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.

A Study Just Measured How Much of LinkedIn Is Fake.
The Number Is Ugly.

LinkedIn built its whole reputation on being the place where real professionals share real expertise. New research says a lot of that expertise never came from a person at all.

AI detection company Pangram scanned over a million posts across five major platforms between April and June. LinkedIn topped a list nobody wants to top. More than 41% of its long-form posts, anything over 250 words, were flagged as fully AI-generated.

The scale gets worse when you zoom out. LinkedIn made up roughly a third of everything Pangram scanned, but accounted for nearly two-thirds of all the AI-written content it flagged. X and Substack had their own AI problems, but nothing close to LinkedIn's share.

LinkedIn is trying to fight back, testing reach penalties for posts that read like generic AI output with zero original perspective. Fair enough, except the same platform hands every user an AI Enhanced Writing Tool right there in the compose box. The company punishing AI writing and the company selling it are the same company.

The call is coming from inside the house.

Read more →

Can Your Poker Face
Fool AI?

For as long as poker has existed, professionals have leaned on one skill above the cards themselves: keeping a poker face. ESPN's broadcast just made that skill a lot less reliable.

Just another installment of the World Series of Poker during the long summer months... but this year with an AI twist. The bluff. Independent engineer Luke Geel built the tool: an AI model that studies a player's posture, facial movement, and blink rate to estimate whether they're bluffing or holding something strong. ESPN is now deploying it live during the broadcast.

The system never sees anyone's cards. It was trained on months of historical footage and outcomes, hunting for patterns between behavior and result. Producers only show the read after a hand ends, once a player has folded or been eliminated, specifically so the graphic cannot influence a hand still being played.

Geel says the system was harder to build than he expected, and that seasoned pros have likely trained themselves out of obvious tells already. But players have already started asking whether the tool could scout future opponents, or flag their own unconscious habits at the table. The tell used to belong to whoever noticed it first. Now it belongs to whoever's running the AI detection.

Read more →

OpenAI's First Hardware Product Isn't A Phone.
It's a $230 Keyboard. With No Letters.

Every rumor about OpenAI's hardware ambitions has pointed toward something bigger: a screenless companion device, a Jony Ive collaboration, a new way to talk to AI entirely. The first thing OpenAI actually shipped is a keyboard.

The Codex Micro, built with keyboard maker Work Louder, costs $230 and is designed for people running AI coding agents through Codex. It has 13 mechanical keys, a joystick, a rotary dial, and a compact aluminum body. Six keys light up in RGB to show what your Codex agents are doing: thinking, running, waiting on approval, or stuck on an error.

The controls are built around managing agents rather than typing. The joystick launches workflows like reviewing a pull request or debugging code. The dial adjusts how much "reasoning" an agent spends on a task, essentially a physical knob for token usage. Command keys accept changes, reject them, or start a new chat, all without touching a screen.

This keyboard doesn't exist on its own. OpenAI shut down Sora, its video app, this spring. It shut down Atlas, its standalone browser, this month too, folding both into ChatGPT and Codex instead of running them as separate products. At the same time, it rolled out GPT-5.6 in three tiers, Sol, Terra, and Luna, built specifically around agentic coding, cybersecurity, and long-horizon tasks. Taken together, it looks like OpenAI settling into a lane: agentic AI, coding agents, and the tools built to manage them, instead of chasing every consumer category at once. A keyboard built to monitor your Codex agents fits that lane exactly.

Read more →

New York Just Told AI Data Centers,
Hold Up, Wait a Minute.

If your electric bill has crept up the last year or two, a data center humming somewhere near you might be part of the reason. A single large AI data center can pull as much power as a small city, and the water and noise that come with it land on whoever lives closest. New York just became the first state to say "not so fast."

For up to a year, nothing new gets built if it pulls 50 megawatts or more, enough electricity to power roughly 15,000 to 20,000 average homes. Anything not already approved stays on hold while the state works out how to measure the impact before signing off on the next one.

New York is pointing at cost, not fear of the technology itself. Bigger bills, strained water supplies, and noise complaints are the reasons on record, and the state says these projects still have to clear local zoning like anything else gets built. It's also floating the idea of making data centers help pay for the grid upgrades they require, and pulling their tax breaks.

This is not a one-time story. It's another chapter of a fight that's starting to show up in state capitals and city council meetings everywhere AI infrastructure is getting built. We'll keep tracking it as more places make their move.

Read more →

The Best AI Tool You Use
Might Not Be A Chatbot. It's
The Connector.

Most people treat their AI tools like separate apps: ChatGPT for writing, Gmail for email and communication, calendar to track your day, Otter or Granola or Read.ai for taking notes in a meeting. Used one at a time, each of those is useful on its own. Used together, something more interesting happens.

For example, meeting intelligence tools have quietly gotten very good. Granola and Otter transcribe and summarize calls automatically. Read.ai goes further, scoring engagement and flagging who talked over whom. On their own, they save you from taking notes. Connected to Claude or ChatGPT through a plugin or a connector, they start acting like a research assistant that already sat in every meeting you had this month.

The transcript by itself was never the valuable part. What matters is being able to ask a single question across every transcript you have. What have three different clients said about pricing this month? A meeting tool connected to an AI model that can search across all of them can answer that, then go further: surface the reasons behind each one, draft the reply, etc, all in one motion. That kind of connected flow is already possible. Most people just haven't wired it up yet.

Meetings are just one example. The same pattern is showing up everywhere plugin marketplaces and connectors go. Your calendar, your CRM, your file storage, your project tracker: none of them were built to talk to each other. AI is becoming the thing that finally does.

FORMA's Take: stop asking which AI model is the best one. Start asking which of the tools you already use are not connected to your LLM yet. Plugin marketplaces and connectors are becoming AI's version of the App Store: the model is the phone, and every tool you plug into it is another app that makes the phone worth carrying. These marketplaces are getting more robust by the day, with new connectors landing constantly. Take ten minutes and go browse yours. There's a good chance the tool you already use every morning is one connection away from talking to your LLM directly.

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.