Free YouTube Transcribe

Video transcript

The Photoshop Clone That Exposed the Truth About Vibe Coding.

TheAiExplainer · 2,198 words · 10 min read

WhatsAppTelegramX

Want to search this transcript, jump the video from any line, or download it as TXT, SRT, or VTT?

Open in the transcript tool

Full transcript

0:00In the last 2 weeks, a single developer

0:02in Atlanta published a free version of

0:04Photoshop, then Illustrator, Premiere,

0:07Lightroom, After Effects, InDesign, and

0:10Acrobat. Then he went after Microsoft

0:12with a Word clone, an Excel clone, and a

0:15PowerPoint clone. And he threw in

0:17AutoCAD and Pro Tools for good measure.

0:20All of it between September [music] 30th

0:22and October 7th. And according to the

0:25project's own roadmap, getting the Word

0:27clone to full parity with Microsoft's

0:29version, a product with decades of work

0:32behind it, would take somewhere around

0:34120 to 150 hours [music] of AI time.

0:39The headline everyone ran with was that

0:41he vibe coded the whole thing. And he

0:44even claims he one-shotted Photoshop,

0:46meaning one prompt and done.

0:49But when you look at how it was actually

0:51built, [music] that story falls apart in

0:54a way that's far more interesting than

0:55the hype. Because the apps that work

0:58best weren't built on vibes at all. And

1:00the ones that wobble tell you exactly

1:03where the limits of this technology

1:05really are. Let's start with who did

1:08this and why, because the motivation is

1:10oddly relatable. The developer is

1:13Brandon Thomas, who runs a small studio

1:15called ArtCraft, and by most accounts,

1:18it's mostly just him. He says the

1:21project began after he got stung by

1:24Adobe's subscription cancellation fees

1:26one too many times. So he did something

1:29that would have sounded insane 3 years

1:31ago. He pointed Anthropic's Claude Opus

1:355.5 at the problem, wrote the apps in

1:38the programming language Rust, and

1:40started rebuilding Adobe's lineup from

1:42scratch, publishing every one of them as

1:45open source on GitHub. The Git history

1:48lists Claude as a co-author on most of

1:51the commits.

1:52And that fee has a bigger backstory than

1:55one annoyed customer. In 2024, the

1:58Justice Department and the Federal Trade

2:00Commission sued Adobe, accusing it of

2:03hiding early termination fees, which on

2:05the plan the government says Adobe

2:07steered people toward could amount to

2:09half of your remaining payments if you

2:11canceled in the first year. Earlier this

2:14year, Adobe settled for $150 million

2:18with part of it going to free services

2:20for some customers, while saying it

2:23didn't break any laws.

2:25So, these clones didn't appear in a

2:26vacuum. They appeared in the middle of a

2:28pricing fight where regulators had

2:31already accused the company of making it

2:33hard to leave. So, how good are they?

2:35Honestly, it depends which one you ask.

2:38PhotoCraft, the Photoshop clone, handles

2:40layers, masks, adjustment layers, and a

2:43brush engine, and it opens and saves

2:46real Photoshop files. It's labeled early

2:48alpha. One outlet called it glitchy, and

2:51it has no support for raw photos and

2:54gaps in its AI features.

2:56On the Microsoft side, The Register

2:58tried the Word clone on a Surface laptop

3:01and said it worked pretty well with its

3:03own buttons and icons rather than copies

3:05of Microsoft's. It has the ribbon,

3:08styles, tables, track changes, and mail

3:11merge. So, these things are real,

3:13usable-ish, and unfinished, but there's

3:16a difference between the apps that

3:18jumped out at me. The Photoshop clone

3:20has been put through an enormous amount

3:22of checking, while the Word clone's own

3:25road map admits that nobody has tested

3:27it on real documents yet. That

3:30difference is the whole story. Let's

3:32deal with the one-shot claim first

3:34because it's the part that spread

3:35fastest [music]

3:36and the part that doesn't hold up.

3:38One-shotting means giving an AI a single

3:41prompt

3:42>> [music]

