Sabitlenmiş Tweet
shmidt
17.9K posts
shmidt
@shmidtqq
predicting the future with code / ai × prediction markets / tg: shmidtqq
ai & web3 alpha ➜ Katılım Ekim 2021
557 Takip Edilen9.5K Takipçiler
@slash1sol tejas really said here's everything you need to know, no paywall
English
This is the most complete breakdown of AI agents I've seen, and it's free.
Tejas AI packed the entire landscape into one video: how LLMs actually work, the core agent loop, tool calling, memory, RAG, vector databases, embeddings, MCP, agent architectures, multi-agent systems, guardrails, and a week by week learning path.
The part that stuck with me: an agent isn't a smarter chatbot. A chatbot answers. An agent takes a goal, plans, calls tools, sees the result, and loops until the job is done.
If you're building anything with AI, start here.
Bookmark & Watch ↓
Yarchi@undefinedKi
English
my reaction when i paid $20/mo for ChatGPT and bought every $60 game like a good little consumer... then found a GitHub repo running a full AI locally (yeah, PewDiePie built it) and a bot that claims free Epic + Prime + GOG games while i sleep.
Ridark@ridark_eth
English
@ridark_eth the beauty thing hits different, that's the trap everyone falls into
English
MOST LOOPS YOU BUILD WILL COST MORE THAN THEY RETURN
A loop earns its setup cost only when all four are true. Miss one and it burns money instead of working.
- The task repeats, at least weekly. A one-off is cheaper to close with one good prompt.
- Verification is automated. No tests, no linter, and the agent grades its own homework. It is a generous grader.
- Your budget can absorb the waste. A loop burns tokens whether or not it ships anything.
- "Done" is objective. Exit code 0, not "looks good".
Try it on your own work.
"Triage failing tests" passes all four. "Bump dependencies" passes. "Make the interface feel nicer" fails the fourth and always will.
No agent will ever hand you exit code 0 on beauty.
Pass all four, build the loop. Fail one, write a prompt.
Timestamps:
0:21 - what a loop actually is
2:55 - loop for sub-50ms page load
6:54 - overnight doc sweep
9:31 - logging coverage
10:32 - production error sweep
12:10 - full product evaluation
13:54 - where loops break
Save and watch the clip.
shmidt@shmidtqq
English
THE CREATOR OF CLAUDE CODE SAID HE DOESN'T PROMPT ANYMORE HE WRITES LOOPS AND THE LOOPS DO THE WORK
everyone clipped the quote. almost nobody built what he described
a loop wakes up on a schedule, pulls your context (crm, transcripts, docs), reasons over what actually matters and acts inside guardrails so it can't go rogue. you state the outcome once, it keeps working while you sleep
old way: ask → answer → ask again
loop way: set the goal once → it never stops
prompting was stage 1. agents were stage 2. loops are stage 3 and the gap between people who write prompts and people who write loops is about to get very wide
full breakdown on how this actually works ↓
shmidt@shmidtqq
English
I FOUND A 21,000 STAR GITHUB REPO WITH 2,000 FORKS THAT TURNS CLAUDE CODE INTO AN AUTONOMOUS DEVELOPER WORKING THROUGH YOUR BACKLOG WHILE YOU SLEEP
ralph connects to Claude Code or Amp, reads your requirements, picks the highest priority unfinished story, opens a fresh session, writes the code, and verifies the result.
each iteration handles exactly 1 story. a default 10 iteration limit prevents wasted context, then successful changes get committed before the next task starts.
memory stays across 3 layers: prd.json tracks completed stories, progress.txt stores discoveries and mistakes, and Git history preserves every working change.
one cycle runs 8 automated steps, from creating a branch and choosing the task to typecheck, tests, clean commits, state updates, and the next story.
with 108 developers watching, this is more than another Claude wrapper. it turns a 200 file build into small verified tasks that can keep moving overnight.
shmidt@shmidtqq
English
HEDGE FUNDS PAY QUANTS SIX FIGURES FOR THIS, I JUST BUILT IT ON MY PHONE IN UNDER A MINUTE
I opened an app on my phone. Picked SP500, a 4h chart, and a trend-momentum style, then tapped Generate.
Seconds later it built the whole thing. A fast and slow EMA crossover, a momentum filter, RSI and Bollinger Bands, all tuned and coded for me.
Then it backtested a full year. The run came back at +10.83% with a 2.45% max drawdown, a 70% win rate and a 3.19 Sharpe. A smooth line that barely dipped.
That is a backtest, not a promise. But the whole build was a form and one tap, and the Deploy button to take it live was right there.
I put the full build, the real numbers & the honest caveats in the article. Read it ↓
slash1s@slash1sol
English
1 MESSY DESKTOP, 5 FOLDERS, 0 MANUAL SORTING.
The video is not about organizing files.
It is about what happens when AI gets access to the actual workspace instead of sitting in a chat box.
He gives it a cluttered desktop, and the model does the boring operator work: reads what is there, decides the categories, creates folders for Projects, Screenshots, Photos, Videos and Documents, then moves the mess into a usable structure.
That sounds small until you understand the pattern.
Most AI demos are still answer generation. This is computer control.
The post points at the same shift: useful agents are not just smarter text generators. They need context, permissions, tools and a loop that turns intent into action.
A chatbot tells you how to clean the desktop.
An agent cleans it.
That is the whole difference.
Yarchi@undefinedKi
English
HIS CI WENT RED AT 2AM. HE WAS ASLEEP. IT SHIPPED THREE PRs ANYWAY.
Not a demo. A repo, five text files, $5 a night.
