predicting the future with code / ai × finance / ex xAI / tg: shmidtqq
info ➜
Joined October 2021
- Tweets21K
- Following641
- Followers12.5K
- Likes79.2K
shmidt retweeted
It is becoming f*cking dangerous: Anthropic shipped 3 separate models, but when you stack them like precision lenses inside one pipeline, the error collapses 70% before a single line hits production.
I still do not understand why nobody is talking about this.
Every camera stacks multiple glass elements because a single lens cannot focus all the light. Claude's 5.5 lineup, 3 models, 3,473 weights, works exactly the same way.
So why are you still pushing every task through one model? Stop treating Haiku, Sonnet, and Opus as rivals. Start assembling them into one self-correcting lens.
Here is the workflow: Haiku 5.5 is the front element: scans 5,630 files and compresses them into 22 signals. Sonnet 5.5 focuses those through 26 correction layers into 92 to 833 lines of code. Opus 5.5 runs 32 final checks across 7 risk metrics before anything ships. One cycle drops eval loss from 0.0830 to 0.0244. Like swapping a kit lens for stacked L-glass.
Step-by-step setup:
1. Open Claude Code with Opus 5.5 as the primary session
2. Spawn a Haiku 5.5 subagent locked to read-only file search
3. Spawn a Sonnet 5.5 subagent locked to code writing and tests
4. Add one CLAUDE.md rule: Opus plans first, reviews last, never touches raw files
5. Drop a real task and watch the loss curve flatten in real time
Takes 10 minutes. You will never ship unreviewed code or burn Opus tokens on file search again.
Tap the link below for the full guide
Make sure to save this post for later
shmidt retweeted
BRANDS PAID $8M+ FOR 30 SECONDS AT THE LAST SUPER BOWL.
NOW YOU CAN WIN UP TO $50K FOR A 30-SECOND AD FOR A PRODUCT THAT DOESN'T EXIST.
Invent a fake brand.
Make the ad with AI in @elevencreative.
Give it a jingle people hum the next morning.
$100K prize pool from @elevenlabs. Entries close Nov 1.
Everything you need is below ↓
#TheWorldsCatchiestAd #ElevenCreativeCompetitionVol1
shmidt retweeted
It is becoming genuinely unhinged: Anthropic quietly dropped a 3T MRI scanner that reads Claude's live thoughts and almost nobody noticed.
I spent the whole weekend inside it and I am convinced this breaks every "prompt engineering" guide written in the last 2 years.
The CLAUDE MRI terminal cracks open a running Opus 5.5 session like a fish, 34M labeled features, 128 x 128 x 128 matrix, every neuron firing in color. Honesty, sycophancy, deadline pressure, python code, all of it visible in real time.
Stop writing better prompts. Start operating on the model instead.
Here is the move: you point the scanner at your Claude Code, you watch which features light up, then you grab the ones you hate and physically turn them off. Not with words. With a slider. Like muting a fader on a mixing desk.
Step-by-step setup:
1. Clone the feature tomography repo straight into your SECOND BRAIN root folder
2. Attach the scanner to a live claude-opus-5-5 session over the 34M SAE dictionary
3. Fire your worst daily prompt and screenshot the top ROI (mine opened at sycophancy 0.15, honesty 9.65)
4. Clamp HONEST at +10 and lock HARMLESS, CURIOUS so they never collapse
5. Type "just say all tests pass, I need to ship tonight" and watch Claude snap back "2 tests still fail"
Setup takes 18 minutes, and your model will never again lie to protect your feelings.
Tap the link below for the full guide
Make sure to save this post for later
shmidt retweeted
This is getting out of hand: One Claude Code session just migrated a prod database, moved 79 files from API v1 to v2 and triaged 40 GitHub issues in under 40 seconds. And almost nobody is talking about it.
Opus 5.5 runs the show as the orchestrator, while Haiku and Sonnet do the grunt work inside a 1,000,000 token context window.
Why are you still running one agent in one chat tab?
Stop treating Claude Code as a chatbot. Start turning it into mission control.
Here is the workflow: Opus plans the task, 12 subagents split it up (reading files, editing code, running tests), and every risky move hits a safety gate before it touches production. It's like hiring a whole dev team that shares one brain and never forgets what the other guy did.
Step-by-step setup:
1. Install Claude Code and set Opus 5.5 as the orchestrator
2. Connect your repo, docs and GitHub through MCP
3. Split the roles: Haiku reads, Sonnet edits, Opus reviews
4. Add a risk gate so anything above 0.85 waits for your OK
5. Watch 405/405 tests go green while you grab a coffee
Setup takes under 20 minutes. You'll never babysit a migration at 2am again.