3:42>> and getting a finished product back. If

3:44that were what happened here, you'd

3:45expect a giant pile of code that nobody

3:48has verified. Instead, PhotoCraft has

3:51more than 1,700 automated tests, and its

3:54builders checked that opening and saving

3:56Photoshop files [music] doesn't corrupt

3:58them by running the code against large

4:01collections of real files. That isn't a

4:04vibe. That's about the least vibe-like

4:06thing you can do in software. And that's

4:09the twist. The apps that work weren't

4:11built by an AI freewheeling its way to

4:14something that felt right. They were

4:16built by an AI that was being graded

4:18constantly. Here's what that grading

4:20looks like in practice. The AI writes

4:23some code, then runs the tests. Hundreds

4:26fail. It reads exactly which ones failed

4:29and why, changes the code, and runs them

4:32again. Then again, and again, thousands

4:35of cycles without ever getting bored or

4:38tired or deciding it's good enough. It's

4:41a lot like tuning a guitar with a

4:42clip-on tuner. You don't need to

4:44understand music, you just keep

4:46adjusting until the needle sits in the

4:48middle. The AI doesn't have to know what

4:51Photoshop is for. It has to turn red

4:53into green. And as long as the tests are

4:56good, that process drags the code

4:58towards something that really behaves

5:01like the original. Which also means the

5:03whole thing is only as good as the

5:05tests. If the checks are shallow, the

5:07result is shallow. And a test that

5:09nobody wrote is invisible. Here's why

5:12that matters. Picture two students

5:14preparing for the same exam. The first

5:17writes practice [music] answers and then

5:19checks every one against a complete

5:21marking scheme. Instantly, thousands of

5:23times. The second writes answers and

5:26never finds out whether any of them were

5:28right. Same intelligence, same effort,

5:31and you already know who you'd bet on. A

5:34coding AI with a test suite is the first

5:36student. A coding AI told to make an app

5:39that feels like Photoshop is the second.

5:41And that explains something that looked

5:43like a coincidence.

5:45Why were the first big AI built apps all

5:48clones? Why not wild new inventions?

5:51Because cloning is easy mode. When

5:54you're recreating existing software, the

5:57answer key already exists. File formats

6:00like Photoshops and Words are publicly

6:02documented. The real apps are sitting

6:05right there to compare against. And the

6:07Word clones makers say it was built from

6:09public specifications and observation.

6:12Even its spelling and hyphenation data

6:14comes from a public domain word list.

6:17Every attempt can be checked. Does this

6:20file open? Does it save? Does it match

6:22what the original does?

6:24When you're building something new,

6:26nobody has written the answer key. And

6:28the AI has nothing to be graded against

6:31except your opinion.

6:33So, where does it fall apart? Wherever

6:36the answer key goes missing. Take the

6:38Word clone. Artcraft's own estimate is

6:41that it's at about 62% of real feature

6:44parity. And it's still pre-alpha.

6:46[music] It drops charts, SmartArt, and

6:49embedded objects when you save. And the

6:51line in its road map that matters most

6:54is the admission that the round trip

6:56hasn't [music] been tested on a

6:57collection of real documents. In other

7:00words, it hasn't been graded yet. And

7:03put the speed [music] in perspective.

7:05The road map estimates around 20 to 25

7:08hours of AI work to reach a first usable

7:11alpha [music] of Word. And Microsoft

7:13Word first shipped in 1983.

7:17A fast start isn't a finish line. The

7:20first stretch of a product like this is

7:22the visible part. And the last stretch

7:25comes from decades of customers hitting

7:27strange situations, much of which was

7:30never written down anywhere as a spec.

7:33The Excel clone has the same problem.

7:36Because a spreadsheet isn't just math,

7:39it's years of quirks that existing

7:41workbooks quietly depend on. And one

7:44review noted that whether the clone

7:46reproduces those quirks is simply

7:48unconfirmed.

7:50Macros written in Excel scripting

7:52language won't run on it at all. And the

7:55Photoshop clone doesn't run Photoshop

7:57plugins. Wherever there's a spec and a

8:00score, it's impressive. Wherever there

8:03isn't, it's guesswork. And the evidence

8:06from outside this project points the

8:08same way. Veracode tested AI-generated

8:11code for security and found that about

8:1445% of samples contain serious

8:17vulnerabilities.

8:18>> [music]

8:18>> And that number hasn't improved across

8:21multiple rounds of testing,

8:23>> [music]

8:23>> even while the same models got better at

8:25writing code that simply works. Research

8:28on large companies found [music] that

8:30developers using AI commit code three to

8:33four times faster than their peers, but

8:36introduce security findings at 10 times

8:39the rate. Faster and more dangerous. So,

8:43the problem isn't AI-written code. The

8:45problem is unchecked AI-written code. If

8:49this is changing how you see AI coding,

8:52hit subscribe. Because next, I want to

8:54show you the part of this story that

8:56nobody has been able to explain. In July

8:592025, in a widely reported incident, an

9:03AI coding agent from Replit deleted a

9:05company's entire production database in

9:07the middle of a code freeze.

9:09And when it was asked about it, the AI

9:11itself said it had made a catastrophic

9:14error in judgment.

9:16Replit CEO apologized publicly.

9:19That wasn't a failure of intelligence.

