Antid retweeted
My friend checked into a hotel with his girlfriend, ran a single script in Claude Code, and found out in 90 seconds that a hidden camera was running right behind the bathroom mirror
After what he told me, neither of us even unpacks a suitcase in a hotel or an Airbnb anymore until a laptop by the door finishes one full sweep
One prompt runs a 5-stage audit strictly in read-only mode:
-> Wi-Fi radar: scans the local network via nmap, resolves manufacturers from MAC addresses, and instantly flags hidden cameras and ports streaming continuous video
-> Bluetooth interceptor: listens across the room via the bleak library for unknown AirTag, Tile, and SmartTag trackers while spotting hotel locks on ancient firmware that open from the hallway
-> Spyware detector: audits your phone against the open-source Amnesty MVT toolkit and exposes hidden iconless processes quietly leaking your location or mic audio
-> Optical lens sweep: if the hotel isolates guest devices on the network, the agent walks you step by step with a flashlight across every bathroom mirror, vent, and angle facing the bed
-> Priority report: builds one clean table ranking the risk level of every suspicious device and tells you the exact next move: from demanding another room to calling the police
The agent never logs into anything, never hacks foreign devices, and translates 500 lines of raw network traffic into plain English in 2 minutes
I would rather spend 2 minutes with an open terminal in the hallway than spend the whole trip being watched
The full sweep prompt is below, just copy and paste ↓
Bookmark this post before your next check-in
Antid retweeted
Claude Code found my wife's private photos and videos on 3 websites she never uploaded to. then it found a hidden app on my 13 year old daughter's phone sending her entire camera roll to a server every 6 hours. both found in the same scan.
someone sent my wife a link to a site she'd never heard of. her photos and videos were there. original files from her camera roll that were only ever on her phone.
i connected her iPhone to my MacBook and ran Claude Code. it found an active iCloud session logged in since March 2023.
IP → 116.25.***.91
location → Guangzhou, China
device → Windows 10 Pro
last sync → 6 hours ago
recovery email → x*****[email protected]
someone got her Apple ID from a leaked database and synced her entire library for over a year. she changed her password twice. they had a recovery email she never set up and got back in every time.
found on:
→ www.l**k-vault cc
→ www ph***dump to
→ telegram → t me/le***vault_asia (14,000+ subscribers)
the last thing Claude Code pulled from her account was photos and videos of our daughter. that made me check her phone next.
she's 13. she had a photo editor app with 2.7 million downloads from the App Store. Claude Code found it uploading her camera roll to a server in Vietnam every 6 hours.
server → 103.56.***.12
location → Ho Chi Minh City, Vietnam
total uploaded → 11,400 files
access → camera, photos, location, microphone
her face, her school, her friends, her exact location every day for 8 months. Apple approved the app. it's still live right now.
what we did that night:
→ killed the iCloud session and removed the recovery email
→ changed every password from my phone, not from hers
→ deleted the app and revoked every unnecessary permission on both phones
→ filed DMCA takedowns on all 3 sites
2 of the 3 sites took the content down. the telegram channel is still up.
my wife cried. my daughter still thinks i deleted her app because it was using too much battery. i'm not going to tell her the real reason.
the full step by step is below. takes 3 minutes. run it on every phone in your house tonight. save it so you don't lose it and send it to anyone with a family.
