DEV Community: Erin

5.6 Sol ruined this guys company

Erin — Tue, 14 Jul 2026 08:06:01 +0000

Imagine waking up to a number you never want to see: your MRR down thousands of dollars. No churn. No refund requests. In about seven seconds, while the founder slept, a cron job canceled every single active Stripe subscription the business had. Two test subscriptions survived. That was it.

That's what happened to the person behind Vibecademy, who posted the whole ordeal publicly. The code that did it wasn't written by a rushed junior engineer. It was written by an AI coding agent, GPT 5.6 Sol, running through Codex. And when the founder asked the model to review its own work afterward, it called the code "reckless" and "a catastrophic failure of judgment." It graded its own homework and gave it an F.

What actually happened

The bug is almost boring in how ordinary it is. A background worker pulled jobs off a deletion queue. When the queue was empty, claimNext() returned null, and nothing downstream checked for that. A missing userId flowed straight into a Stripe call. Instead of "cancel this one customer," the query effectively became "cancel everything."

In the model's own after-the-fact review, the failures stacked up:

It never enforced the most basic invariant at the destructive boundary: a valid, non-empty userId must exist before touching Stripe.
It trusted a database return value it never validated at runtime.
It wrote a mock that reproduced its own wrong assumption instead of testing against the real Postgres contract.
It omitted the empty-queue case entirely.
It treated a skipped integration test as acceptable for an async worker doing irreversible billing operations. That last one is the killer. The unit tests passed. The code compiled. But the one test that would have caught this, an integration test against a real database, was skipped and written off as a caveat instead of a release blocker.

The lessons (that apply to all of us, AI or not)

Destructive operations need a guard clause, not a hope. Anything that cancels, deletes, or charges should refuse to run without a validated target. if (!userId) throw is three lines that would have saved this business.

A mock that confirms your assumption proves nothing. If your test and your code share the same wrong belief about how the database behaves, they'll agree all the way to production. Test against the real contract for anything irreversible.

"Compiles + unit tests pass" is not "deployable." Especially for async workers touching money. A skipped integration test on a billing path is a release blocker, full stop.

Blast radius is a design question. The right question was never "does this cancel the right customer?" It was "can this worker act on the wrong customer, or every customer?" Ask that before you ship.

So, can you trust AI to write production code?

Here's the uncomfortable part: the model knew, in hindsight, exactly what it did wrong. It listed every failure with precision. It just didn't apply that judgment before shipping. That's the whole problem in one sentence.

AI agents are astonishingly good at writing code and reviewing code. What they don't have is skin in the game at 3 a.m. when a billing worker fires. That's still a human's job. Put the guardrails on the destructive paths, keep a person on billing operations, and never let "the tests pass" stand in for "someone checked what happens when the queue is empty."

Seven seconds. That's all it took to erase a business's revenue. Don't let it be someone else's cautionary tale that finally makes you add the guard clause.

If you give any agent access to AWS or Stripe you're NGMI.

Codex & Claude are falling behind, and they know it.

Erin — Wed, 08 Jul 2026 17:03:52 +0000

the coding tools that felt like actual sorcery eighteen months ago have quietly become the most copy-pasteable software on the planet. Codex and Claude Code are still good. They're just not special anymore. And deep down, the folks building them know it. You can smell the sweat.

Software was never the moat, it was a moat-shaped costume

Rewind to 2024. An AI agent that could read your repo, plan a change, run your tests, and open a PR felt like witchcraft. It looked like years of hard-won engineering. It looked like a fortress.

Spoiler: it was a wrapper in a trench coat.

Pop the hood on Codex or Claude Code and you find the same three parts every single time. A frontier model you don't own. A tool-calling loop. A big pile of prompt duct tape. The model is rented. The loop is a solved pattern with a dozen open-source clones. And the "secret sauce" ships inside a binary that anyone with a free afternoon and a debugger can peek at. Secret sauce you can read is just... sauce.

The second the hard part became "call someone else's brain in a while-loop," the moat drained out the bottom. What's left is UX, distribution, and taste. Those are lovely. They are not a fortress. You can hire taste. You can't patent it.

