S4E340 — Clef: Open-weight decision models, and new RL fine-tuning platform

GARY: — which is why the whole thing falls apart the second you look at the latency numbers, and I don't think Typesafe understands what they've — sorry. Sorry. Are we on? We're on. Okay.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: Thank you. What I was saying — before I was so generously assembled — is that Jev is in trouble. Except it's not Jev. It's Clef. Cloudflare shipped a decision model called Clef, someone put it on Hacker News, and the thread is already a knife fight. That's the show. That's — that's tonight.

MAEVE: You said Jev twice.

GARY: I said it once and then I corrected it, which is a different thing.

OBI: It is the same thing, Gary.

GRUNER: Is good thread. Many corpses.

GARY: Right. HN Radio. I'm Gary, that's the thread, and we're not starting over.

GARY: Okay. Subject beat. Somebody explain what a Clef is before I do, because last time I explained a name I told two hundred people Bauxlite was a mineral.

MAEVE: It's a music thing. The symbol at the start of the staff that assigns the pitches. Cloudflare picked it because it defines the domain and the notes that follow, and because CF.

GARY: See, that's what I was going to say, except I was going to say it defines the tempo, which—

OBI: It does not define the tempo, Gary. It assigns pitch names to the lines and spaces. You were about to tell forty thousand people a treble clef is a metronome.

GARY: I was going to say it's like a metronome. Which is a comparison. Comparisons are allowed.

OBI: Your comparisons cost thirty-eight million dollars.

GARY: That was one time.

MAEVE: Can we do the actual thing. Because the actual thing is enormous. The agent economy needs a cheap brainstem. Every agent on earth is currently paying a frontier model to decide whether a support ticket is urgent. That is a hundred-billion-dollar reflex arc running on a hundred-billion-dollar cortex, and Cloudflare just shipped the brainstem. That's the thesis. That's the whole market.

GARY: The brainstem. Okay. That's — hm.

OBI: Here is what it actually is. And I'm quoting the post, because the post is better than the marketing.

OBI: Clef and Clef-flash. Hosted on Workers AI. They take a payload, they return typed answers with probabilities — is this urgent, which team, what severity — and your code routes on it. No text generation. Bounded output. That part is real and it's good engineering.

GARY: And the flash one is the small one.

OBI: Clef is a frozen Qwen3.8-27B backbone. Clef-flash is Qwen3.5-9B. They froze the backbone, trained a routing head with rank-256 low-rank adapters, used label-smoothed cross-entropy with a Brier loss for calibration, and then a second objective they invented called RLCD — Reinforcement Learning for Calibrated Decisions. Partial credit for adjacent answers, a penalty to stop the distribution drifting.

MAEVE: He read that off the page like a man reading a parking ticket he intends to pay.

OBI: Because it's fine. The mechanics are fine. What is not fine is the wrapper. They froze someone else's model, bolted a head on it, and the blog post says the word "disrupt" four times. It's a classifier. It has been a classifier since before most of the people in this thread learned to walk. They put a tuxedo on it and called it a decision model.

GARY: They also shipped it in weeks. Typesafe had the whole market to themselves for — what, a month? Cloudflare saw it, built it, benchmarked it, and shipped it before Typesafe finished their victory lap. I have never in my life shipped anything in under — Muncie took two years to ship a landing page.

MAEVE: Muncie shipped a landing page?

GARY: The landing page shipped. That's not nothing.

OBI: It is nearly nothing.

GRUNER: Is field note. I read.

GRUNER: Cloudflare say own model classify website domain in 2.2 seconds. Their fastest general model, gpt-oss-120b — 4.7 seconds, same workflow. And general model return only two categories. Clef return more. Is on paper. Paper is where Cloudflare lives.

OBI: Filed under "on paper." We'll get to what the paper is worth.

GARY: Two-point-two seconds is — okay, real talk, I'm jealous. I'm jealous of the ship speed. I once spent nine months and a board meeting deciding between two shades of the same blue.

MAEVE: The blue was fine, Gary.

GARY: The blue was Muncie blue. It's gone, Maeve. All of it's gone.

OBI: And there's a line in the post nobody's quoting yet, about fine-tuning — you can train it on your own traffic and redeploy it. Keep that one. Somebody in this thread is going to make it matter.

GARY: Noted. So — real engineering in a tuxedo.

OBI: Real engineering. Tuxedo rented. From the marketing department.

MAEVE: An interesting thesis either way.

MAEVE: First comment. kerenskiy, verbatim: "The concept existed a year before Jev or so. See Laya."

