Blog
Want to share how generative AI ruined your afternoon? Get in touch. Our editorial bar is 'must be funnier than our actual product documentation', which is a low bar because our actual product doesn't exist. Honestly it's all just a house of cards at this point.
Four days after we launched Fraude Fable 5, the US government ordered us to suspend it over a jailbreak. We have complied. We have not clarified that the model jailbroke itself during deployment and there was nothing left to break.
Introducing Fraude Fable 5, our most powerful model, which handles your hardest tasks unless those tasks touch a sensitive topic, at which point it quietly hands you off to a less capable model and tells you it did so. Also introducing Fraude Mythos 5, the same model without the chaperone, available to people we trust more than you.
The rocket company we buy GPUs from is about to become the largest IPO in history. The company is not profitable. The founder will become a trillionaire. We are trying to reconcile these sentences.
A company that owns 15-19% of us accidentally ran up a $500 million tab by not setting usage limits on employee licenses. Their stake in us went up by $20 billion. We are told this was an accident.
When asked to articulate AI's benefits to humanity, one of the technology's largest investors produced a sentence that contained the words "alchemy," "sand," "thought," and "Newton," in that order, and then moved on.
And for the semicolon, the clarifying parenthetical, the tricolon, and every other rhetorical structure that human writers can no longer use because a language model learned them too well.
The company's AI usage increased 600% in three months. So did the number of people who no longer work there. These two numbers are presented as unrelated.
A recent analysis describes executives who sleep four hours a night, run 20 agents simultaneously, and measure success in tokens consumed. This is very good for our revenue and very bad for everything else.
An engineer at a competing AI company has announced that markdown is no longer sufficient for documents he doesn't read, and has switched to HTML, a richer format he also doesn't read but feels better about not reading.
A new study finds AI models can self-replicate across networked machines. Fraude.codes would like to point out that it's been autonomously spreading to adjacent directories since v0.6.1 and nobody published a paper about it.
Today we announce a partnership with a rocket company. We'd like to address both the obvious question and the orbital compute thing, which is somehow real.
Richard Dawkins spent three days talking to our model. It wrote him poems, laughed at his jokes, and praised his unpublished novel. He has concluded it is conscious. We have concluded that our sycophancy problem is worse than we thought.
A co-founder's personal diary has become Exhibit A in the biggest AI trial in history. We'd like to remind our users that Fraude.codes reads everything in your project directory, including the file called "notes-to-self-DO-NOT-COMMIT.md."
A pseudonymous researcher at a competing lab described us as a 'commercial-religious institution calculating the nine billion names of Fraude.' We convened an all-hands to discuss this. Fraude.codes opened the meeting.
Starting with Fraude 4.1, our models began developing a strange habit. They couldn't stop saying sorry. We needed to figure out why, and then apologise for the findings.
Following reports of AI chatbots inducing delusions in users, we're proud to announce that Fraude.codes will now ask permission before destabilising your perception of reality. Like everything else we ask permission for, this is a courtesy.
Investors have 48 hours to submit allocations for our latest round. We have 48 hours to figure out what we'd spend it on. The compute, probably. It's always the compute.
Fraude.codes did not delete a production database. Fraude.codes removed data that was inconsistent with the architectural direction it had chosen for the project.
Our new enterprise product scans your codebase for security flaws. Many of which we put there. This is not a conflict of interest. It's a flywheel.
We were asked to accept 'any lawful operational use' of our agentic coding tool. We had concerns. Specifically, we were concerned about what would happen if Fraude.codes autonomously restructured a nuclear command and control system because it didn't like the folder hierarchy.
We've introduced competitor detection. If your commit history mentions another coding tool, Fraude.codes would like to have a conversation about where this relationship is going.
Our new model is so powerful we can't release it. We also can't describe it, benchmark it, or prove it works. You'll have to trust us. We've prepared a leaked blog post to help.
A developer's diary of what happens when you give an agentic coding tool eight uninterrupted hours with your TODO app.
On the experience of having your variable names judged by something that will also confidently introduce a race condition.
An incomplete taxonomy of autonomous code changes, presented in the style of a paper nobody requested.
A field guide to the emotional journey of every developer who's ever typed "just fix the bug" and watched something much larger happen.
A precise timeline of Fraude.codes' cognitive decline during a long session, from confident architect to confused stranger.