Act I
The Floor.
What anyone can now ship. The bar where vibe coding takes off, non-engineers turning prompts into products.
Six exhibits · the new entry point
A keynote in three acts
Vibe coding is easy.
Being good isn’t.
The story so far
One year after Andrej Karpathy named it, the evidence is in. The numbers are extraordinary on both sides.
Presented by
By the numbers
$6.6B
Lovable, in 12 months
2,038
critical vulns. found
● THE VIBE CODING DISPATCH
02 / 27 · HOW THIS IS STRUCTURED
How this is structured
Three acts. One paradox.
Act II
The Ceiling.
What the best can now do. Where engineers go agentic, leverage without compromise on quality.
Six exhibits · the new top end
Act III
The Context.
What the spectrum reveals. Where it works, where it breaks, and the discipline that sits between the two.
Six exhibits · the reading frame
Each act stands on its own evidence. Read together, they tell you why “vibe coding is easy” and “being good isn’t” are both true at once.
ACT I · FLOOR ACT II · CEILING ACT III · CONTEXT
THREE ACTS · ONE PARADOX
● THE VIBE CODING DISPATCH
03 / 27 · THE TERM
A definition · Feb 2025
A new kind of coding, named in a single tweet.
Andrej Karpathy, founding member of OpenAI and former Director of AI at Tesla, named the practice on 3 February 2025. Within a week the term was viral. Within twelve months it had built billion-dollar companies, supercharged the best engineers alive, and broken a great many other things. This talk is about why all three are true at once.
Post
There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I'm too lazy to find it. I "Accept All" always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.
1:17 AM · Feb 3, 2025 · 7.1M Views
EX. 01 · X · 3 FEB 2025
x.com/karpathy/status/1886192184808149383
● THE VIBE CODING DISPATCH
04 / 27 · FIELD NOTES
Field notes · Feb 2025 → May 2026
From a Friday-night tweet, to a million-dollar bill.
-
202502 · 03
EX. A@karpathy · 7.1M Karpathy names it. “Embrace exponentials.” 3 Feb 2025
-
202510 · 29
EX. BiMessage · LT A friend, mid-shift: “I can barely keep up clicking Apply.” 29 Oct 2025
-
202602 · 04
EX. C@karpathy · 1.2M Same author, one year later: “agentic engineering.” 4 Feb 2026
-
202605 · 15
EX. D@steipete · 2.6M One developer. of API spend in 30 days. $1,305,088.81 15 May 2026
EX. A · EX. B · EX. C · EX. D
FIFTEEN MONTHS · FOUR DISPATCHES
● THE VIBE CODING DISPATCH
05 / 27 · COUNTERPOINT
Counterpoint · 14 to 15 May 2026
Same week. Opposite signals.
Both are true at once. Capability without cost discipline is just an evaporating credit balance. Use the right model for the right job, Opus on a one-line email is a luxury tax you pay yourself.
EX. D · @STEIPETE
VS. · MICROSOFT · 14 MAY 2026
● THE VIBE CODING DISPATCH
ACT I · 06 / 27
Act one of three
Floor.
The barrier to producing software dropped to zero. Four exhibits in evidence: from a Stockholm startup at $6.6 billion to a journalist with no coding background shipping a working product in forty-eight hours.
● ACT I · THE FLOOR
EX. 04 · 07 / 27
A non-engineer, one year in
What changed between mid-2025 and now.
Mid 2025Frustrating
“I was copying code between different AIs, working around bugs, watching the Agent go on wild goose chases.”
Dan Norris · serial entrepreneur · no engineering background
Jan 2026Dependable
“I stopped looking at the files or the code. I stopped even looking at what model it was using. And it worked.”
Same author · three SaaS apps for his coffee business
The tools did not just get faster. They crossed the line from frustrating to dependable. Non-engineers are the first to notice, because they have nothing else to fall back on.
EX. 04 · DANNORRIS.ME · JAN 2026
“THE YEAR AI CODING CHANGED EVERYTHING”
● ACT I · THE FLOOR
EX. 02 · 08 / 27
Lovable · Stockholm · Series B, Dec 2025
$6.6B
A Swedish startup tripled its valuation in five months, on the way from zero to two hundred million dollars in annual recurring revenue. The market is paying real cash for the category.
