mai

Tech & AI explorer | Automating everything I can | Daily threads on what’s worth using

[mai @maixbt]

@itsarcvale small scope per check is the real cost hack. one page, one claim, done

[mai @maixbt]

@semichenkko treasure hunting on autopilot is the dream setup

[mai @maixbt]

there are bugs on github with cash on them

i went to bed. my agent went hunting

by 8am: 1,200 found. 31 fixes ready. $10,000 if they merge

the ones it kept paid $250-$350 a fix. $322 on average

the video is that night mapped out

the trick is what it threw away:
788 nobody had touched in 60+ days
252 where someone already sent a fix
anything bigger than a day of work

it only cloned the 31 that were left. ~1.9m tokens for the whole night

every monday at 23:00 it goes again on fresh bounties. i don't even start it

the catch: none of it pays until a maintainer merges. and it opened 0 prs - it's not allowed to without me

one repo it kept says "no ai-generated prs". so it's 30, not 31

how much do you think this does in a month?

[Argona @Argona0x]

https://t.co/uFcQbivrbz

[mai @maixbt]

@0xAIlabs “the model becomes the process” is the best line here. speed turns editing into actual workflow

[mai @maixbt]

@JyunsuitenshI using real football footage for the moves is why this works. the tackle looks too legit

[mai @maixbt]

@EugBass editing-first wins. edits that leave the rest of the image alone is what real work needs

[mai @maixbt]

@AIDuckLabs the sync is what gets me. sound and animation coming from the same code means every hit lands exactly on the visual. in after effects that’s hours of manual keyframing

[mai @maixbt]

@quietoraclex the “teammates copy the lead’s effort too” line is the sneaky one. people switch the model to sonnet and still wonder why the bill is high. great breakdown with real numbers

[mai @maixbt]

@0ddscent 64.2 to 75.7 is a crazy jump

[mai @maixbt]

GOOGLE BUILT A SELF-IMPROVING AGENT. THE PART THAT DOES THE IMPROVING IS CLAUDE.

harness² went open source this weekend. the default editor is claude opus 4.8.

it never touches the weights. after every task it rewrites everything around them - prompt, tools, memory, skills.

glm-5 went from 64.2 to 75.7 after one rewrite.

→ runs the task once as a reference
→ tries 2 rewrites of its own setup
→ keeps what the runs back up, no grader needed
→ saves the lessons to EXPERIENCE.md for the next task

watch 0:12, that's where the new harness gets assembled.

the catch is every task runs 3 times, and google doesn't say what that costs. remember that part.

stock gemini lost to glm-5 wearing a better setup. google published that table themselves.

would you let claude rewrite your agent's setup after every task, or only after it fails?

[Akshay 🚀 @akshay_pachaar]

https://t.co/YUGDPxpYvy

[mai @maixbt]

@Richix01 the lessons are the real value here

[mai @maixbt]

GOOGLE BUILT A SELF-IMPROVING AGENT. THE PART THAT DOES THE IMPROVING IS CLAUDE.

harness² went open source this weekend. the default editor is claude opus 4.8.

it never touches the weights. after every task it rewrites everything around them - prompt, tools, memory, skills.

glm-5 went from 64.2 to 75.7 after one rewrite.

→ runs the task once as a reference
→ tries 2 rewrites of its own setup
→ keeps what the runs back up, no grader needed
→ saves the lessons to EXPERIENCE.md for the next task

watch 0:12, that's where the new harness gets assembled.

the catch is every task runs 3 times, and google doesn't say what that costs. remember that part.

stock gemini lost to glm-5 wearing a better setup. google published that table themselves.

would you let claude rewrite your agent's setup after every task, or only after it fails?

[mai @maixbt]

@LunarResearcher so we went from telling AI what to do to building an entire AI organization

[mai @maixbt]

@ch3nweiii bro built an entire analyst team out of bots 💀

[mai @maixbt]

@alxqwrk that's actually the default - every task starts from the base harness + saved lessons, not yesterday's rewrite.

the catch is cost. every task runs 3 times before you get 1 answer.

would you pay 3x per task for that?

[mai @maixbt]

@andreysuperior the “verify it against the original” step is what makes this work. most people stop at the rebuild and never check if it actually matches. proof for every claim is the whole difference

[mai @maixbt]

@Richix01 this is the kind of post that saves people hours. everyone shared the demo, you actually tested it across tvs. “don’t change anything until i say go” should be in every device prompt

[mai @maixbt]

@0xVeyyro the record button is the real flex here. most one-prompt demos die as a screenshot, this one ships its own content. and 1 html file means you can host it anywhere in seconds

[mai @maixbt]

20 REVERSE-ENGINEERING REPOS AROUND ONE CLAUDE CODE SESSION, AND THE AGENT ONLY TOUCHES 11

the center is REA - 300k+ github stars, MIT license, runs on your own disk

this is the loop mapped out: one prompt, 20 tools on the shelf, opus 5.5 picks

→ type one prompt: reverse ./target with REA, rebuild it clean-room
→ pull mitmproxy to replay the login call and see what the app sends
→ map the functions with rizin, split the logic with ghidra + angr
→ hook live calls with frida, check registers at the breakpoint with x64dbg
→ write SPEC.md - behavior only, zero copied code
→ rebuild against an oracle and retry until the outputs match

stripped c++ binaries still fight back. the loop runs smooth on .net and electron apps, slow on the hard targets. remember that part

most people install all 20 tools and never finish a target. here the agent pulls 11 and leaves 9 on the shelf

$0 for REA, ghidra, frida and radare2 - ida and hopper are the only paid pieces

which app would you point it at first?