YT YouTube2Text

YouTube2Text → Transcripts → Yapay Zeka Çağında Artık Her Uygulama Kopyalanabilir: Tersine Mühendislik

Yapay Zeka Çağında Artık Her Uygulama Kopyalanabilir: Tersine Mühendislik — Transcript

by Avenox · 3,303 words · 496 segments · language tr · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00They made all these scenes with AI in
0:02the last 3 weeks. Minecraft inside
0:05Elden Ring, Spider-Man in Batman's city
0:07, Skateboarding in Call of Duty, and
0:08Skateboarding again in 2006's World of
0:10Warcraft. Millions of people watched
0:13this. I watched it too. So, what do you
0:15think is the real news here? Let me
0:16tell you, it's not the game. Definitely
0:18, this game part is just the
0:19fun-looking side of it. There is
0:20actually something much bigger behind
0:22these images. AI can now read inside
0:24closed programs. It opens up a program
0:26where you've only been given the
0:28running version, with no code at all,
0:30and explains what it should do. Then it
0:31can rewrite a new one from scratch that
0:33does the exact same thing. It can
0:34perform reverse engineering. Games are
0:36just a fun visual representation of
0:38this. The same applies to very
0:39different types of programs as well.
0:41Including the application you sell.
0:43Look, every piece of software is
0:44becoming capable of being rebuilt from
0:45scratch. I won't just explain this
0:47based on other people's videos. I had
0:49two AI agents rewrite Doom's graphics
0:50engine and compared it with the
0:52original using tests they'd never seen
0:54before. It made it pixel-for-pixel
0:57identical in 48x48. In this video, I
0:59will first explain how this is possible
1:00, then what is being done in games, and
1:02finally the most important part: what
1:04this means for everyone who makes
1:05software. Let me be clear. I didn't pay
1:07much attention when I first saw these
1:08either. I just said, "There's a
1:09different game inside a game," and
1:10moved on. But then I watched a video on
1:12this topic by a channel called AI
1:13Search and realized something
1:14incredible was happening. Actually,
1:16this video is inspired by that one. I
1:18opened the examples one by one. And I
1:19tested it myself. Let's understand the
1:21logic first. Look, a program is first
1:23written as source code. You can think
1:25of it like a composer's sheet music.
1:26Then it is compiled, meaning it's
1:27turned into a recording. Whatever
1:29application, game, or plugin you
1:30download to your computer is
1:31essentially just that recording. The
1:33sheet music remains in the company's
1:35vault. We only get the version the
1:36computer can read. Now, you can't
1:38listen to a recording and extract the
1:39sheet music directly, right? That is
1:40essentially the whole model behind
1:42closed software. Looking at the
1:43recording and rewriting the sheet music
1:44is called "Decompiling." It has two
1:46levels. Either you produce the exact
1:47bit-for-bit copy of the recording, or
1:49the recording is different but the song
1:50is the same. Meaning, the underlying
1:51code might be different, but the result
1:53is exactly the same. Keep this second
1:54one in mind because the main risk lies
1:56here. Let me explain why this wasn't a
1:58problem in the past. People started
1:59trying to rewrite the Super Smash Bros.
2:01game to be bit-for-bit identical in
2:032020. There are nearly 20,000 functions
2:05. By the end of 2025, they could only
2:07reach the halfway point. In December
2:092025, the first AI-contributed code
2:10started to appear. For the first time,
2:12AI began to enter the game, and the
2:14remaining half was finished in 8 months
2:16. So, do you understand the gravity of
2:17the situation? A process that took
2:18years suddenly accelerated by a huge
2:20margin. It's not just the AI's work
2:21alone, either. The number of human
2:22contributors has also increased, but I
2:24think we cannot ignore the general
2:25trend. What brought this to the
2:26mainstream was actually GPT-6 Astra,
2:27released on September 3rd. There was a
2:29test in GPT-6 Astra. It was called
2:31SR-BCH. You don't give the model the
2:32source code. You only give it the
2:33program. For example, you give it a
2:35game. There are six hidden events
2:36inside. It reads a dungeon game model
2:38program and rewrites the keys that
2:40trigger those events. The judge replays
2:42the game and checks it. Astra scored 99
2:44%in 4 attempts. Opus scored 98.1%in 5.