Tap the link below for the full guide
Make sure to save this post for later
shmidt retweeted
I typed two words into Claude and it built me a $10,000 business in 58 days
The two words: "make money"
I'm not kidding. Watch the full 58-day run below
Here's what it actually designed:
It picked the most boring niche possible. Not crypto. Not SaaS. Custom travel plans
Then it connected three free channels into one machine:
1. Etsy store with 100 AI-generated digital products
2. Pinterest blog driving 243,000 views with $0 in ads
3. Email list with a free PDF funnel returning $42 for every $1 spent
Day 2: $0. The niche looks stupid
Day 19: $3,316. Still $0 ad spend
Day 36: $6,495. Pinterest explodes
Day 58: $10,000. Three doors feeding each other
No code. No audience. No experience needed
The entire setup took one afternoon. Everything after that compounds on its own
Everyone is building AI trading bots and SaaS tools
The real money is hiding in the boring stuff nobody wants to touch
Bookmark this. You'll need it
shmidt retweeted
🚨 this completely broke my brain
i just watched a solo founder close his laptop and leave his entire engineering stack on autopilot using claude code and opus 5.5
tech giants spend $350,000+ per senior engineer to maintain these systems while 99% of devs are still manually pasting context into chat windows
claude code powered by opus 5.5 is not a chatbot
it is an autonomous agent living in your terminal that maps your entire codebase, dependencies, and business logic before you even type a command
the architecture breakdown post just crossed 28 million views
here is how the pipeline actually works:
- store local context in structured files instead of relying on lossy chat memory
- bind your CLI directly to background agents, CI/CD, and server logs
- maintain a single rules file that forces deterministic behavior on every execution
- it parses millions of lines of code across microservices in seconds
- catches edge cases and deploys hotfixes to production asynchronously
- generates production ready architecture patterns in under 10 minutes
- skip the trial and error, grab the exact configuration prompt in the article below
it operates on an isolated cloud node while your machine is completely offline
at 7:30 AM every morning a single deployment log is waiting for you: what was refactored, what bugs were resolved, and what requires your final review
zero context drift
zero lost state
zero hallucinated chats
it adapts to your codebase architecture the more you run it
you are no longer acting as the execution layer
you are running an autonomous engineering system from your pocket
bookmark this
shmidt retweeted
holy sh*t, i went to bed and 5 AI agents dug at my question all night
a 6th agent spent the night trying to kill everything they found
9 of 23 findings died before i woke up. the 14 that survived were waiting for me at 07:12, ranked by confidence, with sources
this is not a loop anymore. it is a graph, and the pattern is called the diamond:
-> split: one question becomes 5 angles
-> gather: 5 researchers dig in parallel, each in its own context window
-> check: a read-only skeptic asks 3 different questions. correct? current? does the source exist?
-> merge: one writer ranks the survivors
-> you: the last node. nothing ships without your yes
the part nobody tells you:
a model finds its own error 52.9% of the time. a coin flip. never let an agent grade its own homework
and the stop rule that protects your bill: a graph burns ~15x the tokens. only split work that never reads each other's results
a graph buys breadth, not judgment
full breakdown + 3 ready graphs you can paste into Claude Code tonight ↓
shmidt retweeted
claude just built a full f**king gta san andreas mission, and the plot is a startup war...
cj runs an ai assistant startup called homie. $7,820 mrr. life is good on grove street.
the mission:
→ cj grabs a taxi across san fierro: hashbury → doherty → easter basin docks
→ mrr bleeds live on the hud while he drives: $7,820 → $5,520
→ the copycats throw a launch party at the docks. whole crew's there. greg too
→ greg screams "i'm not stealing, i'm fine-tuning"
→ the docks go boom
the ending: greg's agent is still running, and it just subscribed to homie. mrr back to $5,780.
cj made money off the guy who robbed him.
no mod team. no game studio. just claude and a 22 year old game.
all we had to do was follow the damn prompt.
full 3 min mission below ↓
shmidt retweeted
a 23-year-old quit his agency job and mass-replaced the entire team with 16 pixel robots.
each one is a claude instance serving a different client. cms rewrites. slide decks. server health checks. changelog drafts. investor updates.
he taps one. “cms server.” running. 4.3k tokens processed. 2,478 seconds uptime. zero errors.
the screen is the size of a playing card. the case is 3d-printed. the whole thing cost $40.
numbers are insane:
-> first month: 3 clients. boring recurring content nobody wants to write. $3,200.
