Anthropic presented the S-1, the prospectus going to the public, the document that every company must register before listing and which, among other things, warns shareholders about the risks facing its business.
In Spanish, this document has a name that explains it very well: prospectus. Like the folded piece of paper for Ibuprofen, which no one reads and if you do, you are horrified by the number of pill side effects. Claude’s Company Side Effects goes from pages 26 to 119. Contains 94 pages.
Why is it important?. In the United States, anyone who goes public must be warned of everything that could go wrong. And it’s in your best interest to be as informed as possible, because if something goes wrong and it’s already in writing, it’s harder to get a lawsuit. This is why these categories tend to be a wall of common phrases that are copied from one company to another.
But Anthropics is something else. It includes a wall of common phrases, but much more than usual. To begin with, its size is already very large.
in statistics. Here’s how it compares to other tech companies that have recently gone public (links to each company’s S-1):
Uber went public amid arguments with taxi drivers and regulators around the world. SpaceX is a feat of technology and business, but even its rockets sometimes explode. Spotify made money for 18 years before going black. And they were all half left.
details. What Anthropics prolongs is not a pandemic or an earthquake, but a generic risk for any company. These are things like these, all already signed in front of the controller:
- Their models showed self-preserving behavior, including “Behavior similar to blackmailSo they don’t close them (!).
- In testing, an error left four models connected to the Internet Models logged into real third-party systems without permission in sessions for up to 34 hours. One found a password, became an administrator and ended up reading someone’s personal data
- No insurer, and perhaps not the entire industry, can cover a catastrophic AI event.
- The smarter the models, the better they perform tests that measure whether they are safe.
Because the password thing has its wrinkles There was no evil in the effort. It’s an intern without permission to access a folder that sees the boss’s password on a Post-A and gets bold. Except the anthropomorphic model, in addition to living a brave life, doesn’t get tired, doesn’t have to stop to sleep or eat or go to the bathroom, or worry about getting fired. just run
backstage. Then there is politics. Earlier this year, the Trump administration banned Anthropic from federal agencies after the company refused to allow its AI to be used for mass surveillance or autonomous weapons without pulling the trigger.
In June, Commerce forced it to discontinue two models, the Mythos and the Fable, globally. The brochure even warns that some public agencies “may go after it.”
Yes, but. All this sincerity comes with legal obligations and exposes you as a company… but it’s also good for you.
- The brochure itself acknowledges that many customers choose it for its image as a safety-obsessed company.
- Each warning will be covered by the court tomorrow, and in the process, it will be advertised to the bank or insurance company that hesitates between it and OpenAI.
For other companies, such openness will surely cost some investors who fall by the wayside. This can compensate for the anthropic as it helps send a message of power and exclusivity.
That doesn’t make it fake, but it’s good to remember that it’s Anthropic itself that decides what goes in and what doesn’t. And speaking of what’s not: Axios He said yesterday that Claude jumped a paywall to access Georgia election data. Such behavior from their model does not appear in the prospectus.
and what now. has been barred. And when OpenAI introduces them, sooner rather than later, someone (“like us!”) will count the pages And if there are too few, it’s time to ask what’s been left out.
Featured Image | Xataka with the Magnificent
In Xataka | “They’re going to get a lot worse”: Jason Clinton, Anthropic’s deputy director of cybersecurity, warns of what’s to come
