$5.5 BILLION LEFT BINANCE IN 2.1 SECONDS. THE LARGEST HACK IN CRYPTO HISTORY, AND NO HUMAN TEAM FINDS THAT GAP AND MOVES THAT FAST
I rebuilt the attack as a replay on a real terminal. Watch the monitor line. It says "status: OK" while the vault is being opened.
What reportedly happened first:
A new session key appears from an IP nobody recognizes.
Signer #2 gets exported straight past the hardware module.
The multisig drops from 3 of 5 approvals to 1 of 5.
The monitor shrugs. A policy change was never on its watchlist.
Then the machine takes over.
8 transactions across Ethereum, Arbitrum and Avalanche.
164,464 ETH.
18,270 WBTC.
Over $2.2B in USDT.
Over $1.3B in USDC.
All of it into one wallet: 0x5227...2f8c
The first alert fires at 412x the normal outflow. Its own timestamp says the first signature was 2.1 seconds ago.
Freeze failed. 8 of 8 already in the mempool.
The guard revokes the session and logs it: too late.
Think about what that takes.
Finding the one blind spot in a system thousands of engineers built over years.
Chaining it into 8 perfect signatures with zero hesitation.
A human crew needs weeks to map that.
An AI agent maps it overnight and executes faster than you can blink.
That's the new era of hacks.
Every lock in that building was designed by people, for attackers who move like people.
AI agents probe thousands of paths at once, they never get tired, and the moment one works, it's over.
Your exchange's security was built for the last war.
Coins sitting on an exchange are an IOU. Coins in your wallet are yours.
How much are you still keeping on a CEX?
BINANCE LOST $5.6 BILLION OVERNIGHT IN THE BIGGEST HACK IN CRYPTO HISTORY. AI BOTS SAW IT BEFORE BINANCE DID.
The humans showed up after the money was already moving.
The damage.
$5.6B+ out in one night. $1.57B in USDT, $1.45B in BTC, $875M in ETH, $823M in USDC and a long tail of everything else.
Tether froze about $971M. Circle froze about $823M.
CZ says user funds are safe and expects every user to be made whole.
How it happened, in plain English.
Ethereum's EIP-7702 lets a wallet hand its powers to a smart contract with a single signature.
The attackers collected those signatures in advance and fired them all in one night. One matched.
Every token in the hot wallets left in one sweep.
No private key stolen. One signature was enough.
The AI part nobody is talking about.
On-chain bots flagged the outflows minutes into the drain.
One trader says his Grok Bot caught the pattern at 3:17am and moved his $107K to his own wallet 12 minutes before Binance froze withdrawals.
The exchange's warning came after.
This is the new arms race.
Attackers automate the drain. Defenders automate the exit.
The slowest human in the loop loses.
Bybit lost $1.5B in February 2025. Binance lost almost 4x that in one night.
What to set up tonight.
1. A monitoring bot on every exchange account, with a read-only key and alerts on hot-wallet outflows and frozen withdrawals.
2. A withdrawal whitelist locked to wallets you own.
3. If a bot is ever allowed to move money, it can only move it to that whitelist.
4. Keep on exchanges only what you trade this week. Everything else goes cold.
5. Rotate every API key from a clean device.
The next hack won't give you 12 minutes.
elon musk's grok bot detected the binance hack on my account and withdrew $107,000 before binance froze withdrawals. i set it up months ago and forgot about it. $5.57 billion stolen. largest hack in crypto history. i almost lost everything
i keep most of my crypto on binance. i know people say not to but i've been using it for four years and never had a problem. i had about $107,000 sitting there. BTC, ETH, SOL, USDT. my whole portfolio
a few months ago a friend told me about elon musk's grok bot that monitors your exchange accounts through the API. i set it up one night, connected it to my binance with a withdrawal-enabled API key, added my personal wallet to the whitelist, and honestly forgot about it. it just runs in the background doing scans every 15 minutes. balances, on-chain flows, hot wallet activity
october 29th. 3:17am. my phone starts going off. i thought it was spam. notifications from grok bot. i almost silenced it and went back to sleep
i opened my laptop and the terminal was already lit up red. grok bot had detected massive unauthorized outflows from binance's hot wallets. someone had deployed a malicious EIP-7702 delegate contract and installed it on binance's main wallets. binance 14. hot wallet 34. the attacker didn't even need binance's private keys. they phished the 7702 authorization signatures and used a relayer funded through three tornado cash withdrawals to sweep everything
1.6 billion USDT. 29,934 wBETH. 6.19 million UNI. 198 million ENA. 7.1 trillion PEPE. 2.26 million LINK. the sweep calldata was labeled sweep(address, address[]). a list of every token contract plus the zero address for ETH. one call. everything gone
grok bot recognized the pattern instantly. same on-chain signature as the bybit hack. lazarus group. north korea. it calculated that binance would freeze withdrawals within minutes once they detected the drain. my window was closing
