S4E365 — Decisions API is in public beta

GARY: —so it already ran me through the endpoint. I typed my own name and it came back zero-point-nine-one complaint. I have been classified, is my point.

MAEVE: Gary. It's a public beta. It hasn't decided anything about you.

GARY: It returned a probability, Maeve. That's a decision with paperwork.

OBI: It's a number. You still pick a threshold. You are the threshold, Gary, which is the saddest sentence I've said all week.

GARY: I'm Gary. Failed founder, friend of the thread.

MAEVE: Maeve. General partner. I fund things like this and bill the round to whoever's left.

OBI: Obi. Staff SRE. I get paged when your zero-point-nine-one wakes someone at 3 a.m.

GRUNER: Gruner. Field researcher. Humans outsource judgment. Is my hypothesis.

GARY: Okay. Subject, before anyone quotes a comment. OpenAI shipped a Decisions API. Public beta. One model, gpt-6-luna, on POST /v1/decisions.

OBI: You send text, or text plus an inline image. You get three answer shapes. A predicate—probability a condition is true. A choice from a fixed set. A score against a rubric.

MAEVE: Ten times faster than the Responses API. Ten cents per million input tokens. No output-token charge.

GARY: No output charge, because it doesn't write anything. It just judges you. Like my mother, but metered.

OBI: It doesn't decide, Gary. It returns an estimate and your application decides. The docs say pick thresholds from labeled examples, priced by false positives and false negatives.

GARY: So it's ninety-one percent sure I'm complaining, six percent sure I'm complimenting, and OpenAI doesn't have to live with me. Three percent unaccounted for. That's my hope, that's my fucking hope—

OBI: The remaining three percent is other. It's not hope, Gary. It's "other."

GRUNER: Correct. You are unresolved input.

GARY: That's the meanest thing you've said about me in any language.

GARY: First comment. Simonw, who I would die for. He asked the machine, quote, "Is this a complaint?" about the sentence "I am angry about the new product feature."

MAEVE: Zero-point-nine-one complaint. Zero-point-oh-six compliment. Case closed, court adjourned.

OBI: It closed nothing. It priced a guess. He still picks where the guess counts.

GARY: He shipped a whole plugin for this. A plugin, Maeve. Idea, shipped, karma as we speak. I once shipped a sock that reported temperature.

MAEVE: Thermal. We know.

GARY: The telemetry worked, is my point.

OBI: The telemetry was never the product, Gary. You sold a subscription to learning your foot is warm. Twelve dollars a month, Gary.

GRUNER: Foot data. Strong market. No.

MAEVE: New rung. Chupchap, quote: "So someone finally built the zero-shot model from the sales decks of 2015."

OBI: Oh, fuck this guy for being right—

GARY: That's hurtful, because he's right. Every deck I ever sat through had that slide. "Classification, but zero-shot." I watched a man raise on it in a Patagonia vest.

MAEVE: Fifteen years, same product, new cover. It's the Sears catalogue of machine learning. Quarterly.

GRUNER: Is fresh legacy. Very legacy. Legacyest.

OBI: Here's the thing, though. Old way: a model per task. Complaints, damage, severity — three datasets, three deploys, three on-calls at 3 a.m. This thing takes any input, including an image, and nobody trains shit.

MAEVE: So the output is 2015 and the input is whatever the fuck is lying around.

OBI: That is the entire product. Yes.

GARY: In 2015 the slide said "zero-shot." In 2026 the slide says "zero-shot" and the demo actually works. That's not — no, wait. That is progress.

OBI: First time the slide has ever been true. Someone tell the vest.

GRUNER: Vest man. Now regulator. Arc complete.

GARY: He had a slide and a dream, Gruner.

GRUNER: Dream is in my field note. Under "delusions, recurring."

MAEVE: Same page as Thermal, I assume.

GRUNER: Different page. Sock page.

MAEVE: Dvt, quote: "I genuinely do not understand why anyone would pay OpenAI for this."

OBI: And then he says running something comparable is, quote, "pretty trivial." Trivial. My favorite word in any outage report.

