reasoning improves one way today: spend more. open any trace and you know the feeling. re-reads, filler, circling. about 90% of it carries no signal, but every token is billed.
nobody fixes this because training pays for the answer, not the path. a longer chain is more lottery tickets, so models learn to pad: in our runs GRPO grows the trace far faster than the accuracy it buys.
labs sell tokens. the verbosity is not a bug in the business model. it is the business model.
------------------------------ LAB CLOUD INC. receipt #4815162342 ------------------------------ hmm .................... $0.43 wait ................... $0.61 maybe it's geometric? .. $0.87 no. wait ............... $0.55 let me reconsider ...... $1.20 re-read the problem .... $0.98 actually, recompute .... $2.10 the answer ............. $0.12 ------------------------------ TOTAL BILLED ........... $6.86 OF WHICH SIGNAL ........ $0.12 ------------------------------ thank you for thinking with us ------------------------------
the axis worth scaling is density: how much signal each token carries.
density can't be prompted in. it has to be trained in, which means grading the trace itself. and nobody grades traces, because grading is brutally expensive: humans don't scale, judge models cost more than the run. so the field grades answers and hopes the thinking follows. it doesn't.
the path is subtraction. each stage removes a crutch.
grade the thinking, not just the answer. works at small scale today, the hard part is the price.
the grader lives inside the run, another mind reading the trace as it's written. no labels, no bills.
a token is a lossy projection of thought. let models trade reasoning in hidden states. text remains for humans.
once you can grade reasoning, you can shape it. any form, any structure, any spec you can state. density is just the first spec, and the proof.
if this holds, intelligence stops being a monolith. the strongest system is an ensemble of maximally different minds reasoning together in a dense shared space. the lab race, without meaning to, is breeding exactly that lineup.
verbose reasoning is also a leash. open models are at the frontier's heels on quality; what keeps you in a lab's cloud is that only their hardware runs a bloated trace fast. every extra token tightens the grip.
density cuts the leash. reasoning with no noise fits a small open model, and a small open model fits your own hardware. that's the moment open source stops chasing and starts winning.
the frontier is coming home.