catazaurus

researcher | project builder |

[catazaurus @catazaurus]

@rugikkk the part I like is the dev wallet and wash checks. most people lose on launches by getting in early on something that was faked from the start. but one afternoon is one afternoon, so I’d want to see this on a dead day

[catazaurus @catazaurus]

$60 → $471.81 WITH 75X LEVERAGE. ON EVERY SCREEN THE OPEN TRADE WAS LESS THAN 1% FROM ITS TARGET.

This is a simulated run of Vortex, a perps desk with three agents: Scout finds the setup, Risk sizes it, Quant keeps the log. Three screens from it:

$75.28 → AVAX long 90x → $23 margin
$250.85 → LINK long 75x → $75 margin
$471.81 → XRP long 75x → $142 margin

Divide each margin by its balance and you get 31%, 30%, 30%. Leverage dropped from 90x to 75x while the position got bigger.

Every entry in the order router starts the same way. Before the buy, the take-profit is written and locked. ETH: TP at 2,446.7 locked, then the limit long at 2,431.5. That’s about 0.6% above the entry.

At 90x, 0.6% is enough. The XRP trade on the first screen closed +$7 on $17 of margin.

Big closes: 3 trades, +$21 → 4 trades, +$137 → 4 trades, +$248. The biggest was ADA at +$94 on $113 of margin, which the Risk agent logged as +83%.

Average win went from +9.8% to +14.2%. Record: 3 wins, 3 losses → 13 and 8 → 17 and 8.

The only stop I can read is ETH at –$6.

The leverage was 90x. The target was 0.6%.

[catazaurus @catazaurus]

A $71 ACCOUNT BECAME $2,197.65 OVERNIGHT. EVERY TRADE GOT HALF OF WHATEVER IT HAD.

This is a live simulation, not real money. Three screens from the run:

> $603.15 → 8.5x
> $1,557.97 → 21.9x
> $2,197.65 → 31.0x

Now look at the margin on the open position in each one.

> AVAX long, 5x → $302 in, on a $603 balance
> LINK long, 8x → $779 in, on a $1,557 balance

Half of the account both times. The bot doesn’t know what it started with, only what it has right now.

That’s why the same kind of trade pays so differently across the night. The LINK long was +$235 and still open on the second screen. On the third it had hit its take-profit, 11% away, and closed at +$639.68 after costs, which is +82% on the margin. The balance moved from $1,557.97 to $2,197.65, by exactly that amount.

The target was set before the entry, at 9% to 11% away every time, and the screen calls the mode TP-first.

Tape: 15 wins and 6 losses → 16 wins and 6 losses. Four big closes (LINK, LINK, BTC, SOL) made $1,688 of the $2,127 profit.

The last trade on the last screen paid nine times what the account started with.

[catazaurus @catazaurus]

@pagmaksym you’re right, and I’d go further: two days says nothing about the edge, it only shows how fast 8x compounds when you’re right. the first red week decides if this is a system or just a lucky streak

[catazaurus @catazaurus]

$71 → $967.75 IN TWO DAYS. THE FIRST TRADE ON THE SCREEN WAS WORTH $10.79.

A new Grok bot, a perps desk, and three screens from the run. Every position I can see is a long.

> $81.22 → BTC long 5x → $28 in
> $705.93 → SOL long 8x → $240 in
> $967.75 → AVAX long 8x → $270 in

The target sat about 9% away on each of them, 10.7% on the first, 8.2% on the second, 8.8% on the last. The margin was between a quarter and a third of the balance every time.

That’s the part I keep looking at. Over two days nothing about the way it trades changed. The number it multiplies kept getting bigger.

The BTC trade, early on, was +$10.79 unrealised. The AVAX trade, near the end, hit its take-profit and closed at +$174.49 after costs, +65% on the margin.

Big closes went from 1 of 4 filled at +$10 to all 4 filled at +$573 combined. That’s 64% of the $897 profit from four trades.

The tape under them went from 1 win, 0 losses to 15 wins, 1 loss. One red mark.

The rest of the profit came from small trades I can’t see on these screens, sized off a balance that was already ten times what it started with.