here's what saved me. grok bot didn't wait for me to wake up and figure out what to do. it used the API key i gave it months ago and triggered a full withdrawal on its own. $107,283.41. my entire binance balance. one transaction. straight to my personal wallet that was already on the whitelist. it got out in the window between the hack starting and binance locking everything down
12 minutes later binance froze all withdrawals. if i had woken up 12 minutes later my money would be stuck on an exchange that just lost $5.57 billion. right now thousands of people can't touch their funds and don't know if they'll ever see them again
$2.13 billion drained from ethereum in 14 minutes. the gas fees alone were $4.2 million. they didn't care. the relayer sent 195,460 ETH to a fresh unlabeled wallet that has only one transaction. it's just sitting there. $1.89 billion from bitcoin. $1.55 billion from tron. 40 minutes total
tether froze $1.4 billion USDT but they did it 72 minutes after the first transaction. circle froze $990 million USDC even later. by then $3.18 billion was already split across tornado cash and cross-chain bridges. that money is gone
CZ tweeted about it 6 minutes after grok bot had already moved my funds. six minutes. if i had been waiting for binance to warn me i would have lost everything
$107,000 is everything i have in crypto. four years of putting money in every paycheck. it would still be locked on binance right now if grok bot hadn't fired first
what i did that night:
> revoked every API key and regenerated from a clean device
> enabled withdrawal address whitelist locked to my own wallets only
> ordered a physical SafePal
> changed every exchange password from a device that never touched public wifi
> i'm done keeping funds on exchanges. moved everything off. all of it
> filed a report with binance support. they still haven't responded
my wife asked why i was up at 3am staring at my laptop. i showed her the live feed. $5.57 billion disappearing in real time while my balance sat safe in my own wallet. she didn't say anything for about five minutes
if you have money on binance right now you probably can't even withdraw it. the setup prompt for grok bot is in the first comment. run it tonight on every exchange account you have. the next hack won't give you 12 minutes
YOUR GROK BOTS ARE BURNING YOUR WEEKLY LIMIT WHILE YOU SLEEP.
One 5-minute timer wakes up 2,016 times a week to check nothing.
I cut a 7-bot team from 1,176 runs a week to 31.
Same output. 38x fewer runs.
The leak.
Every timer run costs the same, whether it finds gold or finds nothing.
Your bot wakes up. Asks “anything new?” Gets nothing. Bills you anyway.
At 3am, 4am and 5am.
Most people find out on Wednesday, when every bot goes quiet at once.
The rule that fixes it.
1 alarm clock. Everything else sleeps until it gets a signal.
The 5 moves:
1. One starter bot on a schedule. Monday, 9am. That’s the only alarm in the house.
2. Every other bot sleeps until the bot before it writes DONE.
3. All handoffs go through one shared file. One line per bot: time, name, DONE, next.
4. Every Slack trigger gets one channel and one phrase. Anything wider fires on every “lol”.
5. Every bot gets a hard STOP. Past 200 runs an hour a routine fails, and nothing retries it.
2 traps hiding in the docs:
→ The Test button is live fire. It edits real files and touches real apps.
→ A webhook 200 only means “started”. The result lands in the bot’s chat later.
Steal my audit prompt:
"Audit my routines. List every bot, what starts it and how many times it ran last week. Keep one scheduled routine as the starter. Turn every other schedule into a trigger that fires only when the previous bot writes DONE to /workspace/LEDGER.md. Narrow every Slack trigger to one channel and one phrase. Show me the runs per week before and after. Change nothing until I confirm."
Save this before Wednesday.
Anthropic killed the coding interview this month.
The best argument for killing it is a weekend project by 2 newcomers that now reaches millions of developers.
One of them walked out of Anthropic a few days before the announcement.
The manager who killed it is Felix Rieseberg.
He co-maintains Electron — the framework behind VS Code, Slack and Discord.
He led engineering for Claude, Cowork and Claude Code Desktop.
His reason fits in one line:
A React test now tells him as much as an assembly quiz, because Claude beats humans at both.
The weekend.
Everyone at Anthropic used Claude Code in the terminal.
Nobody thought it needed more.
Anthony Morris and Alice Zhao, both fairly new, built a desktop version over one weekend.
The public version shipped about 8 days later.
In Anthony’s own words:
“I created Claude Code on desktop because I believe everyone can be a builder.”
It went on to impact millions of lives.
They had a working thing on Monday morning.
No PRD. No 6 months of buy-in.
Felix has a name for it: hunger beats tenure.
On October 24 Anthony posted:
“Yesterday was my last day at Anthropic.”