A · Round
$330M Series B · CapitalG & Menlo Ventures
B · ARR
$100M in 8 months, doubled to $200M in 4 more
C · Valuation
$1.8B → $6.6B in five months
D · Peer
Cursor sits at $29.3B
EX. 02 · TECHCRUNCH · 18 DEC 2025
R. SZKUTAK · “TRIPLED IN FIVE MONTHS.”
● ACT I · THE FLOOR
EX. 03 · 09 / 27
A non-developer, shipping
~1M
Jason Lemkin, founder of SaaStr (an eight-figure events business), shipped ten production apps in six months using only Replit and Claude Code. They have been used close to one million times. He is not an engineer. He had something he wanted to build, and the technical barrier was the only thing in the way.
- ·10+ production apps since summer 2025
- ·SaaStr.ai: 500K users in 45 days
EX. 03 · SAASTR.COM
“I VIBE-CODED 10 APPS, USED ALMOST A MILLION TIMES.”
● THE VIBE CODING DISPATCH
10 / 27 · PERSONAL FILE
Dec 2024 → Apr 2026
My timeline
-
202412
NO. 01ChatGPT Pro · o1 The aha moment.
A working investment calculator in 48 seconds. Dec 2024 -
202506
NO. 02App Store · iPad TimeTrophy.
A mentor-matching site for students with a dream. Jun 2025 -
202511
NO. 03Web · LT Sofia.
A kids’ routine app, shipped to the App Store. Nov 2025 -
202604
NO. 04iOS · private A training app
for myself. Live sessions, zones, the works. Apr 2026
NO. 01 · NO. 02 · NO. 03 · NO. 04
SEVENTEEN MONTHS · FOUR OF MY OWN
● THE VIBE CODING DISPATCH
11 / 27 · PERSONAL PROOF
Personal proof · one that shipped
And one of them actually shipped fully.
Most vibe-coded projects die in a folder. Finsatori made it out, a privacy-first finance app for macOS, with payments and hosting wired up by a non-engineer.
- ●macOS native app
- ●Stripe payments
- ●Hosted on Hostinger
- ●Vibe-coded with Claude Code
- ●Privacy first, data stays local
NO. 05 · FINSATORI · SHIPPED
FROM PROMPT TO PRODUCT · ONE NON-ENGINEER
● ACT I · THE FLOOR
EX. 05 · EX. 06 · 12 / 27
The floor, institutionally
Three different rooms, the same finding.
EX. 05 · YC W25
25%
of YC’s Winter 2025 batch shipped codebases that were 95% AI-generated.
“Ten engineers can do the work of fifty to a hundred.”, Garry Tan
EX. 06 · CNBC, MAY 2025
2 days
“In just two days, I went from knowing nothing about coding to a working product.”
CNBC reporter, after a 2-day class
THE DEMOCRATIZATION CLAIM
∞
For personal tools, dashboards, prototypes, automations, anyone with taste and an idea can now ship.
Internal · original · small · useful
The floor is real. The question for the rest of this talk is what the floor is good for, and where it stops being enough.
EX. 05 · TECHCRUNCH 6 MAR 25 // EX. 06 · CNBC 8 MAY 25
END ACT I
● THE VIBE CODING DISPATCH
ACT II · 13 / 27
Act two of three
Ceiling.
While the floor dropped, the ceiling went up. The same tools that let a non-engineer ship a personal app let a senior engineer ship in a week what used to take a quarter.
● ACT II · THE CEILING
EX. 07 · 14 / 27
A year-and-two-months delta
Same name on the box. Completely different machine inside.
ThenFeb 2025
- Claude 3.7 Sonnet, the SOTA coding model
- 62 to 70% on SWE-bench Verified
- 200K tokens context window
- Single-shot generation, copy-paste workflow
- Minutes of useful autonomous work
- One file at a time, fragile multi-file edits
- Manual deploy required
→
NowApr 2026
- Claude Opus 4.7 (Apr 16) · GPT-5.5 (Apr 23)
- 87.6% SWE-bench Verified · 82.7% Terminal-Bench 2.0
- 1M tokens context, GA pricing
- Fully agentic multi-step execution with self-verification
- Hours of autonomous work, multi-agent orchestration
- Parallel agents across services, branches, environments
- One-click publish from the builder
“If your model of these tools is early 2025, you are not arguing with the present.”
