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.

AI Releases Are No Longer
Just Product Launches.
They Are Negotiations.

Two AI companies. Two new models. Two very different experiences with the same government. And a third company watching both and quietly rethinking its timeline.

A few weeks ago, the US government ordered Anthropic to pull Fable and Mythos offline with almost no notice and no detailed explanation. Anthropic had no way to restrict access to just foreign users, so to comply, both models went dark for everyone. Days after launch.

OpenAI took note. When GPT-5.6 launched on June 26 -- three models: Sol, Terra, and Luna -- they did something similar to what Anthropic had done with Mythos: voluntarily limited early access to a small group of trusted partners. The difference is what happened next. Anthropic got forced offline entirely. OpenAI, by briefing the government ahead of launch and cooperating proactively, kept its models running. Restricted, yes. Dark, no.

The lesson the industry is drawing: get ahead of the government, or the government gets ahead of you. Now look at Google. Gemini 3.5 Pro was promised for June. It got pushed to July. The official reason is more testing time. But the rumor circulating in AI circles is that Google watched what happened to Anthropic and decided it would rather have the government vet the model first than face a forced shutdown after launch. That has not been confirmed. But it does not need to be confirmed to be completely plausible.

The government did not write a law. It did not go through Congress. It just made its preferences known, and the entire industry is now reshaping its release strategy around them. That is the new normal taking shape in real time.

Read more →

AI Is Biased. But Not
the Way You Think.

Research keeps surfacing the same two findings, and they are starting to add up to something worth understanding.

First: do AI models lean left? Well, sort of. But not because anyone made a conscious decision to build them that way. The Washington Post ran major models through a battery of political questions drawn from a 2025 Stanford-Dartmouth study. Most leaned left. Prior research found users perceived OpenAI's models as left-leaning at roughly four times the rate of Google's. This pattern has shown up consistently since at least 2023, across dozens of studies and multiple model families. No one at these companies sat down and said make it left-leaning. What happened is more interesting than that, and we will get to it in a moment.

Second: do AI hiring tools discriminate? Yes, and the numbers are hard to ignore. A new Stanford study, the largest independent analysis of AI-powered hiring algorithms ever conducted, found that 26% of Black applicants and 15% of Asian applicants applied to positions where the AI system screened them out at rates that trigger federal discrimination standards. Same qualifications. Different outcomes. Again, no one programmed the system to do this.

So what is actually happening? Here is the explanation that makes both findings make sense.

AI models are trained on human-generated data: decades of writing, publishing, hiring decisions, online behavior, and news coverage. The internet is not a neutral archive. It reflects who had access, who got published, who got hired, and who got to set the terms of public discourse. AI does not invent bias. It inherits it. It runs a compression algorithm on human output and hands it back to us at scale.

The models are not putting a thumb on the scale. They are holding up a mirror. What makes that uncomfortable is that we do not always like what we see.

Read more →

Agentic AI Made More Apps
Than Ever. The Problem Is
No One Is Downloading Them.

Agentic AI did not just change how we use software. It changed who can build it. And the data is starting to show what happens when building gets easy but demand stays fixed.

App releases on the iOS App Store were up 80% in the first quarter of this year compared to the same period last year. By April, that number hit 104% across both the App Store and Google Play. The driver is agentic AI coding tools that let people build and ship mobile apps faster than ever before, many of them for the first time.

The problem is on the other side of the equation. User adoption stayed flat. Same number of phones. Same number of hours in a day. Same amount of attention to go around. The supply of apps exploded. Demand did not move.

What agentic AI changed was the cost of building. What it did not change is whether anyone wants what got built. There is a line circulating in product circles right now that captures this moment precisely: shipping an app is easy. Getting a thousand active users is the hard part.

For years, the barrier was technical. You needed developers, time, and money to get to market. Now the barrier is something older and much harder to code your way around: whether you built something people actually want.

Read more →

Two Governments. Two Kids.
Opposite Answers on AI
in the Classroom.

Norway and China looked at the same technology, the same age group, and the same question: what do we do about AI and children in school? They arrived at answers that could not be further apart.

Norway pulled the plug. Starting in August, students in grades one through seven, ages six to thirteen, will have no access to generative AI tools during school. Older students can use it only under direct teacher supervision. The Prime Minister's reasoning was simple: children need to learn to read, write, and do math first. Let them skip those steps and you may not get them back. Norway already banned smartphones in schools in 2024. Test scores went up. Bullying went down. They are applying the same logic to AI.

The same week, China's State Council released a five-year blueprint making AI a mandatory core subject for every student from first grade through university. Third graders learn the basics. Fifth graders study intelligent agents and algorithms. The stated goal is direct: build a generation that does not just use AI, it builds it. The country that gets this right wins something that compounds for decades.

Two countries. Two philosophies. One says AI is the threat to how children develop. The other says the real threat is a childhood spent without it.

There is no wrong answer yet. The data will arrive about fifteen years from now, when both of these generations enter the workforce at the same time. Every country watching from the sidelines is already choosing a side, whether they realize it or not.

Norway: Read more →     China: Read more →

Your AI Can Access New Information.
That Is Not the Same as
Being Trained on It.

There is a distinction most people miss when they use AI tools, and once you understand it, it changes how you interpret everything the model tells you.

AI models are built on training data: a massive snapshot of the internet, books, and other text sources, frozen at a specific point in time. That snapshot closes months before the model ever reaches you. The process of training, fine-tuning, testing, and releasing a model takes time. The gap between when the data cuts off and when you start using the model is typically four to six months at minimum, often longer.

Here is where the confusion comes in. Many AI tools now have web search built in. So yes, the model can look up something that happened yesterday. It has access to current information. But access is not the same as knowledge.

When a model searches the web, it retrieves information in the moment, the same way you would Google something. But its underlying knowledge, the instincts it draws on, the confident baseline answers it gives without being asked to look anything up, those are baked into the training. And that training ended months before you opened the app. The model can read today's headline. It has not lived the context that makes that headline make sense.

Think of it this way. A model without web search is like a brilliant colleague who has been on sabbatical for the better part of a year. A model with web search is that same colleague, now able to pull up articles on their phone mid-conversation. They can read the page. But they have not absorbed it the way they have absorbed everything they learned before they left.

FORMA's take: for anything time-sensitive, verify. Use web search when the model offers it. And treat confident answers about recent events the way you would treat a strong opinion from someone who just got back from six months off the grid. Smart. Helpful. But worth a second look.

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.