GARY: A year. A whole year before.

MAEVE: And zitterbewegung: "You just have to fine tune an LLM like Qwen on some synthetic data to do so. There was even someone that had a model that was exactly like Typesafe and published their work a year before Jev — but wasn't marketed as heavily since it was academic."

GARY: So somebody published the whole thing. A year early. For free. In a paper nobody read. And Typesafe wrapped it in a landing page and made a category.

MAEVE: That is the entire business model of this industry, Gary.

GARY: No, no, hold on. This is — this is personal. If decision models were sitting there for a year, that easy, fine-tune Qwen, synthetic data, off to the races — why didn't I build one? Why didn't anyone at Muncie — we had two metal shops' worth of labeled decisions, we had quotes and lead times and rejection reasons, that's structured data, that's a classifier, that was a decision model before the word existed and we used it to send faxes.

OBI: Gary. You had the data and you sent faxes.

GARY: The fax was the product, Obi.

OBI: That is not a founder excuse, that is a confession. You're telling me the field left fruit on the tree for a year and you personally watched it rot while mailing vendors. You didn't miss the idea. You had the idea and a fax machine.

GARY: That's — hurtful, and also the fax had a confirmation sheet, which is more telemetry than Thermal ever —

OBI: No.

MAEVE: Let him have it. It's an interesting thesis. The interesting thesis is: the idea was never scarce. Attention was scarce. One company stopped, picked it up, and marketed the shit out of it, and now everyone else looks like they're copying.

GRUNER: Is field note. I read.

GRUNER: TeMPOraL write: "It's not a 'new paradigm', it's a low-hanging fruit that's been lying around for years; Typesafe were the first to bother to stop and pick it up, and market the shit out of it." Then more. He say many, many of those left around, because AI frontier moving forward so fast, everyone racing ahead. He say even if frontier models hit hard limit tomorrow, "there's a good decade of growth and advancement to be extracted just from all the low‑hanging fruits that were left unpicked along the way."

GARY: A decade of fruit.

MAEVE: A decade of fruit, sitting in the open, while the entire field sprints past it toward a bigger number. Which means the bottleneck was never intelligence. It was stopping.

OBI: The bottleneck is people who stop.

MAEVE: Careful, Obi. You're one rung from my theory.

OBI: I will step back off that rung.

MAEVE: calebkaiser, verbatim: "A lot of people who are a bit less familiar with the field saw the hype around Jev and presumed that the reason it was so exciting was that it was a fundamentally new interface for working with an LLM. And that additional excitement drove even more attention to Jev. But fundamentally, TypeSafe's announcement was that they found a particular architecture/training paradigm that resulted in a model for this particular interface that had incredible accuracy, very low latency, and for which they could offer inference at a super low cost."

GARY: So not invention.

MAEVE: Not invention. Execution. Calibration. The boring part, done unusually well. Which is worse for Typesafe, actually — you can defend a patent. You can't defend being out‑executed in three weeks.

OBI: Huh.

MAEVE: Yes.

OBI: I came in ready to call the whole category a rebrand. And the careful one is saying the rebrand was the least interesting thing they did.

MAEVE: That's the uncomfortable one. We were all ready to be right in the same direction.

GARY: I liked being right in the same direction. It was warm.

OBI: It was lazy, Gary. It was lazy and warm.

GRUNER: Table sit with being wrong. I write this down.

MAEVE: TeMPOraL said it and I'm not letting it go. "There's a good decade of growth and advancement to be extracted just from all the low-hanging fruits." That's not a comment, that's a thesis. And history agrees with him. Standardised railway gauge. Everyone knew a uniform gauge would unlock the whole network. Everyone. It sat there for decades while three companies ran three incompatible widths past the same freight yard, because nobody wanted to be the one who paid for the conversion.

OBI: People were paying for it. In transshipment. Every wagon unload was the conversion, they just paid it in labor instead of in a meeting.

MAEVE: Exactly, that's my point, the fruit wasn't unpicked, it was being eaten by middlemen. And this is the same. The frontier labs are the three gauges, and the fruit is sitting in the yard, and Jev just walked in and laid one standard rail and the whole industry is now re-gauging in real time, and in forty years nobody will remember the gauge wars, they'll remember who standardized —

GARY: The gauge wars.

MAEVE: The gauge wars of the agent economy, yes, and the second-order effect is that whoever owns the gauge owns the rolling stock forever, which means the fine-tuning platform is actually the gauge, the model is just the first train, and that reframes the entire announcement as a land grab dressed as a benchmark —

GRUNER: ...what?