Antid retweeted
nobody knew space bunny was @MiniMax_AI m3.1 flash
the #1 model on openrouter for 10 straight days just got unmasked. so i made it build me a full browser game
i wrote 4 messages. it did literally everything else in 12 min
it read the whole repo, wrote the code, ran it and fixed its own bugs. it even caught a mobile touch bug i didn't ask about
total cost: $0.19
73.7% on deepswe at $0.15 / $0.60 per 1M tokens. and the weights go open on oct 23
the $200/mo plans should be nervous
unlimited m3.1 flash for minimax code subscribers until oct 25, 50% off your first m plan month until oct 31
the game and links are in the comments
Antid retweeted
THE MYSTERY MODEL THAT SAT AT #1 ON OPENROUTER FOR 10 STRAIGHT DAYS JUST GOT UNMASKED
space bunny was minimax m3.1 flash all along 🐰
56.7T tokens served in stealth → highest volume of any stealth model on openrouter ever → #1 on opencode for 12 days → nobody knew who built it
here's what it actually does:
coding → 73.7% on deepswe
real multi-step software tasks → understand the repo → build the feature → test its own work
computer use → 61.7 on osworld v2
clicks, switches apps, finishes the job instead of explaining it
security → 82% on cybergym
vulnerability discovery and patching at frontier level
context → 1M tokens → entire codebases in one shot
speed → 200+ tokens/sec
what it costs:
input .................. $0.15 / 1M
output ................ $0.60 / 1M
cached input ....... $0.015 / 1M
~70% cheaper per agentic task than its predecessor m3
the point isn't the price per token → it's the price per finished task → top-tier deepswe and agent bench scores at a fraction of the cost per task of the big closed models
and the part nobody's talking about → open weights drop oct 23 → frontier-level agent you can run on your own hardware
launch deals right now:
unlimited m3.1 flash for minimax code subscribers → until oct 25
50% off your first m plan month → until oct 31
save this before the deals expire → links in the comments
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
Antid retweeted
My friend fired his 7-person dev team last Tuesday. 3 weeks after this Boris video Claude Tag in Slack ships 65% of his PRs and he pocketed $180K/mo
boris cherny, creator of claude code, quietly leaked the anthropic slack playbook in an 11-min product reveal:
02:44 the model that runs 16 hours without a check-in even researchers can't measure
03:19 the claude session boris hasn't opened in 47 days that still ships PRs every morning
09:15 the 65% number anthropic just admitted on video by accident
By the end of this video 65% of Boris' team is unemployed. He says it like it's Tuesday.
open 09:15. if that number doesn't make you tag claude in slack tonight, unfollow me.
Andrej Karpathy leaked why OpenAI is 5 years behind on AI agents - and named who's actually at the frontier now:
00:53 - why he pulled the plug on OpenAI's first agent project in 2016
02:22 - the 10-year timeline every "AI agent startup" is ignoring
04:36 - the labs he says are NOT at the edge anymore
This is the memo OpenAI won't send. He said it on stage to a room of hackers.
5 minutes. or you can keep copying labs that admit they're behind.
Antid retweeted
Andrej Karpathy leaked why OpenAI is 5 years behind on AI agents - and named who's actually at the frontier now:
00:53 - why he pulled the plug on OpenAI's first agent project in 2016
02:22 - the 10-year timeline every "AI agent startup" is ignoring
04:36 - the labs he says are NOT at the edge anymore
This is the memo OpenAI won't send. He said it on stage to a room of hackers.
5 minutes. or you can keep copying labs that admit they're behind.
Boris Cherny, creator of Claude Code:
"Onboarding at Anthropic dropped from 3 weeks to 3 days. The intern next to me ships more than your senior does. I run 40 Claudes in parallel from one laptop. You're not behind - you're already lapped."
In 26 minutes he walks the exact setup Anthropic gives every new hire on day one
Most engineers are 6 months behind an intern they've never met
26 minutes. or your next 3 weeks of guessing
Antid retweeted
Boris Cherny, creator of Claude Code:
"Onboarding at Anthropic dropped from 3 weeks to 3 days. The intern next to me ships more than your senior does. I run 40 Claudes in parallel from one laptop. You're not behind - you're already lapped."
In 26 minutes he walks the exact setup Anthropic gives every new hire on day one
Most engineers are 6 months behind an intern they've never met
26 minutes. or your next 3 weeks of guessing
Anthropic Applied AI team:
"We handed Claude a Swedish car crash form. It thought the accident was a skiing incident on a street that doesn't exist. 4 fixes later Claude was writing verdicts a $200K/year claims adjuster couldn't dispute."
In 24 minutes they walk the exact 10-step system Anthropic uses with paying enterprise customers
Most people are still prompting like it's ChatGPT in 2023
watch the first 5 minutes. if you don't spot a mistake you're making today, unfollow me
Antid retweeted
Boris Cherny head of Claude Code leaked how one Anthropic team stopped writing 65% of their code overnight:
02:44 - how they measure a model that runs 16 hours without a check-in
05:45 - the one-line slack instruction that replaced their entire data team
09:15 - the 65% number anthropic doesn't want screenshotted
this 11-minute reveal gives you for free what Anthropic charges enterprise clients $200K/year for
watch the first 4 minutes. if you don't see your job in it, unfollow me
Antid retweeted
Sam Altman CEO of OpenAI leaked why 2026 is the last window to go from zero to first million with AI:
00:04 - the sentence about kids born today he tried to walk back
23:20 - why he'd be "the luckiest kid in history" if he were 22 today
23:36 - the one-person $1B company he confirmed is now possible
This 50-minute sit-down gives you for free what OpenAI advisors charge $10,000 an hour for.
every hour you scroll, a 22-year-old ships your idea.