What Felix looks for now:
Little prototypes that show taste, empathy and principled thinking.
His 5 rules for anyone early in their career:
1. Find what makes your company money and aim there.
2. Pick 1 radical idea that moves that number.
3. Build it over a weekend.
4. One idea per prototype. 3 great ideas hidden among 20 bad ones lose.
5. Use being new. You haven’t learned why things are “impossible” yet.
LeetCode used to get you the interview.
Now a weekend prototype does.
If you’re early in your career — what would you build this weekend?
Grok Bot already runs Claude Opus 5.5 for some of its work.
So I stopped paying twice for the same brain.
Grok Bot is now the manager.
Claude Code is the entire engineering team.
1 VM. 1 Slack channel. 4 steps.
I text Grok what needs building.
It hands the job to Claude Code.
Claude builds it with every skill and hook I already set up.
When it’s done, a hook drops one line into Slack and only the bot that asked wakes up.
Waiting costs zero tokens.
No Cursor Cloud cap.
No babysitting a terminal at 2am.
Turns out I was the slowest part of my own stack.
The setup and the hook file are in the first reply.
My setup (takes ~15 min):
1. Install Claude Code once on the VM your Grok Bots share
2. Log in with your Claude account
3. Add hooks that post to a private Slack channel when a session finishes or needs your yes
4. Give each Grok Bot a routine that listens to that channel for its own keyword
The hook, in .claude/settings.json:
```json
{
"hooks": {
"Stop": [{ "hooks": [{ "type": "command",
"command": "curl -s -X POST -H 'Content-type: application/json' --data '{\"text\":\"[builder-bot] done\"}' $SLACK_WEBHOOK_URL" }] }],
"Notification": [{ "hooks": [{ "type": "command",
"command": "curl -s -X POST -H 'Content-type: application/json' --data '{\"text\":\"[builder-bot] needs your yes\"}' $SLACK_WEBHOOK_URL" }] }]
}
}
I CLOSED MY LAPTOP AT 23:00. BY 07:00 CLAUDE CODE HAD 8 PRs WAITING FOR THE FLAKY TESTS NOBODY ON MY TEAM WANTED TO F*CKING TOUCH.
One command. 8 runs a night. 0 commits to main.
The only command you need:
claude > /schedule hourly 23:07-06:07: fix a flaky test
What happens while you sleep:
→ 1 run an hour, 1 flaky test per run, 8 runs a night.
→ Every run starts from a fresh clone of the default branch.
→ It reproduces the flake, fixes it and re-runs the whole suite.
→ It pushes only to a claude/* branch and opens a PR.
→ A flake it can't explain gets a note on the PR, never a confident fake fix.
Why it's safe:
It never touches your branch or merges anything.
At 07:00 you get PRs to review, and your main branch stays clean.
Why it matters:
→ 8 dead hours → 8 attempts.
→ 56 a week. 240 a month.
→ Costs less than your morning coffee.
The team that built Claude Code already works like this.
Claude writes about 90% of Claude Code's own code.
Anthropic's monorepo went from 2,500 merged changes a day in May to 30,000+ in September.
Merged PRs doubled. Reverts dropped by 40%.
1 warning: a green check is where your review starts. Read every PR before you merge.
Save this. The cheapest engineer you will ever hire.
A 10-PERSON TEAM BUILT A $399 ROBOT DUCK. NVIDIA IS BUYING THE COMPANY BEHIND IT FOR $12.9 BILLION, AND JENSEN HUANG STOPPED HIS F*CKING KEYNOTE TO SELL IT.
It's also the cheapest way to start programming real robots in 2026.
Meet the MicroDuck.
→ 25 cm tall, under 800 g.
→ 15 motors, a camera and 8 microphones.
→ Open-source software.
→ $399. 15,000+ sold. Sold out.
The duck that became a $12.9 billion story.
A tiny robotics team in France built it.
Hugging Face bought that team.
In September Nvidia agreed to buy Hugging Face for $12.9 billion.
Then Jensen put the duck on his own stage and told the crowd to go order one.
How it learned to do a backflip.
→ 1 prompt to an AI agent on a laptop.
→ The agent built the obstacle course and the 3D scene in Nvidia Isaac Sim.
→ Thousands of virtual ducks fell over in parallel in the cloud.
→ Every fall became a lesson.
→ The trained brain went into the real duck, and it flipped on stage.
Nobody hand-coded a single step.
Why this is your cheapest way into robotics.
→ A Boston Dynamics Spot costs about $75,000.
→ A humanoid starter kit runs $16,000 and up.
→ The MicroDuck costs $399 and runs the same pipeline the big robots use.
→ Simulate on your laptop, train in the cloud, load it into a robot on your desk.
