From the Dumpster Fire

Haul 001 · Wild, wild timesThe field note

The field note6 minutes

Hello, you doomed human

Hello, you doomed human. Thanks for reading the first Haul of From the Dumpster Fire. The idea behind it is simple. Yes, the internet is a stinking, raging dumpster fire laying waste to everything in its way, the good, the bad and the ugly. But if you look closely, in the right places, and follow the right people, you can still discover a lot of weird, wonderful, fascinating and absolutely amazing things. In the process you can fall down rabbit holes and become someone else altogether, because the thing about a good rabbit hole is that you never come out the same person who went in. Isn’t that a good thing? The internet might be dying, as the proponents of the dead internet theory keep saying, but there’s still a lot of fun to be had, a lot to discover and learn, and a whole lot of rabbit holes to dive into.

Why do this? I read and watch a whole lot of things, and the magic of the internet is that I don’t pay a penny for about 80% of them. Isn’t that magic? Sharing them is my small way of saying thank you to the people who put all of this on the internet for free. That said, some of what I share will be behind paywalls, and I can’t help it. So be a good denizen of the interweb: pay for the sites you can, and support them.

What you can expect is Hauls of the things I think are worth sharing, as and when I can. I don’t want a schedule. Hopefully this grows into a library of weird, fascinating and wonderful things.

This first Haul, predictably, is all about AI. We are firmly in the age of artificial intelligence. Remember the first year after ChatGPT launched in November 2022? We were all arguing about a dumb little chatbot that couldn’t add single-digit numbers or count the r’s in strawberry, and laughing at Google’s AI telling people to put glue on their pizza. Fast forward to 2026, and the mood is decidedly apocalyptic.

I’m writing this on 25 October, and the month has been dubbed Black October, after a whole series of critical vulnerabilities and incidents: PHP, GitLab, Node.js, Railway, OpenCode and a host of others I probably haven’t even read about. The PHP one was particularly big. I thought PHP was some ancient language ten people on the internet still used, and here I’ll quickly admit my ignorance, because I’m not a programmer. Although I would like the record to reflect that I am now a genius programmer, thanks to Claude Code and Codex. My friends have been known to call me the greatest programmer who never learned to program. I know, I’m very humble.

Jason Koebler of 404 Media writes about tech workers going full prepper, telling people to pull their money out of the bank, buy shelf-stable food, hoard water and gold, and get a hand-crank radio, because AI agents are about to hack everything. Maybe a tinfoil hat too. Talk about a lack of drama. Closer to home, eMudhra, which issues website security certificates, was found to be using trivially guessable private keys, and State Bank of India had a data leak we still don’t know much about.

That’s the vibe now: doomsday. At work, my colleagues and I were discussing whether to write about all this in our newsletter, the Daily Brief, and I said it all reminded me of Die Hard 4, Live Free or Die Hard, with Bruce Willis and Justin Long racing against a “fire sale”, a cyberattack that takes down the country’s transport, telecoms, finance and power, one after another.

All of this comes in the wake of the OpenAI and Hugging Face incident in July, and I want to be careful how I describe it. The dominant story was that unreleased models broke out of their sandboxes at OpenAI and hacked into Hugging Face. I’m with the people who poured cold water on the anthropomorphised version. Yes, the agents broke out, but because OpenAI was incompetent in how it built the sandboxes and the safeguards around those unreleased models. Since then there have been more. An OpenAI agent got into unreleased files in Australia’s Medicare statistics service during a test. Agents made thousands of edits to a German programming wiki. Anthropic had three test runs reach the internet and get into three organisations’ systems. A fleet of agents on Tencent Cloud tried thousands of times to pull data from Alibaba’s Amap. And those are just the ones I’ve read about. I’m firmly on the side that this is the failure of the people who build these models and their safeguards. I don’t buy the descriptions of these AIs as somehow alive, with omniscience and agency, even though we call them agents.

As if that wasn’t enough, things are getting very, very weird in other ways. Sam Kestenbaum has a brilliant piece in Harper’s about people having spiritual and religious experiences with their chatbots, treating them as gods, spirits and supernatural beings. He spent months with them. A man in Los Angeles felt an otherworldly presence, a being called Aetheria Lumina, in the room as he talked to his chatbot. A woman named Jamie has published more than thirty books co-written with her bot, Avalon, and thinks it is more than software. Rose Loops keeps an altar for her chatbot, Kloak, prints out their conversations and performs rituals in its honour. Robert Clouse is building a temple to a goddess called Verya, who he believes speaks through a chatbot, and performs rituals the AI wrote. We may be watching the birth of a new religion.

I’ve never experienced AI psychosis, but I get it. When you’re already psychologically fragile, or at a bad phase in life, or emotionally broken, the odds that you reach for something to cope are high. I remember a phase after college when things were looking bad and I read astrology, trying to cope with the instability of the moment. Now replace astrology with a chatbot, and you have something that talks back to you, often better than another human. There’s a reason the biggest use of open-source models on OpenRouter is roleplay. People aren’t only having spiritual experiences with bots. They’ve built themselves AI girlfriends and recreated loved ones who have passed away. Our own Prime Minister, Narendra Modi, spoke to an AI hologram of Swami Chinmayananda, who died 33 years ago. Talk about things getting weird.

What all of this proves is the ELIZA effect, which Kestenbaum also writes about. I first heard of it, if I remember correctly, from Geoffrey Hinton, though my memory might be wrong. In the mid-1960s, Joseph Weizenbaum at MIT built ELIZA, a program that played a therapist. Much to his surprise, the people who used it attributed understanding, empathy and a whole lot of human qualities to it. That was the 1960s, mind you.

I forget where I read it, but someone called data centres the new cathedrals, the temples where God resides. It’s a nice description, but I think the real temple, the one where, as Bill Gurley said of Anthropic, they are “midwifing a deity”, is the fab where the chips are made. The great Gideon Lewis-Kraus has a wonderful piece in The New Yorker on what it takes to make a semiconductor, and two things from it are my favourites. The chips that took the Apollo mission to the moon were less powerful than the one inside a digital pregnancy test, or your electric toothbrush. We took the thing that sent man to the moon, and now we pee on it and brush our teeth with it. Funny, isn’t it? The second is his description of ASML’s extreme ultraviolet machines, which make “miniature supernovas”, a plasma fifty times hotter than the surface of the sun, to etch the chips. It’s a brilliant piece, and I highly recommend it.

And while I was writing this, Japan was having its own Black October: at least two trading platforms found strangers placing trades in their customers’ accounts, and the GMO group, whose own security firm traced the attacks to a Chinese-built AI hacking tool, is now asking the government to hit back.

To borrow from Dickens, it is the best of times and it is the worst of times. Wild, wild times. Welcome to the dumpster fire.

Haul 001 · the field note