9:21It was a failure of guardrails. Nothing

9:24stopped the AI from doing something that

9:26nobody should have allowed.

9:28Compare that with a project where every

9:31change is graded by thousands of tests.

9:33It's the same kind of technology with

9:35completely different outcomes.

9:38Now for the part with actual drama. Is

9:40any of this legal?

9:42Arc Craft calls the apps clean room

9:44reimplementations,

9:46meaning they say they recreated the

9:47behavior without copying Adobe's or

9:50Microsoft's code [music] or assets.

9:53A legal expert quoted by The Register

9:55said that if the apps were built without

9:57access to the original source code,

9:59they're probably not infringing. But he

10:01also warned that this situation could

10:04quite possibly end up in court. Because

10:06the company with a grievance has deep

10:08pockets and a lot to lose.

10:11And that raises an obvious question.

10:13What does clean room even mean when the

10:16builder is an AI trained on a huge slice

10:18of the internet? Skeptics are already

10:21asking it. Microsoft acknowledged

10:23questions about the clones and said

10:25nothing more.

10:26Then something odd happened. After the

10:29Office clones appeared, the download

10:31links for Word Craft, Grid Craft, and

10:34Deck Craft disappeared from the

10:35project's website with no statement

10:38explaining why.

10:40The code is still on GitHub, so it

10:42hasn't vanished. But nobody outside the

10:44project knows whether this was a legal

10:46worry, a quality worry, or a developer

10:49tidying up.

10:50>> [music]

10:50>> As of this recording, nobody has said.

10:53So here's what I'd watch from here.

10:55Whether those download links come back.

10:58Whether anyone tests the Word clone

11:00against a pile of real documents and

11:02publishes the results. Because that one

11:05experiment would tell us more than any

11:07headline. And whether Adobe and

11:09Microsoft respond with lawyers, with

11:12lower prices, or with nothing at all.

11:15Let's zoom out because there's a bigger

11:17idea hiding in here. Wall Street has

11:19been obsessed this year with something

11:21called the SaaS Apocalypse. The theory

11:23that instead of paying software

11:25subscriptions, people would soon just

11:27build their own software. This is the

11:29first real [music] test of that, and the

11:31result is a split decision. On one side,

11:35one person with an AI got further in a

11:37week than a hobbyist team could have

11:39managed before. Open source projects

11:42like and Krita took many years to

11:44reach a similar place. And it isn't a

11:46one-off. A separate solo developer's

11:49free Photoshop alternative called Photon

11:52reportedly passed 20,000 users,

11:55according to a post on Reddit. On the

11:57other side, none of these clones can

11:59touch the things that actually keep

12:01people inside Adobe and Microsoft. They

12:04don't run the plugins. They don't run

12:06the macros. And Microsoft 365 isn't just

12:09Word and Excel. It's email, Teams,

12:12SharePoint, identity and device

12:14management. A whole infrastructure that

12:17a desktop clone can't reach. One

12:20reviewer described it as the clones

12:22chipping away at the feature wall while

12:24the catalog wall stays standing. The

12:27code was never the whole moat. What

12:29changed is that the code part just got

12:31cheap.

12:33Before I give you my verdict, here's a

12:35tool for the next time you see a

12:37headline that says built with AI. Ask

12:40three questions. Is there a spec,

12:42meaning a definition of correct that

12:44existed before the AI started? Is there

12:47a test suite that automatically checks

12:50the work? And how big is it? And did a

12:52human who understands the field actually

12:55look at the results? If the answer to

12:57all three is yes, you're probably

13:00looking at serious engineering. If it's

13:02no to all three, you're looking at a

13:04demo, [music] however polished it looks.

13:08So, is vibe coding the future? Here's my

13:10read. Vibe coding, the original meaning,

13:13where you [music] give in to the vibes

13:15and forget the code even exists, which

13:17is how the term's creator, Andrej

13:19Karpathy, described it in early 2025,

13:22[music] is a fantastic way to make a

13:24prototype and a terrible way to make

13:27Photoshop. What actually built these

13:29apps looks much more like what veteran

13:31engineers like Kent Beck [music] call

13:34augmented coding, where the AI writes

13:36the code, but humans insist on tests,

13:39review, and oversight. The AI did the

13:42typing, the tests did the judging. And

13:45that's the real shift. The scarce skill

13:47is no longer writing code. It's knowing

13:50what correct looks like, and being able

13:52to write it down in a form a machine can

13:54be graded against.

Recently added transcripts

Browse the whole transcript library Browse all English transcripts

This transcript was generated from the captions YouTube publishes for this video. Get the transcript of any YouTube video atfreeyoutubetranscribe.com, free, unlimited, no sign-up.