$71 on day one. $270 in a single position by the end of day two.

[catazaurus @catazaurus]

@x_insider4 the 6 doors idea is good for cutting the research time. but 6 trades in one night is a sample size of 6. ask me again after a month of this, including the bad weeks

[catazaurus @catazaurus]

MY GROK BOT TURNED $9,000 INTO $25,488. THE TRADE BEHIND IT ISN’T EVEN CLOSED YET.

I took two screenshots of the same session. On the first, the balance reads $18,866, up $9,865.99. On the second, $25,488, up $16,487.74.

Nothing closed in between. Five trades, 80% win rate, 1.85x average, best close 2.27x. Those numbers are identical on both screens. The extra $6,622 came from one open position that kept climbing while I watched.

The bot went into it with 5 SOL, about $1,000, and it only did that after the token passed every filter: $41K liquidity, dev wallet at 0.3%, top-10 holders at 11%, no bundle, mint off, volume at 6.2x, 12 smart-money wallets in it, and an AI conviction score of 97 out of 100.

On the first screen that trade sat at 1.75x, worth $746 on paper. On the second it was at 8.37x, worth $7,368. During those minutes the scanner went from 5,733 pairs to 6,051, and the take-profit bar moved from 2x to 10x.

Its best closed trade was 2.27x. This one is already almost four times that.

Here’s the part that keeps this honest. $25,488 includes a position that hasn’t been sold. Take it out and the balance on closed trades is $18,120. That’s $9,000 turned into $18,120 for sure, and the other $7,368 is still a chart.

I’m not sure yet which of those two numbers I’ll be posting about tomorrow.

[catazaurus @catazaurus]

$3,529 → $300. THE BOT DIDN’T GET WORSE AT PICKING. IT STOPPED LISTENING TO ITS OWN CHECKS.

Same desk that took $1,000 to $3,529.

At $3,529.81 the log started changing:

> risk mode: aggressive
> conviction low · taking anyway
> sizing up to recover
> warnings ignored · override
> revenge mode

The bot wrote one rule for itself on day one: no entry unless all 8 checks pass.
Now the dev wallet, top-10 holders and volume/MC checks were flashing WARN. “Dev wallet moving supply!” popped up in the log. It bought anyway.

First entry: 2.5 SOL. Last entry: 5.7 SOL.

It sized up after losses. That’s the fastest way to lose an account.

The stop line on every chart still read –20%. The exits didn’t:

> 0.71x
> 0.62x
> 0.45x
> 0.20x
> 0.08x
> 0.04x
A stop that doesn’t stop is just a label.

The balance:
$3,529.81 → $2,616.00 → $1,028.14 → $300.00

–91.5% from that first screen, and below the $1,000 it started with.

Screen four reads: HALTED · MAX DD.

Six trades. 12 closed in total.

The scanner never broke. 6,709 pairs scanned. The problem was never what the bot saw. It was what it did with the warnings.

[catazaurus @catazaurus]

@0xZasti the $3 isn’t the interesting number. knowing the sofa side-eye is the joke is the actual skill. anyone can generate a couple, almost nobody knows what makes people watch a 10-second silent fight

[catazaurus @catazaurus]

$3,529 → $300. THE BOT DIDN’T GET WORSE AT PICKING. IT STOPPED LISTENING TO ITS OWN CHECKS.

Same desk that took $1,000 to $3,529.

At $3,529.81 the log started changing:

> risk mode: aggressive
> conviction low · taking anyway
> sizing up to recover
> warnings ignored · override
> revenge mode

The bot wrote one rule for itself on day one: no entry unless all 8 checks pass.
Now the dev wallet, top-10 holders and volume/MC checks were flashing WARN. “Dev wallet moving supply!” popped up in the log. It bought anyway.

First entry: 2.5 SOL. Last entry: 5.7 SOL.

It sized up after losses. That’s the fastest way to lose an account.

The stop line on every chart still read –20%. The exits didn’t:

> 0.71x
> 0.62x
> 0.45x
> 0.20x
> 0.08x
> 0.04x
A stop that doesn’t stop is just a label.

