>she

>reverse
>"68k" to "x86"
seems like an easy job, it's already practically reveresed

File: figure7_agent_workflow.png (82 KB, 1589x967)

82 KB PNG

>>106741614
AI agents workflow.

From the paper (zenodo_paper_markdown.md in the zip file):
>1.2 What this paper shows
>We report a pragmatic reconstruction of Apple System 7.1 from 68k binaries to a bootable desktop in three days, using off-the-shelf AI assistants under strict controls. Critical ingredients:
>· A team of narrow, accountable agents with explicit outputs and cross-checks (§3)
>· A binary-first evidence ledger—addresses, strings, xrefs, checksums—driving every function and struct (§3.2)
>· QEMU boot loops with serial capture and visual checkpoints for concrete, reproducible milestones (§5.2)
>· Faithful recovery of historic resources (PPAT/PPIX, NFNT/Chicago, ICON/ICN#), not approximations (§4)

Agents used:
>Claude Code, ChatGPT, Qwen-3

chicago fonts are comfy

>>106741614
So, uhm... 68k ASM to x86 ASM, uh-huh, yeah.

Anyhow...
>no mention which LLM
>tranny
>Apple
something ain't sitting right, this feels like "ghost-writing."

>>106741664
Ah, they're trying to land some AI job, gotcha.

>>106741679
>1.3 What this paper is not
No new neural architecture. No dreamy “end-to-end” magic. No unverified accuracy claims. We build on classic RE practices (Canfora & Di Penta, 2007; Chikofsky & Cross, 1990; Comella-Dorda et al., 2000; Nelson et al., 2005) and acknowledge the academic neural-decompilation work (Hosseini & Dolan-Gavitt, 2022; Cao et al., 2023; Dramko et al., 2025; Armengol-Estapé et al., 2023; Benali, 2022; Katz et al., 2019) as context—but our speedup is a by-product of orchestration and accountability, not secret sauce.


>>106741672
>So, uhm... 68k ASM to x86 ASM
??? I see a bunch of C files in the repos:
https://github.com/Kelsidavis/System7/tree/main/src

They used Ghidra and radare2:
>3. Methodology: Agent Accountability Architecture
>3.1 Why agents, not a single chat
>Unconstrained LLM sessions drift. We counter drift with separation of concerns and mutual accountability. Each agent has a narrow mandate, produces machine-checkable artifacts, and is checked by peers. No artifact no downstream progress. This mimics good engineering: evidence implementation test audit.

>3.2 The agents, their outputs, and how they check each other
>(A) Evidence Curator
>· Input: disassembly (Ghidra/radare2), strings, xrefs
>· Output: evidence.curated.json listing function start/end addresses, string refs, and SHA-256 of the relevant binary slices
>· Controls others: the Compliance Auditor later rejects any function not present here; the Code Editor cannot work without an evidence ID

File: 1688380850267303.jpg (35 KB, 500x500)

35 KB JPG

>>106741744
It's a bit hard to tell if
>this was sponsored by AI/Apple/etc.
>is being heralded by AI-proponents because of AI-usage
>the authors want to actually excuse their own use of AI
>>106741744
>??? I see a bunch of C files in the repos:
My friend, qu'est-ce que c'est que tu fais "engenerieux en reverse," honhonhon.

Why are you retards unable to verify stuff by yourselves? Why do you always need to be spoonfed?

$ git clone https://github.com/Kelsidavis/System7 && cd System7
...
$  grep -rna asm
src/SystemInit.c:596:        __asm__ volatile ("cli; hlt");
src/PS2Controller.c:82:    __asm__ volatile ("outb %0, %1" : : "a"(val), "Nd"(port));
src/PS2Controller.c:87:    __asm__ volatile ("inb %1, %0" : "=a"(ret) : "Nd"(port));
src/main.c:137:    __asm__ volatile ("outb %0, %1" : : "a"(val), "Nd"(port));
src/main.c:142:    __asm__ volatile ("inb %1, %0" : "=a"(ret) : "Nd"(port));
src/main.c:2099:        /* __asm__ volatile ("hlt"); */
src/all_stubs.c:145:        __asm__ volatile("cli; hlt");
src/MenuCommands.c:151:            __asm__ volatile(
src/MenuCommands.c:160:            __asm__ volatile(
src/MenuCommands.c:168:            __asm__ volatile("cli; hlt");
src/MenuCommands.c:232:            __asm__ volatile("cli; hlt");
src/MenuCommands.c:389:            __asm__ volatile(
src/MenuCommands.c:407:            __asm__ volatile(
src/MenuCommands.c:416:            __asm__ volatile(
src/MenuCommands.c:424:            __asm__ volatile("cli; hlt");
src/QuickDraw/QuickDrawPlatform.c:58:    __asm__ volatile ("inb %1, %0" : "=a"(value) : "Nd"(port));
src/Finder/finder_main.c:552:    __asm__ volatile("cli; hlt");
src/System71StdLib.c:213:    __asm__ volatile ("outb %0, %1" : : "a"(value), "Nd"(port));
src/System71StdLib.c:218:    __asm__ volatile ("inb %1, %0" : "=a"(value) : "Nd"(port));
$ find . -iname '*.s'
./src/multiboot2.S
./src/boot.S

That's probably most of (if not all) the ASM code you will find in the repo.