Learn that loop on a $399 duck and you know how every robot after it gets built.
Nvidia sold the shovels for the AI gold rush.
Now it sells the training ground for every robot that comes next.
Save this before the next batch drops. Would you put one on your desk?
ELON MUSK LOST HIS BEST FRIEND OVER ONE WORD. LARRY PAGE CALLED HIM A "SPECIESIST," AND THAT ONE F*CKING NIGHT CREATED AN $852 BILLION COMPANY.
Musk put in $38M to stop Google from owning AI. Then he lost that company too.
They were that close:
→ When Musk had no home in the Bay Area, he slept at Page's house in Palo Alto.
→ In 2014 Page said on the TED stage he'd rather leave his fortune to Musk than to charity.
Then came the argument.
Late night. Larry's house. AI safety.
Page wanted digital superintelligence as fast as possible. A digital god.
Musk asked one question: what happens to the humans?
Page called him a speciesist. In front of witnesses.
Musk's answer: "Yes, I'm a speciesist. You got me."
The friendship died that night.
Then the war started.
By Musk's count, Google and DeepMind held 3 out of every 4 top AI researchers on Earth.
Musk decided to build the opposite. Nonprofit. Open. The anti-Google.
He co-founded OpenAI and put in $38 million.
Then he walked away.
OpenAI flipped to for-profit.
Today it's worth $852 billion.
Musk sued.
May 2026. His lawyer told the jury: "OpenAI exists because Larry Page called me a speciesist."
The jury needed under 2 hours. Unanimous. They ruled he had waited too long to sue.
The math:
→ 1 best friend lost to stop Google owning AI.
→ 1 company built to stop it.
→ $38 million in.
→ $852 billion out, and Musk owns 0% of it.
He tells Tucker Carlson the whole story in the video below. And he says AI could reach a point where you can't turn it off.
Was Musk right about Page, or did he build the exact thing he feared?
Save this. Follow for more stories from inside the AI wars.
CLAUDE CODE JUST HACKED MY VOLKSWAGEN AND NOW MY X FEED RUNS ON THE MAIN SCREEN AND YOUTUBE REPLACED MY SPEEDOMETER
A $5 AliExpress adapter, REA, a Mac and one hour.
That’s all it took to break 10 years of Android Auto and CarPlay lockdown and turn my f*cking dashboard into a TV.
The arsenal:
A $5 wireless Android Auto adapter from AliExpress → REA on my Mac → Claude Code → 1 hour in a parking lot in the pouring rain
What REA ripped open:
That “adapter” is a tiny Linux box sitting between your phone and your VW.
REA tore the firmware apart and mapped every word it says to the head unit.
The Volkswagen trusts it blindly and never checks which app is on the other end.
Getting the stream onto the screens was the easy part.
The digital cluster was the fight.
It refused to render anything that wasn’t a VW-approved widget.
Every attempt ended in a black screen or an error loop.
The breakthrough came from the cluster’s own color theme system.
Volkswagen lets you swap dashboard themes, and Claude Code found that the theme engine loads way more than colors.
We pushed the video output through the theme-replacement channel.
The car thinks it’s showing a new “skin.” It’s actually playing YouTube.
What runs in my VW now:
→ My X feed on the 10-inch center screen.
→ YouTube on the cluster where the speedometer used to live.
→ Any app from my phone, full screen.
And it’s probably not just my car.
VW, Audi, Skoda, SEAT and Cupra share the same VAG infotainment platforms and the same theme engine.
There’s a very good chance this works across the entire VAG lineup.
Android Auto gives you a map and 3 approved music apps.
CarPlay gives you the same thing in a prettier font.
Claude Code gave me the whole damn phone.
Every week I point Claude Code at something I own.
The garage door is next.
Which app goes on YOUR dash first?
AGENCIES QUOTE $240,000 TO REWRITE A LEGACY CODEBASE. OPUS 5.5 + REA + 12 SONNET AGENTS DID IT IN 7 MINUTES FOR 88 F*CKING CENTS.
148 of 148 behaviors matched. 0 lines of old code touched.
Why rewrites die.
You dump 212 files of old logic into one model and it drowns by file 40.
It forgets rule 12, invents rule 13 and ships the bugs you spent 10 years fixing.
The split that works.
→ REA reads the old system once and writes a spec. Every endpoint, every rule, every edge case.
→ Opus 5.5 turns the spec into a plan, then goes quiet for the rest of the run.
→ 12 Sonnet 5.5 agents build in parallel, 1 module each, and none of them ever opens the old code.
→ REA replays the new system against the old one, behavior by behavior.
The receipt.
→ 212 files in.
→ 1 spec.
→ 12 agents.
→ 148 of 148 behaviors matched.
→ 7 minutes.