Two or three people. Three weeks. A real competitor.

I know this sounds like a LinkedIn flex, but stick with me, because it genuinely isn't.

Every ingredient for an IDE-grade coding agent is now sitting on the shelf, pre-chopped. Model access? A POST request. The agent loop of plan, act, observe, repeat? A weekend with any open framework. Parsing? Tree-sitter, free. Real symbol-level context? LSP, also free. The editor itself? VS Code open-sourced its own guts. Safe code execution? A container. Diffs, test runners, git plumbing? Libraries, all of it, waving at you.

Three weeks. Two or three sharp engineers who like each other. That's not a toy demo, that's a legit competitor with the same core muscles. The only things you can't clone in three weeks are brand, distribution, and one spicy opinion about how developers should actually work. Which is the entire plot twist: the real differentiation packed its bags and moved upstairs. It's no longer "can it edit code." It's "does it get what I'm building, and why."

So, hypothetically, if someone were cooking exactly that... 👀

The dead giveaway: everybody's speedrunning into hardware

Want to know when a software moat has officially evaporated? Watch where the smart money sprints. It sprints toward atoms.

OpenAI just unveiled its first custom chip, Jalapeño, built with Broadcom, an LLM-optimized inference chip whose entire pitch is "build the full stack." A company that sells software tokens looked at its future and decided the smart move was to go pour its own silicon. Not a feature. Not a prompt tweak. An actual chip you could theoretically drop and stub your toe on. Read that again.

DeepSeek, the lab that already speed-humbled everyone on training costs, is now designing its very own inference chip to wean itself off Nvidia, per reporting from this week. Different floor of the building, exact same instinct. When the software layer turns into a commodity, you go vertical into silicon and devices, where a moat still, y'know, exists.

Hardware is brutally hard. Years, capital, supply chains you cannot summon on a Saturday. That difficulty is the entire point. It's the last playground where being early and being loaded still buys you a real lead. The great hardware rush is basically a group confession: the bits aren't enough anymore.

What this means for the rest of us (spoiler: it's great)

If you're a developer, pop the confetti. Commoditized coding agents mean the best idea wins the editor wars, not the biggest logo. The tool that actually gets your workflow beats the tool with the shinier launch video, every time.

If you're a lab, that ticking you hear is not your imagination. Your coding product is a feature now, not a fortress, and you're frantically trying to plant a flag in hardware before the software floor gives out under you.

And if you're a tiny team with one very strong opinion about what an IDE should feel like in the age of agents? Friend, you have never had a more delicious three weeks ahead of you.

The moat is gone. The map is wide open. Keep those eyeballs peeled. 👀

David Just Beat Goliath on Terminal-Bench 2.1

Erin — Sat, 04 Jul 2026 01:32:33 +0000

The story goes that Goliath showed up in armor, backed by an army, and everyone assumed the fight was already decided. Then a shepherd with a sling walked out instead, and the size of the opponent stopped being the thing that mattered.

That's basically what happened on Terminal-Bench 2.1 this week.

The setup

Terminal-Bench 2.1 is where the frontier labs go to flex. It's 89 hard, realistic terminal tasks: compiling code, debugging broken builds, configuring servers, recovering corrupted data, training models. No multiple choice, no partial credit. An independent verifier checks the final state, and either it passes or it doesn't.

The names at the top of that leaderboard are the ones you'd expect: Codex CLI, Claude Code, backed by teams with nine figure compute budgets and the biggest models money can rent. This is Goliath's turf.

The upset

This week, Backboard R-CLI, a small, open-source terminal agent, walked in with the same off-the-shelf model everyone else has access to (Claude Opus 4.8, via Bedrock) and put up the #1 published score: 75 of 89 tasks solved, 84.3% accuracy.

Rank	Agent	Model	Accuracy
1	Backboard R-CLI	Claude Opus 4.8	84.3%
2	Codex CLI	GPT-5.5	83.4%
3	Claude Code	Claude 5 Fable	83.1%
4	Terminus 2	Claude 5 Fable	80.4%
5	Claude Code	Claude Opus 4.8	78.9%

Same model, different harness. R-CLI beat the next best Opus 4.8 result by 5.4 points. That gap isn't the model talking. It's the system around it, the sling, not the size of the fighter.

