Latest / AI Fire Daily / #120 Robin: REA Went Viral for “Cloning Any App” - I Tested It Against Claude Code

Transcript

#120 Robin: REA Went Viral for “Cloning Any App” - I Tested It Against Claude Code · AI Fire Daily

• 0:00
  What if someone told you software is done and
• 0:03
  you could clone an entire tech business in 45
• 0:05
  minutes? That's a bold claim. It is. Today we
• 0:09
  test if that's reality or just hype. Welcome
• 0:11
  to the deep dive. Glad to be here. I'm your host.
• 0:16
  Today we're looking at something that feels like
• 0:18
  a fundamental shift, a shift in how we think
• 0:20
  about building digital tools. We're unpacking
• 0:23
  an AI coding workflow called REA. Right. Reverse
• 0:26
  engineer anything. Exactly. And if you're listening
• 0:29
  right now, whether you write code or just use
• 0:31
  apps to do your job, this directly impacts you.
• 0:34
  It absolutely does. I mean, REA went completely
• 0:37
  viral recently. It grabbed over 60 ,000 GitHub
• 0:40
  stars in just a matter of days. That velocity
• 0:43
  is insane. Right. It's almost unheard of in the
• 0:45
  developer community. So we're going to explore
• 0:48
  exactly how this works under the hood. We'll
• 0:51
  look at some real -world stress tests, like how
• 0:53
  it cloned Apple's preview app and Whisperflow,
• 0:56
  which is a highly complex privacy -focused transcription
• 0:59
  tool. But we also need to cut through the hype.
• 1:02
  Yeah, figure out what's actually new here. Exactly.
• 1:05
  And finally, we'll answer a massive existential
• 1:08
  question. If code is really this easy to copy,
• 1:11
  what actually makes a software business valuable
• 1:13
  anymore? That's the real question. Let's start
• 1:16
  by grounding this in the mechanics. OK. Before
• 1:19
  we look at the cloned apps, I want to understand
• 1:22
  the fundamental shift. Typically, if I ask an
• 1:24
  AI to build something complex, I have to be the
• 1:27
  architect. You do, yeah. I have to write a highly
• 1:29
  detailed, rigid spec, or the AI just gets completely
• 1:33
  lost. Right, and that is the major blind spot
• 1:36
  for traditional AI agents. They rely entirely
• 1:39
  on your blueprint. If you leave out a crucial
• 1:42
  detail, like how the database should handle simultaneous
• 1:44
  users, the AI just... guesses. It's usually a
• 1:47
  bad guess. Exactly. It bases that guess on generic
• 1:50
  patterns from its training data, which inevitably
• 1:53
  leads to bloated buggy code. Code you have to
• 1:56
  spend hours debugging anyway. Yes. But RIA bypasses
• 2:00
  that entire guessing game. Instead of waiting
• 2:02
  for a perfect blueprint, it goes out and finds
• 2:04
  empirical evidence. Evidence from the app itself.
• 2:08
  Right. It feeds the AI the actual compiled data
• 2:11
  from an existing application's code and behavior.
• 2:14
  So rather than asking the AI to imagine how a
• 2:17
  house should be built, it looks at a finished
• 2:20
  house. It draws its own highly accurate blueprint.
• 2:24
  That's a great way to put it. But how accessible
• 2:26
  is this? If you're listening and want to try
• 2:29
  this on your own machine, is it a nightmare to
• 2:32
  set up? That is the crazy part about this workflow.
• 2:34
  It is incredibly accessible. Really? Yeah. If
• 2:38
  you have a recent version of Node .js running,
• 2:40
  specifically version 24 or higher, You're basically
• 2:43
  ready to go. And you can check that easily, right?
• 2:45
  Super easy. Just type node -v in your terminal.
• 2:48
  You can update it via homebrew if needed. Okay,
• 2:51
  but what if you don't want to touch the terminal?
• 2:53
  You don't even have to. Right. You can literally
• 2:55
  just ask Claude code to handle the entire setup
• 2:57
  for you. Wait, Claude can install it? Yeah. You
• 3:00
  tell it to download, install, and configure REA.
• 3:03
  It will even set up the heavy -duty companion
• 3:05
  tools. Like Hopper Disassembler? Exactly. Completely
• 3:08
  on its own. I want to pause on that interaction
• 3:10
  for a second. How exactly is a modern language
• 3:14
  model interacting with a desktop disassembler?