- On-call engineer for CI: $180K a year
- The loop: $5 a run
- Time to build: one evening
But here is why your version will fail.
A prompt makes YOU the loop. Write, wait, fix by hand. It all halts the second you step away.
A loop runs itself and wakes you only to escalate.
The difference is not the model. It is whether something can fail the work without you in the room.
Read that again.
The model that wrote the fix will never call it bad. That is how you wake up to a green suite and a deleted test file.
Done is exit code 0. Never an adjective.
Tomorrow two people wake up. One opens three finished PRs. The other opens the repo he closed last night.
Same tools. Same models. One built the gate.
Save this before you build yours.
shmidt@shmidtqq
English
Claude (free plan) + YouTube = $77,000/month.
Stickman channel: 14 videos, 2 months, 14.5M views. Hand-drawn stick figures, zero paid tools.
Whole pipeline in 12 minutes. One step at 2:39 most people do backwards and that's the one killing their views.
Save this alpha and watch tonight
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
A $5 LOOP JUST DID THE JOB OF A $180K ENGINEER
CI went red at 2am. By morning: three ready PRs and an honest list of what it could not solve.
You wrote zero prompts.
The numbers are insane:
- On-call engineer for CI: $180K/year
- Contractor for triage: $4K/month
- The loop: $5 per run, capped by a flag
- Time to build: one evening
- Skill required: writing markdown
The difference between a prompt and a loop is not the model. It is the gate.
A prompt makes you the loop. Write, wait, fix by hand. Everything halts the second you step away.
A loop runs itself and wakes you only to escalate.
One person + five text files + exit code 0 = an engineer who works while you sleep.
Save and watch the clip.
shmidt@shmidtqq
English
@shmidtqq If you learn to use artificial intelligence correctly, you will gain a huge advantage
English
ONE WRONG PORT AND A $599 MAC MINI M4 THAT WON'T BOOT STAYS A DEAD SLAB OF ALUMINUM FOREVER, HE SHOWS THE ONE PORT THAT ACTUALLY REVIVES IT.
00:02 he pinches a braided usb-c cable and lines it up against the mac mini's rear panel, skipping the power connector and the hdmi side completely, then pushes it into the single thunderbolt port sitting right next to the power button before flipping the unit over to hold that button down.
this isn't a settings menu or a software reset. it's the one hardware level port that talks to the chip before macos even tries to load, so a mac mini that won't show an apple logo, won't boot, won't respond to anything becomes recoverable through one cable and a second mac instead of a trip to a genius bar appointment.
apple's $599 mac mini m4 ships with 4 rear ports and exactly 1 of them is dfu capable, apple configurator then pulls a restore image that runs past 15gb, and the whole sequence needs the power button held for roughly 10 seconds while that single cable stays connected the entire time.
the interesting number here is not the 15gb restore file or the $599 price tag. it is 1, as in exactly 1 port out of 4 that brings the machine back, plug into any of the other 3 and the same $599 box just stays bricked.
Gipp 🦅@gippp69
English
me before knowing this GitHub list: 💸💸
me after:
> $80 games → free-games-claimer: a bot claims them free
> Claude → odysseus: full AI, local, by PewDiePie
> paid AI watermarks → remove-ai-watermarks: free
> 1Password → Vaultwarden: free
> TeamViewer → RustDesk: free
saved: $1,000+/year
my wallet: healing
Ridark@ridark_eth
English
@Bober_smart this is the playbook right here, find what people need and execute
English
A 27-year-old guy from China bought 100 Mac minis and turned his apartment into a server hub
Instead of furniture, the space is filled with neat racks where a hundred Mac minis hum like a living organism, processing gigabytes of data in a continuous stream
Many small AI startups in Asia cannot afford to train models on NVIDIA H100 servers: it is insanely expensive
He provides his Mac minis for running and fine-tuning small, highly specialized models. He takes on the dirty work of data labeling and quality assurance by using his network of agents that check each other's work
> His initial investment was $59,900
> He was able to recoup his costs in 3 months
Right now, his average income is $25,000–$30,000 per month
He is not a genius: he simply found a niche with high demand and created an offer
Bober_smart@Bober_smart
English
@Lummox_eth this is the move, separating concept from execution unlocks real leverage
English
An architect partner changed the whole business model with one sentence : “You sell the concept, I stamp the buildable drawings.”
That split turned 1 solo operator into a legal $15,000 to $40,000 offer.
The studio handled pre-design, style exploration, floor plans, render sets, and walkthroughs. The licensed partner checked feasibility, code, structure, and the documents that actually go to permitting.
That boundary made the offer stronger, not weaker.
Clients got speed upfront and professional sign-off where the law required it.
The trick is not replacing architecture.
It is selling the expensive waiting room before architecture starts.
Shadow Nick@doublenickk
English
this had to be kept secret
researchers just published the exact framework for building a self-improving trading agen
tthe accurate way to make an LLM actually learn from real market feedback instead of hallucinating strategies
bookmark this before someone takes it down
arxiv.org/pdf/2402.03755
Horizon@horizon_trade_x
English
@vijn_crypto @Polystreetdotai no code trading agents hitting different, this changes the game
English
Finally, anyone can build their own trading agent
The best prediction market traders already use them
Everyone else is still doing it manually
@Polystreetdotai made it possible to create trading agents without writing any code.
Trading agents could transform prediction markets the same way algorithmic trading changed traditional finance
Give it a try.
Polystreet@Polystreetdotai
What if anyone could create their own trading agent? The most profitable prediction market traders already have agents that trade for them. Everyone else is still trading manually. Until now.
English