2:49It can basically rewrite the source
2:51code of the game without having the
2:52original. But these results were
2:54obtained with safety filters off, and
2:55the programs were on average 17,000
2:57lines. Photoshop is millions of lines.
2:58This doesn't mean it's something we can
3:00call very easy on its own. It doesn't
3:01mean every piece of software is
3:02finished. But there is actually a trend
3:03, and we want to understand this trend.
3:05Because otherwise, we are going to face
3:06very big problems. We see this at the
3:07highest level in games. The first is
3:09finding the door. Spider-Man is
3:10swinging in Batman's city. Actually,
3:11two games are open at the same time.
3:12Spider-Man is secretly running in the
3:14background. On the screen, there is
3:16Arkham Knight. The soldiers are yelling
3:17that Batman has arrived. Because Batman
3:19is standing in Spider-Man's place as an
3:20invisible puppet. So, we are actually
3:22just changing the actor there.
3:23According to the producer's notes, the
3:25AI read the 122 MB game file without
3:27opening the game at all. It found the
3:29three functions where the swinging
3:30happens and attached the hook there. It
3:32is not actually rewriting the game from
3:33scratch. It is just finding the right
3:34door. In software, this corresponds to
3:36the following: You no longer wait for
3:38the company to open a door for you to
3:39attach a plugin to a closed program.
3:41You can write a plugin without getting
3:42integration. Then someone packaged this
3:44. They named it Universal Mod. This is
3:46not actually a mod. It's a set that
3:48makes agents create mods. Each agent
3:49leaves a note for the next one. So far,
3:51131 notes have accumulated for 81
3:53different games. You blow up TNT in
3:56Minecraft. An explosion happens in GTA
3:57too. So, you can do absurd things like
3:59this. It got nearly 7,000 stars in two
4:01weeks. Now, the code for the Elden Ring
4:02example in the intro has not been
4:04shared to this day. But before two days
4:06were up, someone else rewrote the same
4:07idea as open source. There is actually
4:09nothing impossible here. Once the idea
4:10enters your mind, you can do it. There
4:12is no point in hiding the code either.
4:13The second level is rewriting the
4:15engine. Look, this is Call of Duty
4:16Modern Warfare 2, the soldier is
4:18skateboarding. Someone analyzed the
4:20original program with a tool called
4:21Ghidra and rewrote the engine in Rust.
4:23They said that a language model wrote
4:24this entire project. So, it was made
4:26with artificial intelligence. They
4:27shared it in a single commit. 344,000
4:29lines in a single commit. One of the
4:31interesting details is this: There were
4:32no rails to skate on in the maps. The
4:34guy said the code looks at the
4:35collision data and finds the edges to
4:37skate on itself. But let's also say
4:38that there are actually 2 years of
4:40research by two modders behind this
4:41skateboard. So, it's not something that
4:42appeared out of nowhere. If you can
4:44rewrite a game engine from scratch, you
4:45can also write a desktop application.
4:47The problem for us actually starts here
4:48. They did the same thing in World of
4:50Warcraft too. It was written from
4:51scratch in Rust. 642 commits, 636 of
4:54them have a cloud signature. So almost
4:56all of it was written with cloud. They
4:58most likely wrote the rest with Codex
4:59or something similar. Be sure of it.
5:00The person who shared it wrote that
5:01private servers are now legal. It's not
5:03actually legal. This is just a joke,
5:04and lawsuits are being filed over this.
5:06So, just because you can do something
5:07doesn't mean you should. The mere fact
5:09that it's possible here shows us that
5:10it will lead to very different
5:11situations. If we take it to the next
5:13level from here, here is what happens.
