THE CHART IS BLEEDING. THE BALANCE KEEPS CLIMBING.
Watch the GROKBOT + JEV screen. Then tell me which number you were staring at.
Top right: token charts. Red candles. Sharp drops. The kind of moves that make you refresh your wallet twice.
Bottom left: the balance.
$43,944.
$44,347.
$46,494.
Keep watching.
Another chart rolls over. Another token takes a hit.
$47,321.
Then $51,013.
Then $51,701.
A $7,757 difference between those displayed balances.
Here’s the part that should make you curious.
The red candles are obvious. The reason behind that balance movement isn’t.
What positions were open? Where were the entries? How much risk was on the table?
That’s where the real story is.
Anyone can post a green number. Show me the trades that built it.
Watch it again. This time, follow the balance.
Your backtest is lying to you and you don't even know it.
This is the $200K quant stack — rebuilt with Claude in one evening.
Full code. Copy it tonight or keep trading noise. https://t.co/9p9cTcxDnE
$300,000,000 FOR A MAN WHO WALKED OUT OF OPENAI
BEZOS, ERIC SCHMIDT AND NVIDIA BACKED HIS LAB. ROBOTS COOK WHAT THE AI INVENTS. 2,418,006 CANDIDATES. ONE KEPT. NO HUMANS.
William Fedus ran post-training on the model that made OpenAI famous. In March 2025 he quit.
Six months later his startup, Periodic Labs, closed that seed round.
His co-founder came from Google DeepMind. Their model there predicted 2.2 million new crystals. 380,000 held up. DeepMind called it nearly 800 years of human discovery.
A prediction on a server is worthless until someone mixes it, heats it and tests it. That has always been done by hand, one sample at a time.
Their lab does that part without people. The model proposes a material. A robot mixes and heats it. The result goes back into the model, and the cycle repeats overnight.
The screen below shows what such a search looks like. 2 runs. 90 days. It is an illustrative schematic, marked on screen as not a real run.
One candidate survived: a three-layer hexagonal lattice the model had never seen in training.
The rest was thrown out. Unstable. Duplicate. Too costly. Toxic.
Their first targets are superconductors and materials that stop chips from cooking themselves.
Fedus had one of the best seats in AI. He left it for a search that runs without him.
The whole 90-day search is compressed into 24 seconds below. The one surviving candidate turns red around second nine.
Your backtest is lying to you and you don't even know it.
This is the $200K quant stack — rebuilt with Claude in one evening.
Full code. Copy it tonight or keep trading noise. https://t.co/9p9cTcxDnE
AI DOESN’T WATCH CHARTS IT WATCHES THE MONEY MOVING UNDER THEM.
It scans liquidity, order flow, volume and momentum across the market, then waits for the cleanest imbalance.
No signal = no trade.
Humans react to candles.
AI reacts to what creates them.
That’s what you’re seeing in this video.
A live map of where pressure is building before it becomes obvious on the chart.
This is still early.
AI DOESN’T WATCH CHARTS IT WATCHES THE MONEY MOVING UNDER THEM.
It scans liquidity, order flow, volume and momentum across the market, then waits for the cleanest imbalance.
No signal = no trade.
Humans react to candles.
AI reacts to what creates them.
That’s what you’re seeing in this video.
A live map of where pressure is building before it becomes obvious on the chart.
This is still early.
Your backtest is lying to you and you don't even know it.
This is the $200K quant stack — rebuilt with Claude in one evening.
Full code. Copy it tonight or keep trading noise. https://t.co/9p9cTcxDnE
Nick Leeson kept adding to his Nikkei futures after the Kobe earthquake in 1995. Barings lost £827 million.
Brian Hunter stayed in his natural gas bets at Amaranth in 2006. The fund lost $6.6 billion, about two thirds of its money, in a few weeks.
A gold bot in the video below does the same thing on a much smaller scale. The run is a backtest. One day of gold, 1 July 2025.
It sells gold, and when gold rises it sells again, a little bigger each time. The panel calls it DCA with a multiplier of 1.68. Every order has a stop loss of 0.000. So there is none.
Ten seconds in, the terminal shows five sells stacked on each other. Here are three lines of it as Table 1.
0.01 lots sold at 3316.78.
0.08 lots sold at 3338.77.
0.13 lots sold at 3346.06.
The last order is thirteen times the first one and carries 43% of the whole stack.
Gold moved less than 1% against the first sell. The open loss on screen reached 3,621.97 while the balance stood at 100,804. That is about 3.6% of the account, and the screen still says "Profitable".