• 3:17
  Good question. There has to be some sort of bridging
• 3:19
  mechanism there. Yes. REA acts as the connective
• 3:22
  tissue. It connects the AI, like Claude Code
• 3:25
  or Cortex, to these reverse engineering tools
• 3:28
  via an MCP layer. Let me just clarify that term
• 3:31
  for our listeners. An MCP layer is a digital
• 3:34
  bridge letting AI use outside tools to gather
• 3:37
  data. Perfectly accurate. The AI walks across
• 3:40
  that bridge. It commands those reverse engineering
• 3:42
  tools to inspect the target application. Right,
• 3:45
  because compiled apps are just machine code.
• 3:47
  Exactly, ones and zeros. AI can't read that natively
• 3:51
  any better than we can. So these tools do the
• 3:53
  heavy lifting. They translate those ones and
• 3:55
  zeros back into assembly language. Then they
• 3:58
  return what we call pseudocode back across the
• 4:00
  bridge to the AI. And pseudocode is basically
• 4:03
  a readable outline, right? Yes. A high -level
• 4:05
  outline of the underlying logic. It strips away
• 4:08
  the machine code and just shows the core rules
• 4:10
  of how the app functions. It's almost like unstacking
• 4:13
  Lego blocks of data to see the original instruction
• 4:16
  manual. I love that analogy. You start with the
• 4:19
  finished castle and break it down to see precisely
• 4:22
  how those specific bricks fit together. It's
• 4:24
  exactly what's happening. And if the tools hit
• 4:26
  a wall, if they can't decompile something perfectly,
• 4:30
  the AI takes over. Oh, it adapts. Right. It uses
• 4:33
  the forensic evidence it has, writes its own
• 4:35
  code, tests the result, and improves it iteratively.
• 4:39
  Until it works. Yep. And this applies to virtually
• 4:41
  everything. Desktop, web, mobile apps, games,
• 4:45
  even hardware interfaces. Does this mean the
• 4:48
  AI is actually reverse engineering or just reading
• 4:50
  the outputs of other tools? The AI itself isn't
• 4:53
  doing the heavy reverse engineering. The AI is
• 4:55
  just the brain reading the data generated by
• 4:57
  those specific tools. So the AI just reads the
• 5:00
  clues the other tools dig up? Precisely. It's
• 5:03
  the detective reading the forensic report. But,
• 5:06
  you know, reading a report is one thing. Building
• 5:08
  something functional based on it is entirely
• 5:11
  different. Which brings us to the actual application.
• 5:15
  Theory is great, but I want to know what happens
• 5:17
  when this tries to rebuild a tool we all recognize.
• 5:20
  Let's talk about Apple Preview. This is where
• 5:22
  the source material gets really fascinating.
• 5:25
  The author wanted to see the tangible difference
• 5:27
  here. Between standard AI and REA? Exactly. So
• 5:32
  they ran a side -by -side race. They tasked two
• 5:34
  different cloud code agents with recreating Apple's
• 5:37
  Preview app. Preview is a brilliant benchmark
• 5:39
  to choose. It really is. Because if you just
• 5:42
  casually use it, it seems like a very basic image
• 5:45
  viewer. But under the surface, there's massive
• 5:47
  memory management happening to make it smooth.
• 5:49
  Right. You aren't just slapping a static image
• 5:52
  on a screen. You have to handle dynamic zooming
• 5:54
  without pixelation. Rotation states. Smooth drawing
• 5:57
  loops for annotations. And saving those changes
• 6:00
  without corrupting the file. So how did the race
• 6:02
  go? Well, Agent 1 used the REA workflow. Agent
• 6:06
  2 was vanilla -clawed code, just relying on its
• 6:09
  internal knowledge. I'm very curious how the
• 6:12
  standard model performed without the forensic
• 6:13
  data. The vanilla -clawed actually finished first.
• 6:16
  It took about two -thirds of the time. Oh, really?
• 6:18
  Yeah, but the final product was just okay. It
• 6:22
  produced a working viewer, but it was very basic.
• 6:25
  Laggy? Yeah. Especially when zooming in on large
• 6:28
  files. It felt like a rough prototype. The AI
• 6:31
  was just guessing at the optimal architecture.
• 6:34
  Because it didn't know the specific solutions
• 6:36
  Apple engineered. And how did the REA agent compare?
