Your ‘App’ Could Have Been a Webpage (so I fixed it for you…)
I was asked to install a mobile app to get access to the itinerary of a trip to Disneyland I'm taking with the kids. Fuck that noise: couldn't it have just been a Web page? An hour of reverse-engineering later, it is, and it's superior in virtually every way t…
People often explain OpenAI and Anthropic’s velocity with culture, talent or capital. I think there is a more mundane compounding effect at play: dogfooding.
Lean Software Production
As AI accelerates software development, I see something emerging that I think needs a name. Excuse me for getting fancy, but I propose we...
The PhD Metagame: Try Even a Little at Conferences - Maxwell Forbes
The bar is extremely low. Why not meet just 1 new person each day?
Why 'com_maxfps 250x' uncaps Quake Live's frame rate - Marco Nett
A Ghidra dive into a buggy integer parser that lets a clamped cvar run the engine at ~1000 FPS.
Who is walking who?
One good way to annoy a neuroscientist is to compare an LLM to the brain. It’s appealing though! There are similarities! In infancy we take a complex fusion of sensory inputs and learn to make pred…
Stop Telling Me To Ask An LLM
I already did.
Error Log Pattern Analysis is Hard (and Important)
Why building a catalog of error log patterns in a legacy system is harder than it looks, and why it matters more than ever.
The Restraint Problem
Agent leverage doesn't fix your bottleneck, it moves it. When building gets cheap, volume detaches from value, and the scarce skill becomes restraint. What one operator learned running a company on a fleet of agents.
What loss.backward() actually does
Backpropagation from scratch with microcrad — a tiny autograd engine in C that shows you every step.
If you have used Linux for a while, you have heard the warning. The AUR will eventually run untrusted code on your machine. It is true, and it is also why people look at Nix and wonder if its flake ecosystem could do the same job more safely. The pitch
In defense of not understanding your codebase
As a software engineer, how well do you have to understand your own codebase?
Welcome to LTT Labs - your go-to destination for all things tech. Explore comprehensive test results, insightful commentary, and the latest analysis in hardware.
Stop Making Threads
And you may ask yourself, "How do I work this?"
Someone drained a legacy OpenAI API key I made back in 2023 and used it to run their own AI agent training pipeline. Here is what I found digging through the logs.
Staying in the game
I’m four months into my interim CEO role at Bluesky, and it’s time for an update. Most importantly, as of today, the interim part of the title is gone. I’m loving the mission and …
Anyone can create write an article here! Simply make a PR to the sherbert repo or ask us on Discord! Changes usually take effect within ~10min of being approved.
Book Summaries — Brett Fisher
Accretive Editing | Justin D Fuller
Accretive Editing | Justin D Fuller
How my 2yo taught me constraint solving
My two-year-old son used a Brio wooden train set to explain constraint solving to his silly Dada.
Every legendary designer was once a junior that someone took a chance on. Jony Ive became Jony Ive because Apple invested in his talent over many years—no...
On Design (working thoughts)
Design-oriented software engineer, runner, and occasional writer.
The last time you wrote one it was a <button> and an FTP upload — spare twenty minutes and get up to speed on everything the web did while you looked away.
Heat now kills three times more Europeans than roads. A heat pump is a two-way air conditioner: cooling in July, heating in January, no imported gas.
Apple sues OpenAI, accuses ex-employees of stealing trade secrets - 9to5Mac
Apple has filed a lawsuit against OpenAI today, accusing the company of trade secret theft. Specifically, Apple alleges that its...
Conway’s Law applied to dependency management designs.
Abel Jansma is a scientist studying emergence, higher-order interactions, information theory, and complex systems across biology and artificial intelligence.
IEEE-754 NaN is weird. and because of that, it's often accidentally left unaccounted for. i found two instances of this leaking into programming language design. that is, the semantics of these languages hold implicit assumptions which break with NaN.
A Hacker News thread said a 27 billion parameter model is all you need for real coding work. I’ve run that model against real codebases. It isn’t, and here’s the math on why.
Building intuition about LLM parameter counts
The percentage of parameters that are used by different parts of an LLM can be quite unintuitive. Here's a visualiser for the GPT-2 architecture.
ymichael's personal site