The Arrival of Fable.
Anthropic shipped a new frontier model this week called Fable. I swapped engines immediately.
When Fable arrived, I did what I normally do when I test a new AI product. I designed a small, hyper-personal project to learn on.
Soccer was my first love. I played all the way through college. I'm looking forward to enjoying as much of the World Cup this summer as I can. With the tournament kicking off this week, the project picked itself.
When I think about where to focus my efforts these days, especially in this era of unprecedented change, I think about it like a Venn diagram of me. Stack enough layers and the work becomes impossible to copy. I was excited to overlap my content production experience, my AI expertise, and my love of soccer to test Fable.
Not-so Friendly
What I decided to build was a World Cup edition of the Data Bakery, the public-data toy I made last week. Pick any match and the tool finds a surprising, educational, and delightful connection between public data from the two countries. Forty-eight nations, every matchup. Opening day.
The idea was to educate people about the countries playing, demonstrate how AI works in a fun way, and make content to promote CGA Creative on LinkedIn.
Most organic business content does not work on LinkedIn. The algorithm buries it. I'm hoping that by creating engaging tools and games like this, timed perfectly with cultural moments, I'll occasionally cut through and get CGA some attention, and maybe a viral moment or two.
I've managed content calendars for a long time, and the hard part has always been the looking forward: scanning ahead routinely enough to see what's coming and make something for it in advance. Gen AI enables fast production, which changes the math. You can make real work inside the window a moment is still live, the move Ryan Reynolds and Maximum Effort made famous with their overnight Peloton ad and the wave of fastvertising after it. I've taken one swing at this already, the March Madness mascots. It didn't go viral but it got great attention.
To ensure I keep up with culture, I built Zeitgeist, an AI early warning system for those moments, tournaments, elections, days of recognition, so I see them coming with enough time to produce fast, on-moment work. (It's linked at the bottom.)
By lunch I had built v1 of The Friendly, and I was running tests: Mexico v South Africa, Qatar v Switzerland.
Forty-eight nations, lined up and ready.
The results that came back were terrible. Every pairing produced one of two failures: facts so obvious they were just trivia, or comparisons so forced they meant nothing. The tool was doing exactly what I asked and none of what I wanted. I kept pushing, sure the next prompt would fix it.
I wrestled with Fable all day and eventually killed the project. I had been so caught up in the idea of this powerful new tool, and so sure it would accomplish the task, that I never stopped at the beginning to ask it the one question that mattered: is this idea, in the way that I've imagined it, feasible? It turns out it would have known. What I was trying to build had a fatal flaw, and the AI could have told me on minute one. The fatal flaw is irrelevant. The point is I never asked if we might encounter one. If I had, I would not have wasted an entire day.
Before you build, ask the model whether the thing can be built. Duh.
It is now baked into the protocol here at CGA Creative. Every project starts with a feasibility study: name the objective, name the hardest case it has to handle, and prove one real output before building the machine around it. New engines change what is possible. They do not change the order of operations.
An hour of video, read in seconds.
In one hour of video, there might be 85,000 individual frames, but only 152 that matter to a machine.
The way I make money at CGA Creative is by solving business problems with AI. This week I was working with my family's insurance agency, which has operated essentially the same way since 1901. I ran my usual discovery Zoom with Lou, the CEO. He shared his screen and walked me through his day, the task-management software the agency runs on, and the pain points of running the place.
Afterward I had to get that session into my CGA OS so the system could process it. The transcript was easy, I just paste it in. The screen share was the hard part. How do you hand a machine an hour of someone clicking through software?
I produce video, so I think in frame rates, and it struck me that a Zoom call has almost no meaningful frames. I wondered what it would take for Claude to actually see the video. Could it just screenshot 24 frames every second? That is where the conversation started, and it whittled down fast: we don't need 24 frames a second. We only need the roughly 150 where the screen actually changes. And 150 frames is absolutely capturable.
That is how the Video Analyzer got built. Now CGA OS watches the video and exports a contact sheet of the meaningful frames that are quick to process. I am excited about the compounding value of this tool. When you need to understand a video fast, stop watching it. Turn it into something a machine can read.
The real output. One hour of the call, only the frames that matter.
Inner world models
There is a lot of talk these days about world models as the next frontier. People are now working on developing machines that understand physics, space, cause and effect.
I keep thinking about a different layer: the inner world model. Values. Memory. Aspiration. Emotion. Taste. The model of a person, not a planet. Researchers poke at pieces of it under names like theory of mind, but nobody has made it the bet, and nobody is racing to build it the way a billion dollars is now racing toward the physical kind.
I've started a small experiment: can a machine hold one person's taste, what they love, what they notice, what they'd pick? Not to automate them, just to find the edge, how much of a person fits before the rest slips through. Early days.
I'll report back.
The rest of the week
A friend told me to throw a new frontier model at my hardest problem. So I pointed Fable at the Historical Resurrection Project overnight and woke up to complete pre-production packets, dossiers, scripts, and shot lists, for 250 Untold Stories of America. Can I make 250 documentaries by July 4th? That's a new world question.
I rebuilt the Intelligence page at cgacreative.com/intelligence. Six tools, each demonstrated by a live animated tile, not a screenshot. The page itself is the proof of the work.
The Data Bakery from last week's issue learned a new trick: a visualizer that turns your pairing into a graphic you can post anywhere. One of the best ways you can support my work is by trying the tools and sharing the outputs online. Please give it a try. It would mean the world to me.
I built Zeitgeist, the cultural early-warning radar I mentioned up top. Every tournament, election, and day of recognition worth building for, mapped out through 2027, so the next on-moment piece is never made the morning of. cgacreative.com/zeitgeist
Grounded.
By Friday, Anthropic pulled Fable.
The U.S. Commerce Department issued a national-security export-control directive barring the model, and its bigger sibling Mythos, from any foreign national, including non-citizens working inside the U.S. The scope was wide enough that Anthropic disabled both models for everyone to comply. Reporting traces it to another company that claimed it had jailbroken Mythos; Anthropic calls that a misunderstanding and says it is working to restore access. The older Claude models keep running.
So the engine I dragged out to the plane on Monday was grounded by the government by Friday. I got one week with it. This issue is, for now, a postcard from a model that is no longer available.
What a week in AI.
Pass it on
Please share this with your friends interested in AI and creative work.
If you or someone you know runs a small business with a problem worth solving, I'd love an introduction. It's the best way to keep me flying!
Erich
Founder · CGA Creative