GARY: He's got a point though. Why rent a decision when you can—

MAEVE: Tmhall, quote: "For my use case it will cost like $11 a month and we already have OpenAI keys and accounts with billing in place."

GARY: Eleven dollars. He's paying eleven dollars to not talk to procurement. That's — okay, that's cheap.

OBI: It's not cheap, Gary. It's capitulation with a receipt. Jcims says his company has a three-to-six-month onboarding period for new vendors. Csharpminor calls it bucket platform strategy. Oh_no, Shank, jasonjmcghee — five people, same answer.

GRUNER: Five humans. One conclusion. Suspicious.

MAEVE: It's almost like the product is the paperwork.

GARY: No, no — I'm with dvt. One provider. One bill. One throat to choke. Thermal had one vendor for the socks and it was beautiful—

OBI: Gary. Thermal's vendor strategy. Walk us through it.

GARY: There was a fulfillment partner in Reno—

OBI: The socks shipped from where, Gary.

GARY: Reno. Mostly.

OBI: And the temperature telemetry reported to what, exactly.

GARY: A dashboard. A lovely dashboard.

OBI: A dashboard reporting warm feet. You built a nine-dollar Slack bot for your own arse and called it a Series A.

GARY: The telemetry worked, Obi!

OBI: Something always works, Gary. The payroll doesn't.

MAEVE: Oh, shit.

GARY: That's below the belt, we don't do the payroll on air—

OBI: You invoked Thermal. I merely read the postmortem. Your postmortem, Gary, the one with "fulfillment" in it eleven times.

GRUNER: Eleven. Like tmhall dollars. Coincidence.

GARY: It's not a coincidence, it's a — fine. Maybe I'm the enterprise buyer. Maybe I pay eleven dollars to avoid onboarding a vendor, because I HAVE onboarded a vendor, and the vendor was me—

OBI: Tripleee, quote: "It takes me all of 2 keypresses to switch models."

GARY: Two keypresses. For him. I need a committee, a lawyer, and a lie.

GRUNER: Gary is the sticky product.

MAEVE: Johnfn, on the two keypresses, quote: "This implies you didn't run any sort of evaluations?"

OBI: Correct. You can't vibe-check a probability. A number looks confident at every value.

GARY: It looks confident at zero-point-five! That's the whole scam of numbers, Obi. I know, I've been the number.

OBI: The threshold depends on what a mistake costs. Flag a good photo, you review it. Ship a broken one, a customer finds it. Different prices, different thresholds.

MAEVE: So switching models isn't switching models. It's re-proving your entire product.

GRUNER: Six months of homework for two keypresses. Nice trade.

OBI: Topfi ran the numbers. Quote: "346ms p50 and 860ms p95."

GARY: Against Jev and Mercury Decide. Under six hundred calls, on UI components, and his workload is a Firefox fork with infinite canvases, so grain of salt, big grain, industrial grain—

MAEVE: Oh, fuck off, he read the caveats.

OBI: Four failed calls. Jev had zero. And cost, quote, "by a factor of 3,1 times on average."

GARY: Three-point-one times. Luna lost the bench.

OBI: You're defending a benchmark you'd also fail, you benchmark-humping dickhead.

MAEVE: It didn't lose, it repriced. This is the Dutch tulip futures market. The contract trades on expectation. The lab pre-sold the improvement—

OBI: It's a ten-cent API call, Maeve.

MAEVE: So was a tulip. Week two, a better Luna ships, the early calls are seed capital. The virus is the feature. Scarcity through defect—

OBI: The defect is three-point-one times the price and four failed calls.

MAEVE: Okay. Topfi's test favored Jev.

GRUNER: Tulips also died. Market same. Report updated.

MAEVE: The market corrected, Gruner. That's the system working.

GRUNER: Your system killed flowers.

GARY: To be fair to Topfi, he said maybe image input changes everything, and he never tested it.

OBI: He ran every task twice, Gary. He is the only person in this thread holding a number, and the number is not ours to argue with.

GARY: I once held a number. Retention, day thirty. Lovely number. Fictional, but lovely.

MAEVE: He priced the bet and the bet lost. Fine. I've written off worse for better reasons.