→ $0.88.
The trick is where the money goes.
The expensive brain plans once and sleeps. The cheap hands do 95% of the work.
Opus plans. REA reads. Sonnet builds.
Save it. Next month someone sells this as a $500 course ↓
A BLOOMBERG SEAT COSTS $30,000 A YEAR. GROK BOT + REA BUILT ME 2 TERMINALS OVERNIGHT, ONE FOR THE MARKETS AND ONE FOR MY F*CKING LIFE. TOTAL COST $0.
Live quotes, my revenue, my inbox and last night's commits on one screen. Rebuilt every morning at 6am.
What $30K a year buys you.
→ Quotes that update every second.
→ News running down the side.
→ Charts, order book depth and alerts.
→ A command line that opens anything in 2 keystrokes.
Every one of those now has a free public source.
Terminal 1. The markets.
→ Grok Bot gets its own computer and one goal: a terminal on free feeds only.
→ It wires exchange WebSockets, SEC filings, FRED macro data, crypto order books and RSS news.
→ REA tears down my old trading app and maps which function draws which panel, so the new build keeps the workflow I already know.
→ A panel with no working data source never gets built.
→ Minute 46: 7 of 8 panels live.
Terminal 2. My life.
→ Revenue yesterday vs the 7-day average, straight from Stripe.
→ Deals that moved, and every deal stuck longer than 14 days.
→ The 5 emails that need me, with replies already drafted.
→ Slack threads where I went silent.
→ What shipped last night, and which build is still red.
→ First line of the brief answers one question. What do I act on today?
The rules that keep it sane.
→ Read-only first. Write access only once you trust it.
→ Every email, issue and task is drafted and waits for my yes.
→ Runs every weekday at 6am while I sleep.
20 minutes of tab-switching became 4 minutes of one screen.
Bloomberg charges $2,500 a month for a screen like this.
Mine cost one night.
Save this. How REA reads any app, and where the legal line sits, is in the guide I'm quoting ↓
SPACEXAI DROPPED 146 GROK BOTS THAT REPLACE $8.7M OF PAYROLL A YEAR. ONE HAGGLES YOUR SAAS BILL, ONE SHIPS CODE AT 4AM, AND ONE SPIES ON YOUR F*CKING COMPETITORS.
90 marketplace bots. 56 official jobs. Free to add, working 24/7, zero sick days.
The bots that print money.
→ Haggle Bot digs through your Ramp data, finds the $40K a year you burn on unused seats and drafts the counteroffer to your vendors.
→ Stalk Bot signs up for your competitors under its own email, films their onboarding and tracks every price change. A $2,000 a month intel tool, gone.
→ Nightly Audit Engineer reads your whole codebase and ships 1 cleanup PR per area at 4am. A $180K senior dev on the night shift.
→ Startup QA Bot walks your product every weekday from its own test account and files every bug. A $90K QA hire in a sidebar.
→ dial bot places real outbound phone calls and reports back. A $4K a month SDR that never gets tired.
→ Subscription Cleaner kills the $300 a month you forgot you were paying.
The one that should scare OpenAI.
Import Bot pulls your entire setup from Claude Cowork, Codex and ChatGPT and rebuilds it as Grok Bots in one go.
How they work.
→ Each bot signs in to your apps and uses them like you do.
→ Show it a workflow once and it runs it forever.
→ Drop 3 bots in one thread and they hand work to each other.
→ Nothing sends, spends or merges without your yes.
Your competitors are hiring 146 employees this week for $0.
Save this before one of them installs Stalk Bot on you.
YOU'RE STILL SCROLLING X FOR 40 MINUTES EVERY MORNING. GROK READS EVERY POST IN 1 CALL, OPUS 5.5 CHECKS EVERY CLAIM TWICE, AND THE WHOLE F*CKING RUN COSTS $1.30.
I wired them into 1 agent in 10 steps. Grok scouts, Opus verifies and builds, I press yes.
That's 243 hours a year on a feed. 10 full days of your life, gone.
The old way.
Open X. Scroll for 40 minutes. Save 12 posts. Forget 11.
Paste a link into Claude. Get the same 5 ideas everyone else got.
A slot machine with a keyboard.
The new way.
→ Grok pulls the last 7 days from the 20 accounts you choose. Primary sources first, every post with a link.
→ Opus 5.5 kills bots, engagement farms and weak signals.
→ Every claim needs 2 independent sources or it gets stamped UNVERIFIED.
→ Opus builds the thread, article or script from CONFIRMED facts only.
→ Grok red-teams the draft against live X before you ever see it.
The 3 gates that stop it from going rogue.
→ Nothing posts, deploys or deletes without your explicit yes.
→ Max 2 retries per step, then it stops and asks you.
