Software Costs Millions to Make but Claude Cracked It in Seconds. (Pt - 1)

Value of the Information & Strength of the Argument

The video provides valuable insights into the mechanics of AI reverse engineering, breaking down why certain projects succeed and others fail. The argument is well-structured, using concrete examples and data to support claims. It avoids hype and honestly addresses limitations, such as the cost discrepancies and the dependence on verification oracles. The analysis of the MW2 case is particularly illustrative, showing that without an answer key, even a massive token budget leads to non-functional code. The discussion of SRE-Bench’s nuances is also valuable, highlighting that benchmark scores can be misleading.

Scientific Rigor, Source Quality, Title Accuracy

The video cites numerous sources, including GitHub repositories, arXiv papers, community forums, and independent analyses. These sources are relevant and support the claims made. The title, while catchy, is justified by the content: it does show that Claude (AI) can quickly generate a Photoshop alternative, but the deeper message is about the conditions required for successful reverse engineering. The video maintains a critical stance and does not overstate capabilities.

The title is attention-grabbing but the content actually addresses the claim by showing that AI can generate a Photoshop alternative quickly, while also explaining the conditions and limits of such successes.

8/10

The video presents a balanced and critical analysis of AI reverse engineering, citing specific sources and explaining benchmark nuances. It does not overhype claims and acknowledges limitations, making it fairly reliable.

• 00:00$85,207 of tokens, no playable map
• 00:36Why it looks like everything's open
• 01:2799.2% within four tries: the footnote
• 02:30What a compiler throws away
• 03:41Guess and check: the answer key
• 05:25Melee: six years of answer key
• 06:13PhotoCraft grades itself against Photoshop
• 07:18Modern Warfare 2 had no grader
• 08:21426 times: the subsidy under the demos
• 09:00A crash is an answer key too

• MirrorCode paper — Referenced as a source on AI matching decompilation.
• SRE-Bench paper — The benchmark behind GPT-6 Astra's 99.2% claim.
• Adobe Generative Fill internet requirement discussion — Shows Generative Fill runs on Adobe's servers.
• Gambiconf article on matching decompilation — Explains the compiler-as-answer-key approach.
• ArtCraft apps on GitHub — Home of the seven AI-built Adobe alternatives.
• PhotoCraft repository — The Rust-based clean-room Photoshop clone.
• Melee decompilation guide — Details how Melee reached 100% matching decompilation with AI.
• Kotaku article on TheFlow and PS5 bug — Example of AI-driven bug reporting and its consequences.
• CSA research on AI vulnerability discovery economics — Economics of AI finding bugs like in FFmpeg.
• Simon Willison's analysis of GPT-6 Astra — Independent evaluation of GPT-6 Astra's benchmark figures.
• MW2 decompilation post — Developer's account of the MW2 decompilation attempt.
• MW2 decompilation audit — Audit of token usage and results.
• ArtCraft launch post on Reddit — Announcement of the ArtCraft apps built with Claude Opus.
• Vals comparison of AI models on SRE-Bench — Independent benchmark measurements.

• PhotoCraft repository — Directly supports the claim that such a clone exists and works.
• SRE-Bench paper — Original benchmark details confirm the within-four-attempts nuance.
• Melee decompilation guide — Supports the answer-key explanation for Melee's success.