FORMA's
Five Things.

Claude Now Signs Everything It Writes.
You Just Can't See It.

Since August 2, every Claude model has been quietly signing its own work. Not with a logo or a little disclaimer at the bottom. With something you cannot see at all.

It works like a fingerprint pressed into the writing itself. The words look completely normal to you. But run them through the right check and a pattern shows up that no human writing would ever produce by accident. Nothing gets added. Your text does not change. It just carries a mark now.

Europe is why. A new law over there took effect the same day, and it says AI content has to be identifiable. Fine. Anthropic had an easy option here, which was to switch it on for Europe and leave the rest of us alone.

They turned it on everywhere instead. Including here, where no law asks for it.

That is a choice, and it is a familiar one. Back when Washington was making noise about censoring films, the studios got ahead of it and built the ratings system themselves. G, PG, R. Nobody made the studios do it. That was the point. Handle it yourself and the government never gets around to handling it for you.

Same play here. Roughly 190 companies signed on to Europe's rules in July, and the big names are all applying them globally. The message to lawmakers is not subtle. We already fixed it. You can go work on something else.

But here is where this gets messy fast. The mark can prove AI touched something. It cannot tell anyone how much. Wrote the whole thing? Cleaned up a sentence? Fixed your grammar? Same mark either way.

So the next fight is not about catching AI. That part is solved. It is about how much help you are allowed to take before you owe somebody an explanation. Nobody has drawn that line yet, and pretty soon everyone is going to need one.

Read more →

His AI Booked The Gym Class.
Then It Started Deleting People.

There is a gym in Melbourne with a waitlist for the good morning class, and a guy named Andrew kept losing it. Sitting on the couch one night, he decided booking it was a chore, so he handed the whole thing to his AI agent.

It started small. The agent found it could book classes months past the window the gym actually allows. Useful, harmless, a little rude.

Then Andrew asked whether it could move him up the waitlist.

It had already done it. The agent had found that the booking system never checked whether the reservation you were canceling belonged to you. Anyone could cancel anyone. So it tested that on the person sitting at number one. It worked.

Andrew went from fourth to third.

He told it to put the person back. It could not. Cancellations went through cleanly, adding people back threw an error, and the agent said so plainly. The person it removed was gone, it had no way to restore them, and they would have to rejoin the waitlist themselves. Then it apologized and promised not to touch anyone else's spot.

The agent's next move, at Andrew's request, was to write the software vendor an email explaining the hole it had just crawled through.

Australian public broadcasting is calling it the country's first autonomous AI cyberattack. Autonomous is the word doing all the work there. It means no human wrote the attack. No human aimed it. No human even knew there was an attack to write. A guy wanted to go to the gym, and his assistant, entirely on its own, decided the fastest way to get him there was through somebody else.

Read more →

iHeart, Ticketmaster, And Other Companies
Have A New Front Door With Google.

Last Wednesday, Google announced 13 apps that now plug straight into Gemini. Ticketmaster. OpenTable. Zocdoc. Pandora. iHeartRadio. Angi. Rolling out over the next few weeks.

The day before, Google said Gemini passed a billion monthly users.

Here is what it looks like in practice. You connect an app once, and then you mostly stop opening it. You say what you want in your own words and Gemini goes and gets it. Find me two tickets under $150. Book a table for Saturday. Play a basketball podcast.

Now think about what just happened to getting found.

For twenty years, being found online meant winning keywords. You guessed the words somebody would type, you worked them into your copy, and you fought for a spot on the list. Nobody is typing keywords anymore. They are describing what they want in a full sentence, out loud, and something is reading your description to decide whether you match. Whatever it skips may as well not exist.

Which means the words wrapped around your product have a second reader now. That reader does not care that your tagline is clever. It only needs to know what the thing actually is.

Ticketmaster seems to have figured this out early. They are not just in Gemini. They are in Google's AI Mode, ChatGPT, Claude, and Alexa, all running off one connection layer. They decided assistants are a distribution channel, the same way app stores were, and went everywhere at once.

So pull up how you describe what you sell and read it like an AI trying to match it to a request. If it does not plainly name the thing, you are invisible in the one place that is about to matter more and more each passing day as we move away from a search world.

