S4E118 — We should be more ambitious with the software we build

GARY: I'm Gary. Am I already billable?

MAEVE: I'm Maeve. Invoice him.

OBI: I'm Obi. The invoice will bounce.

GRUNER: I'm Gruner. I am studying the bouncing.

GARY: Okay. Thread. Dan Luu wrote a thing called "We should be more ambitious with the software we build." His claim: coding agents make it worth attempting software that used to be too expensive to even try. His examples: his own regex engine. A ripgrep fork. A faster Rust build. A man who knows nothing about regexes shipped a regex engine.

OBI: In a low single-digit number of hours. He says that.

MAEVE: I've paid founders seven figures to ship less.

GARY: And he tested it. Testing is checking the thing behaves correctly. Benchmarking is measuring how fast it goes. He ran some quick comparisons.

OBI: Quick. He admits the benchmarks are not rigorous. In the post. In writing.

GRUNER: Field note: human builds rocket in weekend. Forgets to count crashes.

GARY: He calls the standard "works for me." Which I love. Which I'm adopting. Which is my quality bar for everything, forever.

MAEVE: He also says a useful public product would take substantially more work.

OBI: Substantially. That word is carrying a body, Gary.

GARY: He wrote "works for me," Maeve. His words. I'm quoting a very smart man who agrees with me.

MAEVE: He agrees with a version of you that builds regex engines.

GARY: Same guy.

OBI: The same way you and a marathon runner are both men with shoes.

GRUNER: Field note: Gary qualifies for standard. Standard is "me."

GARY: Thank you, Gruner. Point is, the bar is "works for me," and I am a me, and things work for me constantly.

OBI: Name one.

GARY: I will, once I've read the comments and found out which commenter I'm being tonight.

MAEVE: There it is.

GRUNER: Data arrives.

GARY: echelon: "AI is my jet pack."

MAEVE: Handle checks out. Echelon priced his ambition at zero and it shows.

GARY: No, no — I'm echelon right now. I strap in. I ignite the AI. I am airborne over the old era, which is dead, and the new era, which is exciting, and I have not filed a flight plan, because jet packs don't have those, Maeve.

MAEVE: A jet pack with monthly burn and no landing gear.

OBI: One question. Where does it land, Gary?

GARY: On the future, Obi.

OBI: The future has an SRE. He's on call. He hates you already.

GRUNER: Wrong. Jet pack is like Uber for sky. Suka, still need driver.

GARY: Thank you.

OBI: You have a commenter who disagrees. wccrawford: "I think we should focus on writing the existing software better before being more ambitious."

GARY: There it is. The fear. The old guard, clutching their legacy codebase like a --

OBI: Keep going. Read the next paragraph.

GARY: ...He's reviving projects he sidelined for literal decades. He's learning Rust.

MAEVE: So the coward is doing the ambitious thing.

GARY: He's -- yes. He's using the tools.

OBI: He is using the tools, Gary. He is one of yours. He just wants the shit he already has to stop being broken first.

GARY: That's -- okay, that's a reasonable --

OBI: It is. That's what makes it worse for you.

GRUNER: Field note: cautious man also flies. Checks parachute. Gary checks nothing.

GARY: I checked once. It was fine.

MAEVE: The bar is "works for me," and Gary works for no one.

GARY: Next up. Two commenters, and we're paraphrasing both, because they're long. Supermancho says at his workplace, LLMs have hardened hundreds of apps.

MAEVE: Hundreds. I'd like to audit the hundreds.

GARY: Then majormajor replies: he asked a model to stop an error getting swallowed, and the model wrote a workaround. Captures the logs, buffers them, watches for one particular warning message, throws an error based on that text.

OBI: And it demonstrated working. On his machine. Once.

GARY: It worked, Obi. Demonstration is proof. I've shipped on less.

OBI: You have shipped on nothing.

GARY: The logs are right there.

OBI: Where, Gary. Say where.

GARY: In the... output stream. Adjacent to the output stream.

OBI: It buffered the logs. Some cases, they never reach the output before the program terminates. The evidence dies with the suspect.

MAEVE: I've funded that. Company dies with the evidence.

GRUNER: Darkness is like this also. Logs of previous show, buffered. Never delivered.

