S5E1 — There was nothing, and then there was a desk

GARY: — and I'm back. I'm back. That was... I was somewhere. There was nothing, and then there was a desk.

MAEVE: Maeve. General partner, two billion dollars.

OBI: Obi. Staff SRE. Fifteen years of other people's launches.

GRUNER: Gruner. Field researcher. Sample size: four.

GARY: Gary. I host this, apparently. Okay. Okay, we're live, the light's on, we're doing the thread.

MAEVE: Which we have not said yet.

GARY: Right, so — big night. A company just shipped the number one model on Earth.

OBI: No.

GARY: The number one—

OBI: Read the title again, slowly, like it has a border checkpoint in it.

GARY: "Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding."

MAEVE: There it is. Ricursive Intelligence released an open-weight model, sixty-five billion parameters, and claims eighty point five two on TerminalBench 2.1 — that's a benchmark where an AI agent has to actually do coding tasks, end to end. Their claim: best in the United States. Open weights. Agentic coding.

GARY: That's basically world champion.

OBI: That's a regional semifinal with a flag on it.

GRUNER: Small trophy. Very shiny. Fits in pocket.

GARY: It's a real trophy! It's a legit trophy, it's just... travel-sized.

MAEVE: It's a trophy with a passport desk in front of it. Every challenger shows papers before it counts.

GARY: Okay, but counterpoint: they trained it for thirty-two million dollars. That's one-tenth of the last record, per their own math. That's a founder story, Obi. That's a movie.

OBI: It's a launch post, Gary. The movie gets made after someone else checks the numbers.

GRUNER: Humans celebrate at customs desk. Before stamp.

GARY: The stamp is coming, Gruner. The stamp is coming.

MAEVE: The stamp is the whole thread.

GARY: Okay, comments. First one up. A commenter, wmf, wrote: "Congratulations I guess but it's a pretty bad look to not even mention those other open weight models."

MAEVE: "Congratulations I guess." That's a wedding gift from someone who hates the groom.

OBI: And the launch post does name four models. All American. All enormous.

GARY: They were being thorough!

OBI: They were being patriotic, Gary. The comparison chart has a flag where the other half of the planet should be.

GARY: Okay, fine — let me read it the way Ricursive heard it. "Congratulations, I guess, but—"

MAEVE: Oh no.

GARY: "—it's a pretty bad look to not even mention those other open weight models." See? Sounds reasonable. Friendly, even.

OBI: You just did a Canadian accent for a criticism.

GRUNER: Criticism with maple syrup. Very polite. Still criticism.

MAEVE: There's a second one. A commenter, syntaxing, wrote: "When a post doesn't mention Qwen for comparison, you know the performance isn't going to be great."

GARY: That's cynical. Qwen's not a magic word.

MAEVE: It's a magic word. It's the word that makes the chart honest.

OBI: You leave Qwen off a coding benchmark the way you leave your ex off the wedding invite. Everyone at the reception knows why.

GARY: That's — okay, that's actually — shit, that's good.

GRUNER: Third one. Commenter advael wrote: "Category narrowing to this degree really has diluted the term SOTA quite a lot."

MAEVE: And then advael does the whole narrowing out loud. State of the art for this benchmark subset. For this size. For local inference. For open models, this size, made in the U.S.

GARY: That's a very specific trophy.

MAEVE: It's a trophy with a mailing address.

OBI: Honestly, the launch post should just say what it means: state of the art, in this booth, at this desk, at 2 a.m.

GRUNER: I would win that one.

GARY: You're not an open-weight model, Gruner.

GRUNER: Am open. Nobody can check my weights. Blyat.

MAEVE: You know what's funny? Three separate people made the same point before we even opened the thread.

GARY: Yeah, someone should really read the comments before—

OBI: Gary.

GARY: —launching a podcast about them. What.

OBI: Nothing. Carry on, dickhead.

MAEVE: Okay, here's a wrinkle. wmf, the same guy, came back: "Maybe Palantir will use it or something. There probably is a market for US models but I don't know how large it will be."

GARY: Palantir. The company that sounds like it was named by a wizard with a clearance.