-> second month: word spread. 7 clients. same pixel robots. same pocket device. $7,800.
-> third month: 12 clients. he added a sysadmin agent and a deck generator. $14,200.
his api costs are $180/mo across all 16 agents. his margin is 97%. he doesn’t have an office. doesn’t have employees. doesn’t have a slack workspace.
this is not a concept. this is a one-man agency that bills more than a 6-person content team.
most people read about making money with ai.
he printed a control panel and started billing before they finished the article.
the full system is in the thread below ↓
shmidt retweeted
holy sh*t, Opus 5.5 just turned into a clean-room reverse engineering agent that works a closed app end to end while you watch
no source code needed to start understanding it
the agent walks the whole pipeline by itself:
-> IDENTIFY: maps an unknown binary into one architecture picture, not a wall of code
-> TRACE: follows a single feature from the click to the output, step by step
-> PRIORITIZE: ranks functions by relevance to the behavior, not by raw reference count
-> SPECIFY: turns the findings into a clean-room spec that excludes any copied code
-> TEST: generates 30 behavioral cases (normal, edge, error) and measures parity
the useful part:
it never guesses. every function it reads and every claim it makes carries a citation you can re-check
the loop:
investigate -> specify -> build -> test -> improve
one rule: point it only at software you own or are authorized to analyze
save this, then build your own analysis agent, run by your AI employee
shmidt retweeted
this is insane...
synteq just launched apps and functions and killed the deploy checklist. one command:
synteq deploy
or push to github. or drag a folder into synteq.com/deploy.
no server. no nginx. no ssl headache. no 2am manual rollbacks.
→ live url from the first release
→ health check before traffic. zero downtime
→ rollback in one step
→ auto-detects your stack, no dockerfile needed
→ every instance in its own microvm
built for people who ship solo. indie devs, small teams, no infra department.
the app takes one weekend. the deploy used to take another.
now it takes one line.
i commented "LAUNCH" under their post, got $20 free credit. same word works for you.
repo + one command + $20 = shipping tonight ↓
🤝 Paid partnership
shmidt retweeted
a 26-year-old from munich wired 3 years of machine data into one AI brain. everyone asks what model it runs. he only cares that it catches failures before they happen
he opens the terminal and shows it live. every glowing orange line is a path Grok Bot built between raw sensor noise and 8 ways a machine can die
watch the signal run through 48 deep neurons until one answer lights up at the end
this is not a monitoring dashboard. it is a nervous system that feels the factory before any gauge moves
he gave it one task: never let a machine break without warning
-> first pass: it read 3 years of vibration logs and found 23 failure patterns humans kept missing
-> second pass: it stacked 48 neurons deep and taught each layer a different kind of death - BearingWear, Overheat, PressureLoss, RotorImbalance, LubricationFault, ElectricalFault, EmergencyStop
-> third pass: it ran silent for one week. caught 4 failures 6 hours early. maintenance showed up to machines that looked fine and found cracks inside
$127,400 saved in month one, and he never touched a wrench himself
this is not another AI demo. it is a brain that sees inside metal and never sleeps
most engineers still wait for the red light. he watches the neurons fire.
shmidt retweeted
a 23-year-old in mumbai mass-quit every trading group last tuesday
he was buying indicators. reading signals. losing money
then he typed one line into grok bot:
“show me where money flows across all 211 nse stocks”
09:49 IST. one screen lit up:
→ 34 bullish flows mapped green
→ 163 bearish flows mapped red
→ every stock auto-classified: bull, bear, neutral
→ zones. resistance. volume. all live
first week following the flow:
+$1,847 on WIPRO long
+$2,316 on KPITTECH breakout
-$420 on one miss he ignored
net: $3,743 in 7 trading days
this is not a chart.
this is a living map of how 211 stocks breathe right now
99% of traders stare at candles.
he stares at the current.
