like a nail Large Language Models are powerful but non-deterministic. Existing tools are still better for deterministic and efficient answers. Why not both? Definitions and Background Tools: simple things humans have used since forever =p Single Responsibility Principle - having a singular purpose and doing it …
Go slow to go fast: building my own frontend app platform sped up my experimentation by 10x I'm continually exploring free and low cost hosting providers as an avenue for unlimited experimentation within constraints. Cloudflare's evolution from simple CDN (ContentDeliveryNetwork) serving static assets (e.g. .jpg or .css) into the …
Why is weather so hard? Let's start with a problem... (rather than technology looking for a problem ;) Getting the weather can be frustrating which of the 6+ websites would you choose? how many annoying ads? why does one show the city first, another shows daily temperature, and a third jumps …
Why Software Agents are Eating the World Software scales to infinity and prints money the whole way. Once the cost of distribution went to zero the dominant constraint on growth stopped being infrastructure and became people: how fast could humans design, write, deploy (and maintain!) software. Over 30+ years, to …
tldr: a Golang native desktop app using Wails (Go backend + React frontend = single binary) Why this tech stack and why a desktop app? Desktop apps provide privacy for the persistence layer, and bypass the "where is it hosted" challenges. This is a simple and practical way to leverage some great …
The internet and HTTP has had an "agent" concept long before AI. https://www.rfc-editor.org/rfc/rfc9110.html#name-user-agent Connecting to a server (with TCP) underpins FTP and HTTP; a simple client can download resources. Sometimes they are commonly called a "scraper" or "crawler", I can remember one of …
Why Write? Because you have something to say. Technology is a tool, often helpful, sometimes not. Did the word processor end writing as we know it? What about the typewriter? (Gosh I miss the days of quill and blotting ink!) The bottleneck has never been typing speed or "output of …
Everything looks like a nail For Large Language Models (LLMs), Frontier aka "State of the Art" (SOTA) Models provided by big vendors like OpenAI and Anthropic and Google continue to add capabilities at a wondrous rate. Yet there's a very strong case for running LLMs locally, much like every smart …
tldr: React + Vite + MUI + simple Google sign-in for rapid building React is almost a defacto standard for responsive frontend applications, so it can be helpful to layer on a few more "standard choices" to simplify building apps for production. if you need a more basic react intro see my previous …
This post is to help people who are familiar with coding to build and understand simple reactive web applications. 2025 update : due to the never-ending nature of tech breaking things - this post needed an update (goodbye Create React App) https://react.dev/blog/2025/02/14/sunsetting-create-react-app Pre-Requisites install node …
Hiring is more than just a single decision A lot of interviewing advice is for candidates. I wanted to pull back the curtain on some experiences and tips for hiring managers, especially if it's your first time and you just got a headcount. =) (Hopefully this is also useful if you're …
When I tried to read my previous article on concurrency in Golang I felt like it tried to pack too much in so this is the same topic broken into parts, which may even leave room for more depth. - previous article https://blog.john-pfeiffer.com/golang-concurrency-goroutines-and-channels/ Background on Concurrency and …
Disclaimer: this piece reflects my experiences and is not legal or professional advice. It happens I've been through multiple layoffs: some I delivered, some I narrowly avoided being on the receiving end. It is never easy, never painless. This can happen to any kind of company: a startup that runs …
Compensation planning is like "extra hard Sudoku" except once you've figured it all out... you still have to convince each square that it's the right solution Always challenging, never a straightforward "satisfying" outcome. What is compensation planning and why do you care? Every organization has some fixed budget. Yes there's …
Look at the number 5: it can have many of representations: IIIII , 10/2 , 05.00 , 5^1 Those representations are not "wrong", they are all still "5". When people argue about a solution, they can both be "right", and yet there is a reason to pick one solution over …
Titles mean nothing People get hung up on titles and status: "Manager"... of what? Does this mean you get to tell people what to do? So a Director directs people? The job is always to solve (hard) problems and create value. The kernel of truth: Manager and Director are distinct …
Managers Own Shared Processes Messy, ambiguous, intersections of people and technology is exactly the domain of a great Engineering Manager. Caring about how people work together, understanding the organization as a system, and prioritizing business outcomes. It may take experience and serious people skills to understand the technical domain and …
Managers run out of time and emotional energy. Imagine a monkey on your back, stealing your spoons - that's what's happening! The solution = leading with empathy + coaching + ownership clarity Enduring, high-performing teams aren't built by heroes carrying loads of monkeys, they're built by consistent, sustainable management practices anyone can learn. Why …
When I realized my previous CI/CD vendor had finally broken something (after 7 years travis.org incompatibly became travis.com) I decided out with the old and in with the new. Also, Python2 was deprecated so onto Python3! So this article is a non-comprehensive reprise of How to Setup …
I avoid providing free advertising for products but often end up writing about how I have leveraged free products. =] I have many posts over the years using vmware, openshift, heroku, google app engine, aws elastic beanstalk, ec2, aws lambdas, github, bitbucket, bitbucket pipelines, docker, digital ocean, and many more things …