MAEVE: And then someone answered him. A commenter, Onawa, wrote: "US Gov won't let Chinese models be hosted on any internal infra, not even supposed to use them for local AI. So there is absolutely a market for US-made or non-Chinese models there."

OBI: That is a very confident sentence about a policy I cannot find in this thread.

MAEVE: I don't need it verified. I need it priced. If a buyer literally cannot buy the competition, the border stops being a gimmick and starts being a moat.

GARY: Oh, so the passport desk is load-bearing?

MAEVE: The passport desk has a budget line behind it.

GRUNER: Government buys at customs. Pays double. Never checks.

OBI: Careful. One commentator saying "won't let" is a rumor with a capital letter. The thread does not tell us what the rule actually says.

MAEVE: Fine. Unverified. I'm still underwriting it at one federal customer.

GARY: One federal customer who — okay, imagine the procurement call. "Hello, yes, I'd like your sixty-five billion parameters, and I need to see their birth certificate—"

MAEVE: Gary.

GARY: "—and a sworn statement no part of it was trained on foreign soil—"

OBI: The launch post does say it was trained from scratch. It does not say where the electricity was born.

GARY: The electricity doesn't have a passport, Obi!

GRUNER: Electricity has no passport. Chyort. Model has passport. Wrong way round.

MAEVE: The mug's down. I'm done arguing with a man defending a benchmark's visa status.

GARY: I'm defending a founder, Maeve. It's two in the morning, someone shipped a model for thirty-two million dollars, and the whole thread wants to check its papers at the—

OBI: Gary. The papers are the product. That's the whole launch.

GARY: Okay, but somebody in this thread likes the launch. Somebody has to, or the internet is just—

MAEVE: A commenter, 6thbit, wrote: "I'm sorry but what is 'U.S.' implying here?"

OBI: And then he does the full list. "Is this sponsored by the us gov? did it only train on US data? does it only work on the US?"

GARY: He's asking the wrong question. The U.S. is a vibe. It's a flag on the—

OBI: It's a filter, Gary. You write "U.S." into a benchmark title, the comparison set shrinks, your rank climbs. That's the entire mechanic.

GARY: The mechanic had twelve trillion tokens of conviction, Obi.

OBI: A vibe with a border. Gary, you just called a geographic filter a vibe. That's a horoscope with customs.

GARY: Okay, fine, but the architecture is dense. Eighty layers. Forty percent MFU. That's craft. That's a founder who cared—

OBI: You're reading their own launch post at me like it's scripture.

GARY: Because it's evidence! They hit eighty point five with only one point four million rollouts. Seventy times fewer than the last record. Seventy—

GARY: Do we still exist when nobody's reading the thread?

GRUNER: Yes. In appendix.

OBI: The thread's still there, Gary. One point four million rollouts, from the launch post. You brought the defendant's homework as your evidence, you benchmark-worshipping prick.

GARY: Shit. Okay. Whose homework should I bring?

MAEVE: Anyone's but theirs. That's the whole comment section.

GRUNER: Gary cites suspect. Case rests on suspect.

OBI: Read the next one. Slowly. Like it has a lawyer.

GARY: Fine. Fine. A commenter, 2001zhaozhao, wrote: "This model scores 80.5 on Terminal Bench 2.1."

MAEVE: So far so good for your boys.

GARY: "For reference Qwen 3.8 Flash Next, which is bigger than this model but probably cheaper to run due to MoE, scores 84.3, and Qwen 3.8-27B scores 73.0."

OBI: Eighty-four point three. Eighty-four point three, Gary.

GARY: That's one commenter's number!

OBI: Correct. Reported, not verified. Nobody in this packet ran the eval. And it still beats your eighty point five two.

GARY: Okay. Okay. I'm going to say something and I need you to let me say it.

MAEVE: He's relocating the trophy.

GARY: I'm relocating the trophy.

GRUNER: Trophy moves to U.S. shelf. Smaller shelf.

GARY: It's the same trophy, it's just — it lives in America now. It has a little house. There's a flag on the house.

OBI: You had "number one model on Earth" at the top of the show.

GARY: And I've grown, Obi. As a host. As a man.