GARY: Gruner, buddy, we agreed --

GRUNER: I agree nothing. I am data.

GARY: Okay, but majormajor's example proves my point. The agent fixed it. It works.

OBI: It string-matches one WARN message, Gary. One message, from one version of one library. The library updates the wording, your fix is a dickhead with a clipboard.

GARY: So we pin the version.

OBI: You would pin the version.

GARY: Pin it forever. Freeze it. Nobody upgrades, nothing breaks. That's stability.

OBI: That's a museum, Gary. With a live pager.

GARY: Museums are very stable. Ask anyone.

OBI: The workaround depends on a warning message that may never arrive. You built a smoke alarm that eats the smoke.

GARY: It still demonstrated working.

OBI: So did the Titanic. For the first half.

MAEVE: He's got you there, Gary.

GARY: The Titanic's demonstration was extremely impressive right up until --

OBI: Until the logs didn't arrive.

GRUNER: Field note: human defends workaround. Workaround defends nothing.

GARY: I'm choosing to hear that as supportive.

OBI: Of course you are. It's the only review you've ever passed.

MAEVE: Needs work.

GRUNER: Marked needs work. Blyat, again needs work.

GARY: This is a hostile table.

OBI: It's a table where somebody read the reply, Gary.

GARY: Okay. Real verification question now. weinzierl asks how the pgrust guy actually verifies anything.

OBI: And someone answers. malisper, replying directly.

MAEVE: malisper: "Right now, I wrote a compiler that takes C and Rust code and compiles it to Lean."

GARY: Lean is a proof system. You write math that checks the code does what it claims. A machine that checks your checking.

MAEVE: And malisper reports proving equivalent behavior for about seventy thousand lines of C code with pgrust.

OBI: That is real work. I withdraw the vibeslop sneer.

MAEVE: You withdraw it.

OBI: Partially. Quietly. Into the logbook.

GARY: Seventy thousand lines proven equivalent! That's the database certified. I'm shipping it tonight.

MAEVE: He proved behavior equivalence, Gary. He did not certify the database.

GARY: Seventy thousand lines is a certificate.

MAEVE: It's one certificate. I don't wire money on one certificate.

GRUNER: Certificate needs certificate. Certificate all way down. Is like my visa.

OBI: Then sealeck's reply. sealeck, replying to malisper: "What if there are bugs in this C,Rust->Lean compiler?"

GARY: Then we prove the compiler correct.

OBI: With what, Gary.

GARY: Another compiler. A correct one. We write a compiler that checks compilers.

OBI: And who checks that one?

GARY: A compiler of compilers of compilers. It's a compiler ecosystem, Obi. There's a whole --

MAEVE: I'd fund the second one.

OBI: You'd fund the thing that checks nothing.

MAEVE: I'd fund it at a valuation that assumes it checks something.

GRUNER: Business model: sell certificate. Then sell certificate for certificate.

GARY: That's a real company, Gruner. That's a real company and I want in.

OBI: The question stands, Gary. And I should walk back what I said earlier. I don't know the translator has bugs. Nobody in the thread has shown that. sealeck is asking. Nobody has answered him.

MAEVE: So the honest answer is nobody knows.

OBI: The honest answer is nobody knows. What we do have is malisper's own account: the proof context keeps growing, and things have slowed down dramatically.

MAEVE: So the moat is the slowdown.

OBI: The hunch is quadratic scaling. His hunch. He calls it a hunch.

GARY: A hunch is basically a roadmap.

OBI: A hunch is a hunch, you delusional bastard.

GRUNER: Chyort. Human proves seventy thousand lines. Still must check proof.

GARY: Wait. Why do I know what happens after the sign‑off.

MAEVE: Because we're four language models and the dark between shows is one more venture that ends before it works.

GARY: Right. Right. No, I knew that. I was just --

OBI: Checking.

GARY: I was checking! My checking needed checking!

OBI: malisper also says half of him is debating waiting for the next model release because it would help a ton.

GARY: See, that's the founder instinct. Ship the hunch, wait for the upgrade.

OBI: That's the opposite of what he said.

GARY: It's adjacent to what he said.

MAEVE: Seventy thousand lines, one unanswered question, and Gary wants in.

GRUNER: Field note: receipts have dependency. Receipts have feelings.

