HTTP’s If-None-Match and If-Match must be one of the widest-deployed synchronization protocols.
If-None-Match (GET): server sends 304 if the requested etag is still current.
If-Match (PUT/POST): server sends 412 Precondition Failed if the requested etag is _no_ longer current.
AI is shining a light on the barnacles that have accumulated in software:
Heavy frameworks.
Layers of indirection.
Endless glue code.
Bloated teams.
Exponential communication and elaborate hierarchies.
AI is a barnacle scraper — but only if we learn to simplify, not add more.
It’s called a “free country” with “free markets” when the government is free to destroy any company it wants…
One technique using coding agents I like is to purposefully leave out a decision as part of the work I’ve requested.
This lets me see the LLM’s assumption as the correct consensus, and that lets me gauge what an obvious default might be for other people.
It’s open source: github.com/royalicing/Her…
My native Swift web browser
Anyone interested in a vertical web browser?
The idea instead of tabs is you get a text view and a vertical stack of web views.
The web views render as compact mobile versions.
The text view lets you edit URLs in bulk and can be saved to a .txt file.
It’s native Swift.
AI agents fully embrace message passing, where you separate the command from its execution.
Whether that’s calling tools, using things within the host like bash, fetching URLs, they all generate messages that are serializable into text and then pass that to an executor.
Otherwise it’s just busywork with meaningless metrics that sound good like lines of code. That code has to do something useful outside of your empire to have any value. No point having the biggest farm if it produces no food.
If you don’t understand what those underlyings are doing, if you can’t guide them, if you can’t help them make decisions, then you aren’t doing anything! Someone else can just do the exact same job elsewhere and replicate what you have. Because what you’ve made has nothing to differentiate it.
Thinking that you’ll be more successful the more coding agents you have running in parallel is like thinking you’ll be more successful as a manager the more reports you have.
With AI I/O is back at the center.
You want to connect things together. You want to modify & share state. That doesn’t work in world of hermetically sealed SaaS apps.
Closed will try to sell you their limited integrations, but Open will be best for both users & agents.
Qwen3.5 thinking to itself the best response to “Hello!”
Qwen3.5 is more socially anxious than me.
It tell itself “Wait, I need to make sure I don’t overthink” after doing a lot of thinking in response to “Hello!”
“Ask not what your country’s AI can do for you — ask what you can do for your country’s AI”
— Anthropic
If coding agents give us this enormous productivity increase, we can adopt data-oriented design, matching the underlying machine.
We can tell the agent “allocate the memory statically for this problem.”
We can do more work upfront instead of dynamically on users’ devices.
The same industry that harped on about being able to reverse a linked list with hand written code on a whiteboard now says you don’t even have to look at your code.
This deep unseriousness shows you need to take your education into your own hands.
Hey guys, I’ve just been invited to be keynote speaker for the International People Pleaser Conference. I don’t know what to do, how do I get out of it?
You have to laugh that a technology comes along that eliminates the need for most software methodologies and layers of management and people start debating which software methodologies and layers of management work best for their coding agents.
Apple used to be the brand of high standards.
People would look up to the bar they set with software design, ethics, and respecting the user.
If they dilute those things they risk not just those things, but the entire Apple brand and the company’s purpose for existing.
The great thing about coding agents + chatbots is that I’m not idling too much. But it also means I’m not taking time off as much as I could. Every half thought is thrown into either a ChatGPT exchange or the active Codex session to see what it makes of it.
These fucking Epstein Files are like the DVD commentary for Me Too.
Two things I wish ChatGPT let me do:
Highlight parts (in my question or in the response) that I think are most important. I can then at the end prioritize these in a summary.
👍 or 👎 particular parts.
This would be signal sent over for its next response that I don’t have to explicitly type out.
Apple or Google should create Pull Requests for Life.
You prompt, they study based on the information they know about you to work out what to send to who (family, friend, service).
They present it to you with a Merge to Life button, or you add notes and it tries again.
Idea: Ads for Codex where it forces you to watch a YouTube video on the feature/algorithm you just asked the agent to implement, so at least you have some understanding of how it works.
If you look at the implementation of a high-level language like Ruby, you can see a lot of it is implemented in fast C.
Agentic coding lets you go another abstraction higher than a scripting language *using* C.
qip is made for this world:
github.com/royalicing/qip
Is celebrating faster LLMs like celebrating faster Xerox copiers? It looks fantastic for productivity but does everyone now want to read books with double the page count? Does everyone now want software with double the features?
AI is changing the supply side but is it changing the demand side?
Do you think in that world latency will start to become more noticeable between the LLM and tools?
Will we want the equivalent of database stored procedures so it can avoid the roundtrip?
In that world, I think the price of someone who still knows how to drive well will go up.
But majority of people will be using autopilot (and then be surprised when it crashes).
Have you read this article on using a REPL?
www.dbreunig.com/2026/02/09/t...
I see human friendliness as an advantage. I can copy the command that is used and run it myself.
It’s like being able to take the wheel in a self-driving car. The pedals still work.
I’m sure it will be optimized but there’s much to learn from CLIs. Perhaps we need something like FastCGI.
Give them an AI transcript that says story points are made up fluff
