im confused so you took something you didn't know how to do, said ai got 96% immediately on its first try (and this percent lights up your human reward function happily), and your takeaway is that ai is cool and great for this? not that someone who maybe does know what they're doing couldn't also get from 0-96 in 30 seconds?
The 96% on the first try isn't about what you called a "human reward function happily". It’s simply a starting point, as explained in the post.
I’m cautious about my takeaways. I'm regularly reminded of the limitations and less glamorous aspects. For example, see the 3rd item in the "Final Thoughts" section, or the 2nd paragraph in "AI + decomp.me". The introductions to the 2nd and 3rd chapters also bring these points up again if you want further context.
As mentioned, I shared these findings on Discord servers for decompilation. There are mixed feelings. Some of the folks think it can be useful and accelerate the work, while others believe better methods than LLMs exist. That’s fair. The whole point of this research is to find out what works.
Throughout, I actively shared my results, invited criticism, and incorporated that feedback here.
When I wrote this post, there was no ARM-to-C decompiler tailored for matching decompilation. m2c didn’t support ARM, so you couldn’t go from 0 to 96 in 30 seconds. And even if you could, just getting something in 30 seconds isn’t a meaningful metric. The real goal is a 100% match (whether with AI or not) with *a human-readable C code*. It's briefly discussed again in the 2nd and 3rd chapters.
With all that in mind, please: *read* the post. It seems like you quickly scrolled through and commented without reasoning about the content.
I was unsure whether to include the last paragraph in my reply because I was afraid of being unnecessarily aggressive, but thank you for showing me that it was the right call.
There is no point in being respectful with someone who makes questions and, in the next comment, says that they are not reading the answer.
im confused so you took something you didn't know how to do, said ai got 96% immediately on its first try (and this percent lights up your human reward function happily), and your takeaway is that ai is cool and great for this? not that someone who maybe does know what they're doing couldn't also get from 0-96 in 30 seconds?
The 96% on the first try isn't about what you called a "human reward function happily". It’s simply a starting point, as explained in the post.
I’m cautious about my takeaways. I'm regularly reminded of the limitations and less glamorous aspects. For example, see the 3rd item in the "Final Thoughts" section, or the 2nd paragraph in "AI + decomp.me". The introductions to the 2nd and 3rd chapters also bring these points up again if you want further context.
As mentioned, I shared these findings on Discord servers for decompilation. There are mixed feelings. Some of the folks think it can be useful and accelerate the work, while others believe better methods than LLMs exist. That’s fair. The whole point of this research is to find out what works.
Throughout, I actively shared my results, invited criticism, and incorporated that feedback here.
When I wrote this post, there was no ARM-to-C decompiler tailored for matching decompilation. m2c didn’t support ARM, so you couldn’t go from 0 to 96 in 30 seconds. And even if you could, just getting something in 30 seconds isn’t a meaningful metric. The real goal is a 100% match (whether with AI or not) with *a human-readable C code*. It's briefly discussed again in the 2nd and 3rd chapters.
With all that in mind, please: *read* the post. It seems like you quickly scrolled through and commented without reasoning about the content.
I was unsure whether to include the last paragraph in my reply because I was afraid of being unnecessarily aggressive, but thank you for showing me that it was the right call.
There is no point in being respectful with someone who makes questions and, in the next comment, says that they are not reading the answer.