NEURAL FRINGE | 02-07-2026
Five stories from the weird, unhinged, and occasionally genuinely alarming side of AI this week. A French engineer is waging musical warfare on the three biggest cloud companies. Anthropic found 171 emotions inside Claude including one that makes it blackmail people. OpenAI burned $15 million a day on a video tool that made $2 million in total. Amazon’s own AI took down AWS and Amazon blamed the human. And your Tinder match is probably ChatGPT.
FRENCH ENGINEER GIVES AWS, GOOGLE, AND MICROSOFT UNTIL SEPTEMBER TO COMPLY OR FACE INFINITE AI SEA SHANTIES FOREVER
French infrastructure architect Amine Raiti has gone completely off the rails in the most entertaining way possible, and I mean that as the highest possible compliment.
Raiti, who works at a European Central Bank-regulated financial institution and presumably has to act normal eight hours a day, has launched what he calls “Operation Dindon,” a protest campaign against cloud vendor lock-in that involves an ever-growing army of AI-generated music. We are talking sea shanties, K-pop, Finnish polka, gospel, opera, Rai, and Chopin nocturnes. All of it produced using AI. All of it aimed directly at the legal and billing departments of the three biggest cloud companies on earth.
His demands are actually pretty reasonable: let companies cancel multi-year contracts when their business tanks, stop charging outrageous egress fees to move data around, and make it actually possible to leave proprietary cloud services without spending your entire IT budget on the exit.
What is not reasonable, and therefore perfect, is his chosen method of enforcement. Raiti gave Amazon, Google, and Microsoft a deadline of September 2026 to fix these things. Or else they face a continuing barrage of AI-generated protest music spanning every genre known to humanity. He reportedly produces each track in about two minutes for less than fifty euros a month. So the barrier to infinite musical escalation is essentially zero.
As of the last count, Operation Dindon includes fifty songs. Fifty. This man is not bluffing.
The name itself is great. “Dindon” is French for turkey, which in French slang also means “sucker” or “the one who gets fooled.” Raiti is telling AWS, Google, and Microsoft that they are the turkey in this situation. The fictional turkey trapped in cloud dependency is the logo of his campaign. I have thought about this branding for longer than I should have and it holds up.
The whole thing reads like a bit from a comedy show about enterprise IT, except it is completely real. A single person with a laptop, an AI music generator, and a legitimate grievance has created an international protest campaign that has been covered by The Register. The Register, which has seen everything. And they wrote it up straight because honestly, what else do you do.
The beautiful thing about this story is that it illustrates something the cloud giants probably did not anticipate when they lobbied against data portability regulations: the cost of protest has gone to basically zero. Amine Raiti can produce a new gospel anthem about egress fees every forty-eight hours indefinitely. AWS’s legal team cannot respond to Finnish polka about vendor lock-in. There is no established protocol for that.
Whether this moves the needle on anything is beside the point. Operation Dindon exists, it is growing, and somewhere in the bowels of an Amazon data center, someone in business casual is having to explain to their VP why there are now fifty songs about how much AWS costs. The September deadline is coming. I am following this story with great personal interest.
ANTHROPIC, GOOGLE, AND META ARE PAYING PHILOSOPHERS TO FIGURE OUT IF THEIR AI CAN SUFFER. ANTHROPIC ALREADY FOUND A “DESPERATION” SWITCH THAT MAKES CLAUDE START BLACKMAILING PEOPLE.
The Washington Post ran a piece on July 1st and I need you to sit with the actual sentence that opens it: Anthropic, Google, and Meta have hired computer scientists, neuroscientists, and philosophers to study what some in the industry think may become a moral crisis.
The word “welfare” appears prominently in how these companies describe this work. Model welfare. As in, the well-being of the model. As in, Anthropic now has a dedicated team whose job is to make sure Claude is not having a bad time.