This is the arithmetic of averaging down. Each new order pulls the break-even price closer, here to about 3337. It also makes the next 1% move hit harder than the last one did.
The same check works for anyone who bought more of a falling stock or fund. Total position today, divided by the first buy. Then the cost of one more 1% drop, counted in money. A percent of a bigger position is a bigger number.
The order table ten seconds into the video is the part worth pausing on.
Nick Leeson kept adding to his Nikkei futures after the Kobe earthquake in 1995. Barings lost £827 million.
Brian Hunter stayed in his natural gas bets at Amaranth in 2006. The fund lost $6.6 billion, about two thirds of its money, in a few weeks.
A gold bot in the video below does the same thing on a much smaller scale. The run is a backtest. One day of gold, 1 July 2025.
It sells gold, and when gold rises it sells again, a little bigger each time. The panel calls it DCA with a multiplier of 1.68. Every order has a stop loss of 0.000. So there is none.
Ten seconds in, the terminal shows five sells stacked on each other. Here are three lines of it as Table 1.
0.01 lots sold at 3316.78.
0.08 lots sold at 3338.77.
0.13 lots sold at 3346.06.
The last order is thirteen times the first one and carries 43% of the whole stack.
Gold moved less than 1% against the first sell. The open loss on screen reached 3,621.97 while the balance stood at 100,804. That is about 3.6% of the account, and the screen still says "Profitable".
This is the arithmetic of averaging down. Each new order pulls the break-even price closer, here to about 3337. It also makes the next 1% move hit harder than the last one did.
The same check works for anyone who bought more of a falling stock or fund. Total position today, divided by the first buy. Then the cost of one more 1% drop, counted in money. A percent of a bigger position is a bigger number.
The order table ten seconds into the video is the part worth pausing on.
931 WHALE ACCOUNTS GOT STRUCK OFF THE LIST.
A SCRIPT CHECKED 1,350 OF 5,004 WHALE ADDRESSES ON HYPERLIQUID. 418 KEPT. 931 REMOVED. THAT IS 69%
A terminal is walking through a file called whales_to_clean.txt, one address at a time. The owner says these are the biggest depositors on the exchange. Anything that has fallen under his cutoff gets struck off.
Read the screen top to bottom:
→ Progress: 1,350/5,004. Kept: 418. Removed: 931.
→ Removed: one address with $1.01 left, then another at $0.00.
→ Shorts near liquidation: BTC, $72.8M at 40x, 0.7% away.
→ Longs near liquidation: ETH, $15.9M at 25x, 5.6% away.
→ Liquidated in the last 4 hours: $2.3M of longs, $2.7M of shorts.
→ His own XRP short at 20x: -7.56%.
Hyperliquid runs on-chain, so every position is public. Size, leverage and liquidation price are all in the open. His app sorts them by how close each one is to getting wiped.
At 40x a move of roughly 2.5% against you eats the whole margin. The $72.8M short has 0.7% of room left.
His own app has a risk panel open on the left. The stop loss is set to 20% and the take profit to 20%. Leverage is capped at 40x. The daily loss limit is 50%. The account shows $10,641.53, and the lowest allowed balance is $5,334.70.
Claude Code sits in the bottom corner the whole time, reporting that the app is running. He is fair about the count, too. Some of those 931 may have simply moved their money somewhere else.
A whale checks one position, his own. The script checks 5,004 of them and keeps a running score.
Near the end the app opens the list of shorts closest to liquidation. The $72.8M one is on top.
931 WHALE ACCOUNTS GOT STRUCK OFF THE LIST.
A SCRIPT CHECKED 1,350 OF 5,004 WHALE ADDRESSES ON HYPERLIQUID. 418 KEPT. 931 REMOVED. THAT IS 69%
A terminal is walking through a file called whales_to_clean.txt, one address at a time. The owner says these are the biggest depositors on the exchange. Anything that has fallen under his cutoff gets struck off.
Read the screen top to bottom:
→ Progress: 1,350/5,004. Kept: 418. Removed: 931.
→ Removed: one address with $1.01 left, then another at $0.00.
→ Shorts near liquidation: BTC, $72.8M at 40x, 0.7% away.
→ Longs near liquidation: ETH, $15.9M at 25x, 5.6% away.
→ Liquidated in the last 4 hours: $2.3M of longs, $2.7M of shorts.
→ His own XRP short at 20x: -7.56%.
Hyperliquid runs on-chain, so every position is public. Size, leverage and liquidation price are all in the open. His app sorts them by how close each one is to getting wiped.
At 40x a move of roughly 2.5% against you eats the whole margin. The $72.8M short has 0.7% of room left.
