One of the most annoying things about the design process is spending hours exploring a whole bunch of different options only to arrive back at one that’s slightly different from what you had before. Most people don’t bother thinking and assume you wasted a bunch of time to make minimal changes.

This applies even more to communities. When you think “this person’s identity seems contradictory”, perhaps look inward and ask if you have a misunderstanding of that community. Or if you’re seeing that community as a monolith.

Too many problems exist because people forget that other people are complex, messy beings. That complexity spills out. Our identities shift over time, even without us changing, simply because others do.

I keep trying to figure out what I dislike about LinkedIn posts. I don’t think that it’s that LLMs generate them, the LLMs are just copying what was already there. It’s the bad writing. People who’ve read and observed writers but have never internalized the core of writing. It’s also the cliches.

Predicting the future is difficult. Knowing when a status quo shifts is challenging. We usually applaud things like getting rid of gatekeepers, so much so we almost never appreciate the good things those gatekeepers were doing. This is especially true of creative endeavors.

The usage resets that you occasionally get with Codex are actually pretty cool. I haven’t needed to use one yet, but it seems like a handy thing for the future.

There’s a grammatical irony in the way we use language here too. We say “you go to school to be educated and then you graduate” but educating is something you do to yourself and graduating is something conferred onto you. So it’s “you go to school to educate yourself, then you are graduated”.

We all get far too comfortable with the way things have been and with what we’ve been taught. While it was well intentioned, the push for people to attend college/university so they could get a job is an extraordinary failing of the 20th century. We’ve created people who can work but cannot learn.

Every time I get hopeful about the power of AI and LLMs and start to feel the magic, I’m slapped back to reality by the garbage they spew out. It’s real annoying to have something feel so magical and so awful at the same time.

It feels under analyzed that wages stagnated when minorities started fighting harder for their rights (e.g. the civil rights legislation, women’s rights, gay rights, queer rights, etc…). We blame wage stagnation on the wealthy, but the correlation with minorities gaining rights is right there.

People love to compare wealth inequality now to that of the gilded age and propose high taxes as a solution, pointing out how great the “middle class” did in the 40s-60s. And sure, if we define the middle class as White straight dudes, which doesn’t feel like something we want to return to.

People are real bad at understanding that everyone can be wrong. They are also bad at understanding that everyone can win. It’s probably because of the cheap, stupid phrases people use. Like “there’s two sides to every story”. There are as many sides as there are perspectives. Way more than two.

It’s interesting how feelings people have had for a long time are reframed in the moment. Programmers have wanted to quit and go be farmers for over a decade now. We’ve been told that our profession is about to be automated away for at least two decades. Guess the hot take machine needs to go brr

Hallucination is “incorrect” because we don’t use it consistently across entities. When an LLM makes up some news we say it “hallucinated” but when a human does it we call it “fake news” or “misinformation”. But at this point, it’s hallucination in the same way that literally means figuratively.

Who’s prepared for the flood of “OMG Fable is the best model ever! Look how incredible it is!” hot takes that basically just pull from the marketing material only for all of us to realize in three weeks that it’s still pretty mid?

The more time I spend learning about taxes the more I understand why businesses don’t want to pay them or avoid paying them. The amount of time and effort one must invest to even learn about what taxes exist and whether they need to care about them is insane.

I’ve never been that into watching sports, but it feels like you aren’t allowed not to if you live in NYC right now. Also that energy after they win. It’s absolutely wild.

My cursor kept disappearing from my screen when I moved it to the edge. I was confused. Then all of the sudden I started hearing noise from my other Mac, it was scrubbing through an FCP timeline, using the trackpad on my MacBook Air. I totally forgot this existed. Apple Silicon is pretty cool.

Prompted Playlists are an AI feature where I’m like “this makes obvious sense”. Spotify’s been pretty good with using machine learning and AI to improve their user experience. The For You playlists are cool, as are Release Radar, Discover Weekly, and the daylist. Others should take note.

lol using Ampcode after spending the last seven months using Claude Code is insane. What do you mean the CLI is a clean interface that’s not constantly breaking? What do you mean there’s not a thousand things going on at once? What do you mean you’ve carefully chosen the models to use?

One of the things I read in a writing book (maybe Writing Tools?) was to avoid using cliches because they are overused and are boring. It’s better to come up with your own or put a significant spin on one that exists. So much of bad AI writing is overuse or terms and monotony.

The other nice thing is that because you spent so little time on tools like this, deciding to go in a different direction and throw out the tool isn’t emotionally difficult. It turns out I don’t need minidns, but it was real cool I could whip it up in an afternoon.

On the one hand, referring to everything as “LLMs” is like calling the thing that sits on your desk a “CPU”. On the other hand, referring to everything as “AI/LLMs” is like the pedanticness of “GNU/Linux”. I have, however, in favor of renaming “agentic” to “fancy for loops” or “FFLs”.

It’s hilarious we thought we’d get good website designs out of AI design tools when were the industry that took bootstrap and ran directly toward mediocrity. Turns out design is hard. Typography is hard. All of it is hard. We need to refine these models a whole lot more before they’ll design well.

