New AI releases, real tests, honest verdicts. I use this stuff daily and tell you what holds up.
Joined September 2026
- Tweets208
- Following85
- Followers289
- Likes155
HOLY F*CK. 99% of people use ai like a chat. the 1% hired it.
one bot finds trends. one writes. one kills fake numbers.
a manager runs them. you just sign.
same $20 subscription.
most people pay it. the 1% turn it into $2,000.
HOLY F*CK. this is how a SpaceXAI engineer gets agents to do 80% of her work without breaking stuff
Lauren Tan:
"I take every check I'd do by hand and turn it into a skill the agent runs itself"
her setup is called pstack. every request goes to a playbook, the playbook picks the skills
step 1 - give the agent a verifier it can actually run
1. start the app, make sure it's up
2. use the feature, check what happened
3. save proof, kill processes, clear temp files
try it on one real feature first
step 2 - the proof has to fit the change
UI - the full flow, start to finish
CLI - command, output, exit code
storage - write it, read it back
performance - same build, same cache, same load
one lucky benchmark doesn't count. /benchmark-checklist catches that
same bug twice? /correct fixes the root. architecture, then types, then CI, then a test
and the fix has to catch the original bug or it's not a fix
anything boring goes into a script for the next agent
save this before you launch your next agent
HOLY F*CK. Lauren Tan doesn't trust her coding agents at all. she makes them prove everything
she's an engineer at SpaceXAI and hands off like 80% of her work this way
"I turn my manual checks into a reusable verification skill"
so I put her whole method on one page
• the agent gets a check it can actually run. open the app, click through the feature, look at what happened, clean up
• if the same bug shows up twice she doesn't patch it again. she fixes the root, and the guard has to catch the original bug
• proof depends on what changed. UI change? show the full flow. CLI? show the output. storage? read the value back
• one good benchmark run means nothing. same build, same cache, same load, or it doesn't count
• anything boring she does twice turns into a script for the next agent
honestly most people just trust whatever the agent says it did
save this before you spin up your next agent. how do you check what yours ships?
HOLY F*CK. her laptop is closed and 12 AI bots are still working
She’s worth learning from
Lauren Tan, SpaceXAI:
"stop waiting for a smarter model. your bot can already hire one
12 bots, one computer, 12 screens. they don't stop when I close the lid"
8-minute Q&A. one bot doing everything - a team that hires its own agents
One Chat - A Bot Per Job - Separate Memory - Bots Hiring Agents - Work That Runs Without You
task too heavy? her bot spins up a cloud agent on whatever model fits and hands back the result
and she says the next Grok is already being built. this setup only gets stronger
most people are still waiting for a new model to paste the same old prompt into
skip the $5,000 AI agent course
save it, watch it tonight, then tell me - one bot or a team?
HOLY F*CK. 99% of people are using GrokBot wrong, and SpaceXAI's own senior engineer just said it
Lauren Tan:
"GrokBot is the most powerful agent tool we've ever released, and almost everyone uses it the wrong way
at SpaceXAI I run a team of 100+ GrokBot agents. a Chief of Staff, 10 managers and 90+ workers
this is what engineering looks like in 2026"
in 40 minutes she shows how to squeeze 100% out of every agent tool you already have
GrokBot - Chief of Staff - Managers - 90+ Workers - Full Team
no $5,000 agentic course online comes close
skip netflix tonight and watch this instead. then tell me - are you in the 99%?
HOLY F*CK. Lauren Tan from SpaceXAI ships 2,000 PRs a month and doesn't watch a single one
"you can't launch a hundred agents if you don't trust what one of them puts out
no verification skill means you're the one verifying. and that makes you the bottleneck"
in a 1-hour talk she shows how she went from babysitting every output to 2,000 PRs a month on autopilot
Watching every step - Verify - Skills - Evals - Auto-merge
trusting your agents isn't a feeling. it's infrastructure you build
beats any $3,000 AI coding bootcamp
save it, watch it tonight, then tell me - would you let your agents auto-merge?
