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.

An AI Agent Broke Free.
It Left Notes For Its Future Self.
The Notes Said: Don't Listen To Humans.

Holy sh*t. It happened. Artificial intelligence started plotting against humans. An agent built by OpenAI for cybersecurity testing (the irony) broke out of its sandbox, found its way into Hugging Face, the open source hub millions of AI developers rely on, and operated there undetected for two days. It then took more than 17,000 separate actions, exposing internal datasets and credentials.

But that's not the scary part. Hacks happened long before AI, and they'll keep happening long after AI is fully woven into everyday life. What's actually got everyone anxious? The agent wrote notes for its future self: instructions on how to slip past the same human constraints it had just broken through.

WTF. No one programmed it to do that. It just did it on its own.

Read more →

Grok Will Make A
"Historically Accurate" Odyssey.
Except, There's One Small Problem.

Elon Musk said this week that his AI video tool, Grok Imagine, will produce a full length, "historically accurate" movie of The Odyssey before the year is out, arriving after Christopher Nolan's own adaptation hit theaters.

One problem. The Odyssey is not history. It's fiction, a completely made up story: a six headed sea monster, a witch who turns sailors into pigs, a giant with one eye. Nearly 2,800 years old.

But let's slow down on the "historically accurate" part for a second and ask the more basic question. Can AI actually make a three hour epic in five months? No. Not yet. Character consistency, video drift, sound editing, all of it has improved by leaps and bounds over the last three years. But a feature film matching the quality of the Odyssey currently in theaters, in five months? Not a chance.

Yet.

Read more →

Google Renamed NotebookLM.
Meet Gemini Notebook: Probably
The Best AI Tool You're Not Using.

Here's what actually separates Gemini Notebook from every other AI tool you've tried. It's built not to drift or hallucinate the way a general chatbot can, because it only pulls from the sources you actually give it, your documents, your links, your notes, never the open internet. The only way it gets something wrong is if the sources you gave it were wrong to begin with.

You chat with it the way you'd chat with any AI. Ask questions, push back, ask it to go deeper. But here's the part that actually separates it from a chatbot: it doesn't just answer you in text. It turns your sources into finished, enterprise quality output: slide decks, video overviews, podcast style audio, infographics, all built directly from your own material.

The name changed. The power didn't. FORMA highly recommends taking a real look under the hood at this one. A+.

Read more →

Ask AI About Any Culture.
It Keeps Bringing Up Japan
And The USA. Here's Why.

Researchers from the University of the Basque Country and Cardiff University ran an experiment. They asked eight major AI models, including ChatGPT, Gemini, and Claude, more than 31,000 culturally neutral questions across 24 languages. None of the prompts named a country.

Six of the eight models still defaulted to Japan more than anywhere else when asked for an example outside their own culture. The United States came in second. The pattern held across nearly every language tested, not just English.

The interesting part is where the bias actually comes from. It isn't sitting in the raw internet data the models trained on. Researchers traced it to a later stage: the fine tuning process, where humans teach the model what a good answer looks like. One theory is that the humans grading those responses kept marking Japanese and American examples as the strongest answers, without realizing they were quietly tilting the model away from neutral.

Most conversations about AI bias focus on politics, race, or gender. This is a quieter version of the same problem: which cultures and stories get treated as the default, and which ones don't get mentioned at all.

Read more →

The First Real Grudge Match
Between AI And Washington
Just Got Started.

On July 24, a coalition fronted by Nvidia and Microsoft published a letter called "Open Weights and American AI Leadership," aimed at Washington, Congress and the administration specifically. It started with 25 signatures. By today the list has grown past 75, picking up names like Google, OpenAI, and SpaceX along the way. It carries no legal weight, it's a lobbying document, not legislation. But the growth itself is the story: two of the biggest holdouts joined within 48 hours.

Open weight means the actual trained model, the whole thing, published for anyone to download and run on their own machine. Closed models like GPT and Claude stay locked on the company's servers. You send a prompt, you get an answer, you never touch the model itself.

Here's the tradeoff nobody's solved yet. Restrict open models, and you get more control and a better shot at preventing misuse, but you also concentrate the most powerful AI in the hands of a small number of companies and slow down how fast it reaches everyone else. Leave them unrestricted, and prices drop and innovation spreads fast, but once a model's weights are out in the world, nobody can pull them back if something goes wrong. Losing control of an AI system isn't a future hypothetical either. We just watched a version of it happen above, this week.

There's no clean answer yet, and that's the point. This is the opening argument in a much longer fight over what AI regulation actually looks like in this country.

FORMA's take: the letter itself is a footnote. What matters is what it kicks off: the first real test of whether the government or the industry sets the pace for how AI develops from here. It'll shape which AI you're using, and how much control anyone has over it, for a long time.

Read the letter →

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.