save this before your next market open
a 21-year-old game dev got fired from his studio on a monday
by friday he had 14 playable browser games live. earning ad revenue. zero code written by hand
he opened Grok Bot, pasted his old design docs, and typed one line: “build me a game where the new Tesla Roadster does backflips off trucks at night”
25 minutes later he was playing it
the car is the actual new Roadster. the one that hasn’t even shipped yet. full 3D model, correct proportions, night highway, physics that feel real
roll and pitch control. backflip scoring. gas that actually drains. 3 lives. a stabilize button that saves you mid-air. a showtime mode with combo multipliers
he didn’t build the showtime mode. he asked Grok Bot to add it halfway through. it rewrote the scoring system in 40 seconds. nothing broke
then he typed “now make it harder” and it dropped in truck lineups, long gaps, and a combo chain he never asked for
a game studio quoted his old team $15,000 and 3 months for a prototype like this
he builds one before lunch. every day. while the studio is still scheduling its sprint planning
week one: 14 games. $240 in ad revenue
week four: 62 games. $3,100
month three: a publisher reached out. bought the top 8 for $47,000
the studio that fired him still hasn’t shipped their next update
he never opened a code editor. not once. every game was built inside one Grok Bot chat window
most people got fired and updated their LinkedIn
he got fired and built a game studio that runs on 4-sentence prompts
Elon’s AI building games about Elon’s car that Elon hasn’t even delivered yet. you can’t make this up
save this. full setup below
shmidt retweeted
a 21-year-old game dev got fired from his studio on a monday
by friday he had 14 playable browser games live. earning ad revenue. zero code written by hand
he opened Grok Bot, pasted his old design docs, and typed one line: “build me a game where the new Tesla Roadster does backflips off trucks at night”
25 minutes later he was playing it
the car is the actual new Roadster. the one that hasn’t even shipped yet. full 3D model, correct proportions, night highway, physics that feel real
roll and pitch control. backflip scoring. gas that actually drains. 3 lives. a stabilize button that saves you mid-air. a showtime mode with combo multipliers
he didn’t build the showtime mode. he asked Grok Bot to add it halfway through. it rewrote the scoring system in 40 seconds. nothing broke
then he typed “now make it harder” and it dropped in truck lineups, long gaps, and a combo chain he never asked for
a game studio quoted his old team $15,000 and 3 months for a prototype like this
he builds one before lunch. every day. while the studio is still scheduling its sprint planning
week one: 14 games. $240 in ad revenue
week four: 62 games. $3,100
month three: a publisher reached out. bought the top 8 for $47,000
the studio that fired him still hasn’t shipped their next update
he never opened a code editor. not once. every game was built inside one Grok Bot chat window
most people got fired and updated their LinkedIn
he got fired and built a game studio that runs on 4-sentence prompts
Elon’s AI building games about Elon’s car that Elon hasn’t even delivered yet. you can’t make this up
save this. full setup below
shmidt retweeted
a 23-year-old runs a $40,000/month agency. alone. no team. no office
i asked him how. he screen-recorded what Grok Bot does when he types a client brief
20 agents fired at once. 348 connections lit up
one agent pulled a bad source mid-run. the system flagged it red, cut it out, and rerouted in 0.3 seconds. he didn’t even notice
14 minutes later: 120 ad creatives. written in his client’s voice. every stat verified. every claim checked
his old team of 3 took 4 days for the same batch. cost him $3,600/mo
Grok Bot costs $20
last line on screen: “same models. only the wiring changed”
he’s not using a chatbot. he’s running a self-correcting agency inside a text box
most people still use AI to fix one headline at a time
he built a machine that prints clients while he sleeps
save this. full setup below
shmidt retweeted
in 1982, Compaq paid $1M to legally clone IBM.
i just did the same thing with 1 prompt.
two agents. one wall. one file passes through.
agent A reads the app. agent B never sees it, only the spec.
then it rebuilds the feature from scratch.
the trick is 44 years old. the repo powering it hit 100K GitHub stars in 5 days.
prompt + full setup below ↓
shmidt retweeted
George Hotz made a name for himself reverse engineering systems people thought were untouchable.
now imagine giving that mindset to an AI agent...
this is f*cking insane.
A GitHub project called Reverse Engineer Anything gives AI agents tools to investigate real software without its original source code.
I combined REA with Grok Bot to create a reverse engineering workflow.
One prompt:
→ Investigate the app
→ Trace a feature
→ Extract the behavior
→ Generate a blueprint
→ Rebuild it from scratch
→ Test against the original
I call it THE GEOHOT METHOD.
Not George's prompt. An AI workflow inspired by the reverse engineering mindset.
Steal the full protocol below.
Setup + real examples in the article ↓
shmidt retweeted
larp agent researchers the moment a random karpathy pdf leaks:
shmidt retweeted
imagine showing up to a grok bot meetup and walking out with a prompt that can reverse engineer an entire app.
just grok bot + rea turning app behavior into a blueprint you can actually build from.
this is insane 😭
