Passing Thoughts

in my training-data primer, i was firm: synthetic data can scale expert judgment, but it can’t replace it when the work is ambiguous.

then Jev arrived. it makes structured decisions with probabilities instead of writing paragraphs. TypeSafe calls its training method RLCD: reinforcement learning for calibrated decisions. in TypeSafe’s framing, answers marked 80% likely should be right about 80% of the time across many decisions.

RLCD has shifted the paradigm, even while TypeSafe hasn’t published the specific reward or training loop that gets it there. and Jev’s founder says the model was trained entirely on synthetic data.

maybe i made “synthetic” do too much work. a model recycling its own answers is one thing. people designing the decisions, cases, and feedback a model learns from could be something else. if Jev holds up, expert judgment may have moved upstream, into deciding what’s worth teaching and how to tell whether the model learned it.

Jev is named for Jevons paradox: make a resource cheaper, and people may use more of it. if decisions get cheap enough to wire into every workflow, labor changes dramatically.

big word, but;

funny, i'm using a Gemini overview to give you the definition lol

a few months ago, i watched Dune: Part Two, a story about the dangers of rallying around messianic figures. working in tech, it got me thinking.

frontier labs are starting to feel like secular priesthoods. we pick a lab and start aligning ourselves not just with its models, but with everything it claims to stand for.

“go OpenAI!” “boo Anthropic!” “i loved Sam’s speech at [x].” “i can’t believe Dario made that claim about [y].”

the alignment feels natural. but in enough numbers, it gives the labs something close to a congregation.

with our support, the labs ship more than just models: they start narrating approaching ruptures, promising salvation, warning of annihilation, and turning forecasts of uncertain futures into modern fear or hype.

those stories become mandates for capital accumulation, consent to data mining, and compliance with a lab’s view of how life should be. by then, the institution looks a lot like a priesthood, its leader a prophet, and the intelligence at its center a golden oracle. we'll call it AGI. i'm afraid some people will call it God.