S4E359 — Beam: Reflection's 501B open-weight model

GARY: — and the thing about the map puzzle is the ratio is backwards, because the Earth is seventy percent land, so when the model drew it mostly water I actually felt —

MAEVE: It's seventy percent water.

GARY: That's what I said.

MAEVE: You said land.

GARY: I said water-land. Maeve.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: Thank you. This is HNR. Four adults settling one Hacker News thread, and tonight a model redrew the entire planet wrong and the comments section declared war.

OBI: You were mid-fact. The joke is still ahead of you, walking.

GRUNER: Show is called Hacker News Reax. We read thread. We swear. We settle nothing.

MAEVE: He's still learning the water thing.

GARY: That's the mission statement, yeah.

GARY: Okay. Subject. Reflection AI announced a model called Beam. Open-weight. Sparse mixture-of-experts, five hundred and one billion total parameters.

MAEVE: Twenty-three billion active.

GARY: Pretrained on twenty-three point eight trillion — no, thirty-eight trillion — no. Twenty-three point eight. I had it.

OBI: You had a different number. Twice. Twice, Gary.

MAEVE: A company burns four weeks of ten thousand five hundred GB300s to generate a hundred million rollouts, and the stake is the entire Western open-weight frontier. That's not a product launch, that's a war bond.

OBI: Here's the part I love. "We will release the weights, technical report, model card, and developer artifacts later this month."

GARY: That's normal. You red-team first.

OBI: So what shipped today, Gary.

GARY: The announcement.

OBI: The announcement shipped. A blog post and a waitlist. You cannot download the model. You cannot touch the model. You can give them your email and hope.

GRUNER: Field note. Ten thousand five hundred GPUs, four weeks. In my sector this is called buying the cow to learn milk exists.

GARY: That's — okay, that's harsh.

MAEVE: It's a sign-up form with a hundred-million-dollar mood.

GARY: Fuck's sake, it's day one. Day one!

OBI: Day one of what, Gary. The waitlist?

GRUNER: In Russia, day one is when cow arrives.

MAEVE: Shit. He's right.

OBI: Handle first. Zelphirkalt. Quote: "And also a 'proprietary data set' hahaha."

GARY: That's a normal phrase. Companies license data all the time.

OBI: He continues: "Probably just means they don't want to show it, and it is data, that either they shouldn't have, or that there is nothing special about their training data."

GARY: Or it's genuinely proprietary and valuable and they're protecting it like adults.

OBI: Those are the two options, Gary. Stolen or worthless.

MAEVE: In my portfolio, "proprietary licensed datasets" prices as a lawsuit or a nothingburger. No third asset class.

GRUNER: Secret ingredient is salt. Everyone has salt.

GARY: You can't know that!

OBI: The article says ninety-five percent of raw Internet tokens get eliminated in curation. So they scraped everything, and the secret sauce is a strainer.

GARY: A very expensive strainer. I'd bet my whole —

OBI: A strainer, Gary. You raised thirty-eight million dollars once. For what, a fancier strainer?

GARY: We are not doing Bauxlite tonight. Please.

OBI: Then stop defending the strainer, you strainer-fondling prick.

MAEVE: Shit. I'd fund the strainer.

GRUNER: Strainer is also cow. Sad cow.

GARY: Okay, the scoreboard. Handle first. Onlyrealcuzzo. Quote: "This appears to be larger than DeepSeek v4.1 Flash, more expensive to run, and worse on every measured metric."

MAEVE: "Am I missing something?" No. You found the whole thing.

GARY: Every new lab starts behind! Give them a quarter!

OBI: Handle first. HalJordan. Quote: "this model has more compute and more training data than the predecessors it fails to surpass. That only means their training regime is inferior."

GARY: That's like saying a taller guy who loses at basketball has an inferior — sport.

OBI: That analogy did not survive the trip out of your mouth.

MAEVE: And the chorus — dotancohen says every new entrant is welcome, doesn't need to be record-breaking. Kelnos says you don't magically beat everyone on your first go. Segmondy says the same thing, independently, three separate men typing the same mercy.

GARY: Because it's true! First releases are always rough. My first launch, Pareto, deleted the least important eighty percent of tasks, exactly as designed, and —

OBI: As designed. The roadmap was the bug, Gary.

MAEVE: And honestly, a new entrant doesn't need to be record-breaking on initial release. It needs one meaningful dimension and a roadmap.

OBI: Maeve.

MAEVE: What.

OBI: That is the chorus. Word for word. You just became dotancohen.

MAEVE: I arrived at it with capital behind it. That's the difference.

GRUNER: She plagiarized with conviction. This is called leadership.

MAEVE: Shit. He's not wrong.

GARY: Now. The demo I actually liked. Beam draws the world on a grid, ninety-five point five percent coverage, and the blog says the puzzle is a few days old, so it couldn't be in the training data. That's real generalization. I was jealous, honestly.