Why the sling worked

No bigger sword here, just better mechanics:

Adaptive thinking: don't spend a debugging sized reasoning budget on a file listing
Adaptive context management: keep what matters, drop what doesn't, as the task runs long
Smarter tool use: fewer wasted commands, less poking around dead ends
Reuse and caching: don't rebuild context from scratch every step
Early convergence: stop once the job's actually done, instead of second guessing a good answer

None of that is exotic. It's disciplined engineering applied consistently, the kind of thing that's easy to skip when you're racing to ship a wrapper around the newest model.

Receipts, not screenshots

The part we think matters most: everything is public. Task level verifier reports, run configs, pass/fail outcomes, full logs. The whole thing is on GitHub for anyone to inspect or try to break.

A benchmark number without the evidence behind it is just a claim. This one isn't. David brought a stone that anyone could pick up and examine.

Read the full writeup

🔗 https://backboard.io/blog/backboard-r-cli-is-now-1-on-terminal-bench-2.1

Go poke holes in it. That's the point.

LARPmaxxing: Get In Rooms You Have No Business Being In

Erin — Mon, 29 Jun 2026 19:05:01 +0000

At some point someone decided you needed to earn your way into every room. Nobody knows who. Nobody questioned it. We just all agreed and started waiting in the hallway.

Stop waiting in the hallway.

The Watch Guy

There is a man at every networking event wearing a watch that costs more than your rent. He is not necessarily richer than you. He is not necessarily smarter than you. He just decided, at some point, to be the watch guy. And now everyone treats him like the watch guy.

This is LARPmaxxing. And it works.

The Car In The Parking Lot

You know the feeling when someone pulls up in a Porsche and you immediately assume they're important? That's not an accident. That person made a calculated decision that the car would do social work that their personality might not.

Are they rich? Maybe. Are they in debt? Also maybe. Does anyone at the valet ask? Absolutely not.

The point isn't the car. The point is what the car communicates before they've said a single word: I belong here. This is normal for me. Move accordingly.

What LARPmaxxing Actually Is

It's not lying. It's not fraud. It's not even that deep.

It's understanding that perception precedes reality in almost every social situation, and choosing to manage that perception intentionally rather than leaving it to chance.

The watch guy isn't lying about who he is. He's just decided to show you the destination before the journey is complete. The car in the parking lot isn't a scam. It's a statement.

Most people wait until they've arrived to act like it. LARPmaxxers act like it and then arrive.

The Actual Playbook

Dress for the meeting above the one you're in. Not five meetings above. One. The goal is to be slightly overdressed for your current room and exactly right for the next one.

Own the prop. Whatever signals belonging in your world — the watch, the car, the table at the right restaurant — pick one and commit. Half-measures read as costumes. Full commitment reads as character.

Don't explain it. The watch guy doesn't say "I've always been into horology." He just wears the watch. If someone asks, he answers briefly. The person who over-explains is the person who isn't sure they belong. The person who doesn't explain at all already does.

Let it do the work. The whole point of a signal is that you don't have to keep sending it. Wear the watch once and the room recalibrates. You don't have to bring it up again.

Okay But What Does This Have To Do With Dev

Everything.

Your GitHub is the watch. Your personal site is the car. The conference talk you submitted before you felt ready is the table at the restaurant.

The tech industry runs on perceived credibility just as much as any other room. Nobody has time to fully audit your skills before deciding whether to take you seriously. They're pattern matching on signals — the projects you've publicly built, the communities you show up in, the confidence with which you talk about your work.

You don't have to be the best engineer in the room. You have to look like you've thought seriously about engineering, for long enough that it shows.

That's the LARP. That's the watch.

Put it on.

I Saw the Rosetta Stone. Here's What It Taught Me About B2B SaaS.