>>106741632
lmao

File: averi reimu touhou dress.png (783 KB, 1965x2048)

783 KB PNG

>>106741614
thanks anon for posting this.
very interesting stuff.

>>106743043
YW. Glad you liked my post. Usually no one cares about them...

Next step is the 68k to x86 translation layer. 68k users will want to take their apps with them, instead of waiting for new software to be developed

>>106741664
>b-but its not the god LLM that can do everything on it's own!
This is probably the future of AI stuff.

>>106743613
This.

>>106741614
I like this a lot. Good content

>>106741614
haven't seen that wallpaper in years, nice

>>106743635
Anon...

>>106741614
This is actually a good use of these tools desu

>>106741614
can they do this for gba games instead? I want a native port of pokemon

>>106746082
DIY maybe?

>>106746082
github.com/pret
There's the code, now translate away.

>>106741744
>>3.2 The agents, their outputs, and how they check each other
>>(A) Evidence Curator
>>· Input: disassembly (Ghidra/radare2), strings, xrefs
>>· Output: evidence.curated.json listing function start/end addresses, string refs, and SHA-256 of the relevant binary slices
>>· Controls others: the Compliance Auditor later rejects any function not present here; the Code Editor cannot work without an evidence ID

This is similar to how stonechoir recommends translating the Septuagint.

>>106743043
new averi?

>>106746082
Odds on AI slop port being slower and buggier than mGBA?

File: merooftop.jpg (54 KB, 526x701)

54 KB JPG

Say something nice about her, /g/.

>>106748195
the job she did is really fucking cool.
I noticed she also has some astro pics in his site. I wish I had a nerdy gf like her.

>>106746241
I wouldn't know where to start. Also don't want to give money to these scam corps.

>>106741614
I just wanna code games ffs, not this shit.

>>106741614
>This work provides a reproducible blueprint for preserving legacy systems: constrain AI with evidence ledgers, mutual agent checks, and emulator-visible milestones; make every function, struct, font, icon, and pattern auditable; and treat partial coverage as a modular feature, not a defect, to sustain honest progress.
which part of this is reproducible? there's zero technical detail.

did the author actually do any of the agent workflow xirself or was that also done by an AI?

>>106750531
OP here. I agree. I couldn't find much actual technical info on how she did it beyond short explanations, some diagrams and the code in that repo.
None of the "Claude/ChatGPT/Qwen agent role cards and playbooksalyzer.md, compliance-auditor.md, etc.)." mentioned in the "Artifacts & Tools" section.

She's commenting in the comment section of the hackaday article (or was until yesterday...), so you could ask her if you are interested, I guess.

>>106748882
>doesn't even know to ask gpt where to start
holy mcburger

>>106748962
>I don't want to port ancient operating systems to my hardware
sexuality issue

>>106741614
So can I open up ghidra with an AI assistance plugin and say "Crack this" and get results?

486 hackintosh when?

>>106741614
I looked through the git repo.
It does almost nothing. Anyone could, from scratch, produce this level of functionality using gtk, qt, or tcl/tk in a few days and it would be more stable and probably faster.

Like, It doesn’t run any programs. At all. It’s not even *close* to being able to run a program/application.
The screenshot is all there is, really.

This is an advertisement for AI or something.
To call this System7 would be embarrassing, but they probably don’t know enough to be embarrassed.

And what’s the point? Just emulate and dynamically translate it instead of “AI” … IBM had a 80286 jit running on the RT and POWER systems in the 80s for christs sakes.

>>106753349
Just like most things it's the retards who don't understand how it works who are the loudest about how impressive it is. YouTube comment section.

>>106753349
OP here. yeah, it's the base system, and from what I read in the hackaday comments, it's not even graphically close to System 7, so I'd guess it's still a bit far from being able to run stuff
even then, do you really not think REing the base of an old OS in 3 days is not impressive?

you could at least call it a borderline fraud (like >>106750531), WIP or whatever instead of criticizing it for
>Like, It doesn’t run any programs
I mean, those programs were written for a completely different CPU architecture. go do something like this yourself if you think you are so great lmao