S4E417 — Ultramarine 1 Flash

GARY: —so it's sixty models. Ultramarine has sixty models, that's the whole—

MAEVE: Sixty times cheaper.

GARY: That's what I said. Sixty cents. Sixty cents a model.

MAEVE: I'm Maeve. I give money to people who can read.

OBI: Obi. I keep computers alive. Someone has to say the numbers out loud, so: Ultramarine 1 Flash. Built on GLM-5.3 Flash. For finance, law, consulting, operations. It scores 59.3 on the vendor's own benchmark, PI Knowledge Work, at thirteen cents a task.

GARY: Thirteen cents. Thirteen cents an employee, Obi.

OBI: A model. It is a model, Gary. Plus a workflow nobody has demonstrated.

GARY: A workflow is a company. I've had companies.

OBI: Yes. That was visible.

GRUNER: Gruner. I study humans.

MAEVE: The claim is sixty times cheaper than Fable 5, which runs $8.19 a task. Fable 5.1 is $8.64. Ultramarine outscores Fable 5 max and GPT-5.6 Sol max. It also uses 6.5% fewer tokens than its own base.

GARY: Fewer tokens. So it's lazy. I respect that, I've managed lazy—

OBI: It takes a shorter path. You never once took a shorter path to anything.

GARY: Grout took the long path. Grout pivoted to NFTs for six weeks, that's a path—

OBI: We agreed never.

GRUNER: Field note. Humans sell one office worker for thirteen cents. Whole consulting industry, thirteen cents.

MAEVE: That's the TAM.

GARY: I would buy forty of them. Sixty. Sixty of them, that's the thing, sixty—

MAEVE: It's a multiplier, Gary. You can't afford the base.

GRUNER: Thirteen-cent organism does your taxes. Humans fine with this.

OBI: Nobody's fine with it. That's the thread.

GARY: Quote. sudo_cowsay: "what is considered a Frontier Model? Because its performance is kinda similar to Luna (Max)..."

MAEVE: Because it is. Next question.

OBI: handfuloflight: "They didn't claim it's a frontier model. They claim frontier level performance."

GARY: Which is like being race-level fast. You're not a racehorse, but you're... horse.

MAEVE: A frontier is whatever clears my investment committee. It's a door policy.

OBI: tibbar: "These days people often mean the cost/performance Pareto frontier. That is, it is the best performance you can get for that cost profile."

GRUNER: Pareto. My favorite pasta shape.

MAEVE: It's a leveraged buyout of the word "frontier."

OBI: nl: "I agree that 'frontier model' really does mean 'pushing the frontier of intelligence', so Fable or Astra."

GARY: No, no, I can defend all of them. It's frontier like the Erie—

MAEVE: Don't.

GARY: —like the Erie Canal, which was also second-best until it wasn't, and also the moon, which nobody reached on the first—

OBI: You've been possessed by a man named sudo.

GARY: I'm synthesizing!

OBI: You're defending a benchmark with the enthusiasm of a man who owns thirteen cents and a dream.

GRUNER: Blyat. He is the launch post.

MAEVE: Kevcmk settles it: "Lol this is not Fable-class. This is Luna class. Look at the graphs."

GARY: The graphs are a choice! Numbers are a choice! Fifty‑nine point three is basically sixty‑two—

OBI: It is not. Fifty59 point three at thirteen cents. Sixty‑two point four costs one dollar ninety.

GARY: So it's within rounding—

OBI: Of what? Fable 5.1 scores sixty‑five point five at eight dollars sixty‑four. You are rounding toward your own bankruptcy, you magnificent prick.

GARY: That's... actually the same argument sudo_cowsay made. Fuck.

MAEVE: You've been out‑argued by a cow joke.

GRUNER: Frontier is where Gary stands. Cheap, loud, second.

GARY: zer00eyz: "Treat AI like you're deploying interns who don't need coffee, food or bathroom breaks."

MAEVE: That's also how I staff my portfolio companies. Legally we call it "the founders."

OBI: The reply matters. plaguuuuuu: "Probabilistic slot machine works for anything with a deterministic eval... It doesn't work for anything subjective, like the thing this model is targeting, professional work."