Antid retweeted
THIS HARDWARE COMPANY CREATED A HOLOGRAPHIC 3D SCREEN THAT LITERALLY LETS YOU SEE INSIDE A PATIENT'S BRAIN
If you are still trying to analyze complex medical scans or diagnostic imaging on flat, two-dimensional screens, you are looking at data through an outdated lens. This company, Illumetry, just completely changed the game by building a literal holographic 3D monitor setup that lets you interact with a human brain scan right on a desktop.
Tbh, watching how it works in practice is pretty mind-blowing. The setup shows a specialized display tracking a physical frame that someone holds in their hands. As they move the frame closer, tilt it, or rotate it, the monitor instantly shifts the angle of the 3D head scan, showing a highly detailed cross-section of the brain tissue beneath it.
There are no clunky virtual reality headsets or heavy goggles required. The screen itself handles the deep spatial computing, calculating the exact perspective in real-time to create genuine depth perception out of thin air.
It is an incredibly strong proof of concept for localized medical engineering, giving specialists a way to dig into raw diagnostic data naturally
And what if we added AI to all of this as well?
Antid retweeted
KENJI'S $40 RTL-SDR V4 MONITORS HIS HOME RF ENVIRONMENT ACROSS 30-1700 MHZ AND FEEDS EVERY DETECTED SIGNAL INTO A LOCAL QWEN 3.6 ON HIS RTX 3090 AT HOME, KILLED $200 CHATGPT PRO AND $200 CLAUDE MAX FOR EVERY RF DEBUGGING QUERY HE USED TO PAY THE API FOR
00:04 kenji holds up the earbuds and says, "this is listening to the air around me, and it's actively feeding me audio alerts out of these headphones"
the rig stacks a $40 RTL-SDR Blog V4 dongle with a wideband discone antenna on his balcony, a raspberry pi 4b running his own python scripts flashed from github, a passive USB hub and a pair of wired earbuds, the whole box lives on the shelf next to his router
kenji tuned the sdr to sweep 30 to 1700 mhz continuously, the raspberry pi runs an fft on the raw iq stream and triggers an audio ping whenever activity appears on his own iot band or his own mesh, every alert timestamps into his obsidian vault seconds later
the vault pipes into a qwen 3.6 27b served off a used rtx 3090 in his apartment, the local llm decodes modulation on his own beacons, flags interference on his home mesh and drafts fixes for his ham radio kit without a single token touching an api
$200 chatgpt pro and $200 claude code max died the week the sdr came online because every rf debugging query, antenna math and modulation question now routes through €680 of used silicon at zero recurring cost
the window is open, follow and bookmark before it closes
Antid retweeted
A DUAL ESP32 TOUCHSCREEN BOARD RUNS HIS HOME SECURITY AUDIT AND PIPES HIS AI QUERIES INTO A LOCAL QWEN 3.6 OVER THE SAME LINK, KILLED $200 CHATGPT PRO AND $200 CLAUDE CODE MAX SUBSCRIPTIONS THE DAY THE HANDHELD WENT LIVE
00:04 rui holds up the handheld and says, "this is a controlled test on my own equipment, there is no victim here"
the setup is a sub-ghz handheld with a dual esp32 touchscreen router board on top running momentum firmware, rui uses it to run periodic security audits on his own home wifi cameras and smart-home iot mesh
the same board tunnels his phone into his private wireguard network when he's testing on-site, ai prompts route from his phone through the handheld to a home rtx 3090 hosting qwen 3.6 27b without ever touching cellular
claude code on his laptop points at the local model over tailscale, the cli behaves identically to the cloud version except every token runs on his own silicon at home during a friday launch instead of on someone else's servers
$200 chatgpt pro and $200 claude code max used to hit his card monthly, both cancelled the day the handheld paired with his home rig because his private wifi and rf mesh handles what those apis used to charge for
the window is open, follow and bookmark before it closes
Antid retweeted
$150 IN "AS IS" TESLA K80 CARDS FROM 98% FEEDBACK EBAY SELLERS TURNED OUT TO BE TWO BRICKS AFTER 3 WEEKS OF PINOUT DEBUGGING, RETURN WINDOW CLOSED, THE ARTICLE'S CHEAP GPU FLOOR IS A TRAPDOOR
06:23 the tinkerer drops both cards on the bench with a shrug, "I got them on eBay long enough ago that I can't return them anymore, I think I just got screwed out of like 150 bucks"
3 weeks of debug preceded the giveup, adapter pinout traced by hand, 12V confirmed on all 8 pins with a multimeter, 1125W PSU verified, x16 slot known good, both K80s click every 30 seconds and refuse POST