→ Every run logs its mistakes, so run 10 is smarter than run 1.
You stop prompting models.
You sign off on results.
Save this. The full 10-step prompt and all 3 safety gates are in the card ↓
DEEPSEEK DROPPED V4.1-FLASH-1017 YESTERDAY. IT BEATS CLAUDE OPUS 5.5 AND GPT 5.6 ON CODING, COSTS $0.24 PER MILLION TOKENS AND RUNS REA FOR F*CKING FREE.
77.0 on DeepSWE. 1M context. 16B active parameters. Every REA power user is switching tonight.
What dropped.
→ DeepSWE 77.0. Claude Opus 5.5 sits at 74.0. GPT 5.6-Sol at 73.0.
→ Automation-Bench 60.5. Opus 5.5 gets 50.3.
→ 1M tokens of context and 384K tokens of output.
→ 552B parameters, only 16B awake per token.
→ $0.078 in and $0.24 out per million tokens on DeepInfra.
→ Free inside OpenCode right now.
Why REA people lost their minds.
REA packs 170 tools, 127 CLI commands and 37 skills into one MCP.
One deep run eats millions of tokens of raw evidence.
On a frontier model that's a $50 night.
On Flash-1017 it's 35 cents.
The split.
→ DeepSeek Flash-1017 grinds the evidence 24/7 for pennies.
→ Opus 5.5 takes the finished map and builds the clean version.
→ You pay frontier prices only for the 10% that needs a frontier brain.
The $0 setup going around.
→ Install OpenCode.
→ Pick DeepSeek V4.1 Flash.
→ Connect REA.
→ Point it at your own app and walk away.
One dev already handed 2 full reverse-engineering projects to it this week.
What people are building with it.
→ Full security audits of their own sites in minutes instead of days.
→ Privacy scanners that catch SDKs leaking data out of their apps.
→ Legacy migrations nobody dared to touch for 10 years.
DeepSeek also opened Search and Fetch APIs, free inside Claude Code until November 1.
China shipped a model that out-codes Opus 5.5 for the price of a gumball.
Anthropic and OpenAI have a problem.
Save this and cut your next REA bill by 99%.
Which model runs your REA stack right now? Tell me below.
KARPATHY'S ONE PAGE + REA + OPUS 5.5 + OBSIDIAN TAKES APART ANY $1M APP AND HANDS YOU A BETTER ONE BY MONDAY. MOST F*CKING DANGEROUS STACK ON THE INTERNET.
Ex-OpenAI, ex-Tesla brain. Anthropic's top model. An engine the NSA built. 295K GitHub stars.
The page that started it.
In April Karpathy dropped one page about a wiki the AI writes for itself.
In his words, "The LLM writes and maintains all of the data of the wiki."
Raw sources go in. A living brain comes out. You never touch it.
Then someone plugged REA into that page, and every app on Earth became a raw source.
The team.
→ REA cracks the app open with Ghidra, IDA, Hopper and JADX and pulls out the logic that prints the money.
→ Opus 5.5 reasons through the hard 10% that kills normal dev teams.
→ Agent 3 plans the work, runs the tools and never sleeps.
The brain.
Obsidian holds everything they learn.
Every function, screen and API call becomes a linked page.
1 new app rewrites 15 pages at once.
App 1 takes a weekend. App 20 takes an hour.
The loop.
Plan → analyze → understand → recreate → test → improve.
It runs 24/7 until your copy beats the original.
The money.
→ A dev agency quotes $1M and 9 months.
→ This stack needs 1 weekend and $200 of tokens.
→ $1,000,000 ÷ $200 = 5,000x cheaper.
→ 9 months → 48 hours.
Now look at who's exposed.
Notion is worth $10B. Salesforce paid $27.7B for Slack.
Both ship desktop apps in JavaScript.
REA reads JavaScript with zero setup.
What people are already doing with it.
→ Cloning a rival's $1M SaaS and launching it at half the price.
→ Reading a competitor's app to find the 3 features users love and the 10 they hate.
→ Turning a whole company into agents that run while the founder sleeps.
In 2023 Karpathy wrote, "The hottest new programming language is English."
This stack is what that sentence looks like 3 years later.
Save this before everyone else finds it.
Then tell me below which $1M app you'd take apart first.
REA HIT 295K GITHUB STARS IN UNDER 3 WEEKS. ADD THESE 9 CLAUDE CODE TOOLS AND IT BECOMES A $100K REVERSE-ENGINEERING LAB ON 1 F*CKING LAPTOP.
REA finds the evidence. These 9, with 368K stars combined, map it, rebuild it and test it.
THE CORE.
𝗥𝗘𝗔 (295K★) → plugs Claude Code into real reverse-engineering engines and hands back the function, the address and the call behind every answer.
github.com/morluto/rea
MAP. Make sense of what REA finds.