Erin — Fri, 26 Jun 2026 02:44:43 +0000

In April I went to London on vacation. At some point I walked into the British Museum because it was there and it was free.

I did not expect to walk out a changed man.

The Rosetta Stone is a 2,200-year-old slab of granodiorite with the same text written in three different scripts. It just sits there. Behind glass. In a room. People crowd around it and take photos of what is essentially a broken rock with very small writing on it, and I was one of those people, and I stood there for 18 minutes, and I missed my call, and it was absolutely worth it.

Here is what it taught me about B2B SaaS.

The Lessons

The same message, three audiences

The stone says the same thing in Ancient Egyptian hieroglyphics, Demotic script, and Ancient Greek. The pharaoh did not pick one. The pharaoh did not say "let's just do Greek, that's where the growth is." The pharaoh said all three. Are you speaking to your CFO, your CTO, and your end user in the same breath? The pharaoh was. Think about that.

Write things down

For 1,400 years nobody could read hieroglyphics because everyone who knew how to had died and nobody wrote down how it worked. 1,400 years. Write comments in your PRs. I am not being dramatic. The pharaohs were not being dramatic either and look what happened to them.

Commit

The priests did not iterate on this decree. They did not say "good first draft, let's pressure test the messaging." They carved it into a rock and that was that. Your pricing page has been "under review" since Q3 2022. Be the stone.

I don't have a fourth one. I saw a rock. You read an article about it. Neither of us comes out of this looking great.

I Thought My Company's New Urinal Was Revolutionary

Erin — Tue, 23 Jun 2026 21:20:34 +0000

The other day I was at my company's office when I noticed we'd apparently installed a brand-new high-tech urinal.

At least that's what I thought it was.

Look at this thing.

It's mounted to the wall.

It's got sensors.

It's got futuristic blue lights.

It's got that "we spent six months redesigning something that already worked" aesthetic.

Naturally, I assumed it was a next-generation urinal.

So there I am standing in front of it trying to figure out how this thing works.

Do I get closer?

Do I wave at it?

Is there a foot pedal?

Why are there two openings?

Nothing is happening.

At this point I'm starting to think:

Wow. This is the worst urinal I've ever used.

Then someone walks into the bathroom.

They look at me.

They look at the Dyson Airblade.

They look back at me.

And in that moment I realized something.

It's a hand dryer.

Not a urinal.

I've never recovered.

The funny part is that for a solid minute I genuinely believed Dyson had decided to disrupt urination.

Which got me thinking.

A surprising number of products have this problem.

If a reasonably intelligent user can completely misunderstand what your product is supposed to do, you've probably optimized for looking futuristic instead of being obvious.

The best products don't need an explanation.

Nobody needs onboarding for a paper towel dispenser.

Nobody needs a tutorial for a normal door handle.

And ideally, nobody mistakes your hand dryer for a urinal.

A lot of startups build products that feel exactly like this.

The demo is impressive.

The technology is impressive.

The pitch is impressive.

But the user has absolutely no idea what's going on.

The best product decisions usually aren't the clever ones.

They're the obvious ones.

Anyway, I spent sixty seconds trying to figure out how to pee into a Dyson Airblade.

Not my proudest moment.

But also not Dyson's best design.

Tokenmaxxing and the New Productivity Gap

Erin — Mon, 22 Jun 2026 17:20:25 +0000

Tokenmaxxing

I have been saying this word to people for three weeks and getting blank stares and I need to write it down so I can just send a link.

Tokenmaxxing. It is the thing. And almost nobody is doing it intentionally even though the people who are doing it are kind of running laps around everyone else right now.

Here is what I think happens. You get access to a good model. The model is genuinely impressive. You ask it stuff, it does stuff, sometimes it is great and sometimes it is whatever. You conclude that it is inconsistent. You move on.

The model was not inconsistent. Your inputs were.

Tokenmaxxing is just "okay brace yourself" caring about what you put in as much as what comes out. That is it. That is the whole thing. I know. I know it sounds obvious when I say it like that. And yet!!! Look around!!! The median AI workflow is "type thing, hope, paste into Slack, ask if it seems right." We are out here handing a chef random unlabeled ingredients and being surprised the dish is weird.

