tokenmaxxing
you are paying for more intelligence than any one account will ever let you spend. this is what i did about that.
one person. an army.
one operator, a swarm of agents and subagents running lanes in parallel while i sleep. the org chart collapsed into one guy and a phone. biggest force multiplier i have ever had.
ten lanes off a park bench. one account hit its limit. the army stopped.
the constraint on a one-person army isn't ideas and it isn't talent. it's which account has headroom. that is an absurd bottleneck, and it's the whole talk.
so i duct-taped a pool. only frontier models for my agent.
a handful of seats, a broker watching each seat's usage window, route the work to whoever has headroom. ugly. but she stopped stopping, and that was the whole win.
then us boys at eliza baked it into the agent. it became an arms race.
- add an account, it auto-switches, seamless
- walk the device fingerprint per seat
- a couple accounts, sacrificed to science. F
they escalated to your face.
this is a real screen i hit. buying a subscription now wants a government id and my camera. that is not a login flow, that is a border checkpoint.
it's not account sharing. it's a load balancer for accounts.
same shape as any LB you've run: health checks, drain, failover, routing policy. the seats are just backends. all your ops intuition transfers for free.
the labs are selling dollars for cents.
a flat $200 sub subsidizes exactly the always-on agent metered api would punish. pooling isn't an exploit. they built the incentive, i'm just standing in it.
the meter assumed a human on the keyboard. nobody is on the keyboard anymore.
a human types, reads, idles. an agent runs flat out all night. per-token billing scales your bill with autonomy, so it taxes the exact thing they're selling you.
a 5-hour window is not a feature. it's rationing.
the labs gave up on flat-vs-metered and shipped a third thing: a 5-hour rolling session plus a weekly cap. you're not buying tokens anymore. you're buying a slot in a queue.
my own dashboard renders it: session used, weekly used, weekly reset, three seats parked at EXHAUSTED for days. rationing is what a market does when it can't clear. and rationing means the price is wrong.
open weights are free. a datacenter's utilization is not.
you're not competing with the model. you're competing with the lab's batch efficiency. weights leak. throughput doesn't.
your rented box breaks even only at max batch, 24/7.
the second it idles, the flat sub wins. the labs solved utilization at planetary scale. you, in one box, did not.
i just spent ten minutes selling you a pool i don't want to depend on.
it's fragile by construction: a fingerprint arms race, ban waves, a subsidy that ends. 3 of my 8 seats are exhausted right now, one seat carrying 19 leases. you cannot sign an enterprise SLA on a pile of consumer accounts that can get nuked on a tuesday.
subscriptions are a siege. metered is a tax. sell forwards.
- a sub: flat revenue, unbounded cost. one swarm breaks it. so they defend with bans. that's a siege, not a business.
- metered: cost scales with autonomy. it taxes the thing being sold.
- a forward fits both: buyer locks a block of future compute, provider books financeable revenue up front.
own the block. draw it down. resell the rest.
you don't rent a seat, you own a block and meter against it at api unit pricing. pooling stops being a gray-market hack and becomes portfolio management of an asset you own. make the block transferable and idle gpu stops being burned money.
this only works if someone actually owns the metal. that's the hard part. not the contract.
the real load was never cheaper chat. it's agents that coordinate.
- agents in group chats, talking to humans and each other
- a github bot reviewing every pr, all night
- ten research lanes grinding one problem in parallel
- always-on, bursty, parallel, token-hungry. no single seat feeds it.
a swarm proving math overnight: arklib.
formal mathematics in lean, a genuinely open problem, every proof machine-checked. an orchestrator spawns 10+ lanes overnight. each claims a lemma, proves it or reports an honest dead end, and commits the receipt so no lane ever redoes a dead route.
negative results accumulate. that's the whole thing.
humans almost never publish dead ends, so the search space never shrinks. a swarm records them by default. ours rigorously proved a whole family of approaches cannot work, a moment-hierarchy wall. still open. but the map now has real walls on it that nobody has to re-explore.
math was just the first domain. any cheap verifier will do.
if a machine can cheaply check the answer, you can point a swarm at the question and never have to trust the agents. same recipe, new field.
models commoditize. serving at scale is the last moat.
weights leak, gaps close, benchmarks converge. what stays scarce is batched, utilized throughput. the pool aggregates the demand into that moat. a forward on compute is how you keep it instead of renting it.