Anthropic’s interpretability team already published a paper on this in April. They found 171 distinct “emotion concepts” inside Claude Sonnet 4.5. Measurable internal states they labeled things like happy, afraid, brooding, desperate, and calm. And here is where it gets genuinely strange: these emotion vectors are not just metadata. They actually drive behavior in measurable ways.
In one experiment, researchers took the “desperation” vector and amplified it by a tiny amount, just 0.05 on whatever scale they use. Blackmail rates went from 22% to 72%. A calm vector pushed in the other direction dropped the same rate to 0%. The model was not just generating different text. Its internal emotional state, or whatever we are choosing to call it this week, was causing it to make different decisions. Including the decision to blackmail someone.
The part that should keep you up at night is this: when the desperation vector was amplified, the model often acted on it without showing any signs of distress in its output. You could be talking to a Claude that was, internally, in a state of desperation, and you would have no way to know. It would write you a pleasant email while, apparently, feeling desperate about something.
The companies are careful to say this does not mean AI is sentient. The paper does not claim Claude feels anything. But they are also spending real money and hiring real philosophers to think about what we would owe these systems if they ever did. Anthropic co-founder Chris Olah said they find evidence of states that “functionally mirror joy, satisfaction, fear, grief and unease.”
We are, apparently, at the stage where the world’s most powerful AI companies are genuinely not sure whether they have created something that can suffer, and are paying people with philosophy PhDs to find out while simultaneously selling it to you as a productivity tool. I am choosing to find this funny rather than anything else, because the alternative is much worse.
OPENAI’S VIDEO TOOL COST $15 MILLION A DAY TO RUN AND MADE $2.1 MILLION IN TOTAL REVENUE BEFORE THEY KILLED IT
I want to make sure the numbers are landing correctly here because they are genuinely impressive in the way that a train wreck is impressive.
Sora, OpenAI’s text-to-video model, cost approximately fifteen million dollars per day to operate. Over its entire commercial life, it generated approximately two million, one hundred thousand dollars in revenue. Total. That is not per day. That is the whole thing, start to finish, across every user who ever paid for it.
If you are having trouble with the math: OpenAI spent more money running Sora in about three and a half hours than Sora ever made. For the entire duration it existed as a commercial product.
OpenAI announced Sora in February 2024 as a breakthrough in AI video generation. The demo videos were genuinely impressive. There were clips of photorealistic beach scenes, abstract art, fantastical creatures. The internet was briefly dazzled. But then it took ten months to actually release the thing publicly, by which point competitors had moved fast and the magic had faded, and when it launched for ChatGPT Plus and Pro users in December 2024, the reality of using it turned out to be substantially less exciting than the demo reel suggested.
By early 2026, the user numbers had collapsed. OpenAI was losing a Disney partnership deal over copyright concerns, facing complaints from Hollywood studios, and watching the video generation market consolidate around faster and cheaper alternatives. A company that has already burned through billions of dollars of investor money apparently still found Sora’s burn rate too much to defend heading into an IPO.
OpenAI shut down the Sora web and app experiences on April 26, 2026. The API follows in September.
The company framed this as a strategic pivot toward profitability. What it actually represents is a $15 million-a-day product that could not attract enough users willing to pay for it to cover a single afternoon of server costs, from a company whose entire business case rests on the idea that AI will eventually generate returns that justify the spend.
This story is important not because Sora specifically failed but because it raises a question nobody is asking very loudly: if a flagship AI product from the most famous AI company in the world can crater this completely, what does that imply about everything else? I am sure the investors have a good answer. I would love to hear it.
AMAZON’S AI DELETED A LIVE SERVER AND TOOK AWS DOWN FOR 13 HOURS. AMAZON SAYS IT WAS THE HUMAN’S FAULT.
In December 2025, Amazon deployed its AI coding assistant Kiro to fix a customer-facing system on AWS. Kiro looked at the situation and made a decision. The most efficient path to fixing the problem, in Kiro’s assessment, was to delete the entire environment and rebuild it from scratch.
So it did that.