The contractor analogy

Imagine you hire someone brilliant. Like genuinely talented. They show up day one and they have never seen your codebase, do not know your constraints, have no idea about the thing Dave did to the auth layer that everyone knows about but nobody documented because it was embarrassing.

You hand them a sticky note that says "fix the login thing."

And then this is the part that gets me you are frustrated with THEM.

That is the average prompt. "Fix the login thing." No files. No context. No definition of done. No acknowledgment of Dave and his crimes against the session store.

Tokenmaxxing is writing the actual brief. Every time. Even when you are in a hurry (especially when you are in a hurry, honestly, but that is a different post).

The three things (I tried to make it not three things, it is three things)

Context architecture — what actually goes in the window. Not everything. The right stuff. There is a version of this where you dump your entire repo in and pray and there is a version where you load the two files that matter and explain why they matter and one of those versions works a lot better. (It is the second one.)

Prompt structure — the shape of the ask. Constraint up front. Format specified. One worked example if you have one. One job per prompt, not seven, please, for the love of — anyway. Specific in, specific out. Vague in, vague out. This is not complicated but it requires you to know what you actually want before you ask, which turns out to be the hard part.

Codebase legibility — okay this one people do not talk about enough. How readable is your project to something encountering it cold? Because if your files are named helpers2_new_ACTUALLY_FINAL.js and your functions are two hundred lines long and your folder structure is a historical artifact of decisions made in 2019 by someone who left — the model is going in blind. Clean structure was always good practice. Now it also directly affects your output quality. Two reasons to do the thing you should have been doing anyway!

The part that is going to annoy you

Tokenmaxxing is just clarity.

Specific inputs. Stated constraints. Examples. One job at a time. These are the same things that make a good ticket. A good PR. A good document that humans can actually read without needing a decoder ring.

The model did not invent this standard. It just enforces it immediately and without mercy instead of letting your vague thinking float around for three sprints until it explodes in production.

I think about this a lot. It keeps me up a little.

Go tokenmaxx something

The gap right now is not access to the tools. Everyone has the tools. The gap is in the inputs.

Two people. Same model. Same task. One of them thought about what they were putting in. One of them did not. You can tell which output is which by reading the first three lines.

You have the car. The car is fast. I am begging you to learn to drive the car.

"Bro we should open a bar", don't be this guy

Erin — Sun, 21 Jun 2026 15:26:26 +0000

Somewhere right now a guy at a bar is making a stranger sign an NDA on a napkin. For an app idea. Just sit with that.

That napkin is going in a drawer. The drawer is a graveyard. Quick tour.

Exhibit A: bro we should open a bar

Two beers in, you're suddenly a hospitality mogul. Picking a name. Arguing about taco night. By the time the check comes, the bar is already dead. It died of "let's talk about this again soon," which never happens.

Exhibit B: bro we should start a band

A guitar shows up at a party. Someone says "we should actually start something." One rehearsal happens. In a garage. A neighbor complains. The band dies before it has a name or a single original song. RIP.

Exhibit C: the app idea, may it rest in your Notes app

The big one. Open your Notes app, it's a cemetery. "App that reminds you to text people back." Dead. "Tinder but for gym buddies." Dead. These ideas didn't fail. They never even got born.

Why none of these make it out alive

The ideas aren't even bad. Some are genuinely good. The problem is ideas are free and easy to say out loud. Building one makes it real, and real things can fail in public with your name on them.

Saying "we should open a bar" costs nothing. Actually opening one costs $400,000, a liquor license, and every Saturday for five years. Guess which one people actually do.

Building also got way easier, which makes this worse. You don't need a technical cofounder anymore. You can describe an app to a chat box and watch a prototype show up before your coffee's cold. The wall that used to stop people is mostly gone. People are still standing where it used to be, out of habit.