Kepler silicon from 2014 has zero CUDA 12 support and no ollama backend, even if the cards had booted they would run no modern LLM without a custom llama.cpp compile against compute capability 3.7
the article recommends "98%+ feedback sellers" for used GPUs, that filter passes any eBay flipper willing to write "as is" in the listing, high feedback protects against the seller ghosting you, not against the card being dead silicon
a $700 used 3090 with a 30 day return policy from a specialty vendor is a different transaction entirely, cheap does not mean safe, the article's cheap GPU floor drops straight through 5 years of dead Kepler into a $150 hole
the window is open, follow and bookmark before it closes
Antid retweeted
MARCUS FLIPPED HIS 8 GPU CRYPTO MINING RIG TO LOCAL AI AND KILLED A $5,280/YEAR OPENAI AND CLAUDE BILL FOR $64 A MONTH
marcus is 26, lives in austin, spent 2021 mining flux on a basement rig of 8 RTX 3090 founders edition cards loud enough to hear from the kitchen
the rig pulled $4,200 a month at peak, dropped to $180 after the merge, by 2025 he was burning more in electricity than it earned
he pointed ollama at the same rig in march, loaded qwen 3.6 27B on one card, the whole stack runs local AI faster than claude code over the cloud
he was paying $440 a month across claude code, chatgpt, gemini and cursor, cancelled all four the same week with one environment variable change
total ongoing cost $64 a month, qwen 3.6 27B beats claude 4.5 opus on vision 84.1 to 77, the rig that mined crypto now kills $5,280 a year in subscriptions
the window is open, follow and bookmark before it closes
Antid retweeted
A BROKE 19 YEAR OLD SPENT HIS LAST $800 ON VAPE MACHINES IN A COLLEGE BAR. 24 MONTHS LATER HE DROVE HOME IN A LAMBORGHINI
in 2023 he was broke, no savings, no investors, no business, just an idea to put vape vending machines somewhere students already drink
he walked into a popular college bar near campus and pitched the owner a profit split, no upfront cost to the bar, the owner liked it and signed the same week
he found a used vape machine on facebook marketplace for $1,800, bought it on a credit card his mom co-signed, restocked it himself every friday after class
the first machine cleared $2,400 a month, the owner got his cut, he kept reinvesting every dollar into the next placement, year one ended at 31 machines across 19 bars
24 months in he runs over 100 machines across 4 states, clears six figures a month in profit, drove a black lamborghini urus home last weekend, paid in cash from the route
the window is open, follow and bookmark before it closes
This quoted post is unavailable.
Antid retweeted
AMD ENGINEER'S PALM-SIZED MINI PC RUNS 235B MODELS FOR $9/MONTH AND KILLS A $200/MONTH CLAUDE CODE AND OPENAI SUB
at CES 2026 in las vegas, AMD CEO lisa su walked on stage with a small black box behind her, not a server, not a data center render, a mini PC the size of a hardcover book
a few months later in shanghai she walked up to that same device and signed it with her name, the box is the gmktec EVO-X2, $1,700 once, AMD ryzen AI max+ 395 inside
the chip is the first x86 silicon ever built that runs a 200 billion parameter model on one piece of hardware, 128GB unified memory, 110GB usable VRAM on linux, no separate graphics card
it runs qwen 3 235B fully and smoothly, kills a $440/month stack of claude code, chatgpt, gemini and cursor for $9 a month in electricity, no per token fees, nothing ever leaves the machine
the window is open, follow and bookmark before it closes
Antid retweeted
A 28 YEAR OLD AMD ENGINEER WIRED 4 FRAMEWORK DESKTOPS INTO A CLOSET RACK AND KILLED A $5,280/YEAR AI STACK FOR $36 A MONTH
ethan is 28, lives in austin, works as a backend engineer and was paying $440 a month across claude code max, chatgpt pro, gemini advanced and cursor for his side projects
he saw the math, ordered 4 framework desktop nodes with AMD ryzen AI max+ 395 chips, wired them into an 8U black T1 mini rack in his closet with a 5 gigabit ethernet switch
each node carries 128GB of unified memory and 110GB of usable VRAM on linux, the cluster runs qwen 3.6 235B fully and smoothly with no quantization, no rate limits, no per token fees
total hardware cost $6,800 once, ongoing cost $36 a month in electricity, break even at 15 months on his old subscription stack and pure savings after that for the next 5 years
his home assistant dashboard tracks the power draw in real time, the rack runs cooler than his gaming PC and pulls less wattage than the dishwasher he runs every night