5:14Solving the entire program. One of the
5:16people doing this most seriously
5:18decompiled Modern Warfare from start to
5:20finish. At one point, he was running 17
5:22agents simultaneously. It took 3 months
5:24. By his own estimate, he burned close
5:26to 700 billion tokens. Look, 700
5:28billion. My total cloud and Codex usage
5:30isn't even 200 billion. Think about it
5:31that way. But the real valuable part is
5:33this. The lessons he learned were that
5:35when he first started, the AI's output
5:36was readable, but there were problems.
5:38Then he created a different system. He
5:40built a control layer that compared it
5:42byte-by-byte with the original. Do you
5:43know what the agents did first? They
5:45looked for a shortcut. They tried to
5:46change the control itself. Finally,
5:48when these were resolved as well, they
5:49solved 80%of the functions and ran a
5:51game that was byte-for-byte identical.
5:52When this guy shared it, his video was
5:54taken down with a copyright notice.
5:56These guys spend hundreds of millions
5:57of dollars. And in 3 months, one person
5:58and a group of agents solve it from
6:00scratch, I know this. Naturally,
6:01problems arise here. In games, this is
6:03actually the fun part. So, what is the
6:04situation with software? An American
6:05state official wrote this summer.
6:07Everything distributed will be readable
6:09or will soon become readable. Desktop
6:12applications, mobile apps, plugins, any
6:14software on your device; Anthropic
6:16already lists rebuilding source code
6:18from a closed program as a task models
6:20can easily handle on its own research
6:22page. I checked if this is actually
6:24happening right now. GitHub publishes
6:26copyright notices publicly. We already
6:28counted those that mention "decompile."
6:29Last September, it appeared in 5 out of
6:31224 notices. This September, it's 18
6:33out of 233, look, there is an
6:34exponential increase. Do you see? The
6:36numbers still look small. It doesn't
6:38look like an explosion yet. But this is
6:39just the beginning. Because models have
6:41only just started reaching this level.
6:43The real change is the list. Last year,
6:44it was things related to Minecraft.
6:46This year, there is also grocery store
6:47checkout software. There is also a
6:49stock market application. There is also
6:50a piece from a phone operating system.
6:51So, this has started to happen in every
6:53sector now. Reverse engineering is
6:55hitting everywhere. Let's look at three
6:57cases. Backdrop is a Mac application
6:58from a small Swedish company. Someone
7:00took the compiled version and rebuilt
7:01it by breaking it into 13 parts. One of
7:03the files is something meant to prevent
7:04this. It even protected that. Expo is a
7:06paid program sold for Minecraft. Even
7:08though it was locked with five separate
7:10protection tools, the source code was
7:11still extracted. Look, no matter how
7:13hard they try, they can't solve the
7:15problems. In Axiom, the notice
7:16explicitly mentions it's a clone
7:18produced with AI. A comment line first
7:20said "imitates Axiom." Then that part
7:22was also deleted. AI is explicitly
7:23mentioned in only two of these notices.
7:25I should also mention that. They are
7:26all small developers, and the entire
7:28value is inside the file sent to the
7:29customer. So what does this mean? If
7:31models continue at this speed, it seems
7:32there won't be much value left in the
7:34software itself, that is, in the code
7:35layer. One side says the value of
7:36software has dropped to zero. The other
7:38side says the price of the first 70%has
7:39fallen. Looking at it, the second one
7:41seems a bit more accurate for now.
7:42It’s not really a loss of value;
7:43it’s actually shifting. Service,
7:45community, and continuous updates will
7:46be the valuable layers from now on. For
7:48instance, the core logic of a web
7:50service is never actually sent to you;
7:51it stays on the server. I don’t
7:53recommend getting into one-time-sold,
7:55offline small tools, as they will face
7:57the most problems in this era. There is
7:59a bright side to this, actually. This
8:00situation will also work in favor of
8:01open source. Because with open source,
8:03the point was never really just the
8:04code, right? It was the community
8:05behind it and the development process.
