Google Search Is Dying. What Comes Next Is Worse | The Walrus
As AI eats the web, the internet’s collective memory is disappearing
GitHub - antirez/h3.c: MiniMax H3 inference engine for Mac computers
MiniMax H3 inference engine for Mac computers. Contribute to antirez/h3.c development by creating an account on GitHub.
Joining Cursor
I’m joining Cursor (soon to be part of SpaceX) to work on Cloud Agents. I’ll still be managing state, only agents this time instead of databases.
The joy of simple computing
Progress, these days, seems to be equivalent to ever-more polished interfaces; more whitespace around buttons, glow, glassy, or frosty effects; smooth animations even on websites. For a long time, I appreciated this: using a sleek, pretty (in modern terms) app…
I Benchmarked Local LLMs on the Laptop I Already Have
Local-model benchmarks all seem to run on 128GB Macs I don't own. I ran five current open models on my 16GB M1 Pro against 21 real tasks, SQL and bug fixes and quick lookups, and priced the same work at API rates. The models did better than I expected, but the…
Dark mode toggles: two states are enough • Lea Verou
Yes, the underlying model must have three states, but one is always irrelevant to the actual user goal. Users do not seek out solutions to problems they don’t currently have. A lot of the hate towards two-state toggles is based on poor implementations.
The dumbest ransomware I've ever seen | Mr.UNIX
A few days ago, I saw a Facebook post from an anonymous person who shared an APK for an app they claimed was an AI teaching app for high school students. The …
The Norway Problem in Yaml
Norway's ISO country code is NO. Under the YAML 1.1 spec, unquoted no (case-insensitive, along with yes, on, off, y, n, true, false) is treated as a boolean and not a string. You can see the problem h
Distorted Reality
When you divide performance by salary, being expensive looks like underperforming. The metric quietly shifts from impact to intelligence per dollar, and everyone reading it ends up believing something that is not true.
Continual Learning Needs More Nuance
(This blog post is derived from this video response on my YouTube channel.) Dwarkesh Patel has been making the case for continual learning for a while, both with guests on his podcast and, most recently, in a blog post and an accompanying video. …
No, local models will not win
Every time a new open-weight AI model is released, people say that local models are the future. Why spend billions of dollars building out datacenters when everyone will just be able to run AI models on their laptops or phones? I think this idea is doomed. No …
Nathan Cooper - IPython is All You Need
"I use IPython as my terminal's shell."
![]()
Trust is all you need
And all we've lost
Claude Code self-improved on business workflows through conversation, not memories
Told only to get better, Claude Code raised held-out success from 34% to 44%. Ablations locate the improvement in the conversation, not the memory files it wrote.
Separating sex and gender is statistical malpractice – Notes from a data witch
Statisticians and academics in Australia should ignore the terrible advice of most LGBTIQA+ organisations and read this post instead. The 2020 ABS standard is almost never the correct way to collect data on sex and gender, and it causes huge practical problems…
![]()
When longing becomes hate
the author muddles throught thoughts about having lost something they never had, what could have been and how disappointment can lead to deep frustration
Patrick Colm Audley — Hacker · Full-Spectrum Technologist · Polymath
Seasoned technologist. AI/ML, security architecture, and capital markets. Thirty years of international startup experience.
Humanity renamed genes because of Excel
Science adapted to the program: gene symbols were renamed so Excel would stop turning them into dates.
How to lie with benchmarks
A look at bad benchmarking
Illinois HB5511: What It Means for Linux and Open Source
Illinois signed HB5511 on July 31. Past the TikTok and Instagram headlines is a mandate for operating systems, with no exemption for open source.
Escaping the LLM Coding Rat Race
There's only one winner in the AI coding rat race: your inference provider.
Agents remind me of sculpting
Imagine taking a big rock and slowly removing pieces from it until it becomes something meaningful or useful.
Tokenisation: Who decides what a token is anyway? — idlemachines
How does a neural network understand text? It's a question so basic, you can be forgiven for forgetting about it. Words and letters are human-readable, but are discrete (even if infinite in combination) and not compatible with the continuous representations th…
Watch out for cache read costs
For long agentic sessions cache reads are most of the bill, not input or output tokens. KV caches have shrunk dramatically while cache read pricing has barely moved.
With Flying Pen, When Mortals Sleep
It was in that moment, reading those words, that I had a new quest. I needed to see this letter. Not a translation on a page, but the real thing.
How I explored Tokyo and found peace there 🇯🇵
On my second trip to Japan, I left Tokyo's tourist trail behind and found calm in daily rituals, quiet neighbourhoods and unexpected encounters.
I Used an LLM to Quit Vaping | samscript
Summarising the books didn't work. Having something to argue with at midnight did. What LLMs were actually useful for when I quit nicotine.
Drive, meaning and status in the face of ASI abundance
What happens to human life when economic necessity disappears?
ChatGPT Already Knows Who It'll Recommend Before It Searches — Suganthan
ChatGPT writes brand names into its own search query before it fetches a single page. I read the traffic across 60 conversations to work out when it happens, what being named is worth, and what actually gets you onto that list.
every company needs a cassandra
an ai agent for the socially expensive work of organizational dissent
I've vibe coded an application to see what it's like
I’ve always wanted to write a simple, custom CLI application that would let me log amateur radio contacts quickly and easily. The focus was on logging satellite contacts and the desire to avoid using a database. Instead, the log entries should be edited, delet…
How I built a private travel journal for iPhone
How Travel Log turns occasional location updates and photo metadata into places, trips, and an exploration map completely on device.
Catching lightning in a VM — Peter Graugaard
How we built an open-source deterministic hypervisor in four months on a $200 budget.
The New Craftsmanship
Some of the best engineers I know have not touched a coding agent.
We're not done with point clouds
If you wait long enough to solve a problem, someone else might just solve it for you.
Agents Can Talk, But They Can't Commit
MCP, A2A, and AP2 standardized how agents find each other and speak. None of them standardize what happens when step four fails and steps one through three already happened.
Guitar lessons with Roku and Claude
Technical PM, homelab tinkerer, hacker.