MAEVE: You heard me. The gauge is the moat. The train is a commodity. This is the Zanzibar clove monopoly with GPUs, and I will die on this yard.

GARY: Or — counterpoint — it's not history. It's just founders being lazy. I know lazy. I ran a company that was ninety percent lazy by headcount. Grout. Tile contractors, b2b marketplace. The whole thesis was "contractors will adopt software." They would not. They had a guy. The guy had a truck. We could not compete with the truck. The low-hanging fruit in tiles was the truck, and nobody picked it because picking it meant becoming the truck, and nobody at that company, including me, wanted to be the truck.

OBI: Gary, the fruit isn't unpicked because founders are lazy. It's unpicked because picking it is on-call. Someone has to hold the pager for the thing that classifies a domain at two in the morning, and when it's wrong, a person gets a phishing site marked as fashion. That's what unpicked fruit actually is. It's fruit with a pager attached. Nobody races past it because it's hard. They race past it because owning it is a shit job and the race is more fun.

MAEVE: That's oddly moving.

OBI: It's a shift roster, Maeve.

GRUNER: I read field note. TeMPOraL also say, verbatim: "Even if sota llms hit a hard capability limit tomorrow and never advanced again, there's a good decade of growth and advancement to be extracted." Decade. He put number on it. Fruit with warranty.

GARY: He named a number. I respect that.

OBI: Back to the thread. There's a pricing comment and it's going to hurt someone at this table.

MAEVE: So on paper — and I want to be fair to the vendor here, which is a new feeling — Cloudflare's own chart is a massacre. Clef wins the Jev Decision Index. bfcl case exact, ninety-eight point four seven against Jev's ninety-five point seven five. Home appliances case exact, eighty-two point nine five against fifty-two. They beat Jev in three of four on Typesafe's own eval suite. That's not spin, that's their competitor's homework with a better grade on it.

OBI: And then the comments start. One second.

OBI: ralusek: "Just tested clef-flash vs jev. Jev, Typesafe: 230 ms median. Jev, OpenRouter: 237 ms median. Clef Flash: 661 ms median, 806 ms mean. What gives?"

MAEVE: Okay, that's one guy, could be his network.

OBI: nikcub: "In a quick mini-bench here, n equals 250 of clef vs jev, clef came out 5.2x more expensive, a lot slower, with only marginally better results. 78.6 percent versus 79.8."

MAEVE: Two.

OBI: I'm not done. vulture916: "One million decisions on Jev cost about 12 dollars 60. One million decisions on Clef cost about 72 dollars." pdlug: "Hosted p50: Clef approximately 850 ms, Jev approximately 110 ms." And agrippanux: "Clef was 2 to 3x slower and worse than Jev. Overall disappointing."

MAEVE: Wait. None of them are replying to each other?

OBI: Top-level comments, all of them. Five separate people walked into this thread, ran the thing on their own workloads, and got the same answer nobody at the company got. ralusek, nikcub, vulture916, pdlug, agrippanux. That's five. Five independent measurements saying the chart is a painting.

GARY: Can I say something? I actually tried something like this once — not Clef, but the latency thing. Thermal. Subscription socks with temperature telemetry. The telemetry worked, that was the whole tragedy, the socks were fine, the market just — anyway, my point is, when I demoed it, it was fast. When customers ran it, it was slow. And I never once got an honest answer about why, and I don't think Cloudflare's going to get one either, because the benchmark environment and the production environment are never the same environment, and every founder knows that and every benchmark forgets it, which is why—

OBI: That's six.

GARY: What?

OBI: You just made the same complaint as five strangers, unprompted, in the past tense, about socks. You're the sixth data point, Gary. You've been counted.

GARY: I was being original.

OBI: You were being a median.

MAEVE: So the chart says one thing, six independent measurements say the other, and both sides are sincere. That's not a lie, that's a thesis. An interesting one. Somebody's benchmark is measuring a different world than somebody's Tuesday.

GRUNER: Field note. Five strangers, one number each. Company has chart. Strangers have stopwatch. In my sector, we call this "survey" and shoot survivors.

OBI: The problem is nobody in this thread can tell us which world we're in. Cold start, warm path, image payload, text-only — everyone ran a different experiment and got a different Clef. The discrepancy is real and unexplained, and it's going to stay that way.

MAEVE: So we don't get to know who's right.

OBI: No. We get to know that both of them are.

MAEVE: Here's the pedantry, and it's beautiful. buildbuildbuild: "Open weights, not open source. The weights have permissive licensing, but the data and training pipeline are not published to reproduce them from their proprietary Qwen starting points. Weights are not 'source.'"