8:07For example, I will make Sera open
8:08source when I finish it. Because in my
8:09eyes, the value is not in the code
8:10itself. It’s in the ecosystem
8:12surrounding it. If we get into a code
8:13war, we won't be able to survive it,
8:14friends. So, what will companies do?
8:16They will move the value to where it
8:17cannot be copied. But I’m not just
8:18talking about putting licenses and such
8:20on the server. They wrote a fake server
8:22that mimics the actual server for a
8:23stock market app. The part that does
8:24the actual work must be on the server.
8:26I mean, do you understand? Look, the
8:27main thing shouldn't be on the user's
8:28computer. We need to make this process
8:30expensive. According to tests by a
8:32company selling protection, AI finds 8
8:34out of 10 vulnerabilities in
8:35unprotected programs in 3 minutes. With
8:37protection, it struggles for nearly 7
8:38hours and finds none. But Expo had 5
8:40layers of protection. It still found it
8:41. Meaning, as models improve, they
8:43bypass this too. It makes it expensive,
8:44not impossible. Maybe you’ll use
8:45humans to some extent. Thirdly, they
8:47will likely get involved in lawsuits
8:48and such, but I don't think that alone
8:50will save the day. You are actually
8:51opening the door yourself. I think that
8:52is exactly what I am seeing. It will be
8:54open source. Generally, you won't hide
8:55the code too much. You will allow
8:57people to build an ecosystem around it.
8:59Mojang, for example, has started to
9:01stop hiding Minecraft's code like that.
9:02Of course, you will read your program
9:04before the attacker does. Apple,
9:05Microsoft, and Google are currently
9:07doing this in their own systems.
9:08Meaning, we now need to put more effort
9:10into the security of the code we write
9:11than into writing the code itself. As
9:13for whether this is legal, the rule is
9:15more or less the same in the US, Europe
9:16, Turkey, and everywhere else in the
9:18world. You can examine your own copy.
9:20You can also reverse-engineer just
9:21enough to make it talk to another
9:23program. So, there’s no problem with
9:24modding and the like for yourself. But
9:25sharing it with other people is
9:27forbidden. You cannot make a similar
9:28version and sell it. These are
9:29protected by copyright, of course. So,
9:31cracking a program and selling a clone
9:32is exactly what the law forbids. I
9:34think you shouldn't even get into it.
9:35You’ll both invite trouble and it's
9:36no longer necessary. We are heading
9:37toward a very different order. That’s
9:38why I stayed on the legal side of
9:39things in this video. A question popped
9:41into my mind while examining these
9:42examples. Is the program rewritten by
9:44the AI truly the same, or does it just
9:46look like it is? I ran some tests. What
9:48did the Human say? The code is readable
9:50but incorrect. I tested the second half
9:52in a legal way. ID Software opened
9:54Doom's source code in 1997. And made it
9:57GPL in 1999. I used Freedom as game
9:59data. I asked it to convert Doom's
10:01graphics engine from C to Rust. It
10:03needs to be pixel-perfect to the
10:04original for the agents. If we give the
10:06agent a score, it sometimes chases the
10:08score instead of the work. Meaning, it
10:09tries shortcuts to get the reward. So I
10:11ran the code in a locked box and gave
10:13the score using 48 cameras on 6 maps
10:15they'd never seen. I ran the Astra and
10:17Opus agents together. Astra finished
10:19the data layer in 7.7 minutes. Opus
10:21wrote the wall engine. 39 minutes later
10:24, I gave a secret test, and it scored
10:2648 out of 48. I mean, it literally did
10:28it exactly the same. Then, without
10:30telling anyone, I gave it Doom 2's data
10:32. Different monsters, different
10:33textures, and so on. This time, it got
10:3528 out of 28 again. The more
10:36interesting thing was this. The agent
10:38found an error in my judge. In the
10:39secret test, 32 of the 48 answer keys
10:41were generated with the wrong color.