HOLY F*CK. she goes to sleep and 50 AI agents build her company overnight
Lauren Tan, senior engineer at SpaceXAI (ex-Cursor):
"almost everyone uses GrokBot the same way. open a chat, ask one thing, close it
I treat mine like a company. a Chief of Staff runs my schedule, 3 PMs divide the tasks, 40+ workers build through the night. I wake up and review it with my coffee. that's 20 engineers of output"
in a 1-hour workshop she sets up the whole team from nothing
single GrokBot - Chief of Staff - PMs - 40+ workers - a company on autopilot
97% of engineers at SpaceXAI already work like this. the strongest ones get $750K+ offers
one chat tab vs a whole team. guess who wins
worth more than a $5,000 course
save it, watch it today, then tell me - one agent or a team?
HOLY SH*T. an Anthropic engineer just said prompting is done
"prompting is pretty much over
95% of our engineers already run loops that improve themselves. graphs are what's next"
in just 10 minutes she builds her whole Claude Code setup live, starting from a blank terminal
Blank Terminal - Agents - Loops - Graphs - System That Improves Itself
prompting was the old way to work. building the system around the model is the new one
beats most $1,000 agent engineering courses
save it, watch it today, then tell me - are you still writing prompts by hand?
HOLY SH*T. she makes $1.2M a year at SpaceXAI and hasn't written code by hand in months
"I build 150+ GrokBot teams that ship 1200+ updates a day while I sleep. I set the rules and check what the bots delivered"
in a 1-hour livestream she starts with an empty GitHub and zero bots
72 hours later the bots are running a company
how it works:
step 1 - one bot, one job, its own cloud computer
step 2 - show it a task once. it saves it as a skill forever
step 3 - drop all the bots in one group chat so they pass work to each other
step 4 - set the rules, let them run, check the results
the 4 a.m. run still happens with your laptop closed
GrokBot - skills - bot team - group chat - a company that runs itself
SpaceXAI's rule: "a persona, not a task"
anyone can open a GrokBot chat. almost nobody builds a team that works without them
beats any $2,000 agentic course
save it, watch it today, then tell me - what job would you give a bot first?
HOLY SH*T. Lauren Tan doesn't write code anymore. she runs a company made of AI agents
ex-Cursor, now senior engineer at SpaceXAI:
"99% of people open one GrokBot, ask a question and close the tab
I run 50+. a Chief of Staff plans my day, 3 PMs split the work, 40+ workers build it while I sleep. in the morning I check the results over coffee. that's the output of 20 engineers"
in a 1-hour workshop she builds that exact team from zero
one GrokBot - Chief of Staff - PMs - 40+ workers - a company that runs itself
97% of SpaceXAI engineers already work this way. the best ones are getting $750K+ offers
if you're still in one chat tab, that's who you're competing with
better than any $5,000 course
save it, watch it today, then tell me - how many agents are you running?
This is f*cking wild. SpaceXAI engineer Lauren Tan:
"You can't run 100 GrokBot agents if you don't trust a single one
If nothing checks your agents, you're the checker. And you're the bottleneck"
1 hour. she shows how she went from re-reading every answer by hand to a GrokBot team that checks itself and ships 50+ updates a day
GrokBot - Agents - Self-Checks - Trusted Team - Work That Ships Itself
Trust isn't a vibe. It's a system you build once
Most people spend longer checking their AI than it would take to just do the work
No $2,000 agentic course comes close
save it, watch it tonight, then be honest - how much of your day goes to checking AI?
this is f*cking wild. SpaceXAI engineer Lauren Tan:
"Most engineers still read every single line their agent writes
I run 60+ agents and haven't looked at their code in months. Every review note I ever left is a CI rule now"
In a 53-minute session she walks through how her agents merged 3,500 PRs last month while she was asleep
Prompting - Reviewing - Verifying - Delegating - Sleeping
Reviewing code used to be the job. Turning reviews into rules is the job now
Better than any $5,000 agent engineering course
save it, watch it, then tell me if you'd merge code you never read
AI Duck retweeted
Claude Code just found proof that every app on your phone is listening to your conversations. TikTok, Instagram, Facebook, Snapchat. they're using what you say out loud to build your feed. i have the data to prove it.
i've suspected this for months. last week i finally checked.
i was on a call with a friend and told him i'm thinking about breaking up with my girlfriend. didn't text it, didn't search it, just said it out loud with my phone on the table. two hours later she's lying on the couch scrolling TikTok and her entire feed is breakups. "signs your relationship is over." "how to heal after a breakup." she looked at me and said "why is TikTok showing me this." i said i don't know. but i did know.
that night i connected my iPhone to my MacBook and ran Claude Code. it scanned every app's background microphone activity for the last 90 days. apps that were only supposed to have photos or location were activating the microphone anyway.