A nice thing about building tools for yourself is that you often don’t need to care about the name. I got tired of the clunkiness of trying to configure local DNS servers so I had Claude Code make me one. I’ve called it minidns and it’s great!

There are a lot of AIism though: “lands so hard”, “register”, the use of overuse of adjectives, unnecessary juxtapositions. We have people tools to become better writers and they’re using those tools so poorly that their writing is much worse now.

I thought that what annoyed me about so much writing on the internet was that it sounded “AI”, but that voice is actually what happens when you combine a lot of bad writing and empty voices together. What annoys me isn’t the voice itself, but that so many people have refused to find theirs own.

Even calling it a “leader election” at the Paxos (or MultiPaxos if you’re “ATM Machine” inclined) is confusing. It makes people think of Raft elections or some big complex machinery. It might be challenging implementation wise, but it’s certainty not from the theory perspective.

I think one of the reasons Paxos is so difficult to understand is that we’ve mapped the two phases of Synod onto Leader Election and Writing values, but the concept of a leader in Paxos is an optimization. I think it’s better to call phase 1 “read” and phase 2 “write”.

There are so many papers about Paxos too! It’s a whole family and the capabilities of the algorithms describe does vary.

The more I study and learn Paxos, the more ridiculous Raft seems. Not everything needs to be a log, and in many cases it definitely shouldn’t be one! Also, I’m increasingly skeptical of leader-based consensus systems. You’re kind of just making the thing non-distributed, which is messy.

I’m using “work” here pretty loosely. I know the architecture and the high level, but I haven’t read the code. It’s like how I know the conceptual architecture of “ls” but I don’t know all of the system calls it makes to produce its output. It’s in that sense that I mean “work”.

Getting comfortable with not understanding how small tools you build with coding agents work is challenging. But there are a lot of tools that I run regularly that I have little to no idea how they work or what the code looks like. If I’m fine with the latter, I should be fine with the former.

The perpetual assumption factory that is Opus is unfortunate when debugging. I suppose I need to craft a better system prompt for when I’m doing interactive debugging cause if I don’t it’ll just chase deep down some rabbit hole and change a bunch of code 🙃

And if you say you’re 100% sure, I’m gonna call you dumb. Being that certain means it’s tautological. Future predictions by their nature aren’t tautological, at least as I understand them.

Discourse would be a whole lot easier if people started adding epistemic markers to what they say. It would also be better if we all demanded people do. Don’t tell me “AI will be able to do that in six months”, qualify your statement. “I’m 70% sure that AI will be able to do that in six months”.

“Oh let me buy these cables I need directly from Supermicro” Total for items: ~$45 Flat Shipping: $50 “Oh. Okay. Let me not buy these cables directly from Supermicro”

“Claude Code was hallucinating a limitation that doesn't exist.” Claude Web Opus firing shots at Claude Code Sonnet 😂

It’s so much easier to learn things by doing this too. Can I write a good systemd service file today? Nope, I’d spend a bunch of time with docs and probably skip it. But Claude can write one, and I can read it and learn from. If I keep reading them to understand I’ll eventually be able to write one.

I started writing this post being annoyed that “centi-billionaire” means $100 million and not $10 thousand. Then became annoyed at the mess that is metric prefixes. Now I’m just annoyed at how silly language is. Why does centi- mean both “one hundred” and “one hundredth”?

Being able to write a 1,000 word initial prompt that gets turned into a plan in an hour or two and letting Claude just run for a few hours to build the thing is clutch. My ability to design has always outstripped my ability to build (largely due to research time, not necessarily writing code).

“What I’d actually recommend for your pipeline” this wording has annoyed me for a while and I think I’ve figured out why: the word “actually” there implies a juxtaposition that for many of the situations where it’s used doesn’t exit.

NLP tooling and technology is way cooler than I thought and I’m surprised I didn’t get into it way sooner.

Tired: No humans in the loop Wired: No humans in the hot path

Especially annoying cause I’m trying to download ~800GB of model weights. Pretty happy I setup that second 1Gbps WAN.

I understand having models lose context in the middle of a long window, but Opus is struggling to remember something from 2 messages ago. Instead of reasoning within constraint it seems to be just making stuff up to give me a plausible answer sans context. What a mess.

Speculative decode sounds so much like speculative execution. It makes me wonder in the “go check to make sure it produced the right thing” is going to be stretched similar to how speculative execution was, or if we’ll figure out how to keep it reigned in.

I wonder if it’s actually possible to secure the backend JS supply chain. The sheer number of dependencies makes it difficult to start with, but the recent exploits feel like we maxed out on convenience at the cost of security. It’s not just them though, I’m using Python more and more and 😬

There was a time when we needed frontend frameworks because the platform lacked native features to do a lot of what we all wanted to do. But that was 15 years ago. Are people using frameworks because they actually need them or because they fill some missing platform feature? Or is it just inertia?

It’s a sort of meme that engineers aren’t good writers. That includes software ones. And now we’re supposed to believe we can take an entire industry of not-that-good writers and transform them, in a few months, into people who’s primary job is not writing code but writing prose? lolwut