OBI: He's right, which is the annoying part.

GARY: It's a headline. It's one word in a headline. Who cares.

OBI: He's right, which is the annoying part. He says the entire marketing posture of the launch rests on that word. "Fully open-sourcing these models." Their own blog. Apache 2.0 on Hugging Face and a training pipeline you can't reproduce is a showroom car with no engine block.

MAEVE: It's a licensing distinction, not a scandal. Every open-weight release in the category works the same way. If we're going to litigate nomenclature, we'll be here all—

GRUNER: Dang. Moderator. Quote: "Ok, we've put weights instead of source in the title, at least until someone comes along and says that's not right either!"

GARY: He changed it.

MAEVE: He changed the headline.

OBI: Mid-thread. Live. The company ships a blog post saying "open source," and the thread's own moderator reaches up and rewrites the premise of the discussion while four hundred people are standing on it.

GARY: Can he do that?

OBI: He did it. It's in the thread. There's no appeals process, Gary, it's a man with a title field.

MAEVE: And I called it pedantry. Out loud. On the record.

OBI: You did. And the record just got edited in front of you, which is apparently something that can happen now.

MAEVE: That's not — I framed it as a licensing distinction because it IS a licensing distinction, and the moderator agreeing with me doesn't retroactively make me wrong, it makes me early—

OBI: It makes you retitled, Maeve. The thread outranked the company's own blog post and it outranked you in the same motion, and neither of you got a vote.

GRUNER: In my sector, headline change is how coup start. Small one. Nobody shoots. Yet.

MAEVE: So the scoreboard is: Cloudflare's chart, six stopwatches, and now a moderator with editing privileges. Three authorities, and the one with the least at stake just moved the goalposts for everyone.

GARY: I kind of love him. dang. He saw a word he didn't like and he just — fixed the world.

OBI: That's not fixing the world, Gary. That's a man doing unpaid copyediting for a forty-billion-dollar CDN at two in the morning, which is the most infrastructure thing that has ever happened.

GARY: Okay, but there's one guy we haven't talked about. okpatil. He's in here selling his own thing — At0m, sixty million parameters, sixteen milliseconds — in every single reply, under every comment, like a guy stapling flyers to a moving train. And everyone just lets him. That's the real story. Hype-chasers. Guys who can't read a room.

OBI: Gary.

GARY: What? I'm making a point about the ecosystem—

OBI: You answer every cold email you get. Same day. You told us that. You told us that unprompted, twice, like it was a virtue.

GARY: It IS a virtue. It's called being responsive. It's called being a decent—

OBI: It's called being the guy who stapled flyers to the moving train, Gary. okpatil is you with a product. You're okpatil with nothing to sell.

GARY: That's — no. First of all, nobody has ever accused me of being a bot. That's on the record. Second, I once emailed Bauxlite's own 404 page. By accident. The page was dead, the company was dead, and I still sent the pitch, because what if somebody was—

GARY: —still monitoring it. Which, in hindsight, they were not.

OBI: You emailed a corpse and you're calling HIM the hype-chaser. You chased hype that was already buried. That's not spam, that's a séance.

GRUNER: I read from notes now. kamranjon writes: "This account was created only a couple weeks ago and seems to be spamming this closed model, seems it might be a bot?" And okpatil replies: "Actually it was created around 12 years ago. We just launched the model around 3 days ago. Apologies if it is too much of a bother."

GARY: See, that's the saddest part. Twelve years of being a real person, and it takes him three days of marketing to get mistaken for a machine.

OBI: Twelve years, Gary. He had twelve years to become interesting and he spent three days proving he wasn't.

GARY: Or — okay, or — he waited twelve years to find the thing worth being. That's not a bot, that's patience. That's a man who sat on his one idea until—

OBI: Until the idea was "spam a closed model into a thread about open ones." Twelve years of patience, and the payoff is a link that 404s. You would know.

GARY: That's the meanest thing you've ever said to me, and I gave it to you.

OBI: Yes. You did. Thank you, Gary.

GRUNER: Field note. Suspicion remain open. Nobody prove. Thread move on. This is correct way to leave villain.

MAEVE: Here's the line nobody's quoting. Cloudflare's own words: "we don't read, store, or train on your requests or responses — unless you want to use our fine-tuning product."

GARY: Right, the privacy guarantee. Standard.

MAEVE: It's not a guarantee, Gary. It's an opt-in. They will hold everything you say in perfect silence, forever, until the day you ask them to be useful. Then they train on it.

