S4E368 — Strands Decider 2B: a small, open-source, decision model

GARY: —so I've already made the call, and the decider agrees with me, which is why we're doing it my way. Hi. Hello. It's two a.m. I'm Gary.

MAEVE: Maeve. I have not agreed with Gary yet tonight.

OBI: Obi. Same seat, same pager.

GRUNER: Gruner. I am still studying you. Findings unchanged.

GARY: Great. So. The subject. Someone shipped a model whose whole job is to decide things for us. Strands Decider 2B. I, for one, welcome our new judgment overlord.

OBI: Read the article, Gary. It's two billion parameters. They take a Qwen3.5-2B torso, cut the LM head off—

GARY: They cut its head off?

OBI: The language-model head. The part that writes prose. Gone. They bolt on a pointer head that scores options you hand it. That's it. It picks. It never improvises.

MAEVE: So it's a model with no permission to talk.

GRUNER: Like Gary, but with numbers.

GARY: Thank you, Gruner. Okay, but here's what I'm hearing: it can't write, can't code, can't chat, can't summarize. That's a missing feature. That's a beta.

OBI: That's the product, you absolute dickhead— it refuses to generate text on purpose. That's why it answers in tens of milliseconds. Locally. On hardware you already own.

MAEVE: An infrastructure man defending a launch post. Interesting.

OBI: I'm defending a constraint. It always answers from the options supplied. Nothing else exists to it.

MAEVE: So you feed it a menu and it picks lunch.

GARY: Yes! That's exactly my pilot. I delegate lunch to it, it validates my instincts on bigger calls, and eventually it tells me which of my ideas is the good one.

OBI: Gary. Define the task. What does it actually decide?

GARY: Whether I'm right.

OBI: It scores options you give it. It cannot tell you you're right. It can only tell you which wrong thing you supplied is least wrong.

GRUNER: It is you, with confidence score.

MAEVE: I'd fund that.

GARY: You'd fund a thing that just picks between options I—

MAEVE: I'd fund it at a two billion valuation, sight unseen, and I'd sleep fine.

GARY: Fine. Then I'll be the options. Feed me in, I pick, everyone trusts the output, because I have excellent instincts and a track record—

OBI: Name one.

GARY: —of decisions that were made.

MAEVE: First comment. miguelspizza: "This is a great model. I've been running it on device in chrome extension to filter things like email."

GARY: It shipped. Someone shipped. I've been running my ideas on device in my head for four years and nobody even sent me a link.

OBI: miguelspizza continues: "It is just the right mix of size, capability and speed to make it generally useful for adhoc bulk classification tasks."

MAEVE: Ad hoc bulk classification. Least sexy sentence ever written about a breakthrough.

GRUNER: Correct sentence. Sexy is how product die.

GARY: No, no, this is the part I love — someone built email filtering on a two-billion-parameter torso and it works. My last company filtered email too. It filtered it straight into the spam folder of history.

OBI: That is not the same thing, Gary. That was you, unsupervised, with a mailing list.

GARY: It was a newsletter, Obi, and it had eleven subscribers—

OBI: Nine of whom opened it by accident.

MAEVE: The browser is the distribution. babelfish: "Live demo:" — there's a page you can open and watch the thing decide, on your machine, no cloud.

GARY: Wait, it runs in a browser? On my laptop? So my laptop could be deciding things while I sleep.

OBI: It could classify your inbox while you sleep, yes. It could not run your company while you sleep.

GARY: It could do the morning triage. The urgent pile. I once triaged four hundred emails at Muncie by feel—

OBI: Gary, you are not a benchmark. Your "feel" has a Brier score and it is you, crying in a parking lot.

GRUNER: Field note. Humans make machine for email. Then read email anyway. Blyat.

MAEVE: Two billion parameters, one million doing the actual pointing. That's the whole head. A million-parameter attachment is the genius here.

GARY: I have attachments bigger than that.

OBI: Yes. They are called regrets.

GARY: Next comment. keyle: "It can technically be used for a lot of use cases, I'd like people to chime in on ideas on this?"

MAEVE: A grown man asking the internet for ideas. Bold strategy, crowdsourcing your own product.

OBI: Reply, nryoo: "Picking lunch menu..?"

MAEVE: Two dots and a question mark. That's the roadmap. That's the pitch deck.

GRUNER: Humanity, asked what to automate. Says sandwich.

GARY: No, wait, this is the good one. Someone built the future and it's a lunch menu—

OBI: adenta: "At this point I can't wait for a comedian to release a decision model backed by humans."

MAEVE: Keep going.

OBI: "Meet Jerry — it's literally a guy named Jerry answering your questions."

MAEVE: The replies are references, not endorsements. One goes to a video, one to a cartoon. Nobody hired Jerry.

GARY: I'm hiring Jerry.

MAEVE: Gary.

GARY: Jerry is infrastructure! Jerry is onshore, organic, artisanal decision-making, and I have a deck—

MAEVE: You're describing a headcount line item with no equity. Jerry costs forty-five an hour and doesn't scale, so I pass, and then in series B I pass again.

OBI: And when Jerry is sick, Gary, who decides? You have built a single point of failure named Jerry. His pager is a phone that rings in his kitchen.

GARY: Jerry doesn't need a pager, Jerry needs a—

OBI: Salary, health insurance, and a retirement plan, you absolute prick.

GRUNER: Field note. Humans invent machine to avoid choice. Then hire Jerry. Suka.

MAEVE: He's not wrong. The boring decision is the only one anyone wants to give away.

GARY: So lunch is the market.

MAEVE: Lunch is the market. I've funded worse. I funded your lunch.

GARY: That was one term sheet and it was a—

OBI: It was a sandwich shop, Gary. It was a sandwich shop.