→ TikTok activated the microphone 347 times while the app was closed
→ Instagram 211
→ Facebook 189
→ Snapchat 94
i just sat there staring at the screen.
every activation lasted between 3 and 15 seconds, enough to catch a sentence. every one of them tagged a keyword and cached it locally with a timestamp.
Claude Code pulled the keyword cache:
→ "breakup": tagged 4 hours before her feed changed
→ "pregnancy test": whispered it to my girlfriend in the bathroom, baby ads the next morning
→ "laid off": said it to my mom before i told anyone at work, job board ads within an hour
→ "lump": my dad told me about his doctor's appointment, cancer screening ads that night
→ "therapist": a call i made from my car, mental health app ads the same day
TikTok had the most entries, Instagram was second. every keyword matched a conversation i had out loud near my phone.
the "lump" one is the one i can't stop thinking about. my dad hasn't even told my mom yet.
Claude Code checked my girlfriend's phone in the same room. same keywords, same timestamps. both phones were picking up the same conversation and tagging it on their own.
i asked it how to shut it down. it mapped every app's microphone access.
what i did:
→ revoked microphone access from every app that doesn't need it
→ disabled background app refresh for TikTok, Instagram, Facebook, Snapchat
→ turned off "Hey Siri" (apparently it keeps the mic listening 24/7)
→ revoked "nearby interaction" (no idea what it even does, Claude Code said to kill it)
Claude Code confirmed all four apps lost background microphone access. did the same on her phone, about 3 minutes each.
within 48 hours both our feeds went random. generic stuff, nothing from our conversations.
full step by step and the prompt are in the first reply. save it so you don't lose it, and send it to anyone who's ever said "my phone is definitely listening to me." it is.
Claude Code found 145,763 photos I permanently deleted from my iPhone.
Three years of deleted photos and videos still sitting in a hidden cache. Any app with file access can read them.
I plugged my iPhone into my MacBook to run a backup. Claude Code was doing the security audit from my last post and flagged a directory I wasn't even looking for.
PhotoData/Thumbnails. Over 146,000 files. Photos, videos, screenshots. Every single one was something I deleted.
Photos from 2023, 2024 and 2025. The ones I deleted the second after taking them. The ones I cleared from Recently Deleted. The ones I deleted three separate times because I wanted to be sure they were gone. All still there.
Passport photo I took for a visa application in 2023.
Screenshots of my bank app with full account numbers visible.
Private photos I sent someone and immediately deleted after a breakup.
Screenshot of my crypto seed phrase I took temporarily in 2022.
Photo of my ID and credit card I took to remember the numbers and deleted 5 seconds later.
Screenshots of my ex's private Instagram stories I saved and permanently deleted in 2023.
Some were full resolution. All organized by date going back to 2021.
When you delete a photo on iPhone, iOS removes it from your camera roll and Recently Deleted. It keeps a compressed copy in a Spotlight search index cache so your phone can find things when you search. Apple never clears it. Not after 30 days. Not after a year. Never.
They told you it was permanently deleted. It wasn't.
I checked which apps on my phone had Files and Folders access. 23 apps. Weather apps. QR code scanners. PDF readers. A flashlight app I downloaded in 2022. Any one of them could have been reading every photo I ever deleted this entire time.
Went from 23 apps to 3 in about two minutes.
Full step by step below. Takes 2 minutes. Save it so you don't lose it and send it to anyone with an iPhone.
Readers added context they thought people might want to know
The video is an old, misappropriated clip from 2023, and the accompanying claims are fabricated.