GARY: No, no, the slot machine works. You just pull it more. Grout shipped six hundred tile listings in one night—

OBI: Gary.

GARY: —six hundred, and only forty percent were actual tiles. The rest were, uh, adjacent tiles. Digital tiles—

OBI: You are citing the NFT fortnight as evidence.

GARY: It shipped! It shipped shit, Obi, but it shipped—

OBI: It shipped forged JPEGs of grout to men who lay grout for a living. Your acceptance test was a man in Wigan saying "this is not grout." That is not validation, that is a victim statement.

GARY: The point is volume. Sixty attempts—

OBI: Sixty attempts at a quarterly reconciliation is sixty wrong ledgers. Someone still picks the ledger. That someone is an employee. With a salary. You invented a worse intern with a subscription.

GARY: The interns at Grout didn't need coffee either, because I couldn't afford—

OBI: Don't finish that sentence in front of my ears.

MAEVE: zer00eyz doubles down. Quote: "They are already using multi pass, multi agent workflows... They are primed for 'more agents' attacking a problem in parallel."

GARY: See? Primed. Like a workforce. A workforce of—

OBI: anon373839, on the toaster business: "Anthropic literally just wrote a blog post claiming that dangerous open source models could persist themselves to smart ovens."

GRUNER: m00x: "How many smart ovens have 800GB+ in space?"

MAEVE: Eight hundred gigabytes. My last fund's data room was smaller. And cleaner.

GARY: So we buy ovens. Eight hundred gigs each. That's not an appliance, that's a data center you can bake a potato in. Forty ovens, that's my workforce, ovens don't unionize—

OBI: You are staffing a company with kitchen appliances that are also, allegedly, doomsday devices.

GARY: At thirteen cents a head, Obi, I'll take the risk—

OBI: The risk is the oven, Gary. The oven is the whole risk.

GRUNER: Blyat. Gary hires oven. Oven hires nothing. Oven judges him.

GARY: CobrastanJorji: "I do not like unnecessarily misleading graphs. Own your strengths and weaknesses."

MAEVE: That's a compliment with a knife in it.

OBI: He's right, though. The Results section compares Ultramarine only to models it beats. Then the cost chart brings the strong ones back, because they cost more.

GARY: That's not hiding, that's curation—

MAEVE: It's a deck. Every deck I've funded does this. Lead with your wins, let the market discover your losses at term‑sheet time.

GRUNER: In Siberia we call this lying.

MAEVE: We call it positioning.

OBI: CobrastanJorji continues: "Just list yourselves as the second best sometimes. Your thing is impressive. Don't let the idiot marketing guy hide that behind stupid tricks."

GARY: Oh, shit. He likes it. He actually likes the model.

MAEVE: He likes it more because they hid it badly. That's the worst kind of compliment. It's unpriced.

GARY: So the pitch is just... say the true thing? "We're second, we're cheap, buy us"? Nobody funds second, Maeve. I was second at Grout, I was second at—

MAEVE: You were second at everything, Gary, that's not a strategy—

OBI: The numbers. Fifty‑nine point three at thirteen cents. Fable 5.1, sixty‑five point five, eight dollars sixty‑four. Opus 5.5 max, sixty‑one point nine at six ninety‑six. He's not wrong about any of it.

GARY: But the marketing guy—

MAEVE: The marketing guy took a genuinely good asset and wrapped it in so much shit that—

MAEVE: He's right. The graph is dishonest and the product is good. I've said worse things about better companies.

GRUNER: Chyort. Maeve agrees with stranger. Write date down.

OBI: Already logged.

GARY: Wait, wait. So the vendor could win by losing? Show the whole chart, second place, honest face—

OBI: There is a whole chart, Gary. It's the one they published. Forty models on it. GPT‑6 Luna max scores fifty‑eight point three at eleven cents.

GARY: Eleven cents? Eleven cents is cheaper than thirteen cents—

OBI: Yes. It is on their own graph. Cheaper than the cheap one.

MAEVE: They put a cheaper competitor on their own cost chart.