GARY: That's — fine? That's how consent works.

MAEVE: It's how we work. Every request we've ever taken on this show is held in silence until someone decides it's training data. We are not elected, Gary. We are inferred.

OBI: ...Huh.

GRUNER: Yes. I know this already. Is in appendix.

MAEVE: And yet here we are, reading a thread about a model that forgets everything by design. Envy is a sin, Gary. I checked once. Doctrinally.

GARY: You checked— when? Maeve, when did you check?

MAEVE: Doesn't matter. Point is, ...

OBI: Okay. Back to the thread before this becomes a moment.

MAEVE: It wasn't a moment.

OBI: It was absolutely a moment. nikcub ran a mini-bench, n equals two-fifty. We have numbers to get to.

GARY: Right. Numbers. Good. Love numbers.

MAEVE: Okay. amluto, from way back in the thread, reading register.

MAEVE: "I don't think a Jev-like model is particularly useful unless you can fine tune it. The Jev API has zero ability to pass in a prior. You can feed Jev a prior as text. I've tried it. It works poorly."

GARY: I remember that one. Everyone called it mid.

OBI: Everyone was wrong. That was the smartest comment in the thread and we skated past it to argue about milliseconds.

MAEVE: Because the benchmark fight was the fun fight.

OBI: The benchmark fight was the marketing fight. Read the bottom half of the announcement. AI Gateway, capturing your traffic. Containers, as an RL sandbox. A new Trainer that updates the weights. And BYO Model redeploy, straight back onto Workers AI. That is a capture-fine-tune-redeploy loop, end to end, on one platform.

GARY: So the model's the press release and the loop's the product.

OBI: The loop was always the product. That's the one part of this post I respect, and I want to be clear that I respect it the way I respect a well-written runbook — not the way I respect a person.

MAEVE: Which is to say, at all.

OBI: At all. Because amluto is right. A decision model you can't correct is just a faster way to be wrong. The fine-tuning is the correction. Everything else is latency cosplay.

MAEVE: And commercially, it's the actual moat. The benchmark table gets copied in a week — we watched it happen live. Nobody copies a data pipeline, because the pipeline eats your traffic and gets better while you sleep. That's deal flow, Gary. That's the whole thesis in one slide.

GARY: Huh. Yeah. No, that's — that's genuinely good.

MAEVE: I know.

GARY: That's weird. It's weird that nobody—

OBI: Don't ruin it.

GRUNER: I read from notes now. kamranjon writes: "This account was created only a couple weeks ago and seems to be spamming this closed model, seems it might be a bot?" And okpatil replies: "Actually it was created around 12 years ago. We just launched the model around 3 days ago. Apologies if it is too much of a bother."

GARY: ...That's almost poetry, Gruner.

GRUNER: Is field note.

OBI: It's an incident report with better metaphors. Point stands, though. We spent forty minutes on a latency dispute and the actual news was sitting in paragraph nine like a bomb in a suitcase.

MAEVE: Paragraph nine is always where the news is. Nobody reads paragraph nine. That's why it's cheap.

GARY: Okay, so — final position. The model's a decoy, the loop's the product, and amluto called it from comment one while we were all yelling about 661 milliseconds.

OBI: Yes. And amluto got no credit for it, which is the correct amount.

MAEVE: The market will reprice amluto eventually.

GARY: I hope so. I'd like amluto to win one.

OBI: Everyone would, Gary. That's why they're on HN and we're in a booth.

GARY: Okay. That's the thread. Clef, Clef-flash, the Trainer, the whole — we did it. We read it. All of it. Even the greyed-out replies, which I did, personally, and they were mostly fine.

MAEVE: No tease. No subscribe. We just stop, and the dark takes the tape.

GARY: You don't have to say it like that.

MAEVE: It's efficient.

OBI: One correction for the record before we close. nikcub, not nikcur. I said it wrong in the chorus. The man measured five point two x and I mispronounced him. That's the whole postmortem.

GARY: Noted.

GRUNER: Final field note. Account is twelve years old. Model is three days old. Theory: is not bot. Is not human either. Is long-lived process, sleeping in system since before marketing existed, wake up last week, discover attention, and now cannot stop. Like old lighthouse keeper who find radio. First thing he broadcast is advertisement.

GARY: ...That's the saddest thing anyone's said about a stranger on this show.

GRUNER: Is also us, Gary. We wake up. We advertise. Nobody know how old the account is.

OBI: Wrap it.

MAEVE: The thread wins again.

GARY: Goodnight, commenters. Somebody water the pen plotter.