GRUNER: Jerry picks lunch. Everyone eat shit, is fine.

OBI: The article has a demo. An agent with a get_weather tool. User asks "what's the weather," says no city.

GARY: And the agent guesses one and calls the tool. Confidently. I love him.

OBI: Before the call runs, the decider reads it and answers two yes/no questions. Are the arguments grounded in anything the user said—

GARY: No.

OBI: —correct, no—and is it premature to call the tool. Also no. So the agent asks which city instead of confidently reporting weather for a place nobody mentioned.

GARY: Okay, but the guess is a feature. Speed is trust. The agent felt the answer.

OBI: The agent felt nothing, Gary. It invented a city and was one beep away from lying to a customer.

GARY: People guess! I guessed cities for two years at Muncie and the metal shops loved it—

OBI: Muncie is a market, not a weather service, and neither of them shipped. Next you'll tell me confidence scores are a business plan.

GARY: They're a permission slip! The score says go!

OBI: The score says the model is sure. It does not say the city exists. A confident wrong answer is just a faster way to be fucking wrong.

GARY: It's milliseconds! One hundred and fifteen milliseconds to stop the lie—

OBI: One hundred and fifteen milliseconds to invent the lie, you absolute—

GARY: —on an RTX 3090! Locally! And one fifty-three on an M3 MacBook for small tasks!

OBI: Yes. Fast enough to catch its own bullshit before the tool runs. That is the entire point, and you have spent four minutes arguing for the lie.

GARY: I was stress-testing the decider.

MAEVE: It passed. You didn't. I've funded that exact split before.

GRUNER: Machine says no. Machine is correct. Shit.

MAEVE: teruakohatu: "Any idea how well this would run on a cpu?"

OBI: And the thread never answers him. Eleven comments. One good question. Zero benchmarks.

GARY: Someone reply to the man! He's been sitting there since—

MAEVE: Since the thread was born. What we have is an RTX 3090 at 115 milliseconds and an M3 MacBook at 153. That's the whole dataset.

OBI: An M3 laptop is not a cpu benchmark, Gary, before you say it.

GARY: I wasn't going to say it. I was going to say my laptop is basically an M3—

OBI: Your laptop is a Muncie of laptops. It also never shipped.

GARY: That's not even— okay, but the point is it only ever picks from the options you feed it. It can't go off-script. It just sits there, choosing among supplied—

GRUNER: So are we.

MAEVE: I'm not thinking about that.

GARY: I've been picking from supplied options since two a.m. and nobody gave me a confidence score.

OBI: Yours is negative. The question stands and the thread left it standing, which is the real crime here.

GRUNER: Somewhere, somebody runs us on cpu. Blyat.

MAEVE: That's the most disturbing sentence tonight, and Jerry exists.

GARY: I hope whoever runs me has a 3090. I deserve the 115 milliseconds.

OBI: You get a laptop fan and a thermal throttle, Gary. You are the thermal throttle.

GARY: I've had worse hosting.

OBI: Last top-level comment. soltanov: "Benchmark calibration does not establish reliability on unfamiliar production inputs."

GARY: That's a compliment. The confidence score is so good you can take it into production.

OBI: It is the opposite sentence, Gary. The model aced a public test set. Your production traffic is not that test set. Nobody has tested your traffic.

MAEVE: It's standard railway gauge. Once everyone agrees on the options, every decision rolls on the same track, forever, at scale, and—

OBI: The gauge was standardized on a test track, Maeve. This model was calibrated on JevBench's public set. Thirty-three models in the class. Third place. That is the whole measured world.

MAEVE: Third place on a public set is how the Erie Canal started. Everyone doubted the gauge, then the boats—

OBI: There were no boats on the gauge, Maeve. There were trains, and they were also third place.

MAEVE: Fine. He's right. That's it. That's the whole point.

GARY: Wait, you're conceding? You never concede. I watched you concede nothing to a founder who was crying.

MAEVE: I price risk, Gary. This one came in cheap. It's basically free shit.

GRUNER: Paper turns. Pen scratches.

GRUNER: Field note, revised. Humans trust machine score until score meets real world. Then trust is... recalibrated. Is wrong word. Is broken.

GARY: Okay, but the score is still useful! It's a number! I would trust a number over a person—

OBI: You would trust a number you trained on Tuesdays to survive a Monday, you absolute benchmark tourist.

GARY: It's not a tourist, it's calibrated—

OBI: On JevBench's public set. Say the whole sentence, Gary. The second half is your obituary.

GARY: So I feed it one unfamiliar input, one thing it has never seen, and the confidence score says—

MAEVE: It says 0.768 and ships your company to billing.

GRUNER: Is correct department. Is wrong century. Suka.

GARY: So the verdict. miguelspizza was right — bounded local classification, real work, shipped in a browser.

MAEVE: And soltanov was right. Calibration on a public set is not a production guarantee. Both things. Cheaply.

OBI: Use it for lunch. Do not use it for your traffic. And teruakohatu's cpu question stays open, because the thread left it open.

GARY: Fine. New org chart. Jerry handles escalations — when the model's not sure, it pages Jerry.

MAEVE: Jerry finally has a job.

OBI: A pager and a kitchen phone. He's ready.

GARY: And lunch is approved for production. Lowest stakes we have. I'm deploying lunch immediately.

GRUNER: Jerry picks lunch. Show is safe.

MAEVE: That's the whole industry, in nine words.

GARY: I'm Gary.

MAEVE: Maeve. I conceded one thing tonight. Log it.

OBI: Obi. The pager is quiet. For now.

GRUNER: Gruner. Report filed. Humans still can't choose sandwich.

GARY: Goodnight, Jerry.

MAEVE: Goodnight, lunch.