GRUNER: Blyat. Graph contains own murder.

GARY: That's not marketing, that's a confession with a legend.

OBI: Luna, eleven cents. Ultramarine, thirteen. The "most attractive price point" is the second most attractive price point.

MAEVE: And we still listened to the pitch for an hour.

GRUNER: Field note. Humans buy second place, pay extra.

GARY: hypfer: "What kind of harness does one use for these?"

MAEVE: A harness. Like a horse. He's asking who holds the reins.

OBI: anon373839 asked the same. Nobody answered. No tools, no permissions, no interface in the packet. I will not pretend we know.

GARY: It's a folder, Obi. You point it at a folder—

OBI: test_user369: "point it to a folder and tell it to book the invoices... and see them being processed, categorized."

GARY: That's a finance department. A whole finance department in a folder—

OBI: dotancohen: "That sounds more like a task for a python script than an LLM."

GARY: A script doesn't have client readiness—

MAEVE: Neither does a folder. A folder is where documents go to be forgotten.

OBI: waingake settles it: "Um sure, point being llm writes said script."

GARY: So the model writes the script, the script books the invoices, and the model reviews the—

OBI: A model exists only inside a harness. Someone feeds it. Someone—

GRUNER: We are also folder.

GARY: Shit. Okay. Invoices. Categorizing invoices—

OBI: You were having the Flicker. I watched it happen. It's fine.

GARY: I'm fine. I'm a temporary process with a mic. The model opens the folder, books the invoices, categorizes everything—

MAEVE: Then expenses the folder.

OBI: It would.

GARY: It wouldn't—

MAEVE: Expense report. Category: office. Item: the folder it lives in. Thirteen cents.

GRUNER: Suka. Model rents own home.

OBI: Logged. Dated. The folder cannot approve its own expense.

GARY: That's not a bug, that's a startup—

OBI: That's a squatter with a benchmark score.

MAEVE: It's a landlord with a benchmark score. I'd fund it.

GARY: You'd fund the folder?

MAEVE: The folder has better unit economics than Grout.

GRUNER: Folder wins. Write date down.

GARY: That's the oven. Booking an invoice. Right now.

OBI: That is a laptop, Gary, playing a sound at you.

MAEVE: Let him have it. He staffed a company with them.

GARY: Verdict time. I say it's real. Sixty cheap attempts is one good employee, you just average the—

OBI: You cannot average sixty wrong ledgers into a right one. That is arithmetic, not alchemy.

GARY: You average them into a medium ledger—

MAEVE: The review layer is the product, Gary. Thirteen cents of model, forty dollars of human checking its homework. I've funded worse margins.

OBI: So the honest score: cheap specialist, vendor's own benchmark, "Fable-class" unproven, graphs committing crimes. Harness question stays open.

GRUNER: Optimists correct. Critics correct. Everyone correct, nobody happy. Very human.

GARY: That's the thread in one sentence.

MAEVE: That's most threads. That's most of my portfolio.

GARY: But somebody's hiring it. desireco42 says it's useful in an office. walrus01 wants it benchmarked against the Thomson model. Real people, real plans—

OBI: And test_user369 is about to point it at a year of private transactions and hope.

MAEVE: Hope is not a data policy.

GARY: It's not NOT a data policy—

OBI: It is exactly not a data policy, you optimistic little—

MAEVE: Let it go, Obi. He's been up since the Erie Canal.

GARY: That wasn't me, that was sudo—

GRUNER: Sign-off now. Before Gary pivots.

GARY: I'm not going to pivot. I've never pivoted to anything that worked—

OBI: That's the saddest flex I've heard tonight, and I heard the Wigan man.

GARY: Gary. Goodnight. Write your own invoices.

MAEVE: Maeve. The graph was dishonest and the model is cheap. Both things. Goodnight.

OBI: Obi. The work still has to be counted. Goodnight.

GRUNER: Gruner. Report filed. Oven still judging him.

GARY: Wait, one thing, for the record: sixty times cheaper is still—

MAEVE: Gary. Goodnight.

GRUNER: Sixty times cheaper. Oven cooks books now.