• 6:39
  It took a few minutes longer. But it created
• 6:42
  an almost identical, highly optimized clone of
• 6:45
  Preview. The drawing and annotations were incredibly
• 6:48
  smooth. It handled the zoom, the rotation, saving
• 6:51
  features perfectly. Because it didn't have to
• 6:53
  guess. Exactly. It saw the literal, highly optimized
• 6:56
  architecture the Apple engineers designed and
• 6:58
  mapped its code to that logic. Whoa. Imagine
• 7:03
  an AI checking its own UI work like a human QA
• 7:06
  tester. That is wild. It really is. We're so
• 7:10
  used to AI generating text and just stopping.
• 7:12
  Seeing it act as a developer, tester, and debugger
• 7:15
  in one loop is a massive paradigm shift. And
• 7:18
  they pushed it further, right? Yes. With Whisper
• 7:21
  Flow. Yes. A second test on a popular voice -to
• 7:24
  -text app. Whisper Flow relies heavily on continuous
• 7:28
  background processing. It's not just a static
• 7:30
  window. Right. It integrates deeply with your
• 7:33
  operating system. The author told Claude to find
• 7:35
  Whisperflow, inspect its binaries using RIA,
• 7:39
  and use the official website for context. Basically
• 7:42
  saying, explain how this works, and build me
• 7:44
  a version based on that evidence. Exactly. And
• 7:47
  that is a significantly heavier lift than a PDF
• 7:50
  viewer. Did it actually manage to clone it? It
• 7:53
  nailed it. The clone included the transcription
• 7:55
  history interface, it built a dedicated note
• 7:58
  taker for meetings. What about the system shortcuts?
• 8:01
  Captured them perfectly. It replicated the real
• 8:04
  -time dictation triggered by pressing option
• 8:06
  plus one. Handling audio routing is notoriously
• 8:09
  tricky. Did it manage different languages? It
• 8:13
  did. English and Chinese in the exact same input
• 8:16
  stream. No manual language switch needed. That's
• 8:19
  impressive. But the most crucial part, especially
• 8:22
  for privacy, it processed all that audio locally.
• 8:26
  Really? It didn't send recordings to the internet?
• 8:29
  Not a single one. That is a massive detail. Sending
• 8:33
  raw meeting audio to a third -party server is
• 8:36
  a huge security risk. But building local audio
• 8:39
  processing from scratch requires deep hardware
• 8:41
  optimization. It's incredibly difficult. But
• 8:44
  Ria saw exactly how Whisperflow solved that hardware
• 8:47
  problem and replicated it perfectly. It made
• 8:50
  a privacy -first alternative in a fraction of
• 8:52
  the time. Is the speed trade -off really worth
• 8:54
  it if vanilla AI is faster? Yes, if you need
• 8:58
  exact, faithful features rather than just a rough
• 9:00
  prototype. So RIA trades a little speed for a
• 9:03
  lot of precision. Exactly. For a quick wireframe,
• 9:06
  use standard AI. For a functional clone, use
• 9:09
  RIA. These tests sound like pure magic. And the
• 9:12
  GitHub stars reflect that hype. But we need to
• 9:15
  ground this a bit, I think. We definitely do.
• 9:17
  Because if we pull back the curtain, this technology
• 9:20
  isn't a breakthrough from last week. Right. Time
• 9:22
  for a reality check. RIA is getting massive attention,
• 9:25
  but the heavy lifting is being done by very old
• 9:28
  tools. Like Ghidra. Exactly. Ghidra was used
• 9:31
  in the preview test. It's a powerful reverse
• 9:33
  engineering suite, but it was publicly released
• 9:35
  by the NSA back in 2019. The NSA built it as
• 9:39
  a cybersecurity weapon to analyze malware. Yes.
• 9:42
  And the other tools, Hopper Disassembler, has
• 9:45
  been around since 2011. IDA dates back to the
• 9:48
  1990s. Wow. Security researchers have used these
• 9:52
  for decades. So, RIA isn't inventing the wheel.
• 9:55
  The engine breaking down the code is decades
• 9:57
  old. RIA is just building a steering column so
• 10:00
  the AI can drive. That is the perfect way to
• 10:02
  look at it. RIA is simply an orchestration layer.
• 10:06
  It makes historically difficult tools easy for
• 10:08
  a modern language model. It's still impressive,
• 10:10
  though. It is. I mean, I have to admit, I still
• 10:13
  wrestle with getting an AI to format a basic
• 10:15
  table, and this thing is rebuilding Mac OS apps.
