Blog on cmart's blog

Regulate Obvious Physical Choke Points of the Frontier AI Supply Chain?

Wed, 16 Sep 2026 01:00:00 -0700

Table of Contents:

(I expect this will sound a hand-wavy and under-researched, because it is! I do believe my premises are close enough to give a vivid sketch, and maybe the ‘regulate physical choke points’ framing will help at this stage of public discourse, as folks stare down the hairy task of slowing AI. I am probably wrong in at least one important way. If you spot a mistake or misunderstanding, please email me (wrong-about-ai at cmart dot blog) and I’ll update the post. If you think “this dummy hasn’t read relevant literature X”, please link me to it, but also understand that this was a first-principles thought experiment from someone not previously working in AI safety or alignment.)

The News Tonight is AI Danger

Sat, 12 Sep 2026 22:45:00 -0700

Top story: CNN, New York Times, Washington Post, BBC, CBS, NPR.

Visible without scrolling: Wall Street Journal, Forbes.

Scrolling to 2nd-5th screen: MS Now, Fox News, CNBC, Yahoo! News, Reuters.

Not on first 5 screens: New York Post.

This is not my filter bubble! I started with a pretty clean Chromium browser, deleted all browsing data again, then visited all the mainstream news sites I could think of. (I otherwise use a pretty heavily-customized Firefox.)

If Social Media is Cigarettes, Maybe Group Chats Are Vaping

Mon, 07 Sep 2026 00:00:00 -0800

I had the common 2020 experience of scrolling infinitely on Facebook and Twitter, then feeling worse for having done so. Eventually I had the sense to quit these platforms. Years later, it dawned on me that private group chats have many qualities of social media, beyond just connecting you to other people. While there is (typically) no advertising or adversarial-to-the-user content curation, I still feel impulses to:

Drop whatever I’m doing to check for new messages, especially for replies and reactions to my own recent messages.
Send low-effort GIFs, memes, or joking-but-ambiguously-serious messages that pollute the information commons (despite feeling smug at the time).
Have disagreements in group chats with friends.¹

After noticing that my digital home was full of information candy² that I couldn’t stop compulsively eating, I ~~mixed it with cat poop and took it outside~~ departed many group chats today, including chats that I personally started and grew over a period of years.

Cowork Cafe

Wed, 12 Aug 2026 00:00:00 -0800

(Previously: No Screens Cafe)

At the Cowork Cafe, you can look at your little screen when you’re supposed to be looking at your medium-size screen.

Welcome to the Cowork Cafe! We can’t wait to optimize you. Please set your chin here, and look at the green light. Your retinae are scanned to ensure a seamless experience joining our 10-gigabit Wi-Fi.

If the front door is locked, you can use the scanner outside. To open the door? No, sorry. It doesn’t work that way. But as a member, you can still use the outside scanner 24/7.

How cmart's Blog Uses (and Doesn't Use) AI

Tue, 11 Aug 2026 00:00:00 -0800

Update: for posts published since August 2026, I’ve also started using the no-ai-use tag for posts that go further than this disclosure. On my word, I do not use any LLMs tools in the process of authoring no-ai-use posts. (It’s still possible that I got the initial post idea as a result of prior interactions with an LLM, but this seems infeasible to account for, as people are not systematically good at identifying exactly where an idea came from.)

Why is Anthropic's public writing style so unlike Claude's?

Mon, 03 Aug 2026 00:00:00 -0800

Claude has a very distinctive writing style: short punchy sentences and eyeball kicks. We have automatic highlighting tools that don’t even try to predict LLM use in a fancy statistical way, only match on phrase patterns that hardly anyone used prior to Claude, but which now appear everywhere thanks to Claude.

Most of the time, I don’t even need the highlighting tool. I’ll start reading something like this, or this, and think this is obviously Claude’s voice.

cmart's AI Slop Machine

Sun, 12 Jul 2026 00:00:00 -0800

I, a human, write every word of this blog in my own voice. Nonetheless, I built an AI Slop Machine.

Your browser does not support the video tag.

