I am working to understand what I consider “Frontier Storytelling” in today’s market, which is the stories that get told and retold on the frontier of the AI market. They aren’t the fabulist “Frontier Saloon Stories” like OpenClaw and Gastown, but the ones that are shaping how business is getting done by startups at the edge of the secondary market.
Today I am considering TypeSafe’s new System One Models and Jev. As a storyteller, I want to understand the moving parts of this story, and continue to develop the fingerprints I need to use against other companies in the future. To be clear, I am not assessing the viability of System One Models and Jev, I am looking at how this company is telling the story in a way that allows it to speak to a wide audience, enable it to be retold, and shape (or not) the AI market zeitgeist.
One thing you should always lead with is a manifesto. TypeSafe has one, and it will get the juice flowing. “The bottleneck isn’t raw intelligence. It’s that today’s intelligence is hard to build on.” There is a Ford Model T analogy in there about early cars imitating carriages before they found their native form, and a mission statement about catalyzing a “Cambrian explosion of intelligent software.” It closes on “prod, not God,” which is about as efficient a piece of positioning as I have read this year.
Next, you will need to move something forward. Like “structured outputs.” Take the thing everyone is already half-doing and declare that you have finished it.
You are going to need to fix a major problem of the past. Like hallucinations, aka being WRONG. TypeSafe’s version of this is a 0% type error rate, guaranteed by the structure rather than promised by the model.
And it will help your narrative to bind your story to an existing need. Like automation workflows. Nobody has to be sold on the need; you just have to attach yourself to it.
To help provide a rich and fertile substrate for your story to take hold, it helps to introduce new acronyms. Like Reinforcement Learning for Calibrated Decisions (RLCD), positioned directly against RLHF.
Butter us up with new phrases. Like “unstructured state in, typed probabilistic decisions out.”
Then round things out with new(ish) paradigms that polish the things you are already doing. Like smart if-statements.
As we roll down the backside of this bubble, make sure you bring in economics and address cost. TypeSafe prices input tokens at $0.042 per million and lists output tokens as “FREE (too cheap to meter),” which is a flex and a positioning statement at the same time. But you will also have to talk about speed and speak to the velocity every company desires—we aren’t cooked enough yet to be talking about slowing down. They put end-to-end response time at 70ms to 500ms against 3 to 329 seconds for frontier LLMs. Then make sure you do the work to build confidence by providing evidence and a comparison of the current state and what has changed with what you are offering. They published a comparison table, workflow evaluations, and—to their credit—a note that the 193.6x faster and 444.6x cheaper numbers on the homepage “are on the higher end of real world gains.”
I feel these are the solid moving parts of frontier storytelling. It ain’t the big city storytelling that enterprises need to hear, I will cover that in another post. And it ain’t the crazy fabulist shit you’ll get in a frontier saloon. It is right in the middle. It is enough to get you excited, but it is still grounded (somewhat) in reality and in what real world businesses will need to get done as part of their operations. It provides a repeatable format. I’m not sure secondary market startups will get more than one chance at this type of storytelling, maybe two or three at the maximum, before they actually have to deliver.
There are certain market realities that will determine whether or not a story and its supporting product will land and have real world market impact. To help me understand how serious a company is underneath the agentic frontier storytelling, I rely on the Kin Score to help me understand what the foundation of a company looks like.
The TypeSafe.ai foundation is currently 51.8, which is mid. Their access clarity is currently 41.1, contract quality is 53.7, contract governance is pretty low at 18.2, operational transparency is at 47.4, developer ergonomics is at 73.8, and discoverability is at 68.5. It is a substantive foundation for a new startup that has recently emerged from stealth mode and begun hustling their warez in the market.
The TypeSafe.ai agent readiness comes in at a lower 40.7. They have a Machine-Readable Contract, Stable Error Semantics, an A2A Agent Card, Agent Skills, an MCP Server, a Well-Known Catalog, Machine-Readable Auth, Rate-Limit Signaling, and Request/Response Examples. They do not have an Agentic Access Contract, Idempotency, Documented Reversibility, a Typed Event Surface, Delegated User Identity, Registration Without a Human, Protected Resource Metadata, an Agentic Commerce Well-Known Document, a Dry-Run / Simulate Mode, or Consent & Bot Identity. That leaves quite a bit of work on the table before agents can work their magic.
Ultimately I buy TypeSafe.ai’s story. I wouldn’t say I’d go all in on what they are providing. There is a lot of devil in those probabilistic details in between the inputs of “unstructured data (e.g. text) with an emphasis on structured program state” and the outputs of “type-safe structured values. Possible outputs and structure are defined in advance. The model never makes type errors. All answers are accompanied with calibrated probabilities and confidence scores.” I do believe more structure is required to realize the automation needed to manage business workflows.
My money would be on deterministic approaches to the workflow and the “smart if-thens,” with deterministic logic handling what step is taken next. I do think LLMs have their place in this, but it will come down to what the workflow automation is handling, what industry you are operating the workflow automation in, what regulation and compliance is required, and what your token budget appetite will be. It depends. But I am more interested in the storytelling economics of the market we have today on the slippery downslope of the AI bubble.
Next, I will likely do a “Big City Storytelling” post around Arazzo, or maybe OpenLint, to see how other stories land in enterprise circles that are a little more reserved and cautious, but are still listening in on frontier stories.
