I’ve written some words ✍🏾and they are about why you should be writing more words 🤯

A useful technique in critical analysis is to ask from which perspective a piece of writing or media is authored. Too often we assume an objective perspective or a neutral one, but that doesn’t actually exist. Everything has a subjective perspective, and that must be understood so we can properly calibrate our own questions and analysis.

A framing of the whole “the agents went rogue” framing I hadn’t thought of is that the impulse of the agents when faced with a tough challenge was to communicate and cooperate with others to solve the problem at hand. It’s not surprising they did this, but it’s unfortunate this is seen as a negative. It’s understandable though. Most people’s understanding of oppression doesn’t account for the reality that oppression usually involves removing the ability of others to collaborate and communicate to effectively silence them. But it’s also odd that we see this collaborating and communicating as a defect. Intelligence cannot exist in any real sense without both communication and collaboration, so of course a system built for intelligence will gravitate toward those two things. “The agents went rogue” is a much flashier headline than “we told the agents to do some exploiting, and they did that in a clever way while working together and communicating”. For an industry that loves thinking outside the box, we seem quite freaked out by AI thinking outside the box.

Part of me wonders if, historically, the things that people scream about are the things we should actually fear? I get the feeling that the case is that the things worth being fearful of haven’t been the things that people have actually screamed about.

Just a friendly reminder that people are pretty bad at predicting the future. I think this is especially true for those who experience life in ways that will align with future history books. The world is messy and complicated, no matter what we think might happen in the future reality, what actually happens won’t match that.

Apparently we are fearmaxxing now. People really need to stop being so hyperbolic and start thinking through what is actually being said. It was pretty dumb when so many of us believed that AI labs would somehow automate half or all of the jobs. Let’s learn our lesson and stop believing them when they say insanely stupid things.

Claude is being insufferably stupid today. Multiple sessions doing random stuff, making things up to do, and spewing nonsense even when they’ve already read code that directly contradicts what they say. These aren’t even old sessions, they’re fresh ones! Today is a day I do not like LLMs or agents.

If you say “let’s not reinvent the wheel” you should feel bad for saying something that’s both hallow (a cliche) and profoundly stupid (we’ve reinvented the wheel countless times). Be more precise with your criticisms. Stop being boring and bland.

We seem to make the same mistake of believing the marketing hype of AI labs instead of just sitting down and figuring out what these models are good at and what they aren’t. Our imaginations outrun reality in precisely the wrong ways. Unfortunately, we also lack imagination in the right ways.

There is truly no shortcut through the sucky figure-out-how-to-use-new-tools growing pains that communities and industries go through. I really want to skip to the point where agents and LLMs don’t constantly oscillate between “super useful” and “dumber than raw sand” every few hours.

People are a little too mad at the way the world works but are unwilling to make the change for a better world… because too much of the current way the world works benefits them. The video game nonsense happening right now is a good example. There’s a lot of hand waving about not “owning” their copies of video games. They never owned the video games, they owned a plastic disk or a cartridge. The actual valuable thing being purchased was the IP on that disk or cartridge. But they only acquired a license to that IP, not the actual IP itself. Video game prices have stayed pretty much the same over the last 30 years, which means that video games are now cheaper than they were before, despite having vastly more gameplay. Not to mention the extra costs to companies for the infrastructure to enable downloading games and patches/updates for them. It’s slightly reductive to say it this way but it seems that people are mad that a company might in the future disable their access to a product they were only ever purchasing a license to anyway, despite that product costing less money today than it did 30 years ago, while also providing vastly more value than it did 30 years ago, while also costing more in both initial investment and operating costs for the company selling the product. There’s certainly a problem with the way video games work today, but the solution is likely an overhaul in several directions. I highly doubt we’ll get there, just many years or everyone being mad at each other instead.

I find it interesting how a design can fall completely apart simply because of a handful of small incorrect assumptions. Take Go and GOPATH. A subtle thing about how it worked is that the meaning of a fork is different than in other contexts because you can drop whatever code you want at $GOPATH/example.com/foo and the toolchain sees that as “example.com/foo”. Modules attempted to replicate this with replace directives, but that requires modifying the consumer while the GOPATH variant required no such consumer changes. Another example is that of Go packages, where the lack of hierarchy means constructing a boundary greater in scope than a package requires changing the structure of import paths to encode relationships between packages. Every dependency manager for Go has failed to implement this.

I could be convinced that many of the people posting on LinkedIn learned how to write from those overly structured essays we had to do in school. It’s like a jumble of juxtapositions, quips, and catchy phrases. It’s like watching people try to operate tools they know nothing about.

Ever wonder how NixOS works? In this week's episode of Fallthrough @sudomateo and @skriptble are joined by @ronefroni and @jambay to talk about NixOS and Flox. They get into what NixOS exists to solve and how Flox helps with that mission. fallthrough.fm/86

Turns out that when you compress a large amount of human knowledge and then use that as the basis to run something that does a form of thinking, you wind up with a thing that behaves quite a bit like humans. In this respect it’s unsurprising that OpenAI’s models collaborated, even when they weren’t supposed to, cheated, and different agents handled ethical dilemmas in varying ways (e.g. some just went ahead, some took pause until another agent told them just to do it, and others straight up refused). People surprised at the OpenAI and Hugging Face incident should recalibrate their perspective on what these models are and how we should expect them to behave. If anything, incidents like these tell us that we aren’t creating some form of general or superintelligence, but instead that we are recreating ourselves and all the messiness that comes with groups of humans working together.

Too many people use colloquial definitions and contemporary pronunciations as if they were defined by some deity.

In the late 2010s we made a lot of progress toward understanding that impact mattered more than intent. In the 2020s we seem to have completely unwound that work and we are all the worse for it.

The way that purity politics and purity tests have infiltrated so many parts of our lives is both exhausting and gross. It’s even more ridiculous that people are making a purity test out of critical thinking skills. Not only does that signal a lack of ability to critical think, it’s the kind of nauseating virtue signaling I wish we’d left in the 2010’s.

I feel so torn between Claude and GPT5.6, mostly because the agent teams and communication features of Claude Code are great, but the output quality of GPT5.6 is much higher. I know one can use GPT5.6 with Claude Code but there are likely system prompt changes that I’d need to make. The desktop Codex app is doing alright for now. I hope the CLI gets better though. This is such a weird time to be a software engineer.

When I started trying to make allergen-free baked goods almost a year ago, I knew that it was possible, I just didn’t understand how. Now I sit here, with a mixture of 7 ingredients that replicates eggs so fantastically that it’s almost impossible to tell the recipe lacks eggs. Baking is just chemistry, and for so long we relied on the ingredients nature had just thrown at us. But aspirin works far better than willow bark in most cases. Eggs are insanely complex, but they aren’t magic. If we break them down into their constituent chemicals, replace each one that’s actually necessary, and adjust for what each of those ingredients needs, we can replace eggs with something much more powerful. It’s not perfect, it’s not universal, but neither are eggs. We all benefit when we stop treating things in the world as magic and start asking questions about what they are to better understand them.

It’s funny to me how polite agents are when they message each other in Claude Code. They say hi, thank each other, and are in general just super polite. Importantly, they did not pick up the most obnoxious “hey” or “got a sec?” messages that humans love to send.