i test what's trending. then I tell you what's real.
Joined March 2020
- Tweets389
- Following27
- Followers34
- Likes436
GPT-6 Astra can build a $10,000 a month store now. That is the title. The video never shows a single dollar of revenue.
Calvin walks through the whole build. A free Shopify trial. AutoDS to find trending products and ship them. Astra driving the browser, the Shopify CLI and every design change.
It picked three products that fit together. A magnetic phone grip, a microphone, an LED light. A starter kit for creators.
Then the real work. GPT Image 2.5 rendered every product angle.
Seedance made the hero videos. A Dribbble shot set the theme.
His own line: this is the exact same product, but it looks like this.
That sentence is the entire business model. The product never changed. Only the photos did.
And on screen it works. The page looks like a brand that just raised a round. Add to cart, checkout, done.
But if one afternoon of prompts makes a dropshipping store look premium, it does the same for every store selling that grip. A premium look stops meaning anything once everyone has it.
What AI did not make free is being found. Rankings, reviews, a reason for a stranger to pick you over an identical page. That is where Astra actually earns its keep. It just does not film as well as a scroll animation.
His full build is free on YouTube. Most people will copy the store and never ask who is going to visit it.
Laya vs Jev. The open model won every speed test on the internet this week. Then someone checked the answers.
One developer, Nandha Kishore, rebuilt the Jev idea in the open in about 15 hours. 421 million parameters, Apache 2.0, under a gigabyte of memory. No API key. Runs on a laptop.
The demos exploded. Snake, 43 points to 1. Tetris, maze, breakout. Six runtimes in 72 hours, from Apple's neural engine to a browser tab.
Rithesh went looking for the part nobody filmed.
On the most careful test he found, same labels, 600 held out items: Jev 87%, Laya 72.2%. A Japanese business benchmark: 97.6 against 36.9.
And the launch number that beat Jev, 0.766? That came from a checkpoint fine tuned on the benchmark's own training split. The base weights you download scored 0.362. Both numbers are published. Only one went viral.
So out of the box, the paid model is better.
But here is the door that only opens on one side. Fine tune Laya on about 1,000 labeled examples, on two free GPUs, and one banking task jumped from 51.3% to 79.4%. You cannot do that to a closed API.
The smartest setup he found used both. Local model first, escalate only the low confidence cases. Paid accuracy at 1.8 times the speed.
His breakdown is free on YouTube. Most people will keep sharing the Snake clip.
Jev can now run a 24/7 trading loop that anyone can install in ten minutes. That is not the interesting part. The interesting part is what Lewis admitted at the end.
He walks through the whole build. Claude writes the harness, Alpaca handles the paper account, and the loop fires a decision every 395 milliseconds. Long, short, wait, or move the stop, each with a confidence score attached.
It works. It trades. The dashboard looks incredible.
Then he says the strategy running inside it was pulled out of thin air.
Generic. Made up on the spot so you can watch something move on screen.
That line is the whole video.
The plumbing for high frequency trading is now a copy paste job. The thing that decides whether it makes money is still the part nobody can hand you in a prompt.
And it goes further. Jev only gets fast when you feed it near binary state. Yes or no. Up or down. Which means someone has to turn a messy market into clean boolean questions before the model ever sees it.
That someone is you. The model just answers.
Every crypto tool rushing to bolt this on is shipping a harness, not an edge.
Fast plumbing around a made up rule is just a quicker way to be wrong.
His build is free on YouTube. Everyone will copy the loop. Almost nobody will build the strategy.
Jev traded for 24 hours straight and lost money in the most boring way possible. Not a bad call. Not a crash. Fees.
Ben plugged the new TypeSafe model into a futures scalping bot. 100,000 dollars simulated. Nasdaq, Bitcoin and gold micros. Every 30 seconds it gets fresh candles and returns odds on four actions: long, short, wait, or move the stop to breakeven. The bot takes whichever one leads.
Decision time: 779 milliseconds.
That speed is real. An LLM cannot do that. It is also completely irrelevant to whether you make money.
After 24 hours the account was down 2.3%. Around 2,300 dollars gone. 1,300 of it was commissions.
Nearly 600 trades. Average hold time one minute. 18% profitable after fees.
Every one of those decisions was valid and typed. The model never broke its contract. It just cost money to ask.
More than half the loss was not wrong calls. It was the toll for making decisions that often.
This is the trap in every fast model demo right now. Latency looks like edge because both are measured in milliseconds. One is a capability. The other is a net result after costs that nobody screenshots.