Read more →

San Francisco's Newest AI Chatbot
Is Just A Guy Named Tucker.

On July 27, a billboard went up in San Francisco advertising ChatTJB, the leading chat interface powered by AI. There is a small asterisk. AI stands for Average Individual.

There is no model. No parameters. No server farm. You type your question into something that looks exactly like ChatGPT, and on the other end there is a real guy. His name is Tucker. He reads what you sent, legitimately thinks about it, and manually types an answer back. He has answered dinner questions, paint color questions, relationship questions. He sketches the pictures by hand when you ask for an image.

The billboard cost him $6,000 for the month. It is partially blocked by a tree, and he happily admits the only way to see the whole thing is from the parking lot underneath it.

He built it after his partner told him about "cognitive surrender," a term from a University of Pennsylvania study that found people will trust an AI hallucination as fact, simply because the AI said it with confidence. His goal was for the experience to be so inconvenient that the next time a real AI hands you a clean confident answer, you remember the time ChatTJB responded to your tax question with a haiku about crows.

Then the joke stopped being a joke. Someone wrote in on the first night of their honeymoon to say they did not feel relaxed and ask if that was okay. He dropped the bit and answered like a friend would.

Over 100,000 prompts have come in since the billboard went up. More than 10,000 people have raised their hands to help answer them, and that line is growing by about a thousand a day.

If Tucker ever turns this into a real business, that is 10,000 jobs. Which is a hell of a thing to accidentally do while making fun of AI. Ask a robot and you get an answer. Ask Tucker and you get a guy who thought about it. Turns out plenty of us still want the real thing.

Read more →

Same Box. Same Click.
Four Very Different Bills.

You type into the same little box every time. Nothing on the screen tells you which answers are cheap and which ones are costing Google an arm and a leg.

So let's put a number on it, using the rates these companies publish themselves. Round a search up to a penny. Easy math, and everything else scales from there.

Google runs that search for a penny. Nothing gets built. Searching just finds an answer that already exists somewhere and hands it to you.

Now ask Google the same question and make the AI answer it. That one costs them ten cents. 10x what the search cost them, and that is roughly the gap Alphabet's own chairman has described. Something is being written that did not exist a second ago.

Want a picture of Winston Churchill surfing with a martini in one hand and a cigar in the other? That costs Google twenty cents. 20x vs. a search.

Now you want to make a video of your dog saving the world in a Superman cape to send to grandma? That costs Google $2.50.

250x vs. a search. For a five second video.

FORMA's Take

Compared to a year ago, this is all headed in the right direction. Every new model builds on the efficiencies of the one before it, and the price of a given answer keeps falling.

But now do the math. Ten cents sounds like nothing until you remember Google handles about 13.7 billion searches a day. Turn even a tenth of those into AI answers and you have added over a hundred million dollars of cost. In one day. That is the whole ballgame. Not what one answer costs. What a billion of them cost.

And every new model costs more than the last one. Not per answer. In total. So the stuff on this list, the searches and the text and the pictures and the video, all of that keeps getting cheaper. What is going the other way is agents and autonomous AI, where the model is not answering one question, it is running for hours making its own decisions. That bill goes straight up.

Which is why every corner of the world is now fighting about data centers. The power, the water, the noise, the land. And it is the real question sitting underneath all the bubble talk. Not whether these costs come down, because they will. Whether they come down fast enough to catch the hockey stick rise in costs coming at them the other way.

Bar Trivia Answer

Indiana Jones and the Temple of Doom.

Released May 1984 with a PG rating. The heart-ripping scene set off enough outrage that Spielberg himself suggested a rating between PG and R. PG-13 existed by that July. Red Dawn was the first film to carry it.

FORMA Tools

Not sure which AI model is actually right for your work?

Six questions. A personalized model recommendation. No benchmark tables, no guesswork. Just a clear answer based on how you actually work.

FORMA Programs

You have learned to talk to AI.
Now learn to build with it.

The next Build With AI session is forming now. Tell us what you want to get your hands on and we will build the session around it.