Congrats @typesafeai !!
What's so cool about Jev is that it breaks the tension between expensive, dynamic intelligence at runtime (w/ opex costs) and cheap, brittle intelligence at compile time (w/ capex costs). More:
- One tension between labs and their customers is the economics of coding agents vs general agents w skills / projects / memory
- We went from fat apps (most of 2025) to thin apps that mostly wrap intelligence, with code as a disposable artifact generated at runtime
- Generating code is a crystallization of inference. It captures business logic that could otherwise live in the model as static procedural logic. Cheap to run, but brittle
- The alternative is regenerating everything dynamically at runtime with PhD-level models w reasoning, memory, tools. More elegant / general, but the most token consumptive: you pay opex every time you use the app vs capex once
- This partially explains why labs have pushed so hard for skills / projects / agents over code first - the general agent approach drives far more token consumption
- Most runtime "intelligence" is small decisions: classify / route / score / pick. A fast, cheap System One model gives you dynamic intelligence at runtime for roughly the cost of compile-time code
- So the future stack should be code generated by models at compile time, Jev-style decisions at runtime, and frontier agents + tools reserved for genuinely open-ended work
I LOVE this model & I'm very excited to see what comes next.
What's so cool about Jev is that it breaks the tension between expensive, dynamic intelligence at runtime (w/ opex costs) and cheap, brittle intelligence at compile time (w/ capex costs). More:
- One tension between labs and their customers is the economics of coding agents vs general agents w skills / projects / memory
- We went from fat apps (most of 2025) to thin apps that mostly wrap intelligence, with code as a disposable artifact generated at runtime
- Generating code is a crystallization of inference. It captures business logic that could otherwise live in the model as static procedural logic. Cheap to run, but brittle
- The alternative is regenerating everything dynamically at runtime with PhD-level models w reasoning, memory, tools. More elegant / general, but the most token consumptive: you pay opex every time you use the app vs capex once
- This partially explains why labs have pushed so hard for skills / projects / agents over code first - the general agent approach drives far more token consumption
- Most runtime "intelligence" is small decisions: classify / route / score / pick. A fast, cheap System One model gives you dynamic intelligence at runtime for roughly the cost of compile-time code
- So the future stack should be code generated by models at compile time, Jev-style decisions at runtime, and frontier agents + tools reserved for genuinely open-ended work
I LOVE this model & I'm very excited to see what comes next.
Oct 11, 2026 · 4:26 PM UTC
9
6
82
8,782
@illscience @typesafeai it's really quite great -- a well-executed idea that should have come a long time ago. it's time we revisit all those assumed deadends.
1
0
0
90
@dpetrou @typesafeai Totally - I said every morning for email triage, and it’s blazing fast and very cheap.
0
0
0
60
@illscience @typesafeai is they identified the use case which was sitting there for years but unsolved
That said, considering the number of copycats that came in the weeks after, makes the moat debatable
Maybe the true moat would be finding innovative use cases that exist but not yet truly solved , and making they faster and better
That said, considering the number of copycats that came in the weeks after, makes the moat debatable
Maybe the true moat would be finding innovative use cases that exist but not yet truly solved , and making they faster and better
1
0
0
48
@avithaldas @typesafeai Yeah, but over and over and again we’ve seen that for every category of model there’s comparative advantage / specialization in different directions, and the market is just way, way bigger and more expansive than it seems at first blush
1
0
0
41
@illscience @typesafeai Would love to see Jev handle a routing step in a real app. Any early access numbers yet?
0
0
1
17
@illscience @typesafeai The 'narrative violation' framing is perfect, cheap dynamic intelligence breaks every assumption we built tooling around. I've been waiting for someone to say this out loud.
0
0
0
10