Small Decisions. Big Appetites.

This week I spent an unreasonable amount of time looking at mac and cheese for science. Meanwhile, the AI companies seem to be looking at each other's product roadmaps. Everybody is hungry. Let's get into it.

JEV Has a Lot of New Cousins

Last week I wrote about JEV, the model that makes small, structured decisions instead of writing you another paragraph. This week it feels like one of the most copied ideas in AI.

OpenAI announced its Decisions API at DevDay on September 29: Luna answering questions from a set of outcomes the developer defines. Then Cloudflare released Clef and Clef Flash on October 1, with image inputs, open weights, and compatibility with JEV's API.

Classification existed long before any of these products. Still, watching the industry suddenly get excited about fast, bounded decisions after JEV's arrival is pretty funny. Apparently the hottest new thing is asking a smaller question.

That part makes sense to me. A workflow often needs to choose a queue, flag a risk, or pick the next action. Making a big language model compose an answer every time adds work the application never wanted. Give the decision a clear shape, get a result, keep moving.

I have been experimenting with this in Cerberus to categorize repository findings. Clef's ability to look at pictures gave me another idea. Naturally, I used it to investigate lunch.

I Did the Hard Work. It Was Mac and Cheese.

I released Food by Clef this week. It is a goofy but very serious experiment: give it a food name, a photo, or both, and let Clef Flash apply the Cube Rule.

The rule categorizes food by where its structural starch sits. A hot dog gets treated as a taco. Lasagna gets to be cake. You can disagree, but you are going to need to take it up with the cube.

I put in the hard hours finding pictures of mac and cheese, stuffed peppers, and layer cakes. I'm not calling myself a hero, but someone had to do it, and I stepped up.

In my own informal testing, I would put the success rate around 99%. Then someone brought me a cinnamon roll that registered as sushi. So there is still work to do in the food court.

Food was my excuse to experiment with the model: see how it judges a picture against a rule, where those judgments get weird, and how quickly it responds. Change the angle, show the filling, describe the shape. Does the ruling hold up?

The speed has impressed me. Even calling it through an API, I have been seeing responses in under a second. Watching models like these play video games or choose the next move makes me think about more serious uses: routing a support request, flagging an operational problem, or helping an agent decide what to do next. Fast judgments could make those workflows feel very different. Lunch was a fun place to start figuring that out.

Try Food by Clef → Bring your most legally ambiguous lunch. See if you can break it. I want to know what else ends up being sushi.

Are the Labs Going Through It?

David Robinson, who led safety-report writing at OpenAI, resigned last week and warned that the company's culture was broken.

That landed alongside a DevDay that left me with mixed feelings. Dots moves into the always-on agent territory occupied by Grokbot, Hermes, and Muse. ChatGPT Space brings shared pages and collaboration into territory that reminds me of Notion and Confluence. The Decisions API joins the category JEV has made exciting.

My reaction to Dots was pretty harsh: this is the big reveal? The whole lineup felt a little like someone asked ChatGPT to make a product for every adjacent market. Useful features, maybe. But I wanted to see where the frontier was going. Instead, I kept recognizing someone else's neighborhood.

That leaves me wondering whether OpenAI is finding its next direction or flailing to own more of the software around us. I use these tools constantly. I can appreciate the engineering and still dislike how the commercial race feels.

Looking over at Anthropic does not settle it. Claude is powerful, but premium pricing and the march toward enormous valuations do not inspire me either. Its May funding round valued it at $965 billion. It is getting harder to hear the original mission over the size of the business.

I don't have the answers yet.

I also don't read this as the bubble bursting. I see people learning how to use AI gaining more room to build, decide, and act. The gap I notice is between having access and knowing what to do with it. That “human 2.0” idea is starting to feel real.

I am skeptical of the companies. I am still excited about what people can make with their tools. Including, apparently, a court for cinnamon rolls.

Quick Links

• JEV: TypeSafe documentation · JEV on OpenRouter
• Clef: Cloudflare's release and technical details
• Food by Clef: Try the experiment · The original Cube Rule
• OpenAI: DevDay recap · Dots documentation
• Catch up: Week 16—agents, decisions, and time back