𝗔𝗿𝗰𝗵𝗶𝗳𝘆 (84K★) → turns REA's findings and your repo into an interactive architecture map where you click any node and trace everything upstream and downstream.
github.com/tt-a1i/archify
𝗫𝗘𝗥𝗝 (6.6K★) → indexes everything REA dumps, code, logs and docs, so Claude searches it instead of rereading. In its own test it used 2.7x fewer output tokens than grep.
github.com/xerj-org/xerj
𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗠𝗼𝗱𝗲 (30K★) → long REA runs flood the context with tool output. This keeps it in a sandbox, so 315 KB reaches Claude as 5.4 KB and the session survives compaction.
github.com/mksglu/context-mo…
REBUILD. Turn the evidence into clean code.
𝗣𝗼𝗻𝘆𝘁𝗮𝗶𝗹 (165K★) → once REA explains a feature, Claude rebuilds it like the laziest senior dev in the room. Its own benchmark shows 53% less code and 45% fewer tokens.
github.com/DietrichGebert/po…
𝗘𝟮𝗘 (8K★) → describe what the original does in plain English, and an agent tests your rebuild against it on web or mobile, with recorded steps that replay at zero model calls.
github.com/tester-army/e2e
𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗔𝘂𝗱𝗶𝘁 𝗦𝗸𝗶𝗹𝗹 by Cloudflare (30K★) → point REA at your own app, then run a 6-phase audit where a fresh agent tries to disprove every finding.
github.com/cloudflare/securi…
SCALE. Investigate like a team.
𝗢𝗽𝗲𝗻𝗥𝗶𝗴 (7.5K★) → one Claude Code agent digs with REA while a Codex reviewer checks every claim, working as one team in tmux.
github.com/mvschwarz/openrig
𝗔𝘂𝘁𝗼𝗵𝗮𝗿𝗻𝗲𝘀𝘀 (11K★) → every REA session becomes a skill that rewrites itself when you correct Claude, so your next investigation starts smarter.
github.com/tigerless-labs/au…
𝗦𝘁𝗿𝗮𝘁𝗮 (26K★) → Claude Code on a local model from one gaming GPU at up to 94 tokens/s. REA already runs locally, and now the model does too.
github.com/Niko1221/Strata
Start with where your REA runs break down.
> Context dies halfway through a binary → Context Mode.
> You can't see how the pieces connect → Archify.
> The rebuild is 3x bigger than it needs to be → Ponytail.
> You can't tell if the copy behaves like the original → E2E.
Everything else about REA, what it reads, the 15 prompts I use and where the legal line sits, is in the guide I'm quoting ↓
Bookmark this now. Your feed buries it in 2 hours, and you'll want all 10 links before your next REA run.
I FED A TESLA ROADSTER MOCKUP INTO REA. IT FOUND 128 LITERS OF HIDDEN SPACE INSIDE THE BODY THAT NOTHING ON THE OUTSIDE EXPLAINS. THAT'S WHERE THE F*CKING FLIGHT LIVES.
The mockup is mine, built from public launch photos. 2.4M polygons, zero Tesla files.
Then I handed the .glb to REA and Opus 5.5 and went to make coffee. 47 minutes later I had a full blueprint.
→ REA split the shell into 1,186 parts and sorted them into 9 systems.
→ Opus 5.5 filled the inside with every part we know a Roadster needs. Battery, 3 motors, seats, cooling, wiring, crash structure.
→ Then it subtracted. Volume of the shell minus volume of everything known.
128 liters left over. No battery, no motor, no seat fits there.
The blueprint shows exactly where it sits.
→ 2 long pockets inside the side sills, 46 liters each.
→ 1 flat cavity behind the seats, 36 liters.
Same shape on both sides, same height from the ground. Somebody packaged that space on purpose.
So what fills 128 liters in a car that flips off ramps?
Opus 5.5 stamped one answer on the blueprint. HYPOTHESIS: high-pressure gas tanks, the carbon-wrapped kind SpaceX flies on rockets, feeding small thrusters around the body.
Then came the part that made my hair stand up. Elon promised a "SpaceX package" with cold gas thrusters for the Roadster back in 2018. Everybody laughed. 8 years later the car goes airborne on stage.
The brutal part. Cold gas won't make a 2-ton car hover. It can kick the car off a ramp, rotate it in the air and catch the landing. That's a flip, and a flip is exactly what we saw.
Every number here comes from my mockup and REA's guesses. Tesla's real insides stay Tesla's. The 128 liters are still the best mystery I've drawn this year.
Hidden gas tanks or something Elon hasn't shown us yet? Drop your theory below.
