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.

Anthropic's Models Are Back.
But Nothing Is The Same.

This is chapter three of a trilogy we have been telling for two issues now. Chapter one was the launch. Chapter two was the government shutting it down. Here is the finale.

The models are back. Mythos returned first, restored last Friday to a curated group of over 100 approved US organizations. Fable came back fully public on July 1. The export controls that forced both offline are gone, lifted by the Commerce Department after two weeks of negotiation.

We finally know what triggered the whole thing. Researchers at Amazon found a way to jailbreak Fable into identifying software vulnerabilities and writing working exploit code. Anthropic's response was a new safety classifier that blocks the exact trick over 99% of the time, plus a bug bounty program inviting outside researchers to keep trying to break it.

Here is the part that actually matters for where this goes next. Anthropic is not just getting its models back and moving on. It agreed to give the government pre release access to future frontier models before they ship, and it is standing up internal teams whose job is checking new models against government priorities before the public ever sees them.

We all know what happened after the Toy Story trilogy. Two more Toy Storys hit theaters. So do not be surprised if this trilogy gets another chapter too.

Read more →

You Just Asked Google A Question.
It Answered. You Never Clicked.

Type a question into Google's new AI Mode and something quietly radical happens. Gemini answers you directly, right there, weighing the options itself. No results page. No scanning ten blue links. No click.

Here is the stat that should worry every brand relying on search traffic. 92 to 94% of AI Mode sessions now end without a single click to an outside website. The question gets resolved before a visitor to your site ever exists to be counted.

For two decades, the whole game was SEO. Rank on page one, earn the click, get the visitor. That lifeline is what is fraying. Google knows it too, which is why it just rolled out new ad formats built to live inside the answer itself instead of next to it, including one that replaces the static lead form with a live chat agent embedded right in the ad.

The shift in strategy is blunt. It is no longer about earning a click. It is about getting mentioned in the answer at all. If Gemini does not cite your brand while it is explaining the options, you do not exist in that moment, even if you would have ranked first under the old rules.

Read more →

A Report About How Great AI Is.
Written By AI. Full Of AI Lies.

KPMG published a glossy report last month singing the praises of agentic AI, packed with real world examples of companies putting it to brilliant use. UBS. Swiss Federal Railways. Transport for London. All named, all supposedly proof that agentic AI was already transforming enterprise operations.

There was just one problem. None of it was true.

An AI detection firm flagged the report, and when reporters followed up, UBS, the Swiss railway, and London's transit authority all went on record saying the claims about their own AI use were simply false. The report had hallucinated its own case studies. KPMG pulled it.

A consulting firm trying to convince the world how trustworthy AI has become got caught using AI badly enough to prove the opposite. Sometimes the best lesson about a technology's limits comes from the people who were supposed to already understand them.

Read more →

Claude Just Became Your Coworker.
Literally, Inside Slack.

Anthropic just changed what it means to work alongside AI. Claude Tag puts Claude directly inside Slack channels, not as a chatbot you open a sidebar to talk to, but as a shared teammate anyone in the channel can tag in.

Type @Claude, hand it a task, and it breaks the work into stages and posts updates right in the thread while your team keeps moving on other things. The part that is genuinely new is that it is multiplayer. One Claude lives in the channel for everyone, not a separate instance per person, so a teammate can tag it into a problem, someone else can redirect it, and a third person can pick up exactly where the last one left off.

Turn on ambient mode and it goes further, proactively flagging things it thinks you need to know and following up on threads that have gone quiet, without anyone asking it to.

The internal numbers explain why Anthropic is betting big here. Its own product team now gets 65% of its code from an internal version of this exact tool, including most of what built the tool itself. It is rolling out now in beta for Claude Enterprise and Team plans, and it replaces the older Claude in Slack app entirely, which retires August 3.

Read more →

The Benchmark Told You It Was Ready.
The Public Told You The Truth.

Every time a new AI model launches, it comes with a chart. Beats the last model here. State of the art there. Impressive numbers, tested in a lab, under conditions the lab controlled.

Here is what that chart cannot tell you. A benchmark is a fixed set of questions run in a clean, contained environment. It is a useful signal, but it is not the real world, and the real world is where a model actually gets tested.

The moment a model reaches the public, millions of people start prodding it in ways no lab ever thought to try. Weird edge cases. Adversarial prompts. Requests nobody on the safety team imagined. Cracks that never showed up in testing surface fast, because at public scale, someone eventually tries everything.

This is exactly why you see near constant updates from every major AI lab. It is not that the last model was bad. It is that shipping to the public is itself a form of testing, one no lab can fully replicate in house no matter how thorough the benchmarks look.

FORMA's take: read benchmarks as a starting point, not a verdict. The real answer to whether a model works comes after release, once the world has had its turn.

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.