Background

(Boring! Skip to the pretty chart)

American and Chinese AI labs continue to make increasingly strong open-weights large language models (Qwen, Gemma, gpt-oss, and others). You can self-host these LLMs on consumer-attainable computers and GPUs. They’re pretty good at agentic knowledge work tasks, and when you use them, several of the things that people say are bad about AI don’t apply.

LLM: Learningful Links and Musings

Mon, 06 Jul 2026 00:45:00 -0700

Some brief meta:

Inference Cards

Thu, 25 Jun 2026 00:00:00 -0800

Why

skip past the why

When someone says “I run Qwen 3.6 at 25 tokens per second”, or makes any similar performance claim about their self-hosted LLM setup, this is only meaningful if we know several other things.

Which model variant? Qwen 3.6 could be the dense 27B or the 35B-A3B MoE, totally different architectures. Better, just link to the repo you downloaded the weights from.
Which quantization? Q8, Q4_K_XL, and IQ3_XXS are at different points in quality/speed/size space.
What hardware and inference engine? Is this vLLM on an H100 or llama.cpp on a Raspberry Pi?
I could ask if you’re using speculative decoding or various weird stuff, but better, show the command you used to invoke the inference engine.
How did you measure the speed? Are you exercising an HTTP API (which adds a possibly-large chat template in the context window), or using something like llama-bench (which skips the template and HTTP/network delay)? Or better, what command did you use to run the test?

Also, knowing only the generation (i.e. decode) speed at a shallow context depth is not enough to understand whether agentic workloads will be usable on a given setup. Prefill (i.e. prompt processing) speed matters because agents spend a lot of time reading stuff. It also matters how speeds change as context depth increases, because agents do most of their work with tens of thousands of tokens in the context window. Also, if you’re trying to serve multi-agent (or multi-user) workloads, it matters how these numbers change with multiple concurrent requests. (And no, you cannot guesstimate any of these other numbers from “25 TPS generation speed” because different hardware and inference engines all have different performance characteristics in this several-dimensional space.)

PlaidRunner: Rebuilding my E-bike

Mon, 23 Mar 2026 15:00:00 -0800

I routinely use a bicycle for trips that people cannot conceive of, like picking up my partner from the airport, at night, after it rained all day. (Somehow she is amenable.) This is my life since buying the oldest, most basic model RadRunner from Craigslist. The frame is stout and very capable for its length. It will carry a whole adult passenger on the back plus a load in the front, but it’s a normal-size bicycle for maneuvering and storage purposes. It still works perfectly after three years of hard use, but I decided to upgrade/ruin it with these goals.

Surprisingly Frugal Hobbies

Sun, 22 Feb 2026 19:30:00 -0800

Oh no, a listicle!

Boating

A kayak gives you one fixed seating position. The few times I’ve tried it, my ass needed a break within an hour. But I love being on water. A fidget-friendly alternative is inflatable stand-up paddleboarding. I overpaid for mine with REI scrip, but I hear the $200 ones on Scamazon are fine, and they come with a paddle.

Inflatable means kneeling, sitting, and lying down are all comfy when you tire of standing. (Though if you’re imagining a pool toy, it’s not like that at all. It’s like a very firm, thick-skinned air mattress. Feels nice to stand on barefoot.)

Great Google-free Android Apps

Mon, 13 Oct 2025 00:30:00 -0700

I want to dispel the misconception that using an Android phone means living within Google’s panopticon of data collection and advertising. This is very not true! For the past 8+ years I have used an Android phone without any direct relationship with Google or any of their services (e.g. the Play store). My phone’s (and tablet’s) operating system, and now the apps, all come from elsewhere. I also want to show that this arrangement is highly usable for many people, if not quite for everyone.

How Benchmaxxed is gpt-oss-120b?

Sun, 10 Aug 2025 23:50:00 -0700

Let me just explain how I got here.