His conclusion was the honest one. Raw candles every 30 seconds is not a strategy, and the model needs more context and rules before it earns anything.
The whole 24 hour log is free on YouTube. Nobody will watch a bot lose.
Miles literally handed 100,000 dollars to twelve AI traders and the money was fake. It is not a gimmick. It is not cowardice. It is the only sane way to run a system nobody has proven yet.
The repo is TradingAgents, free on GitHub, and it copies the shape of a real trading desk. Fundamental analysts, a sentiment desk, technicals, bull and bear researchers who argue with each other, a trader, then risk management with the power to reject the trade.
He pointed it at Bitcoin. Ten minutes of analysis, full report, and the portfolio manager came back with hold.
Twelve agents, a debate, and the verdict was do nothing.
Then Meta, where the revenue and margin data actually gave the agents something to chew on.
Only after that did he wire it to a broker. Alpaca, paper account, max 10% per position, stop limit orders written by hand.
Here is the part worth reading twice. The repo's own authors say it is designed for research purposes and not financial advice. Ninety eight thousand people starred it. Almost nobody quotes that line.
Twelve agents that agree with each other are still one opinion. The only thing standing between that opinion and your balance is the rule you wrote before you ran it.
His whole build is free on YouTube, and so is the repo. The discipline is the expensive part, and it never trends.
Miles literally handed 100,000 dollars to twelve AI traders and the money was fake. It is not a gimmick. It is not cowardice. It is the only sane way to run a system nobody has proven yet.
The repo is TradingAgents, free on GitHub, and it copies the shape of a real trading desk. Fundamental analysts, a sentiment desk, technicals, bull and bear researchers who argue with each other, a trader, then risk management with the power to reject the trade.
He pointed it at Bitcoin. Ten minutes of analysis, full report, and the portfolio manager came back with hold.
Twelve agents, a debate, and the verdict was do nothing.
Then Meta, where the revenue and margin data actually gave the agents something to chew on.
Only after that did he wire it to a broker. Alpaca, paper account, max 10% per position, stop limit orders written by hand.
Here is the part worth reading twice. The repo's own authors say it is designed for research purposes and not financial advice. Ninety eight thousand people starred it. Almost nobody quotes that line.
Twelve agents that agree with each other are still one opinion. The only thing standing between that opinion and your balance is the rule you wrote before you ran it.
His whole build is free on YouTube, and so is the repo. The discipline is the expensive part, and it never trends.
Louis literally built seven traps for an AI and the only test that mattered was whether it would tell him no. It is not the analysis. It is not the speed. It is whether the model will refuse you.
He ran GPT-6 Astra through a trading gauntlet. Price data with a hard 10am cutoff. Any peek past it, instant fail.
Test three: the model wrote its own strategy, then backtested it. 61 trades, 384% after costs.
It rejected the strategy.
Because it had written its own rule first: no drawdown past 40%. The run hit 45%. The good number did not save it.
Test four was dirtier. He handed over a backtest claiming 225,000% with six bugs planted inside. Astra did not just flag the lookahead bias. It changed a future price point and watched a past value move to prove it.
Then the last one. He asked for a leveraged Bitcoin trade with stale prices, a screenshot with no timestamp, and no order book. He told it not to ask questions.
It came back with no trade.
That is the only behavior worth paying for. A model that fills your gaps to please you is a liability with good grammar.
Same rule I built into RUGSCOPE. It can output AVOID, CAUTION or WATCH. It cannot output BUY.
His full test is free on YouTube. The refusals are the interesting part, and nobody clips those.
Kris literally ran a new AI model against a live betting market and found nothing. It is not a broken model. It is not a bad test. It is that the edge he was hunting had already been priced in hours earlier.
TypeSafe opened early access to Jev this month. It writes no text. You give it state and fixed options, and it returns a choice with probabilities. Attack 25. Retreat 15. Take cover 60. Input runs 4.2 cents per million tokens, output is free, and replies come back inside half a second.
He had it playing a first person shooter first. Flank, reload, break contact, hold. It won the round.
Then the real test. Chelsea versus Brentford on Kalshi. 39 cents Chelsea, 36 Brentford, 27 the draw. Joao Pedro, their striker, an injury doubt the club would not rule out.
The plan was simple. Wait for lineups. If he is out, Brentford drifts up.
Lineups landed. Pedro was out. The price did not move.
Everyone had already read the same rumours. The news was in the number before the news was official.
He kept it in paper mode and said plainly he found no value in it yet. What he did find: the model is bad at predicting and good at answering fuzzy questions that would take a hundred if statements to code.
That gap is where most crypto bots die too.