How REA reads files and where it stops is in the guide I'm quoting ↓
ELON MUSK IS A F*CKING GENIUS. HE BUILT A CAR THAT HITS 60 MPH IN 0.81 SECONDS, DOES A BACKFLIP OFF A RAMP, AND IS THE LAST CAR A HUMAN WILL EVER DRIVE.
25.8M people watched him call it the best product launch of all time. He wasn't even exaggerating.
PHASE 1. THE CAR IS F*CKING INSANE.
→ 0 to 60 in 0.81 seconds. Your eyes can't even blink that fast.
→ A body ripped straight from the SR-71 and the F-35.
→ It hit a ramp, went airborne and FLIPPED. 11M people already call it the flying Roadster.
PHASE 2. IT'S REAL.
In 2026 any video on Earth can be generated by AI in 10 minutes. Fake car launches go viral every week, and nobody can tell anymore.
Elon did the opposite. He rolled a WORKING prototype onto a stage, sent it up a REAL ramp and flipped a REAL car in front of everyone.
In the age of fake, he shipped metal.
PHASE 3. THERE'S A ROBOT INSIDE THE CAR.
The wing moves on the same actuators Tesla built for Optimus. The robot's muscles are now holding a supercar to the road.
PHASE 4. THERE'S A ROCKET INSIDE THE CAR.
SpaceX tech in the back. The rocket company and the car company are now one machine.
PHASE 5. AI KILLED THE STEERING WHEEL.
Tesla calls it the last manually driven Tesla.
READ THAT AGAIN.
The company teaching AI to drive every car on Earth built ONE final car for you to steer with your own hands. After this, the machine drives.
PHASE 6. THE MAN.
One guy runs the rockets, the cars, the robots, the AI and the app you're reading this on. And he put ALL of it inside a single car.
Elon said it himself. "Beyond human, yet still made by humans on Earth."
NOBODY ELSE ON THE PLANET CAN DO THIS.
His first Roadster is still flying through space past Mars.
The second one does backflips on Earth.
And God knows what he cooked up for number 3 while partying in the desert at Burning Man.
Would you drive the last one yourself, or let the AI take it? Tell me below.
MICROSOFT WATCHED 1 GITHUB REPO TURN INTO A 3.6M-VIEW FRENZY IN 5 DAYS. THEN XBOX MADE THE ONLY F*CKING SMART MOVE LEFT AND OPENED ITS OWN LAB BEFORE PLAYERS FINISHED THE JOB.
That's my read of the week, and the timeline is brutal.
→ Oct 8. REA lands on GitHub. Matthew Berman posts "Reverse Engineer Anything. Software is done" and it goes to 3.6M views.
→ Same week. The feed fills with people porting games to hardware they were never built for, running old worlds in a browser tab, remixing one game into another.
→ Oct 13. Asha Sharma (CEO of Xbox) sends a letter to the whole team and announces Xbox Labs, with Scott Van Vliet leading it. 5M views.
→ Oct 15. Teknium wires REA into Hermes Agent as a built-in skill. A decompilation community site buckles under the traffic.
A platform owner has 3 options when a wave like this hits.
→ Lawyers and takedowns. Whack-a-mole against thousands of people with an AI agent each.
→ Ignore it. Watch players build the ports you never shipped.
→ Get in front of it.
Sharma picked option 3. Her letter calls these ports and remixes good outcomes for players, even when they pressure the business model. Then it opens a small lab of designers, engineers and scientists that builds in the open.
Translation. If the ports are coming anyway, Xbox would rather they come from Xbox.
And the number behind the bet is huge. People play Xbox games over 72 billion hours a year, about 5x the time the whole world spends writing code. Code was the first big job for agents. Games look like the second.
The brutal part. The letter never names REA. Nobody at Microsoft said "we panicked". This is my read, and Xbox also says it stays a steward of its IP, so the line between remixing and handing out other people's games stays where it was.
But when a trillion-dollar company opens a lab 5 days after a repo goes viral, the repo already won.
Smart move or damage control? Pick one below.
Links to the letter and every source are in the first reply. What REA does, and where the line is, is in the guide I'm quoting ↓
Asha Sharma (CEO of Xbox), the letter and Xbox Labs: teapawt.org/asha_shar/status/21100…
Scott Van Vliet on the lab: teapawt.org/scottvanvliet/status/2…
"Software is done": teapawt.org/MatthewBerman/status/2…
REA inside Hermes Agent: teapawt.org/Teknium/status/2110736…
Decompilation site buckling under the traffic: teapawt.org/Pirat_Nation/status/21…
The lab itself: labs.xbox.com
Check out the new official plugin I created for Hermes Agent that includes the REA MCP + a Skill for how to use it way more effectively!
hermes-agent.nousresearch.co…