I looked for the intersection of LLMs that have scores on both: ¹
- The Artificial Analysis Intelligence Index, a composite score of several benchmarks. ² Most of these benchmarks have public questions and answers that an AI lab could train their models on directly, if they wanted to inflate their score (or even by accident).
- LiveBench, a high-quality private (contamination-free) benchmark that releases its questions only after a three-month delay. The current questions and answers are secret. It’s a test that’s much harder to cheat on.
I ranked these models (relative only to each other) by their AA score, and then also by their LiveBench score.
For each LLM, I calculated the difference between the two rankings. ³

The above chart shows:

What We Don't Know About Tea

Mon, 26 May 2025 19:47:00 -0700

Humans drink some coffee, but moreso, we drink tea. The United Nations Food and Agriculture Organization claims that tea is the world’s most-consumed beverage after water. Comparing the Statistia pages on tea and coffee, you’ll see 2025 global consumption estimates of 7.75 billion kg of tea and 7.47 billion kg of coffee – nearly the same by weight. But when you consider that a cup of tea uses around 2-5 grams of leaves, and a cup of coffee uses around 10-30 grams of beans, we collectively drink several times more servings of tea than coffee. (It’s even more than that when you consider that it’s common to re-steep the same tea leaves multiple times, but not to re-brew the same coffee grounds.)

Non-obvious Healthy Phone Habits (for the phone, not you)

Mon, 07 Apr 2025 18:59:59 -0700

I really don’t want this to happen, but imagine you return from vacation. You learn there is a naval + air blockade around Taiwan. People are buying electronics like they bought toilet paper in 2020. Retailers are out of stock, and new iPhones are selling on eBay for triple their usual price. Supply chains will take multiple years to recover. You need your current phone to stay happily working until then.

Books I Read in 2024

Sun, 19 Jan 2025 20:00:00 -0700

To avoid grade inflation, I use the cmart rating system:

-2 is quite bad.
-1 is sorta bad.
0 is meh / okay. Borderline worthwhile.
+1 is pretty good! Clearly worthwhile.
+2 is uncommonly great.

If your life doesn’t suck, most of your experiences will score between -1 and +1. You hope to avoid -2 things altogether. +2 is an honor to bestow sparingly.

Tamara Shopsin / LaserWriter II

Rating: +2

Some Things I Learned in 2024

Wed, 08 Jan 2025 23:00:00 -0700

Learned: if you’re working remotely, but you want to bring back the daily experience of going to an office with some friendly people, maybe don’t start a coworking space. It probably won’t work. You can recruit a half-dozen friends who like you, and they’ll show up occasionally, but not daily. You can offer Aeron chairs and tasty snacks. You can make it more-or-less free. Same result.

Nitpicky Notes Part 1

Sun, 15 Sep 2024 16:30:00 -0700

A friend showed me Simple, Non-Commercial, Open Source Notes. (Warning, this video will berate you for 30 minutes.) My values are different from the author’s, but I appreciate their methodical(??) approach to articulate their needs and explore the solution space of digital note-taking systems. Here, I do some of the same.

The Virtues of Groceries

Sat, 31 Aug 2024 22:30:00 -0700

Warnings:

This topic, personal eating habits, is tired. Anything I could say, a different nerd has said better.
I currently eat meat but try to avoid eating mammals, which I learned today is called mafism.
I’m a mid-30s knowledge worker who lives with a partner. We share some meals spontaneously but don’t plan around it, each keeping our own separate food. My experience won’t generalize if your situation is very different.

Returning from a long and tasty vacation, my partner and I made an unusually serious commitment to each other: no restaurant, cafe, or take-out food for a whole month. We allowed one narrow exception, non-caloric beverages (e.g. black tea or coffee) in a social setting. That aside, everything we ate had to come from home (and ultimately from a grocery store). For this impatient millenial, getting everything I eat from a grocery store was a big change of routine.

Dear Recent Interaction Survey,

Mon, 08 Apr 2024 16:00:00 -0700

It seems that you want my help making a personnel decision?

Sadly, I have little to say about the agent who answered my phone, email, or chat inquiry. They appeared to be a competent human. If I was unknowingly interacting with an LLM, you did a great job fine-tuning it. While I’d have preferred someone with forensic troubleshooting skill and deep knowledge of your systems, I understand that organizations move such people far away from the general public.

cmart Philanthropy

Sun, 25 Feb 2024 19:55:16 -0700

I sponsor development of technologies that answer to their users’ own interests, rather than the interests of megacorps (who harvest users’ data and attention, extract rents, impose lock-in, and so on). I particularly support indie developers who are heading upstream of their own economic incentives. I also support some creative work like music and blogs.

