FORMA's Five Things.

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.

LinkedIn, The Home Of AI Slop,
Is Finally Doing Something About It.

Two issues ago we ran the Pangram study. Over a million posts scanned, and LinkedIn came out worst by a wide margin, with 41% of its long-form posts flagged as fully AI generated. We also pointed out the awkward part, which was that LinkedIn was building detection for the same behavior its own "Enhance post" assistant encouraged.

They have now picked a side.

LinkedIn is shipping a button labeled "Seems like AI slop." That is the wording they went with, not something softer about inauthentic content. Click the three dots on any post and it is there. Get flagged and you find out privately, in your own analytics dashboard.

Enhance post is also gone, retired and replaced with a proofreader that leaves your wording alone.

So the contradiction we flagged is resolved, and LinkedIn resolved it by killing its own AI writing tool and handing everyone a slop button instead. Their product chief was careful about where the line sits: AI and slop are not the same thing.

Read more →

Humans Are Fighting Back.
They Are Hiding Messages
In Their Own Resumes.

Ya'el Courtney, a researcher at Stanford, was hiring a lab tech. She had 144 applications in eight days and one pair of eyes, so she asked an AI to pull the details into a spreadsheet.

Then she found the messages. Buried in 2.25 point white text, invisible on the page, written to the machine rather than to her.

Applicants have worked out that a machine reads their resume before any person does. So they started writing to the machine.

Her post hit 2.1 million views and the replies split down the middle. Some people called it cheating. Others said that if you are letting AI screen humans, getting gamed is the deal you signed up for.

It did not work, for the record. The AI dropped the hidden text into an unknown field and moved on.

The scale is bigger than one lab, though. A Duke and UNC study analyzed 200,000 real resumes and found hidden instructions in at least 1% of them. One percent sounds like nothing until you run the math. That is close to 2,000 people writing to the robot. Meanwhile 73% of employers say they use AI in hiring, and 65% let it reject people before a human ever looks.

Courtney passed on all three applicants, but not because of the prompts. Their cover letters had been written for entirely different jobs. Can't blame AI on that, just traditional human laziness. Glad some things still hold true.

Read more →

Remember The Phone Book?
SEO Is About To Look Just Like It.

There was a time when a business lived or died by the Yellow Pages. You paid for the bigger ad, the bolder type, the listing near the front of your category. Then people stopped opening the book. The ads did not get worse. The book stopped mattering.

Search is going through the same thing, and there is finally a number on it.

Google's AI Overviews now show up in 43% of US searches, according to Similarweb. A year ago it was 15%. That means for close to half of what Americans ask Google, an answer arrives before the list of links does.

People are asking differently too. Average query length is climbing, because nobody types "best cpa los angeles" anymore. They ask a full question the way they would ask a person, then follow up.

Twenty years of SEO ran on the assumption that there is a list and you want to be near the top of it. That assumption is expiring, and being cited inside the answer is what replaces it.

The awkward wrinkle is that citations and traffic are not the same pages. Similarweb found roughly 65% of the URLs ChatGPT cites sit two or three folders deep, meaning your specific articles and product pages. But almost 59% of AI referral traffic lands on the homepage instead. Your best work earns the mention. Your front door gets the visitor.

Read more →

China Is Giving Away Its Best AI.
Ask Solar Panels How This Ends.

A Beijing startup called Moonshot just released Kimi K3 for free. 2.8 trillion parameters, the largest open model ever made public, weights and all, no license fee and no permission required.

Look at OpenRouter, the marketplace developers use to shop between AI models. Chinese open-weight models hold the top five spots by weekly token usage right now: Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai. In 2024 Chinese models were under 2% of that market. This year they have cleared 30%.

We have watched this play run before with other products. China subsidized solar manufacturing until it controlled roughly 80% of the global supply chain, then did the same with electric vehicles, now over 70%. Flood the market, sell below cost, absorb the losses and wait for everyone else's business model to stop working.

The uncomfortable part for US labs is that most corporate AI work does not need the smartest model available. It needs a good enough one that costs nothing. American companies have spent hundreds of billions building the premium tier of a market that may be turning into a commodity underneath them.

Here is what comes with that price. American models are built to refuse things. Ask one for help with something genuinely dangerous and it says no. The cheap open models refuse far less, and some barely refuse at all. That's the trade-off. One is cheap but has limited guardrails. The other is expensive but has governance. Is it an either or situation? Currently that's where we are, but like everything in AI, what is true today can easily flip the other way tomorrow.

Read more →

Half The AI News You See Never Happened.
We Almost Ran Two Of Them.

Five stories came into this issue. Two of them died on the first search, not the tenth.

The first was a Yahoo job listing demanding 10 years of experience with Claude Code, a tool roughly a year and a half old. Funny, damning, 60,000 likes, and almost certainly not real. The screenshot misspelled the company's own city, and Yahoo's actual listing asks for six years of ordinary engineering experience like a normal employer.

The second was Google's AI supposedly handing out opposite advice depending on whether a husband or a wife was doing the yelling. It traces back to a Reddit post from 2023 that gets dug up every few months. And since AI answers change every time you ask, nobody can reproduce it to check, which is convenient for everyone reposting it.

This is the LinkedIn story up top, one level wider. LinkedIn is a single platform. The thumb scroll is the whole ecosystem, and a good share of what passes under it was built to be shared rather than to be true. The outrage post. The screenshot that lands a little too perfectly. The thing that makes you shake your head and send it to three people. Some of it happened. Plenty of it was assembled.

AI is genuinely making people and companies do strange things. But the strangest ones going viral were not selected for being true. They were selected for the click. That is the whole business model, and it works on us too.

FORMA's take: thirty seconds is the whole price. Is there a primary source, or only a screenshot? Does the timeline work? And if this were false, would somebody have said so by now?

Ask those three, or be the person in the replies going "wait, is this real?" eleven months late.

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.