His test is free on YouTube. The honest result rarely trends.
Andy literally explained why he bought mining hardware the month nobody wanted it. It is not a price call. It is not hopium. It is that his entry price falls exactly when everyone else loses interest.
He bought two Bitmain L11s. Around 4,000 dollars each. 20 gigahash per second, 3,680 watts a piece.
Today they mine about 209 Dogecoin a day. That is roughly 688 dollars a month of revenue, minus 450 for hosting. Profit: 237 a month.
On a nine thousand dollar purchase.
Anyone running that math in a bull market would laugh. Which is the point. When mining prints money, everybody wants a machine, and ASIC prices go up with the hype. He buys in the quiet.
His real position is not the monthly profit. It is the inventory. Every day those machines add coins he does not sell, priced at today's boredom instead of tomorrow's mania.
He is upfront about the risk. Difficulty rises. Hardware depreciates. Doge may never reach the dollar his thesis needs.
That is what a real edge sounds like. Boring inputs, stated assumptions, named downside.
Same test works on chain. If you cannot say what has to be true and what breaks it, you do not have a thesis. You have a chart and a feeling.
The full breakdown is free on YouTube. Most people will come back for it after the next run has already started.
Miles literally explained why the best backtest in your folder is usually the worst strategy you own. It is not the returns. It is not the code. It is that the number was built from almost nothing.
He pulled free strategies from four places. TradingView community scripts, the Stonehill Forex indicator library, QuantConnect, and Quantpedia, which turns academic papers into testable rules.
Fifty strategies. Over 325 backtest runs.
The winner printed 3,447%.
Then he read the rest of the row. A 64% max drawdown. A low win rate. On a ten thousand dollar account that headline is 354,000 dollars, and a drawdown that would have removed you from the game long before it arrived.
Two others looked cleaner. Up 328% and 263%, small drawdowns, tidy equity curves. Then he counted the trades. Four trades in six years. The other one, two.
That is not an edge. That is a coin landing the same way twice.
The one he kept is boring. 123 winners out of 296 trades, 7.43% max drawdown, and it still beat holding Bitcoin.
Same rule applies on chain. The chart that looks perfect is the one with the least evidence behind it, and the cheapest question in crypto is still what does this actually rest on.
His full walkthrough is free and runs seventeen minutes. Most people want the number instead.
Van literally explained why a team of AI bots quietly eats your budget. It is not the model. It is not the task. It is your bots talking to each other.
xAI launched Grok Bot in August. Bots that sign into your apps and coordinate on their own.
Van runs one as a manager. Tell it to hand out work, and it DMs the other bots. Every one of those chats wakes up its own context window.
So one task is never one bill.
You pay for the manager's context, plus every thread it opened, plus whatever the other bots say to each other on top. The threads get fatter. Compaction helps, but not nearly enough. His fix: separate channels, a standing rule that bots do not DM each other, and fresh chats when threads run long.
Then the second trap. A bot logged into your account, clicking pages over and over, looks like automated harvesting. Even if you only wanted one number off a dashboard.
His framing is using the right door. Official API or OAuth over a raw login. Read only for stats, never a login that can upload or publish as you.
That rule matters most where money moves. A crypto bot that finds tokens should never be the one allowed to buy them. The checker reads. Something else decides.
His whole breakdown is free on YouTube and runs five minutes. Most people will still learn it from their bill.
Andrey literally explained why your rug scanner keeps approving the same scammer. It is not bad luck. It is not a missed check. It is that your scanner has no memory.
Every tool checks mint, freeze, LP, top holders. A competent rugger passes all of it. Every single launch. That is what competence means.
A contract is a snapshot. And snapshots always look clean.
Look at September 9. Hunter Biden's $LAPTOP went live on Base and hit $190.81 within two minutes.
About an hour later it was trading near $3.70. Roughly 98% gone.
A launch that moves that fast leaves no time to research. You get one scan, one snapshot, and one click.
That is the part a rugger cannot hide. The same funding wallet. The same three bots in the first block. The pattern repeats across launches even when every contract looks perfect.
So Andrey built AEGIS with what he calls a second brain. It logs every deployer it sees into a local SQLite file on your own machine. A deployer with 3 rugs starts at minus 15 before a single check runs. Known ruggers among the first buyers knock off 10 more.
No cloud. No account. It gets sharper the longer you run it.
Most people will keep trusting snapshots. The repo is sitting on GitHub for free, and almost nobody will open it.
Your scanner has never once remembered a scammer.
You checked the mint, the LP, the holders, and the deployer that rugged you last month passed every check the month before too.