His own app has a risk panel open on the left. The stop loss is set to 20% and the take profit to 20%. Leverage is capped at 40x. The daily loss limit is 50%. The account shows $10,641.53, and the lowest allowed balance is $5,334.70.
Claude Code sits in the bottom corner the whole time, reporting that the app is running. He is fair about the count, too. Some of those 931 may have simply moved their money somewhere else.
A whale checks one position, his own. The script checks 5,004 of them and keeps a running score.
Near the end the app opens the list of shorts closest to liquidation. The $72.8M one is on top.
THE BOT GOT FILLED AT 3 CENTS.
CLAUDE CODE WROTE IT. $1.00 IN AT 3 CENTS. OUT AT 41 CENTS. +1,283.33% ON THE SCREEN
A BTC Up or Down market that settles every 15 minutes. Up at 29 cents, Down at 72.
The bot does not pick a side. It parks a stink bid under both of them. That is a limit order so low it only fills when somebody dumps in a panic.
Somebody dumped. What's on the screen:
→ UP FILLED at $0.03. The Down bid gets cancelled.
→ Up, 33 shares, avg 3¢, cost $1.00.
→ Value $9.50. $10.83. $11.17. $13.83.
→ Emergency exit 2 minutes before resolution.
→ Market sell at the best bid, $0.4100.
→ Cash $131.49. Then $134.77. Then $145.16.
The take profit sits at $0.9 and never gets touched. The bot does not wait for it. It sells 33.33 shares into the bid with two minutes left. A binary held into the bell either pays or goes to zero, and the code refuses to find out.
The count checks out. 33.33 shares at $0.41 is $13.67. Cash went from $131.49 to $145.16. Same $13.67.
He is honest about it on the recording. This is only the seventh trade, and it lost the first three.
A trader holding Up at 42 cents starts dreaming about the full dollar. Claude's bot sees three seconds left to its exit and sells.
The cash counter jumps to $145.16 about 12 seconds in.
THE BOT GOT FILLED AT 3 CENTS.
CLAUDE CODE WROTE IT. $1.00 IN AT 3 CENTS. OUT AT 41 CENTS. +1,283.33% ON THE SCREEN
A BTC Up or Down market that settles every 15 minutes. Up at 29 cents, Down at 72.
The bot does not pick a side. It parks a stink bid under both of them. That is a limit order so low it only fills when somebody dumps in a panic.
Somebody dumped. What's on the screen:
→ UP FILLED at $0.03. The Down bid gets cancelled.
→ Up, 33 shares, avg 3¢, cost $1.00.
→ Value $9.50. $10.83. $11.17. $13.83.
→ Emergency exit 2 minutes before resolution.
→ Market sell at the best bid, $0.4100.
→ Cash $131.49. Then $134.77. Then $145.16.
The take profit sits at $0.9 and never gets touched. The bot does not wait for it. It sells 33.33 shares into the bid with two minutes left. A binary held into the bell either pays or goes to zero, and the code refuses to find out.
The count checks out. 33.33 shares at $0.41 is $13.67. Cash went from $131.49 to $145.16. Same $13.67.
He is honest about it on the recording. This is only the seventh trade, and it lost the first three.
A trader holding Up at 42 cents starts dreaming about the full dollar. Claude's bot sees three seconds left to its exit and sells.
The cash counter jumps to $145.16 about 12 seconds in.
@KobeissiLetter meta added $190 billion in one session. that's more than the entire market value of most companies in the s&p 500, gained by one stock before the close
@tonybalogna@zerohedge yeah, newark and philly too, so it's something shared across both approach facilities. has the faa said anything about the new system, or is that still a guess?
@dionbird23@zerohedge good catch, then it's bigger than one building. all four stopping at once points to something they share upstream, like the radar and telecom feeds that knocked newark out last year
@unusual_whales most us diesel exports leave from the gulf coast, and the northeast is where prices hurt most. the pipeline north is already full and the jones act makes shipping it by tanker expensive, so a ban traps cheap diesel in texas instead of sending it to new york
@elonmusk does build let you run other models through the same harness for comparison? would be the cleanest way to show how much of the gain comes from the harness itself
@unusual_whales amazon makes tens of billions a year selling ad placements inside its own search results. an agent that shops for you never sees a sponsored listing, so every outside agent is a direct hit to that business while amazon pushes its own assistant
@KobeissiLetter with the s&p now worth well over fifty trillion, a trillion dollar day is a gain of under two percent. the headline number keeps getting bigger because the market does, not because the days are wilder