I had REA produce a reverse engineered rebol-view-278-3-1.exe binary

Started by Nick on 2026-11-05 15:00:45
After successfully building and compiling R3 with full View features for WASM (https://rebolforum.com/topic/834), I've had more curiosity about how well new AI tools can re-implement old Rebol binaries. You may have recently heard about the full suite of Adobe products (Photoshop and many other apps), and Microsoft Office being duplicated by reverse engineering efforts using AI tools. This is a trend which I wanted to try my hand at a bit. So, I used REA (https://github.com/morluto/rea) in Pi, with Deepseek 4.1 Flash 1017, to produce a reverse-engineered version of R2 from scratch. There are a lot of edges to clean up, but the initial binary works surprisingly well already (with View, Vid, and many other features functional), after just a casual effort. Claude Opus 5.5 almost certainly would have done a much better job immediately, but I wanted to accomplish this without requiring any money to be spent, and I wanted to see how well Deepseek could complete this task. I'm surprised at how well Deepseek Flash 4.1 1017 did. It's an incredibly powerful model, which is as close to free as can be imagined, if you buy your inference from Cline-pass ($80 per year for an absolutely massive volume of tokens). At the moment, I'm not going to release the version of R2 which was created by the effort above, because I don't want to disrespect Carl and I certainly don't want any legal trouble - but anyone can do this. Just read a little about REA - or just point your agent to the REA repository linked above, and also point your agent to any copy of the Rebol 2 binary you want to target, and tell it to reverse-engineer a copy. I built binaries for both Windows and Linux, all cross-compiled on an Ubuntu VPS. You'll need a lot of iterations to complete this sort of work. I ran through 3 full Pi sessions, using nearly the full 1 million context length each time, to finish the effort above (my input throughout the process was basically to tell it to continue). One of the things I'm interested in doing is combining the knowledge gained in the R2 reverse-engineering process, with my WASM version of R3 that currently runs in browsers, and Marco Antoniazzi's http://www.rebol.org/view-script.r?script=vid1r3.r3, to get all the old R2 code running in a fully open source version of R3 (in an R2 compatibility mode), on any platform, including browsers. I also want to integrate all of Oldes's R3 features, including graphics and everything else (not just the core features) with the WASM/browser implementation Fable 5 and Kimi K3 got running for the original R3. This is all a side project for my own personal interest. AI based tools and workflows, with Python, have long ago fully replaced Rebol for all my production software development needs, but I love using those tools to dig into old Rebol implementations. Perhaps something productively useful will come of it for someone...
Nice.
I've been casually running more sessions to clean up the R2 binary, whenever my models get some down time. 7 sessions so far have been filled - the binary is getting much better :)

Reply

Replies are rate limited to reduce abuse.