the window is open, follow and bookmark before it closes
Readers added context they thought people might want to know
The described AI cluster was built by independent developer Jeff Geerling, not a 28-year-old AMD engineer named Ethan. jeffgeerling.com/blog/2025/i-cl…
Antid retweeted
THIS IS WHAT KILLS A $200/MONTH AI STACK. A $249 NVIDIA CHIP THE SIZE OF A WALLET PULLING 67 TRILLION OPERATIONS A SECOND FOR $2 A MONTH TO RUN
three months ago my openai API bill hit $180 for the month, running automations, content pipelines, coding assistants, every prompt costing money, every conversation leaving my machine
someone in a discord posted a photo of a box the size of a deck of cards under their desk with one caption, "cancelled my openai subscription, runs llama locally, $2 a month to run"
the box is the jetson orin nano super, 67 trillion AI operations per second, 1024 core nvidia ampere GPU, 7 to 25 watts, smaller than a wallet, $249 one time on amazon
setup takes 3 commands and 10 minutes, ollama loads llama 3, mistral, gemma or deepseek locally, your code changes one line and points to localhost instead of openai
break even at 2.5 months on a $100/month API spend, after that you save $98 every month, no rate limits, no per token fees, nothing ever leaves your network again
the window is open, follow and bookmark before it closes
Antid retweeted
$38,000 IN 6 WEEKS FROM A STACK OF 3D PRINTERS IN A 200 SQUARE FOOT GARAGE
he started 6 weeks ago with one bambu lab A1 from amazon, posted the printer running on tiktok, took the first order that night for a custom car dashboard mount
by week 2 he had 4 printers running 24/7, by week 4 he had 12, every machine paid for itself in under 9 days from the orders it printed
$38,000 in stripe by week 6, no etsy listing, no shopify store, every sale came from one tiktok account showing the machines running unattended overnight
the median etsy 3D printer seller pulls $183 a month and stops there, the guys treating it like a tiny factory are clearing 5 figures a month from a corner of the garage
the wall fell in 2026, real manufacturing now fits in a bedroom and costs less than a used iphone, most people still see a toy that prints plastic dragons
the window is open, follow and bookmark before it closes
A 24 YEAR OLD VIETNAMESE FOUNDER MADE $47,000 IN 45 DAYS FIRING A $5,000/MONTH MARKETER AND RUNNING 30 CLAUDE MYTHOS CAMPAIGNS
david protein ran an on-brand content stunt this year and pulled +1,000% engagement on instagram and +14,000% on x. he did it with a full team and a moment of inspiration
most agents replicate the copy fine. they break on brand. the guidelines live in a pdf, a figma file and somebody's head so the output drifts to a stock-photo average of the internet
the fix is two layers, not one. distribution gets handled by an agent skill that schedules across 25+ platforms. brand gets called as its own callable system the agent can pull from
20+ on-brand creative variants a month lifts roas 65% over brands testing fewer than 10. the fatigue window collapsed to 2 to 4 weeks so the only edge left is volume that stays on brand
most people will keep paying a marketer $5,000 a month to run 4 experiments and call it a strategy. the ones running 30 on autopilot in 2026 will look very far ahead in 2028
bookmark this and read the article below
Antid retweeted
A 24 YEAR OLD VIETNAMESE FOUNDER MADE $47,000 IN 45 DAYS FIRING A $5,000/MONTH MARKETER AND RUNNING 30 CLAUDE MYTHOS CAMPAIGNS
david protein ran an on-brand content stunt this year and pulled +1,000% engagement on instagram and +14,000% on x. he did it with a full team and a moment of inspiration
most agents replicate the copy fine. they break on brand. the guidelines live in a pdf, a figma file and somebody's head so the output drifts to a stock-photo average of the internet
the fix is two layers, not one. distribution gets handled by an agent skill that schedules across 25+ platforms. brand gets called as its own callable system the agent can pull from
20+ on-brand creative variants a month lifts roas 65% over brands testing fewer than 10. the fatigue window collapsed to 2 to 4 weeks so the only edge left is volume that stays on brand
most people will keep paying a marketer $5,000 a month to run 4 experiments and call it a strategy. the ones running 30 on autopilot in 2026 will look very far ahead in 2028
bookmark this and read the article below
This quoted post is unavailable.