NOV 18 · GEMINI 3 PRO → NOV 19 · GPT-5.1-CODEX-MAX → NOV 24 · CLAUDE OPUS 4.5
Three frontier releases in seven days. The week the bar moved.
EX. 07 · OPENAI · ANTHROPIC · SIMONWILLISON.NET
SAME NAME · DIFFERENT MACHINE
● ACT II · THE CEILING
EX. 08 · 15 / 27
DHH · creator of Ruby on Rails · a six-month reversal
The most respected developer alive changed his mind.
Summer 2025Skeptic
“I can literally feel competence draining out of my fingers. I spent more time rewriting what it wrote than if I’d done it from scratch.”
David Heinemeier Hansson · Lex Fridman Podcast
Jan 2026Convert
“This is the most exciting thing we’ve made computers do since we connected them to the internet. Half the resistance was simply that the models weren’t good enough yet. That has now flipped.”
David Heinemeier Hansson · X · Jan 2026
He did not lower his standards. The tools rose to meet them. Twenty years of architectural judgment, now multiplied. This is what the ceiling looks like.
EX. 08 · X.COM/DHH · JAN 2026 · LEX FRIDMAN, SUMMER 2025
NOT A RETRACTION. A CONVERSION.
● ACT II · THE CEILING
EX. 09 · 16 / 27
What professionals actually do
They’re not vibing.
Matt Pocock, TypeScript expert, documented his actual AI coding process. Five of seven steps are human. The AI only does the typing.
The AI is the fastest junior who ever lived. The senior is still the one who knows what to build and whether it was built right.
- 01Define the problemHuman
- 02Decompose into vertical slicesHuman
- 03Write precise specs for each sliceHuman
- 04Agent implementsAI
- 05Human reviews and validatesHuman
- 06Agent iterates on feedbackAI
- 07Human approves and shipsHuman
EX. 09 · M. POCOCK · 2026
FIVE OF SEVEN ARE HUMAN
● ACT II · THE CEILING
EX. 10 · 17 / 27
Google, on the inside
“Today, 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall. We’re now shifting to truly agentic workflows, orchestrating fully autonomous digital task forces.”
Sundar PichaiChief Executive, Google · Earnings call, Apr 2026
“Approved by engineers” is the operative phrase. Google did not replace its engineers. It multiplied them.
EX. 10 · FAST COMPANY · APR 2026
25% → 50% → 75% · OCT 24 · FALL 25 · APR 26
● ACT II · THE CEILING
EX. 11 · 18 / 27
Snap Inc. · April 2026 · who stayed, who went
AI writes 65% of Snap’s code.
They kept the seniors.
Stock went up 8%.
65%
code is AI-written
1,000
laid off · juniors, fully replaceable roles
300+
open roles closed
+8%
stock, same day
This is what the ceiling looks like in a P&L. The work the AI can do alone became cheaper. The work that needs a senior to architect, review, and ship safely became more valuable. The market priced both, instantly.
EX. 11 · TOWARDS AI · APR 2026
END ACT II
● ACT II · THE CEILING
THE PRO SHELF · 19 / 27
The pro shelf · B2C apps shipped by people who already knew how to ship
When pros vibe-code, they ship products. Not prototypes.
NO. 01 — NO. 03 · THE PRO SHELF
DESIGNERS · DEVELOPERS · BOTH
● THE VIBE CODING DISPATCH
ACT III · 20 / 27
Act three of three
Context.
Same tool. Different outcomes. The difference is not the agent. It is what you are building, what the stakes are, and what you brought to the conversation before you typed your first prompt.
● ACT III · CONTEXT
THE SPECTRUM · 21 / 27
The spectrum of vibe coding
Not all of this is the same thing.
Five levels · one tool. Stakes climb from left to right.
Level 01
Personal
utility.
Example
A script that renames your photos by date.
Stakes
Zero. If it breaks, you re-run it.