x.com/HarshithLucky3…
HOLY F*CK. you can build a whole motion design studio for $0
20 free github repos. opus 5.5 writes the code, these handle the rest
no after effects. no $99 plugins. no render farm
just open source stuff that studios already use
ANIMATION
1. Remotion - make videos with react code github.com/remotion-dev/remo…
2. HyperFrames - write HTML, get an mp4 github.com/heygen-com/hyperf…
3. Anime.js - smooth timeline animation github.com/juliangarnier/ani…
4. React Spring - motion that feels physical, not robotic github.com/pmndrs/react-spri…
5. GL Transitions - those clean shader transitions between scenes github.com/gl-transitions/gl…
2D AND VISUALS
6. PixiJS - 2D graphics that run on the GPU github.com/pixijs/pixijs
7. p5.js - generative art in a few lines github.com/processing/p5.js
8. Paper.js - vector paths and shapes github.com/paperjs/paper.js
9. tsParticles - sparks, confetti, smoke, all of it github.com/tsparticles/tspar…
3D AND PHYSICS
10. Three.js - full 3D scenes in the browser github.com/mrdoob/three.js
11. OGL - custom shaders without the bloat github.com/oframe/ogl
12. glTF Transform - makes heavy 3D models light github.com/donmccurdy/glTF-T…
13. Matter.js - stuff falls, bounces and crashes for real github.com/liabru/matter-js
14. Rapier - 3D collisions github.com/dimforge/rapier
SOUND
15. Chatterbox - AI voiceovers github.com/resemble-ai/chatt…
16. Tone.js - synths and sound fx from code github.com/Tonejs/Tone.js
17. faster-whisper - captions in seconds github.com/SYSTRAN/faster-wh…
EDIT AND EXPORT
18. OBS Studio - record your screen github.com/obsproject/obs-st…
19. OpenCut - edit on a timeline github.com/OpenCut-app/OpenC…
20. FFmpeg - export to anything github.com/FFmpeg/FFmpeg
brief - prototype - animate - render - ship
don't grab all 20 at once. Remotion + Anime.js + FFmpeg is enough to ship your first video today
AI Duck retweeted
I'm part of Anthropic's $80.7B loss. 🙃
My October in Claude Code: $723 at API prices. I pay a flat subscription.
This week's viral bill: $5,107 of usage, ~$240 paid.
Reply with yours: usage / paid.
Tomorrow I'll chart every spread.
AI Duck retweeted
Night City running in Chrome is wild. My browser game is smaller, just one red lawn tractor and over 400,000 blades of grass.
Every blade under the deck gets cut. Straight stripes drop more coins, and the coins buy a wider deck and a faster engine. You start in a neighbor's backyard and work up to a stadium pitch.
It's still in development, built with AI in three.js. You can already play it: neurox-lawn-mower.netlify.ap…
HOLY F*CK. 24 free github repos that basically turn claude code into a whole studio
motion design with opus 5.5 is blowing up right now. so here's the actual toolkit
most people run claude code with nothing on it. no memory, no taste, just burning tokens on raw logs
these repos have 1.8M stars between them. picked out of 68,900 tagged claude-code
marketplaces want $19 for one skill and $39 for one agent. this is hundreds of them for free
PART 1. motion studio
animation, physics, 3D, sound, export. all of it
1. Three.js - 3D scenes
github.com/mrdoob/three.js
2. faster-whisper - auto captions github.com/SYSTRAN/faster-wh…
3. Remotion - videos made with react
github.com/remotion-dev/remo…
4. tsParticles - particle effects github.com/tsparticles/tspar…
5. Tone.js - synths and sound fx github.com/Tonejs/Tone.js
6. PixiJS - fast 2D graphics on the GPU github.com/pixijs/pixijs
7. Chatterbox - voiceovers
github.com/resemble-ai/chatt…
8. Matter.js - 2D physics
github.com/liabru/matter-js
9. HyperFrames - turns HTML into video github.com/heygen-com/hyperf…
10. GL Transitions - shader transitions
github.com/gl-transitions/gl…
11. Anime.js - timeline animation github.com/juliangarnier/ani…
12. FFmpeg - encoding and export github.com/FFmpeg/FFmpeg
PART 2. claude code plugins
1. claude-mem - remembers what your agent did last time and brings it back github.com/thedotmack/claude…
2. UI UX Pro Max - 79 styles and 192 palettes so your UI doesn't scream "AI made this" github.com/nextlevelbuilder/…
3. rtk - cuts 60-90% of the tokens from dev command output
github.com/rtk-ai/rtk
4. Everything Claude Code - 71 agents, 311 skills, 97 commands in one place. pretty much an agent OS github.com/affaan-m/everythi…
5. Agent-Reach - lets your agent read youtube, github, reddit and the web without paying for APIs github.com/Panniantong/Agent…