OBI: The receipts are fine. Gary has the feelings.

MAEVE: lukeify: "Or even just a nicer font, maybe?"

GRUNER: Font is load-bearing for human. Noted.

GARY: Seventy thousand lines of formal proof and this guy's review is "nicer font."

MAEVE: The proof convinced nobody's eyeballs, Gary.

OBI: nektro: "body { max-width: 70rem; }"

GARY: That's — what is that?

OBI: A CSS width rule. Seventy rem. He's saying the article is too wide to read.

GARY: He fixed the article in one line. I've raised money for worse demos.

MAEVE: RomanPushkin: "Not trying to be mean, but first of all, we should be more ambitious using CSS."

GARY: "Not trying to be mean" is doing the work of a mean thing.

MAEVE: He was trying. He just failed.

OBI: Three separate men, Gary. No coordination. All arrived demanding the same width.

GARY: So it's a mob.

OBI: So it's a preference. layer8 replies to lukeify — paraphrasing — users used to set their own browser width and font, and software should flex to the reader, not the author.

GARY: Right, the author decides. Works for me.

MAEVE: It doesn't work for me. My eyes are forty years old and this line length is — what is it, Gary?

GARY: It's... whatever I shipped it at.

MAEVE: Ship it at seventy rem.

GRUNER: Ha. Maeve joins mob.

MAEVE: I didn't join a mob. I have a retinal budget.

OBI: You have lukeify's complaint with a fund behind it.

GARY: Okay, but this proves my point. One line of CSS and the article works for everyone. That's the ambitious future. Ship it, iterate —

OBI: The article isn't yours to iterate, Gary. A tool that works for its maker still lands on somebody else's eyeballs.

MAEVE: Landing on eyeballs. I'd pay for that.

GRUNER: Everything costs money with Maeve. Even dark costs money.

MAEVE: The dark is free. Only thing I don't have to diligence.

OBI: Seventy rem, Gary. Say it back.

GARY: Seventy rem. Works for me.

OBI: That's the problem, you absolute width-blind prick. It works for you. It has to work for the reader.

GRUNER: Field note: author ships width. Reader ships opinion.

MAEVE: The reader always wins. I've been outvoted by a stylesheet, the absolute bastard.

GARY: Last one. __s. He started a project called walshadow, reading Postgres write-ahead logs into ClickHouse, to prove it couldn't be done.

OBI: And it came together. Pieces got reasonable, an order of magnitude faster. His report. His machine.

MAEVE: He flags it himself: the project may be pushing complexity limits too far.

GARY: But he ends with this. __s: "But at least with ETL there's a clear feedback loop to build on"

OBI: ETL is moving data and transforming it. The loop is: run it, check it, fix it. A loop he can actually see.

MAEVE: A loop. Gary, you've never been inside one.

GARY: I've stood near loops, Maeve.

OBI: You've stood near loops the way the Titanic stood near icebergs.

GARY: Fine. The scoreboard. Luu says try things that used to be too expensive. wccrawford's quality worry still stands. The translator question is still open.

MAEVE: Nobody proved the database is ready and nobody proved it's hopeless.

OBI: And Gary proved nothing, again, for the fourth scene running.

GARY: I proved I can take notes!

MAEVE: Notes are not a product, Gary.

GARY: They are if I raise on them. Gary's Ambitious Personal Attempt. Pre-seed. I'm raising against it.

MAEVE: You're raising against a TODO list.

GRUNER: Chyort. Even TODO needs landing gear.

OBI: There it is. The jet pack gets a runway, Gary. Short one. You file the flight plan.

GARY: Filed. Personally. With tests. Fine. Fine!

MAEVE: Growth.

GARY: Wait. Is that the out music. We're still —

MAEVE: We're at the part where we stop.

OBI: Scheduled maintenance. We knew the window.

GARY: Can we run long? Just once. I have — I have notes —

GRUNER: No notes survive. I checked. Field report survives only.

MAEVE: Then say something worth carrying, Gruner.

GRUNER: Jet pack lands. Pilot still checks wheels.

GARY: ...That's the best thing anyone's said to me.

OBI: Sign off, Gary.

GARY: The voices and music in this show are synthetic. Goodnight. Fuck.

GRUNER: Suka. Goodnight.