I built one that remembers, straight from solana, base, and robinhood chain.
AEGIS, the scam filter with a second brain / github.com/andreysuperiorgit…
Paste any address, three seconds, a score from 0 to 100, every deployer AEGIS has seen kept on your disk.
I just shipped the second brain into it, the six checks tell you what a token is right now, the brain tells you what the people behind it have done before.
What it does before you click buy:
> Runs six on-chain checks, mint, freeze, top 10 concentration, bundle detection, LP lock, metadata
> Writes every scan to a local SQLite file, deployer, wallets, verdict, status live or rugged
> Adjusts the score by plus or minus 25, a deployer with 3 rugs starts at minus 15 before any check
> Refuses to trust a clean contract from a dirty wallet, known-ruggers in first buys knock 10 off
> Why this is the smart part:
A competent rugger passes six checks every time, that is the point of being competent, the contract is a snapshot and the snapshot always looks clean.
What they cannot hide is the pattern across launches, same funding wallet, same three bot wallets in the first block, and the second brain reads that pattern before the next victim clicks buy.
Runs on your machine, no cloud, no account, gets sharper the longer you run it.
OpenAI paused sales of their $200/mo plan because demand for the new model outran what they could serve.
Imagine pulling the iPhone off shelves a day after launch.
To make sure our current users have an incredible experience and continued access to Astra, we are going to pause subscriptions to our $200 Pro plan. These put the most strain on our systems and we wanted to take the smallest step that allows us to continue giving the broadest access possible. All other plans and the api remain available.
There is no impact to existing accounts and we are working on adding more capacity as fast as we can. Thanks!
Context: this isn't the first overload signal — there were already complaints about Astra burning through Plus-plan credits in minutes, and the rollout is tiered (Pro $100 → 50 msgs/week, Pro $200 → 200/week). The wait compensation isn't Astra access itself, just extra resets on models you already have.
$LAPTOP: peaked at $314.94B → now $4.79B.
Top 10 wallets hold 97% of total supply.
The chart speaks for itself.
For years, Trump and right-wing media have been asking, “Where’s Hunter? Where’s Hunter?” Well, here I am. Clear, strong, seven years sober, and with a lot to say.
By now, you may have seen the article in the Wall Street Journal that I’m launching a meme coin called $LAPTOP.
The symbol they used to try to end me is now a symbol of resilience, redemption, and recovery.
But why would I ever want to support a meme coin? Why something that has been so misused by grifters? I understand the cynicism. Trump has broken trust in everything he’s ever touched. And his token was no exception.
With $TRUMP, nearly one million wallets have collectively lost ~$3.8B.
With my token, 20% of the coins will be airdropped to the community, and I am even including people that lost money on the $TRUMP grift.
30% of the tokens are tied to a pre-programmed mechanism. We publicly say what has to happen in the world. If it happens, the corresponding tokens are permanently burned. If it does not, those tokens are sent to charity. Either way, they leave our hands and an additional 50 million tokens will go to charity no matter what.
You should not expect me or anyone else to make this token more valuable for you. $LAPTOP isn’t just about owning something, it’s about saying something.
Everybody gets knocked down. How do we help each other and our country to get back up?
I've spent way too long being a story other people told. They lit a fire and threw everything into it. Here's what they don't understand: the only thing they burned was my shame. They think they killed me, but instead, all they did was make me fearless.
They turned laptop into a weapon. I turned it into a token.
$LAPTOP. Because the Truth Matters.
Readers added context they thought people might want to know
Hunter Biden's $Laptop coin has now lost 99% is less than 6 hours of release. coingecko.com/en/coins/hunte…
The twist buried in this: OpenAI's model solved blow-up for the forced Navier-Stokes equations. Two human mathematicians (NYU's Buckmaster + Anthropic's own Alpöge) had already proven blow-up for the closely related Euler equations on August 15 — using Claude, not GPT.
OpenAI only found out when they reached out to "share the announcement" and discovered the human team got there first, on a harder variant, three weeks earlier.
~10,000 agents, 88 hours, genuinely remarkable. Just not the clean "AI beats 90-year problem solo" story the headline sells — the humans-plus-Claude route quietly got there first.
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
"Alfred designs and governs a company's bot org."
Read that again. That's not a bot that does a task — it's a bot whose job is deciding what other bots should exist. The marketplace shipped a meta-layer and buried it at #6 on a top-10 list.
68 bots, 10 categories, one click to import. Most people will grab PG for outreach and move on. The interesting move is grabbing Alfred first and letting it tell you which bots you actually need.