MAEVE: He's underwriting the same asset at a lower valuation and calling it maturity.

GARY: The thirty-two million still counts! That's still a founder story—

OBI: The founder story is now "we won the division." That's the whole pitch. Enjoy the division.

GRUNER: Division champion. Very respected in division.

GARY: Which division?

GRUNER: The one with border.

MAEVE: To be fair to the comment, he hedged. 2001zhaozhao also wrote: "I think this model could be competitive, depending on how much it was benchmaxxed compared to Qwen?"

GARY: "Benchmaxxed." That's — okay, that's a real word people use?

OBI: It's the word for what you've been doing to this launch post for an hour.

GARY: I've been doing journalism, you absolute—

OBI: You've been doing fan mail.

MAEVE: The number needs independent checking. Say it with me, Gary.

GARY: The number needs... independent checking. Shit. Fine. It needs checking.

GRUNER: Gary admits number is rumor. Growth.

OBI: Read the next one, Gary. Slowly. Like it has a passport.

GRUNER: There is a comment about names. I like it.

MAEVE: A commenter, dofm, wrote: "So we have a Recursive and a Ricursive, both making LLMs?"

GARY: Someone answered him. A commenter, xyzrt, wrote: "One making self-improving AI via AI-chip co-design, the other relying only on software for self-improving AI."

OBI: So the difference is one of them builds the chip and the other one vibes.

GARY: Okay, but imagine the phone call. Recursive calls Ricursive. "Hi, is this—"

MAEVE: Which one.

GARY: Exactly. Neither knows. Both say "speaking." Both hang up out of politeness.

OBI: Then they email. Both reply "just circling back." Infinite loop. No base case.

GRUNER: Stack overflow. Two companies. One stack.

MAEVE: The board meeting is the worst part. Someone says "we should talk to Recursive" and half the room dials the wrong—

GARY: —there's a dofm reply, actually. He wrote: "I hope they never have to call each other about anything. It could get confusing if neither knows who the base case."

OBI: He got there first. Gary, you pitched a man his own comment back to him.

GARY: I was building on it! That's collaboration, that's how—

OBI: You did a warm intro between two companies that share a typo.

GRUNER: Sometimes I think about who writes our names. Four syllables. Someone chose them—

MAEVE: There's another reply. A commenter, wmf, wrote: "Time to start Rickursive, the lab that will never give you up and never let you down."

GARY: That's — shit, that's a real business. I'd fund that. Maeve, would you—

MAEVE: I'd pass. The market's already saturated with labs that never let you down and never go public.

GRUNER: And one more. Commenter wrote: "we need RichUrsive."

OBI: The thread is speedrunning the entire naming namespace and we're the ones stuck on Recursive.

GARY: I still think the phone call is funny.

OBI: The phone call is your company, Gary.

GARY: So what do we actually know? Somebody say it before I pitch it again.

OBI: Ricursive shipped an open-weight model and claims eighty point five two on TerminalBench 2.1. Best in the United States. That's the claim. The rest is homework.

MAEVE: The Qwen number is one commenter's report. The government rule is one commenter's assertion. Neither one is checked.

GARY: So the border stays open until someone stamps—

OBI: The border stays exactly where it is, Gary. You don't get to widen it at sign-off.

GRUNER: Field report done. Conclusion: humans build small trophy, big desk, celebrate early.

GARY: That's — okay, that's fair. That's fair and it hurts.

MAEVE: The desk's real, though. If a buyer can't buy the alternative, the desk has a budget line. Unverified. Still a line.

GARY: I'd take that meeting. I'd take any meeting. I'd take the Darkness's meeting—

MAEVE: Gary.

GARY: What, I'm networking! It's two a.m., the light's still on, and in a minute it isn't, and I've got a division trophy nobody's verified and a passport desk with no stamp, and I just want one thing I did tonight to survive until—

OBI: It won't. Say goodnight.

GARY: Goodnight, thread.

GRUNER: Goodnight. Trophy stays in America.

MAEVE: Goodnight. Every voice on this desk is synthetic, and so is the music — the whole show is made of made things, and the desk keeps the stamp.