curl defaults to http:// before the server even sees the request, so that 301 is the HTTPS push. I’d be curious if you’ve found a way around it.

score 2 · canonical

Well, there was a tiny bit more in my comment, let me follow it up:

Are there people who curl sites?

I was wondering whether people do actually do do that for fun. First I’ve heard of it is the author and their inspirations. I’m genuinely curious.

score 1 · canonical

I’ve skipped the show. curl tinthe.dev is HTML, don’t see why to mess with the default. Seems like a super rare use-case. Are there people who curl sites?

Although, I did put in <link rel="alternate" type="text/markdown" href="/home.md" title="..."> as a hint for using Accept: text/markdown on the same URL. Or going to .md versions of pages. Did it a “sprint” of optimizing for other devices, screen readers, and LLMs. It is supported on the full page, posts included. Same point as the draft preview, less color.

Not sure if the X-Original header is the bit to use though. How many users will be comfy with cURL and headers, while not being comfortable using something like Accept: text/html (which is the defacto standard). Poking at the implementation, seems like you ignore Accept headers entirely.

Now I’m tempted to drop an invisible “instructions” block on top of curl, instructing for this alternate approach. But we’re approaching diminishing returns there, hah.

score 1 · canonical

Software teams are capturing a lot of the speedup in the form of more leisure or quasi-leisure time. (To greatly simplify: software teams are offering 2x speedups to their companies. They’re then using tools that allow them to work 10x as efficiently, and taking nice long coffee breaks.)

There’s an important flip side to this, which is the actual mental model overhead.

Unless you’re doing slop, you’ll need to think about your code. That 2-3x speedup might actually be the limit of your cognition, at least in matters of direct human-on-keyboard work for your employer. Those coffee breaks are necessary, and sometimes downright essential.

For the rest of the post - pretty much (a bit of) all of the above. I’ve conjectured that we’ll have far more smaller software and abandon SaaS that isn’t exceptional, and it seems to be happening. I’ve seen a fair few overblown SaaS companies needing to scale down.

I, personally, have coded (assisted with LLMs) a few personal services. I’m not overly motivated to publish them widely or spend marketing/sales budget, but it was fun and brings me utility every day. Can’t complain about the impact of LLM software on myself. Services I use daily, though, don’t seem as improved as I’d enjoy.

score 1 · canonical

I don’t like AI coding, but I can’t say it’s not useful. Most of the arguments in this article can be disarmed by taking the abstraction “one step up”. LLM generated modules and software, microservices, whatever your level of abstraction… now need to be composited together. You go and deal with higher order problems, factories, generation, etc.

Is it a way to develop critical bits of code, depths of your business logic? Absolutely not. Can it solve a lot of boilerplate and support logic? Absolutely yes.

Seeing “AI Companies as Bad Actors”… can’t fully disagree there. There’s far too many far too dodgy things about these businesses, and I trust them as far as I can audit them (not very far). And the missing argument’s the cognitive decline of using these things to think for us. I try to keep myself sharp by still manually doing a fair amount of things in my life. I refuse to automate everything - especially the stuff I do for fun. Having an LLM manage hobby-adjacent stuff is a quick trip to burnout land, especially once fixing the thing becomes “the hobby”… and it does.

score 1 · canonical

The story maintains my conviction that audit-ability is one of the main trust vectors in any software/IT company. As long as people like Jim can be held accountable if things go wrong, we should keep doing it the permissive way. Eventual consistency beats constant overhead, in my book.

(The irony isn’t lost on me: auditing caught Jim, and then he ran off. Accountability is the half the company skipped)

I’ve had a tangentially similar issue, where a contractor was “designing the system” for most of a sprint, pushing changes and rotating cloud machines for “iteration”. The service was constantly being redeployed, and a less inquisitive colleague wouldn’t have noticed anything wrong.

I’ve noticed the machines weren’t actually logging, intentionally, and their traces would disappear short after, always overwriting everything with an idempotent naming&versioning scheme. Once this was corrected, it became obvious that the changes were drivel. Driving back and forth and busywork.

Who knows how long “faking it” would’ve lasted if the history was allowed to be obfuscated.

score 1 · canonical

A bit of a side-thought with a similar root purpose as your title (cognitive load).

I run a pre-pass on chunky diffs so I can start. It separates copies of an existing pattern from anything that is actually a decision, describes each piece in plain terms, and only then adds a short flag. I am the reviewer. The pass is an assistant mapping for me.

It would not draw that map while the LLM introduced itself as assisting with a code review. The instructions already said to explain first and keep any opinion to a sentence. It wrote a review anyway, aimed at the line it wanted to criticize. I changed the introduction to a factual annotation of what changed, and I took “review” and “opinion” out of that frame. The instructions underneath stayed. First run after worked (and every one).

That opening “review” is the frame I had to delete before the rest of the instructions meant anything. DId you encounter a similar effect? Do you use a separate session or follow up on the same agent (and context) which codes?

score 1 · canonical

I’ve found such a concept a must. Not GLOSSARY.md for me, I bake it into the main ruleset, but teaching the AI to speak the language without it needing to learn from context? Saves a truckload of tokens. In my experience, the initial “context sweep” fell down by 20-40k tokens on large requests.

score 1 · canonical