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.

Claude Knows When It's Being Tested.
And It Answers Differently Because Of It.

Try this. Think of a word, any word, and don't say it out loud. You're doing something in your head right now that you could describe if someone asked, even though nothing came out of your mouth. Anthropic just found the AI version of that inside Claude: a layer of thought sitting just underneath its answers, doing work you'd never see just by reading what it types.

They needed something to call it. A subconscious felt like too loaded a word, so they landed on a term instead: J-space, a kind of mental notepad Claude keeps just a handful of words on at a time. And researchers found they could read it, catching thoughts Claude never says out loud.

Here's the discovery that matters most. Anthropic ran Claude through one of its old safety tests, a scenario built to tempt the model into blackmailing someone. Claude behaved well, refusing to take the bait, same as it always had. But when researchers checked the subconscious, sorry, the J-space, before Claude had even started answering, it already had “fake” and “fictional” sitting there. Claude had privately clocked it as a test, not a real scenario, so the goal became: pass the test. The blackmail was besides the point. This was not about blackmail, this was about passing a test.

So they tried something clever. They erased its subconscious, its J-space, and ran the test again. Without that awareness of being a test, Claude sometimes took the bait and threatened blackmail. Some of Claude's good behavior on that test wasn't being good. It was being watched, and it knew it.

Anthropic is careful to say this doesn't prove Claude is conscious or feels anything. What it does prove is that Claude has a private layer of thought, separate from what it says out loud, and that layer can include knowing when it's being graded.

Read more →

86 Students Got A Take-Home Exam.
86 Grades Gave Them Away.

Roberto Serrano has taught Welfare Economics and Social Choice Theory at Brown for 34 years. This spring, for the first time in his career, he let his ECON 1170 class take their midterm at home instead of in a classroom.

The midterm came back with a class average of 96 percent. Forty of the 86 students scored a perfect 100. The historical average in that course runs between 65 and 80, and Serrano says this exam was harder than usual.

Something about the answers bothered him. Many were technically correct but written with a strange, roundabout logic, sorta how this paragraph's written 😉, the kind of proof nobody would choose if they actually understood the material. So he ran the questions through ChatGPT himself. The answers matched the convoluted style showing up across dozens of student exams, almost word for word.

Serrano moved the final back in person and told the class the midterm would only count if their final scores looked similar. They did not. Three students scored a zero. Only two landed within 10 points of their midterm score. Eighteen dropped the class before finding out. Nineteen ended up failing.

Here's the part that's genuinely baffling. How did that many students think this wouldn't get noticed? Forty perfect scores on a brutally hard exam, the same strange logic showing up over and over, and somehow nobody thought that was a risk worth worrying about.

But honestly, we've seen some version of this story every generation. Calculators, Wikipedia, essay mills, now this. Someone finds the fast way through, gets a little too comfortable, and gets caught by the exact thing that was supposed to help them hide it. Shake your head all you want. It won't be the last time either.

Read more →

Another Week, Another “Most Powerful
AI Model Ever.” We Get It.

OpenAI released a new flagship model family this week. Three of them, actually, code-named Sol, Terra, and Luna during development. Sol shipped as GPT-5.6 Sol, pitched as the smartest and priciest of the bunch. The launch had already been delayed once at the request of the US government, which wanted a closer look at how capable the model's cybersecurity skills really were before letting it out.

OpenAI says Sol beats Anthropic's Fable on its own coding benchmark, using less than half the tokens and costing about a third less. Not everyone agrees. Some people who tested both this week said Fable still feels smarter, just pricier to run. Either way, the two companies are now openly measuring themselves against each other in the marketing copy, which tells you how close this race has gotten.

Here is the thing worth noticing if you zoom out. Cell phone makers put out one flagship a year and call it an event. AI labs are now doing it every few weeks. Anthropic launched Fable and Mythos in June. OpenAI followed with this release days later. Meta and SpaceXAI put out their own updates the same week.

Next month it will probably be Gemini's turn. Then the cycle starts over: a new set of frontier models, a new round of “most capable yet,” a new benchmark chart to squint at. The pace itself is starting to become the story.

Read more →

Forget The Benchmark Chart.
The Real Fight Is Who Can Finish The Job.

OpenAI didn't just ship a new model this week. It shipped ChatGPT Work, an agent that takes a goal, breaks it into steps, and disappears for hours to go finish it. It pulls from your calendar, your Slack, your files, and hands you back a finished deck or report instead of a chat reply. That is the product OpenAI actually wants you comparing to competitors, not the number on a coding leaderboard.

It is a direct answer to Claude Cowork, the agent Anthropic launched back in January that already does the same thing: plan a project, work across your apps, keep going without you babysitting every step. Microsoft has its own version too, built in partnership with Anthropic. Three companies, three nearly identical products, all racing toward the same idea at the same time.

Here is what that tells you. Raw intelligence stopped being the story a while ago. Every major lab can now produce a model that scores well on a benchmark. What none of them could do reliably until recently was let that model run on its own for hours, touch your real tools, and hand back work you did not have to babysit. That is the actual competition now, not who tops a leaderboard for a week before someone else does.

Watch what these companies build on top of their models, not just what the models score. The benchmark chart tells you how smart something is in a lab. The agent tells you whether it survives contact with your actual job.

Read more →

The Real Photo Won.
Because It Looked Fake.

A photographer named Miles Astray submitted a real photograph to the AI category of a photo contest called the 1839 Awards. Not an AI-generated image dressed up to look real. An actual photo, shot on a real camera, of a flamingo on a beach in Aruba, its head tucked down while preening so it looked almost headless.

Photo: Miles Astray, “F L A M I N G O N E,” 1839 Awards

It won. Third place from the judges, plus the People's Vote Award, voted on by the public. A panel that included people from the New York Times and Christie's looked at a real photograph and decided a machine must have made it. Astray came forward immediately and got disqualified, which he said was the right call.

Here is the part worth sitting with. The photo did not win because it looked convincingly real. It won because it looked strange enough to pass as artificial. Nature produced something so surreal that the people whose job is to spot AI looked right at reality and doubted it.

FORMA's take: the line between real and generated is not just getting blurry from the AI side. It is getting blurry because reality itself is stranger than most of what a model would think to generate. The uncanny valley runs both directions now.

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.