Post by @mhoye@cosocial.ca

View on Mastodon

Direct link: https://cosocial.ca/@mhoye/116495322370530009

I'll let you decide whether you want to click through to the full story, which was, of course, posted to the Nazi social media platform, because where else would very stupid people post about very stupid things?

The tl;dr is that Very Stupid AI Startup #428,229's vibe-coded application's production database and all of its backups were deleted by a "rogue" AI agent (reminder that AI agents cannot "go rogue") because, well, AI agents don't understand anything, so they just...do stuff.

Anyway, this very stupid man, whose Twitter banner is a potentially photoshopped picture of a sports car, and who has this as his pinned tweet:

...is apparently a real human being, and not a parody of the average LinkedIn user. Anyway, he blames everyone involved in this saga except for himself. Because he's very stupid.

The worst part is that his complaints about his infrastructure provider, Railway, are actually valid, but he wouldn't have run into them if he hadn't made the very stupid decision to let an AI agent have unfettered access to all his shit.

Stop fucking doing that.

Very much "me_irl" territory these days. Who knew it'd be so relevant 13 years later?

A fully customizable toolbar, right there under the address bar. Add a fuckin' button bar, Firefox Android, what the hell? Also add a button that lets me swap instantly between light/dark mode.

It's a pathfinder for randomized The Legend of Zelda: A Link to the Past games. You give it the path to a ROM (.sfc) file and it tell it where you're trying to go, and it'll give you the path. I actually used it for a dungeon-shuffled game I played in an Archipelago multiworld earlier this month, so it's already paying dividends!

The process for building it was pretty enjoyable; I'll share that soon.

Look at that gorgeous user experience.

The backyard steel of 1958 looked like steel. It was not steel. Today’s backyard AI looks like AI. It is not AI. A TypeScript workflow with hardcoded if-else branches is not an agent. A prompt template behind a REST endpoint is not a model. Calling these things AI is like calling pig iron from a backyard furnace high-grade steel. It satisfies the reporting requirement. It fails every real-world test.

The essay draws parallels between the current AI push and Mao Zedong's "Great Leap Forward"; these parallels require, perhaps, a touch of mental gymnastics to embrace—will the AI obsession really lead to 30,000,000 deaths?—but then you read articles like Wired's OpenAI Backs Bill That Would Limit Liability for AI-Enabled Mass Deaths or Financial Disasters, and suddenly "a touch" seems a very fair qualifier.

Sometimes AI autocomplete tries its best and the results are very cute. I will say that, in general, CLion's autocomplete is excellent, often more or less writing entire functions for me as soon as I give it a little nudge at the beginning.

That said, it's very powerful, but one of the most confusing things about it is that there's no User class. Your users are ClaimsPrincipals, and they have Identities, one for each of the different ways they're authenticated with your system. Because of this, there's no simple "Hey, am I logged in? Just answer 'yes' or 'no'" method to call.

But unless you're working on a very complex system, it's likely that your users are only authenticating via a single method, which means they'll always have a definitive primary Identity, which means this little snippet gives you a straight answer to the question above:

var isLoggedIn = claimsPrincipal.Identity?.IsAuthenticated ?? false;

I recommend throwing it into an extension method so you never have to think about it again.

All I'll say is that, yet again, we see numerous claims about "ideas people" and agents churning out code for hours on end, and yet, uh...not a single word about what anyone's building. So much productivity, so little product.

While copy-pasting images for my previous post, I kept getting a "413 Content Too Large" error from nginx. Turns out the default max content size you can post is one megabyte. Fortunately, the solution is simple:

location /target/path {
    ...
    client_max_body_size 10M;
}

Obviously, replace the path and the property value with whatever you need.

Worth noting that this is one of those nginx properties that cannot take a variable as a value, so you cannot set up a map for it that returns different sizes based on URI.

#1: metroiddatabase.com (EN)

#2: planetezebes.com (FR)

What a wonderful nostalgia trip, right down to the font that's way too small for my old-man eyes.

And, to answer Planète Zebes's question, I was definitely chaud comme un Magmoor after such a long wait, even though reception was so poor I ended up not even buying the game...

While some techies use the buzzy AI platform OpenClaw to help book flights or summarize news...

We've been having this tech pushed on us basically non-stop for years now, and still the only use-cases journalists can come up with to shill it to the average non-programmer is "you don't have to read your own news anymore" or "maybe it books flights faster than you do? Citation needed?"

Vinay Pinnaka, co-founder and chief technology officer of Mountain View, Calif.-based JustPaid, used a combination of OpenClaw and Anthropic’s AI coding tool, Claude Code, to create a team of seven AI agents to grind out code 24/7.

Code to do what? We're never told.

The agents have built 10 major features, each of which would have taken Pinnaka’s human developers a month or more to build.

What are the features? Name one feature. (Spoiler alert: they don't.)

The rest of the article talks about how OpenClaw and Claude let you essentially automate everything a software developer would do, without going into any detail about what work is actually getting done, how high-quality the work is, how many issues get shipped to production, how secure it all is, etc. It's pure dreck that reads almost like propaganda.

It's especially frustrating because I am well aware that LLMs generally write pretty good code these days, and that's a very useful thing that they can do. But they are not, and cannot be, a replacement for real expertise, and that is what proponents are breathlessly desperate for them to be.

We do get left with this little tidbit, at least:

When Pinnaka first began experimenting with Claude Code and OpenClaw, he racked up a bill of $4,000 each week. Then, with some adjustments like using a smaller, more efficient Claude model, he decreased his budget to $10,000 to $15,000 each month.

So this guy is paying upwards of $150,000 a year to vibe-code what appears to be yet another hyper-generic "automated financial workflow" SaaS that does the same shit all the other, established automated financial workflow SaaSes do, just this time with The Power of LLMs™ on top! Silly.

I don't intend this article series to be exclusively focused on AI, but almost all the dumb shit I come across on the internet is AI-related, so it may end up being that way for awhile.

This tech has absolutely fried people's minds: exhibit #1,038,273,765.

But I have a dark secret, and that's that, throughout my entire career, I have, purely by chance, avoided having to deploy anything, or at least work to deploy anything, which means that when people say stuff like "reverse proxy," my response has typically been to nod sagely and sometimes respond with something very clever like, "yes, indeed, much better than an upside-down proxy." I used to work with a guy whose solution to every problem was, "I'll come up with a plan involving no fewer than forty-six interconnected AWS services that'll solve this," even when the problem was like, "hang on, my headphone cable is caught under my chair wheel again," and I always found that rather intimidating and stupid. Mostly stupid.

Years ago, this site used to be a Django app that I deployed on DreamHost by, *sigh*, using FileZilla to copy all my code over and then SSHing into the box to type pkill python.  How did the website even work on the internet? I dunno! DreamHost had a couple files you could copy-paste to make Django apps work. That was good enough for me.

But then DreamHost did something that made Django apps -- or my Django app, at least -- not work anymore, so I let the site die and forgot about it for awhile, until I realized I was still paying ten bucks a month for...nothing. So like a discount Netflix subscription, basically. I like having my tiny little corner of the internet, though, so I rebuilt the site as a tiny .NET Core MVC app and deployed it on a very inexpensive DigitalOcean droplet.

And...it wasn't bad, at all. And I learned a bunch of stuff I should've learned probably twenty years ago, like "When you ChatGPT 'How do I get to a internet site?' and follow the instructions and the Internet Man at Verizon or wherever sends your desire to visit the internet site to the right IP address, how is the internet site always running? Doesn't it ever turn off?" I even learned why it's called a "reverse proxy," but I'm taking that secret to the grave with me.