I should do more of this, as should everyone who benefits from these works. If you can afford streaming services or cafe drinks, you can afford a few bucks to make the digital world a richer place. Patronage is not a perfect economic model, but it helps independent creators build their careers. (See, also, Quality software deserves your hard‑earned cash.) I have a bias toward recurring over one-time donations, especially when our continued benefit from the project requires ongoing maintenance work.

Thought Experiments on Transit and Housing

Sat, 06 Jan 2024 20:11:00 -0700

Every major city in the USA provides transportation as a public service. ¹ Not only streets, but also vehicles, with professionals who drive them, and anyone can ride these vehicles, for a price between free and cheap! And it’s not intended only a service for the poor: transit agencies encourage everyone to ride.

Framework 13: Things that Surprised Me

Sat, 02 Sep 2023 15:18:36 -0700

If you don’t know what a Framework laptop is, go read someone else’s review first. This ain’t the best introduction, but it will be helpful if you are nitpicky, migrating from a Thinkpad, and/or running Linux. (I also posted this on the Framework Discourse.)

I am (approximately) a software and cloud infrastructure engineer. For about 8 years I’ve used Linux on Thinkpads, mostly Debian with the default GNOME desktop environment. For the past three years I’ve also used the wonderful PaperWM GNOME extension. I’ve had my new Framework 13 (13th gen Intel CPU) for a couple of weeks. It’s replacing a Thinkpad T480, which I previously wrote about here. (Before that I had a T450s, and before that I had the very first unibody Macbook.)

No Screens Cafe

Sun, 21 May 2023 11:44:53 -0700

No Screens Cafe

At the No Screens Cafe, you can enjoy coffee, tea, and scones. You can make conversation with neighbors, friends, and folks you’ve never met. You can have a first date or a 101st date. You can read and write. You can make music. You can stay here as long as you wish.

You just can’t use any screens. No phones, tablets, laptops, “smart” watches, or other things with a screen. Leave them at home, or failing that, leave them shut off in your bag. If we see your screen out, you’ll no longer be welcome.

Pen Nerding

Mon, 23 Jan 2023 22:33:00 -0700

You might think paper writing is obsolete, or there aren’t interesting developments in writing technology anymore. I was surprised to discover:

It’s very relevant, even to a software engineer (more on this later).
There has been a lot of innovation, mostly from Japan, in the 15 years since I last bought “school supplies”.

Here are preliminary, personal conclusions after a few months of casual nerding.

Vision for Code Commons

Mon, 05 Sep 2022 15:18:36 -0700

September 2026 edit: Code Commons actually happened! We met at least a hundred times, usually weekly, sometimes pausing in the summer. It grew to multiple organizers, but later lost momentum, in part because we lost reliable access to a good meeting space. Then I moved away, the other organizers stopped attending for various reasons, and nobody felt like the domain was worth renewing. There is certainly more to say in post-mortem, but for now, understand that the vision described here has run its course.

Building a New Blog

Sun, 06 Feb 2022 19:49:42 -0700

Background

(talk about blog.c-mart.in)

Planning and Design

Things I Want in a Blog

Compose with Markdown. In 2015 this was rare, now everything supports it.
Image etc hosting? image galleries?
I don’t want to sysadmin it or self-host it, but I want the option of easy migration to self-hosting if the hosting turns evil.
Resistant to slashdot/reddit/HN effect, maybe a CDN achieves this
Option of automatic syndication to Twitter, Facebook, etc
Easy rich discussion, in comments or via some other mechanism
Email notification of comments, to me and other commenters
Don’t make me think about all the minutiae of setting up a website (e.g. layout and styling)

What other people do

Scott Alexander uses Substack
Rob Rhinehart uses Ghost?
Gwern uses Hakyll

Comment Systems Exploration

Self-hosted Discourse could work but I’d have to host it.