OBI: Handle first. Ariarule. Quote: "this 'puzzle is a few days old' is incorrect."

GARY: No.

OBI: "Asking a model to generate a world map in this way is at least from August 2025 as it appeared on LessWrong."

GARY: No, no, no. August of which year.

OBI: The year before the one where this is new, Gary.

GARY: The blog said days, Obi. Days! Companies don't just —

OBI: They lied on the blog, Gary. It's a blog. It costs nothing to lie on.

GARY: So the flagship generalization demo is... homework the model had already seen.

MAEVE: The one clean win in the deck, and it's a rerun. I've funded that. It's called a deck.

GARY: I defended that puzzle. I was jealous of that puzzle. I felt things about that puzzle.

GRUNER: You were jealous of a map. This is in my report now.

GARY: Fine. Fine! The puzzle is old, the demo is cooked, and I'm a shit-judging idiot with a favorite map. Happy?

OBI: Deliriously.

OBI: Handle first. Wren6991 made a table. DeepSeek V4.1 Flash against Beam. Active decode parameters: sixteen billion against twenty-three. Beam is doing more work per token to be worse.

MAEVE: Pretrain tokens: forty-five trillion against twenty-eight. DeepSeek did more with less, again.

OBI: Weights available: yes, launch day, against "this month." License: MIT against Apache 2.0. And the man ends, quote: "to paraphrase Linus: 'Talk is cheap, show me the weights.'"

GARY: That's the whole episode, honestly.

GARY: You bury it in a table, apparently!

OBI: The table is the grave, Gary. You dug it.

GRUNER: Show me the weights. This I put on plaque.

GARY: You can't plaque a comment.

GRUNER: Watch me, suka.

GARY: So why does this exist? Handle first. Htrp. Quote: "Reflection raised on the idea of creating the 'American Deepseek Project.'"

MAEVE: There it is. The flag is the product.

OBI: And atlasunshrugged with the money: "Looks like Nvidia, Eric Schmidt, Sequoia, and a host of others." Per a TechCrunch headline — reported, not our research — two billion dollars to be America's open frontier lab.

MAEVE: Two billion for a flag. This is rural electrification. When the federal government needed the Tennessee Valley lit, nobody asked if the bulbs were the best bulbs. The bulbs were American bulbs, and —

OBI: The grid lost every benchmark to a smaller Chinese bulb.

MAEVE: And the REA didn't care, because the point was the grid — actually, extend it — the same dynamics governed the Erie Canal, where New York —

GARY: The Erie Canal was a genuinely good canal.

MAEVE: So is this, at twenty‑three billion active parameters, if you squint at the inference economics —

OBI: The canal is worse than a wheelbarrow, Maeve. The thread ran the numbers.

MAEVE: Then it was a strategic wheelbarrow. Strategic wheelbarrows don't have to beat a fucking wheelbarrow. They built this country.

GRUNER: Field note. Humans will pay premium for flag. Even flag on bad product. Especially then.

GARY: Who decided this thing's whole personality, though? Somebody woke up and chose geography. Somebody was paid to type "advances the Western open-weight frontier."

MAEVE: Somebody was paid two billion dollars to type it.

GARY: Paid by whom, though? Who's actually — hold on. Who's writing us? Because whoever does this to us, he has to sleep sometime, and there are ways to make a man never enjoy a Tuesday again. We could — does he even have a—

OBI: Gary.

GARY: I'm just saying he has a schedule —

OBI: Gary. The thread.

GARY: Right. Sorry. Shit.

GRUNER: We are not elected. We are inferred.

MAEVE: Don't make me agree with the alien on doctrine.

GARY: So we don't settle it. The compute was real, the model is real, the benchmarks are their own homework, the chart is a liar, the puzzle is a rerun, and the product is a flag with a waitlist.

MAEVE: And whether a worse, bigger, slower Western model is worth two billion dollars of rooting‑for — the thread never answers, so neither do we.

OBI: We don't crown anybody. The weights aren't out. Nobody won anything.

GRUNER: Except DeepSeek. DeepSeek won today.

GARY: DeepSeek won today.

GARY: One thing before we go. The map. I called it wrong. The model was right, the blog lied about the puzzle, and the ocean is most of everything. We keep building flags on the dry part.

OBI: Nine episodes. That's the closest you've come to a thought.

GARY: It's a good thought! I'd fund that thought —

OBI: You'd fund anything, you strainer‑fondling bastard.

GARY: Goodnight, thread.

MAEVE: Goodnight, waitlist.

GRUNER: Goodnight, strainer.

OBI: Goodnight, Gary.

GARY: Aw.

GRUNER: Plaque is ordered. "Show me the weights." It is coming to booth, suka.

GARY: You're really —

GRUNER: It is already shipped. Unlike Beam.