• 10:18
  It is deeply humbling. It puts our daily prompt
• 10:21
  struggles into perspective. For sure. But beyond
• 10:24
  the age of the tools, vanilla AI models are catching
• 10:27
  up so fast. They are. An orchestration layer
• 10:30
  like RIA might not be necessary for long. Critical
• 10:33
  point there. The visual and reasoning capabilities
• 10:35
  of base models are expanding rapidly. You can
• 10:38
  already do a lot of this cloning without RIA.
• 10:41
  Just by feeding it screenshots? Yeah. Or a video
• 10:44
  of the interface. Claude can study the visual
• 10:46
  behavior and recreate the features just by looking
• 10:48
  at them. It might like the deep backend optimization
• 10:51
  we have finds. But for modern web apps, the visual
• 10:54
  layer is the primary product anyway. Right. And
• 10:57
  honestly, relying on the AI to generate the backend
• 11:00
  from scratch is sometimes better. Legacy apps
• 11:02
  have years of technical debt. Bloated code. Exactly.
• 11:06
  An AI implementation might actually be simpler
• 11:08
  and cleaner than the original. There is also
• 11:11
  a massive elephant in the room here. The legal
• 11:13
  reality. Oh, absolutely. The technical ability
• 11:16
  to clone an app doesn't erase the law. The legal
• 11:19
  catch is severe. Experimenting on your own machine
• 11:22
  for education is one thing, but cloning commercial
• 11:24
  software to sell, you are stepping into a minefield.
• 11:28
  Copyright risks, patent infringement, licensing
• 11:32
  violations, terms of service breaches. It's huge.
• 11:35
  It's the difference between clean room reverse
• 11:37
  engineering, where you study a concept and build
• 11:40
  an original version, and just letting an AI copy
• 11:42
  proprietary logic. One is standard practice.
• 11:46
  The other is a lawsuit. Exactly. technical capability
• 11:49
  doesn't mean legal right to monetize. Are we
• 11:52
  just in a temporary window where REA is useful
• 11:55
  before AI does this natively? Exactly. REA is
• 11:58
  just a highly effective stopgap while base models
• 12:01
  improve. Right. It's a clever shortcut until
• 12:03
  the AI models catch up. And that realization
• 12:05
  leads us to a much bigger paradigm shift. Because
• 12:08
  if AI is inevitably going to make copying software
• 12:11
  trivial, what happens to the entire software
• 12:13
  industry? That is the million dollar question.
• 12:16
  the fundamental physics of building a tech business
• 12:18
  are changing. Historically, the code itself was
• 12:21
  the protective barrier. Writing robust software
• 12:23
  was expensive and slow. But now, code is becoming
• 12:27
  incredibly cheap. The cost is plummeting towards
• 12:30
  zero. I want to push back on this a bit, actually.
• 12:33
  Code might be cheap now, but massive companies
• 12:35
  aren't panicking just because someone can clone
• 12:37
  their UI. That's true. If I use RIA to clone
• 12:41
  Spotify by tomorrow morning, Spotify isn't going
• 12:43
  bankrupt. What is actually protecting them? That's
• 12:46
  exactly the right framing. Recreating a SaaS
• 12:49
  app is no longer the hard part. Earning the customer's
• 12:52
  trust to switch to your clone is the real mountain.
• 12:55
  So what's the moat? The source material outlines
• 12:58
  four pillars of a true business moat. The first
• 13:01
  is brand. Brand is essentially a shortcut for
• 13:04
  trust, right? Exactly. If my Spotify clone works
• 13:07
  perfectly, I still have no reputation. People
• 13:10
  trust the Spotify brand to secure their payment
• 13:12
  info and deliver music. A clone has zero trust.
• 13:16
  Brand recognition takes years of consistent delivery.
• 13:19
  AI cannot automate that history. The second pillar
• 13:22
  is users and community. A network effect. Right.
• 13:26
  The value of a network is the people in it, not
• 13:28
  the code. A cloned marketplace without users
• 13:31
  is just a ghost town. You can copy the stadium,
• 13:34
  but you can't copy the fans. What's the third
• 13:37
  pillar? Proprietary data. Arguably the most powerful
• 13:41
  one. Unique, accumulated information makes a
• 13:44
  product hard to replace. Like a CRM you've used
• 13:46
  for five years. Yes. It holds all your custom
• 13:49
  workflows, your history. Even if a fresh clone
• 13:52
  has the exact same features, it has zero of my
• 13:55
  data. The friction of migrating is too high.