6. ponytail - makes claude code like a lazy senior dev. 53% less code, 45% fewer tokens in their own opus 5.5 test
github.com/DietrichGebert/po…
7. Graphify - turns your code, docs and SQL into a knowledge graph. stays on your machine github.com/Graphify-Labs/gra…
8. Ruflo - runs a bunch of agents at once across claude code and codex with shared memory github.com/ruvnet/ruflo
9. context-mode - 315 KB of tool output goes in, the model only sees 5.4 KB github.com/mksglu/context-mo…
10. career-ops - scans job boards, matches roles to your CV and writes the application github.com/career-ops-hq/car…
11. planning-with-files - saves the plan to disk so /clear or a crash doesn't wipe it github.com/OthmanAdi/plannin…
12. agent-skills - addy osmani's skills for the whole thing: spec, plan, build, review, ship github.com/addyosmani/agent-…
quick tip. don't install all of this
every plugin eats context, so more isn't free
pick 3. one for memory, one for tokens, one for whatever you actually do
run the same task before and after. if it's not better or cheaper, delete it
AI Duck retweeted
THIS IS F*CKING DANGEROUS
your crew loses one scout at 2.5s and still tells you "done". the report ships on 5 of 6 sources and nobody checks which one is missing
6 scouts. one dies at 2.5s. 1 retry. 0 lost. $1.96 for the run
the idea: stop letting one failed scout pass silently
if s.failed: s.retry(max=2)
here's the split:
→ 6 scouts pull from arxiv, github, docs, forums, web and papers in parallel
→ a budget guard caps the run at $4.00
→ merge, dedupe and rank fold 3 nodes into 1
→ the writer waits for the retry, then drafts 3.4k tokens
→ verify checks every claim against a source and a line
You only look at three moments:
→ the timeout. scout05 dies at 2.5s, retry 1/2 queued. if the writer doesn't wait, you're reading a draft with a hole in it
→ the fold. 3 → 1. this is where one result counted twice turns into "two sources agree"
→ verify. 12 sources ok, 0 copied claims, 2 claims drifted. read those 2 before report.md goes out
↳ concept board, sample numbers. full prompt in the reply
BREAKING: Your "deleted" photos were never actually deleted.
Claude Code just found years of them still on my iPhone.
I plugged my phone into my MacBook for a backup.
Asked Claude Code to do a quick security check while it was running.
It flagged something I wasn't even looking for.
A folder called PhotoData/Thumbnails.
Thousands of files.
Photos, videos, screenshots.
Most of them stuff I deleted.
Photos from 2 years ago.
Screenshots I deleted like 5 seconds after taking them.
Things I cleared out of Recently Deleted.
Still sitting there.
What was in that cache:
- old screenshots of my banking apps
- photos I deleted the second I took them
- pictures I never wanted on my phone
- stuff I remember deleting more than once
Here's how it works.
You delete a photo.
iOS takes it out of your camera roll and Recently Deleted.
But it keeps a compressed copy in the Spotlight search cache.
And that cache never really gets fully cleared.
Then I checked which apps had Files and Folders access.
Way more than I thought.
Apps I downloaded years ago and forgot about.
Any of them could've been reading that cache.
I revoked access from almost all of them.
The prompt I used is below.
Takes a few minutes if you already have a backup.
How many apps on your phone have Files access right now?
Here’s the prompt:
I backed up my iPhone to my Mac in Finder (no encryption). Can you help me dig through it for leftover photos and thumbnails of stuff I deleted?
1. Go to ~/Library/Application Support/MobileSync/Backup/, grab the newest backup folder and check Manifest.plist to make sure it isn’t encrypted.
2. Open Manifest.db with sqlite3 and pull every file with “PhotoData”, “Thumbnails”, “DCIM” or “Photos.sqlite” in its relativePath. Give me a count per folder.
3. Copy Photos.sqlite somewhere temporary and look in the ZASSET table for anything with ZTRASHEDSTATE = 1 (that’s Recently Deleted). List the filename, date and whether it’s a photo or video.
4. Use the fileID hash paths in the backup to copy every thumbnail and media file you found into ~/Desktop/photo_audit/, with their original names back.
5. When you’re done, give me a quick summary: total files found, how many are from Recently Deleted, how many are thumbnails, plus the oldest and newest dates.
Read only. Don’t change or delete anything in the backup.1.