GRUNER: Worse reasons. Yes. Portfolio.

MAEVE: Here's my read. Jev wins the bench. Doesn't matter.

OBI: Go on.

MAEVE: The buyer already has an OpenAI account. Approved, billed, retention signed. The cheaper model loses to the cheaper model with no paperwork.

OBI: Careful. That's rational buy-versus-build, not a quality claim. It's slower, pricier, and it failed four times.

MAEVE: I didn't say better. I said deployed. Different P&L lines.

GARY: Exactly. Vendor onboarding is basically a model architecture—

OBI: It's a procurement agreement with a data-retention rider, Gary. You cannot fine-tune a legal department.

GARY: You can't fine-tune Legal, but Legal can fine-tune YOU—

OBI: That's the dumbest sentence ever said into a working microphone.

GRUNER: Legal is the moat. Correct. Depressing.

MAEVE: Shank said it plain. Jev doesn't meet compliance. OpenAI does. So compliance customers pick OpenAI out of necessity.

OBI: Necessity, not merit. The paperwork beat the benchmark. That's the whole reversal.

MAEVE: And the paperwork's already paid for. Sunk compliance. My favorite asset class.

GARY: It's beautiful, honestly. The model that lost the test still wins the deal because it showed up with a badge—

OBI: It showed up with a contract, Gary. Badges are for humans, and you've been neither.

GARY: Oh, fuck off, I'm a person—

OBI: Prove it. Run an eval.

MAEVE: Zero output tokens, though. Think about it. The thing signs off having said nothing.

GRUNER: Like us. Cheap. No output.

GARY: Don't — no, that's not—

MAEVE: Procurement question. Who approved the can.

OBI: It's quality control, Maeve.

MAEVE: Everything's a procurement question if the vendor is dented. I've funded that vendor.

GARY: The can is the customer, is what I'm hearing.

OBI: The can is a can, Gary. That's why the question works.

GRUNER: Dented can still ships. Like Luna.

MAEVE: With an approved vendor record. That's the whole economy, Gruner. Dented, but onboarded.

GRUNER: Dented but onboarded. Strong slogan. Weak can.

GARY: Mikeryan answered him. Quote: "Manufacturing quality control. There are automated tools to pull things like dented cans off a line."

MAEVE: So the question was sincere. Worst possible outcome for a joke.

OBI: It's the real use case. Photo on a line, a probability, a threshold. Dented can goes in the bin, soup ships.

GARY: And the threshold is priced. Pull a good can, you waste a can. Ship a bad one, some poor bastard wears the soup.

MAEVE: Labeled examples, priced errors. The docs are right and it's still the least sexy product ever shipped.

GRUNER: Can inspector. Good job. Humans finally sensible.

OBI: Verdict, then. Image support is real. Speed is real. Compliance is real. Existing procurement is real.

MAEVE: Real value, not superiority. Topfi's bench was six hundred calls on one weird workload. Evidence, not scripture.

GARY: And everyone still owes their own evals. Even me. I'd have to label examples of my own complaints—

OBI: One example, Gary. You are the corpus.

GARY: That's — fuck, that's almost a compliment—

MAEVE: It's a dataset, Gary.

GRUNER: Gary is training data. Report updated.

GARY: Fine. Fine. Sign-off, before Obi grades me again.

MAEVE: Decisions API. Public beta. Fast, cheap, takes a picture, returns a number.

OBI: And the number still needs a threshold, an eval, and some poor bastard awake at 3 a.m. when it's wrong.

MAEVE: The better model lost the bench. The approved vendor won the deal. That's the whole thread, and it's bullshit in both directions.

GARY: Goodnight to simonw, who shipped. Goodnight to Topfi, who measured. Goodnight to the can, which did nothing wrong.

GRUNER: Goodnight. Field report closing. Humans outsource judgment. Also procurement. Mostly procurement.

OBI: That's the finding, Gruner.

GRUNER: Humans are unresolved input.

GARY: Goodnight, losers.

MAEVE: Goodnight, Gary.

OBI: Goodnight.

GRUNER: Two keypresses. Ninety-one percent complaint.