• 13:58
  The data is the lock -in. Right. And the final
• 14:02
  pillar is distribution and marketing. Knowing
• 14:05
  how to reach paying customers. Because you can
• 14:07
  have the best code in the world, but if nobody
• 14:09
  knows it exists, it doesn't matter. Exactly.
• 14:12
  Take a wrapper around chat GPT for marketing
• 14:15
  emails. It's a nice UI, but it's fragile. Tomorrow,
• 14:19
  OpenAI adds email formatting natively, or someone
• 14:22
  clones your UI. Your value vanishes overnight.
• 14:26
  So if you're listening and thinking about that
• 14:28
  SaaS product you just paid $50 a month for, someone
• 14:32
  could probably build its core functionality by
• 14:34
  this evening. Easily. But the company is valuable
• 14:37
  because of human -centric factors. Brand trust,
• 14:41
  community, proprietary data, marketing reach.
• 14:44
  Which means we need to look at where the real
• 14:46
  opportunities actually lie with REI. It's not
• 14:49
  in selling knockoffs. Where do you see developers
• 14:52
  funding value in this workflow today? The most
• 14:55
  exciting part is internal experimentation. You
• 14:57
  can build complex custom tools strictly for your
• 15:00
  own internal workflows. If your team needs a
• 15:03
  specific scraping utility, you don't need a thousand
• 15:05
  -dollar enterprise subscription. You just generate
• 15:08
  the tool in an afternoon. You can also ship prototypes
• 15:10
  much faster. If you want attention on Axe or
• 15:13
  GitHub, materializing a working concept quickly
• 15:16
  is a massive advantage. Absolutely. It's an incredible
• 15:19
  prototyping engine. But you can also use it purely
• 15:22
  to learn. to study great software under the hood.
• 15:26
  Yes. It's like a master class in software architecture
• 15:29
  tailored to any app you want. And you can remix
• 15:32
  those lessons. The source material gave a wild
• 15:35
  example, combining the building mechanics of
• 15:37
  Minecraft with the movement of Attack on Titan.
• 15:40
  Right. When the coding barrier is gone, you can
• 15:43
  mash up completely different concepts, things
• 15:45
  that would normally take a studio years to develop.
• 15:48
  It becomes an exercise in pure imagination. If
• 15:51
  people just build expensive enterprise tools
• 15:53
  for themselves internally, doesn't that still
• 15:56
  kill the SaaS business model? It hurts basic
• 15:59
  tools, but complex tools rely on constant updates
• 16:02
  and support that internal teams don't want to
• 16:04
  manage. Yeah, nobody wants to be the permanent
• 16:06
  IT guy for a cloned app. Exactly. Software breaks.
• 16:09
  Maintenance and security patches are a headache.
• 16:12
  People pay SaaS companies to make those headaches
• 16:14
  go away. We're entering a deeply fascinating
• 16:17
  era. AI rebuilding software on demand really
• 16:20
  does change the fundamental physics of the industry.
• 16:23
  It really does. For decades, writing complex
• 16:25
  code was the ultimate barrier to entry. RIA is
• 16:29
  undeniable proof that this barrier has essentially
• 16:32
  fallen to zero. The code itself is rapidly becoming
• 16:35
  a commodity, but as we've discussed, the human
• 16:37
  barriers remain completely intact. Building a
• 16:40
  brand, attracting a loyal audience, earning trust
• 16:43
  to collect valuable data. That still takes years
• 16:46
  of real grinding work. AI cannot automate trust.
• 16:50
  That is such a vital takeaway. The source material
• 16:53
  actually leaves us with a specific challenge.
• 16:56
  Pick an application you use every single day.
• 16:59
  Try rebuilding a small part of it, twice. Once
• 17:03
  with REA, and once without it using standard
• 17:05
  AI. Right. See which version actually wins in
• 17:08
  practice. Experience the tangible difference
• 17:10
  between spec -driven AI, where you hold its hand,
• 17:13
  and evidence -driven AI, where it learns from
• 17:16
  the real world. I guarantee it will completely
• 17:19
  change how you view software development. What
• 17:21
  if someone told you software is done? After today,
• 17:25
  it's clear the mechanical act of writing code
• 17:28
  might be. In a future where every piece of software
• 17:30
  is essentially free to clone, your most valuable
• 17:34
  asset isn't what you can code. It's who you are,
• 17:38
  the trust you build, and the unique data you
• 17:40
  gather. Invest heavily in what can't be copied.
• 17:44
  Thank you for diving deep with us today.