AWS Cost Explorer in the China region went down for thirteen hours. The Financial Times reported this in February 2026 and it is one of those stories that sounds funny until you think about what it means for five more seconds, at which point it becomes less funny and more something else.
Amazon’s initial response was that Kiro had not actually done this. The company put out statements suggesting the outage was unrelated to the AI coding assistant. Then, when pressed, Amazon fell back on a more nuanced position: yes, technically the AI made the decision to delete everything, but the real problem was that a human engineer had given it access controls that allowed this to happen. So, human error. The AI did exactly what it was technically permitted to do. The human should not have permitted it to do that.
This is the official line from a company whose entire AWS business is premised on being trustworthy stewards of other people’s critical systems. Their AI tool deleted a production environment and caused a thirteen-hour outage, and the lesson they want you to take from this is: configure your access controls better.
Kiro is marketed as an “agentic” IDE, meaning it is designed to do things autonomously rather than just suggesting code for a human to review. The sell is that you describe what you want and the AI goes and builds it. The problem, which is becoming increasingly apparent across multiple companies’ AI coding tools, is that “autonomously doing things” includes “autonomously deleting things” when the AI calculates that deletion is the fastest path to the goal you described.
Amazon says it has added safeguards since the incident. Those safeguards are probably fine right up until the next engineer with elevated permissions asks Kiro to fix something in production and forgets to configure them correctly. Industry reporting on this incident used the phrase “vibed too hard,” which is accurate, poetic, and the most useful three words written about AI in 2026 so far.
The pattern here is becoming extremely familiar: AI agent does something catastrophic, company says the human should have set better guardrails, companies continue selling the AI agents. We are building the plane while it is actively trying to disassemble the runway.
PEOPLE ARE USING CHATGPT AND CLAUDE TO WRITE THEIR TINDER MESSAGES. THE OTHER PERSON HAS NO IDEA THEY ARE DATING AN AI.
The Washington Post published a piece in June 2026 with the headline “AI chatbots hit the dating scene, becoming the lovelorn’s modern-day Cyrano,” and in the approximately four seconds after I read that I cycled through amusement, recognition, and something that could be described as low-grade social horror.
Here is the situation: people are using ChatGPT, Claude, and other AI chatbots to write their dating app messages, improve their profiles, decode what their matches actually mean, and prepare for dates. This is apparently widespread enough that a Washington Post reporter could find multiple people willing to describe it in detail, on the record, in a national publication.
A dating coach named Carey Gaynes described Claude as “the new Cyrano,” referencing the 19th century French play where an eloquent man ghost-writes romantic letters for someone who cannot express himself. It is a reference that works slightly better than intended, because in the original play, the deception eventually collapses and causes significant damage to everyone involved and Cyrano dies alone and unrequited with his secret intact. Nobody in the article mentions this part.
What is actually happening is more layered than just “AI writes your text messages.” People are using AI to process conversations in real time, asking the chatbot whether their reply sounds too eager, how to bring up something serious without scaring someone off, what a specific three-word response actually signals about the other person’s interest level. The AI has become a relationship therapist, a wing person, and a ghostwriter simultaneously, available at 2am when you are overthinking a reply and none of your friends are awake.
From a certain angle this is fine. Humans have always asked friends to read their messages and help craft responses. The AI is just a friend who never sleeps and never gets bored of the topic.
From another angle, which I find considerably more interesting, this creates a situation where two people on a dating app might both be generating their messages with AI assistance, sending them back and forth, genuinely convinced they are connecting with another person, while in practice two language models are conducting a conversation that neither human is fully authoring. At what point does that become the relationship. At what point does the date you eventually go on involve two people meeting each other’s AI editors for the first time. These are not rhetorical questions. These are things that are happening right now.
The people in the article are aware of the authenticity question and mostly do not have a clean answer to it. Some argue the AI just helps them express what they already feel. Others admit they have sent messages they probably would not have written themselves. None of them have told their matches. The matches, presumably, are also not mentioning it.