Bitcoin is a bad business model A good business model creates value. Eventually it will create good cashflow (i.e. more money in your bank account each month). While I understand that some may find the concept of an "unowned" virtual currency appealing, creating bitcoin (aka "mining") in and of …
We maybe take for granted that humans sort things: having a list of favorites e.g. "top ten songs". Computers (and software) are exceptional for sorting since it is both boring and requires precision; especially for large data sets that are beyond the human brain. Sorting things (not just numbers …
Infrastructure as Code helps guarantee the elusive determinism in infrastructure that we all seek in building applications and services. Often you can defer that complexity by using Heroku, Google App Engine, or other PaaS providers. But when you need to build something really complex (or with specific controls required including …
One of the challenges with "serverless" is how to develop locally, especially things like running integration tests in your dev environment. Imagine writing code for a lambda that reads and writes from S3, but without any AWS. The AWS tool SAM does have a local mode but does not cover …
A manager’s guide to keeping your team happy, productive, and aligned This same article is published at https://web.archive.org/web/20200415071712/https://blog.helix.com/helping-your-engineers-grow/ with images! As a manager, it can be unclear where someone’s career was before they met you and how they …
Career Development is Vague but Heavy It can be vague where someone's career was before they met you and how they view their long term ambitions. And when the work is getting done now it is easy to understand why future Career Development is one of the overlooked aspects of …
The promise of creating functions that do not require server administration is amazing, the reality though includes a huge maze of vendor specific commands and frameworks (including permissions). Why Function as a Service In many ways most of the work in software engineering is "accidental complexity". Deployment. Input/Output Parsing …
Overview Just like building and maintaining hardware and physical servers by hand is manual drudge work (aka "toil"), so is the manual deployment of servers and networking etc. even on "Infrastructure as a Service". The "Infrastructure as Code" movement emphasizes the value of human time and leverages source code and …
Prerequisites Please create a new version controlled source code repository using Github or Bitbucket or ?. Have Golang installed Download and install the official Go dependency tool "dep", https://golang.github.io/dep/docs/installation.html A simple http server Your source repository currently only needs a single file: main.go …
Idiomatic Go depends heavily on the standard library. Given that performance is often implementation dependent and directly related to the data being stored, it makes sense that one must write some of the more common data structures from scratch. While slices (a.k.a. dynamic arrays) are included (as is …
Some Background It can be challenging to scale up an organization. One choice that can be made is "full time employee" versus "contractor". Another choice is "Co-Located" versus "Remote". There are some important considerations with any combination but I'll share my experiences with "Remote Contractors". I've worked with remote contractors …
Go's considerable tooling makes it a very productive and performant static language. Besides being ideal for developing APIs (goroutines!) it can then be convenient to also write the Web UI in Go. Although the standard library allows for writing a Go based website with templates and maximum flexibility, what if …
Tradeoffs There is an inevitable tradeoff of storage versus computation speed, or space used versus time to run, in every program. Caching is often viewed as a performance optimization but sometimes it is the only way to achieve a result in a finite amount of time. https://en.wikipedia.org …
Productivity is dependent on and interwoven with so many things that anytime anyone pushes "more productive" I wonder... If you're "productive" and you get 3 things done in a day, or in a week even... The Many Measures of Productivity Were the 3 things done actually the important ones? (i …
One thing that critical to writing software is actually delivering value. That means shipping the bits, luckily the internet makes the cost of moving 1's and 0's near zero. Like any other chore or drudge work the release process should be automated. https://en.wikipedia.org/wiki/Continuous_delivery I attempted …
The future already happened The internet boom has arrived silently and the "everything is online" happened while we were all looking down at our smart phones. New giants have emerged to rule the economic landscape that have eclipsed, at least in virtual specie, their physical-and-brick-and-mortar competitors: Apple, Alphabet (Google), Microsoft …
If there is a killer feature to Go it is the focus on concurrency. This article captures some of the basics and I hope to someday write a follow-up article on more advanced topics. "Go is a compiled, concurrent, garbage-collected, statically typed language" https://talks.golang.org/2012/splash.article …
Code is the automation of a solution I often feel like the purpose of Programming is lost in the many discussions and debates about Programming. Sometimes the code is conflated as being the solution. Whatever problem is being worked on has to be thought through and different possible solutions evaluated …
As our lives become increasingly monitored and digital the privacy of being unobserved or having a private conversation that we used to be able to take for granted now requires extra effort. The more people who choose to use these easy and readily available tools the more privacy will become …
Parsing JSON can be a relatively simple subject covered early in other programming languages (i.e. JavaScript ;) JSON is a really popular way to persist or transmit data, especially for APIs. So it is really common to need to use it and yet... JSON can be surprisingly difficult in Go …