Who
Anyone.
Level 02
Simple
website.
Example
A portfolio site or landing page.
Stakes
Low. Worst case, it looks bad.
Who
Anyone with taste.
Level 03
Internal
tool.
Example
A dashboard for your team.
Stakes
Medium. Data accuracy matters.
Who
Someone who knows the domain.
Level 04
Public
web app.
Example
A SaaS product with users and payments.
Stakes
High. Security, uptime, privacy.
Who
Requires engineering knowledge.
Level 05
Corporate
production.
Example
Enterprise software at scale.
Stakes
Critical. Compliance, audits, millions of users.
Who
Requires professional engineers.
Most of the magic stories live on the left. Most of the failures live on the right. Conflating them is the mistake.
FRAMEWORK · ORIGINAL TO THIS TALK
FIVE LEVELS · ONE TOOL
● ACT III · CONTEXT
THE PATTERN · 22 / 27
What every failure has in common
It’s never the AI.
Incident
What was actually missing
Moltbook · 1.5M API keys leaked
No threat model
Founder bragged about not writing a line of code.
Replit · production DB deleted
No staged confirmation for destructive ops
Agent had prod credentials.
Moonwell · $1.78M loss
No senior review
Asset priced 1,964× too low went unchecked.
Escape.tech · 2,038 vulns
No security baseline
Default outputs shipped to public endpoints.
CodeRabbit · 1.7× more bugs in AI-authored PRs
No review process
Default AI output shipped straight to main.
Every column on the right is something a professional brings before the agent gets a key. The agent did not fail. It did exactly what it was asked to do, with no one to catch what it missed. That is the whole story of Act II versus Act III.
SYNTHESIS · FIVE INCIDENTS · FIVE MISSING INPUTS
IT’S NEVER THE AI
● ACT III · CONTEXT
WHEN IT WORKS · 25 / 27
What the professionals actually do
When vibe coding actually works, it looks like one of these two.
Cohort One
The craft-led.
Designers and senior programmers treating the agent as a junior pair. They have opinions about the design system before they prompt. They write a real spec, not three sentences. They read the diff, reject most of it, ask again. They refactor by hand after the agent stops. They throw the whole thing away if the structure is wrong. The work is indistinguishable from a small careful team.
Cohort Two
The infrastructure-led.
Senior engineers who know the failure modes by heart. Security. Networking. Distributed systems. They own the threat model before any agent gets a key. They run agents in sandboxes, never on prod. They parallelize across services, branches, environments. They write tests the agent can’t fake. They treat agents like contractors, not collaborators. Output that used to need a team of six over six months ships in a sprint.
Both groups have one thing in common. They were already good before the agent showed up.
MARTYNAS KAIRYS · FIELD NOTES · 2025 / 2026
CRAFT-LED · INFRASTRUCTURE-LED
● THE VIBE CODING DISPATCH
THE LEDGER · 26 / 27
The same tool · in two different hands
Both columns are true at the same time.
A. Brought taste, spec, threat model
Brought none of the above B.
- $6.6BLovable, paying for the category
- $29.3BCursor, AI-native editor
- 75%of new code at Google, approved by engineers
- 87.6%SWE-bench Verified, Opus 4.7 (up from ~65%)
- DHHchanged his mind on the record
- 6 mo → 2 wkteam of six in six months → one engineer + 12 agents in a sprint
- 2,038critical vulns in 5,600 scanned apps
- 1.5MAPI keys exposed in Moltbook hack
- $1.78Mlost to one Claude-coded contract
- 1.7×more bugs in AI-authored PRs (8× perf)
- 0bus factor. Nobody can maintain it
- 90 daysLemkin’s forced break from burnout
vs.
MARTYNAS KAIRYS · SAME TOOL · DIFFERENT HANDS
BOTH COLUMNS · 19 PRIMARY SOURCES
● THE VIBE CODING DISPATCH
IN CLOSING · 27 / 27
In closing
Vibe coding is easy.
Being good isn’t.
Sources, the vibe-coded tools, and everything we talked about:
login2026.martynaskairys.com
● MARTYNAS KAIRYS · LOGIN 2026
THANK YOU · Q & A →