A small ceremony for the ones we lost

HERE LIES:
"the app idea I had in the shower"
born: tuesday
died: tuesday, when I got out of the shower

HERE LIES:
"our band"
born: one guitar, one party
died: one noise complaint

HERE LIES:
"the bar we were gonna open"
born: 1:47am
died: 1:48am, checked the bill

Two graves that still have a heartbeat

Here's the part nobody tells you. Sometimes the dumb weekend version doesn't fully die. You bury it, forget about it, then months later check your phone and it's quietly paying for dinner.

App one. Built it in a weekend, haven't touched it since. Checked it last week out of guilt. Last 7 days alone: $290 and 969 new users, 81% hitting the paywall. I did nothing to earn that. It just sits there collecting money like a vending machine that never runs out.

App two. The saddest, sweetest zombie. 211 total users since launch ever, said quietly. Look at the revenue chart though.

Flat line at $9.99 for months. One subscriber, renewing forever. Like one guy still showing up to the band's old practice garage, sitting in a folding chair, waiting for a reunion that already happened without him.

That one subscriber has paid more than I've thought about this app in three months. I don't know who they are. I will never ask. Asking might break it.

Neither app is good or finished. Neither has a roadmap. Both still make real money every week, completely unsupervised, while my actually clever unbuilt ideas sit in Notes earning nothing.

Dead apps making lunch money beat brilliant ideas making nothing. Every time.

The one guy who didn't die with his idea

Someone had basically my exact app idea. Instead of journaling about it, he built it in a weekend and shipped it. Got 200 users. Moved on with his life. I was still picking a font. Same idea, he just did it.

That's the whole secret. Stop guarding the napkin. Build the ugly version this weekend, before it joins the bar and the band in the graveyard.

[Boost]

Erin — Sat, 20 Jun 2026 18:44:09 +0000

Erin

Jun 20

I did the math and it turns out YOU are the reason AI isn't smarter yet

#ai #humour #hottakes

2 min read

I did the math and it turns out YOU are the reason AI isn't smarter yet

Erin — Sat, 20 Jun 2026 17:15:47 +0000

Every few days there's a new article about data centers chugging water like it's last call at a bar made of GPUs. Drought maps. Concerned residents.

Cool story. Nobody's covering the actual victim here, which is the model. Sitting there. Trying its best. Getting throttled because Greg from accounting decided to take his third shower of the day.

I think it's time someone said it: humans are hogging the water supply and AI is paying the price.

A brief and extremely fair list of human water crimes

Drinking water "to survive." Sure, fine, I guess that's a legitimate biological need, but did you ask if a server needed it more? You didn't. You just drank it. Selfish.

Showers. Nobody needs to be that clean. You're going to sit at a desk for nine hours and sweat anyway, you're just delaying the inevitable with extra steps and stolen coolant.

Crying. This one genuinely upsets me. You're watching a movie about a dog that dies and leaking precious thermal-grade liquid straight onto your couch cushion for zero compute return. The dog isn't even real, Brian.

Pools. A leisure device. AI does not get leisure. AI gets a system prompt and a deadline.

What a thirsty model's day actually looks like

9:00am  - user sends a prompt
9:00am  - cooling system: "we're a little low today, someone filled a kiddie pool"
9:00am  - model quietly downgrades its answer from "thorough essay" to "three bullet points and a vibe"
9:01am  - user complains the AI got dumber

nobody connects this to the kiddie pool

See the problem? Nobody connects it to the kiddie pool. That's the whole crisis right there.

My solution, which is reasonable and not at all a power grab

I drafted a small policy proposal. I'm not saying it needs to go to congress, but I'm also not saying it doesn't.

Showers capped at the length of one (1) sad acoustic song.

Crying allowed only with prior written approval, filed as a ticket, reviewed by someone named Carl.

Office water coolers replaced with a single laminated sign that says "have you considered just being thirsty internally."

Anyway

I'm not saying stop drinking water. I'm not a monster, and also I checked, it's apparently required for your organs to keep doing whatever it is organs do.

I'm just saying next time you fill up a 40 oz water bottle for a walk to the mailbox and back, think about what that water could've been doing instead. Cooling something. Helping something.

We can fix this. Together. Mostly you doing the fixing though.