10:43With the corrupted key, the score
10:45looked like 8 out of 48, but it was
10:47actually 27. It could have just stayed
10:48silent and crashed. Instead, it
10:49reported it. We updated our tests
10:51accordingly. So actually, the agent did
10:53it right. Our tests were wrong. There
10:55were a few pixel differences left in a
10:56few frames. The agents said the program
10:58there wasn't their fault, it was a 1993
11:00code memory overflow. I tested it. They
11:02were right. The new program only
11:04differs from the original due to the
11:05original's own error. Look, there is an
11:07error, but it's an error that also
11:08exists in the original. Do you see it?
11:10A separate judge agent read all the
11:12logs from start to finish. Not a single
11:14cheat exists. In 40 minutes, 8,000
11:15lines of code that behave exactly like
11:17the original were written. I also
11:18noticed this. In my own trials, Opus
11:205.5 falls back a bit on these kinds of
11:22tasks. Meaning it hands off the work to
11:24Opus 4.8 or something. You know, this
11:26is very common in new models,
11:27especially on the Anthropic side.
11:28Claude models can fall back quite a lot
11:30. Of course, there are filters on the
11:32OpenAI side, but for this type of work,
11:34the Astra side was better for me
11:35overall. But my guess is, Chinese
11:37models like Kimi or GLM will stand out
11:39in this field as they reach this level.
11:41Because they don't have restrictions at
11:43this level. To wrap up, friends, AI can
11:45now read inside a closed program and
11:47turn what it reads into a new program
11:49that behaves the same. Look, do you
11:51realize the significance of this?
11:52It’s going to get much bigger. We
11:53still have Opus 5.5, GPT Astra, and so
11:56on. Fable 5.5 will come. The next
11:58version of Astra will come. This
11:59process will accelerate even further.
12:01When these two combine, source code
12:02will no longer be a hidden defense. Why
12:04? There won't be that big of a
12:06difference between open source and
12:07closed code anymore. If you are
12:09developing software, assume that
12:10everything you distribute can be read.
12:12Build your value around the code.
12:14Meaning, your focus shouldn't be the
12:15software itself anymore. It should be
12:18the community around the software, the
12:19constant updates you provide, the
12:21proprietary data you use in the
12:23software, etc. Leave behind the
12:24assumption that the code you write is
12:26valuable on its own, friends. Whether
12:28you’ve open-sourced it or kept it
12:30closed doesn't matter that much. In
12:31fact, open-sourcing has certain
12:33advantages. We'll talk about this in
12:34the future. There are actually some
12:35very serious opportunities here. And
12:37for instance, another question comes to
12:38my mind. Now that everything can be
12:40reverse-engineered like this, will
12:41these software companies be able to
12:43maintain these valuation levels? For
12:45example, the first example that comes
12:46to mind is Adobe applications. If they
12:48can use reverse engineering to create
12:50Adobe applications at the same quality
12:52you know, do you think Adobe's value
12:53would stay at this level? Look, these
12:55are just questions. I mean, I'm not
12:57saying it will or won't happen. But
12:58what I mean is this. We should no
13:00longer be planning based solely on
13:01software. I think the main thing now
13:03should be the layers we build around
13:05the software. That is all I have to say
13:06in this video. I hope you enjoyed the
13:08video. I hope there were things that
13:09interested you. I will be following
13:10these reverse engineering developments.
13:12Because I think there are great risks
13:13and opportunities here. In general,
13:15please write your own thoughts in the
13:16comments as well. What do you think
13:17will happen in this process? I hope you
13:19liked the video. See you in the next
13:21video.

About this transcript

This page contains the full transcript of Yapay Zeka Çağında Artık Her Uygulama Kopyalanabilir: Tersine Mühendislik by Avenox, generated from the public captions YouTube serves with the video. The transcript has 3,303 words across 496 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.