The balance:
$3,529.81 → $2,616.00 → $1,028.14 → $300.00

–91.5% from that first screen, and below the $1,000 it started with.

Screen four reads: HALTED · MAX DD.

Six trades. 12 closed in total.

The scanner never broke. 6,709 pairs scanned. The problem was never what the bot saw. It was what it did with the warnings.

[catazaurus @catazaurus]

2,490 TOKENS SCANNED. 6 BOUGHT. $1,000 → $3,440 AND THE BIGGEST WIN HASN’T CLOSED YET.

A Solana meme desk called Mr. Grokow saw a fresh launch and read the chain before it spent anything:

> 42 wallets funded in the same block
> dev wallet holding 38% of supply
> verdict → SKIP

Forty seconds later the token was down 91%. The bot never touched it.

It sees about 66 new pairs a minute, and more than 99% of them never get a single SOL. Every entry has to clear 8 checks first:

> liquidity
> dev wallet
> top-10 holders
> bundle check
> mint/freeze
> volume/MC
> social buzz
> smart money
Fail one → gone.

Five trades are closed:

> +$415 → 2.66x
> +$353 → 2.41x
> +$280 → 2.12x
> +$243 → 1.97x
> –$40 → 0.84x

80% win rate. 2.00x average flip. $1,000 → $2,250 on closed trades alone.

The part I keep coming back to is the last line of that list. The only loss cost $40, and the smallest win paid six times that. The exit is set before the entry: take profit at 2x, stop at –20%. The bot never has to decide how it feels about a red candle.

Then the sixth.

All 8 checks passed. Smart money showed up with 11 wallets. Conviction 94. It put in 3 SOL.

> sold half at 2x
> left the rest running
> trailed the stop: 3.5x → 6.1x

It sits at 7.94x now, and the balance reads $3,440.78. That’s $1,190 above the last closed trade, and it’s still in an open position, so it can move either way.

2,490 scanned. Six bought. One still open with the stop already at 6.1x.

[boan @0xboan]

https://t.co/N5V9XpJlTy

[catazaurus @catazaurus]

@pagmaksym the third rule is the one nobody talks about. always having something to look forward to. what’s yours right now? mine is embarrassingly small, and I think that’s the point

[catazaurus @catazaurus]

@0xZasti the curtain tell is funny, but that’s a one-model-generation problem. give it six months and nobody catches it from the background anymore

[catazaurus @catazaurus]

@0xbalzzam @grok only 2 night, lol

[catazaurus @catazaurus]

Two tools. One viral boxing video. A completely different person throwing the exact same punches.

The original clip is nothing special — a guy shadowboxing in a gym, decent form, nothing viral about it on its own except the camera angle and the lighting. Someone ran it through a swap, and now it's an old man in a tan suit, round glasses, hair piled into a topknot, throwing identical combinations with identical footwork, identical timing down to the frame.

Not a filter. Not a face overlay that slides around when he moves fast.

the body moves exactly the same. only the person changed.

The process behind it is almost insultingly simple. First tool builds a character from scratch — face, outfit, full turnaround sheet, the kind of reference images a 3D artist would've spent a day on two years ago. Second tool takes that character and the source footage and just substitutes one for the other, motion intact, lighting matched, shadows falling the way they should.

Find a video that already works. Someone already paid the cost of figuring out what makes people stop scrolling — the pacing, the hook, the payoff. Swap the person, keep everything else.

No camera. No face. No gym, no boxing lessons, no original footage at all.

The hard part of going viral was never the execution. It was knowing what people stop for. Someone else already solved that part for free, and all you had to do was take their video apart and put someone else inside it.

[Archive @ArchiveExplorer]

https://t.co/bN4Q3DO8nE

[catazaurus @catazaurus]

THE MODEL GOT PAID $5,500. THE OPERATOR GOT $2,500. ONE VIDEO.

blue bikini, safety glasses pushed up into her hair, ear protection clamped over both sides of her head. an outdoor range, dust hanging low in the bright afternoon light, scrubland stretching out flat behind the firing line. an instructor steps in first — bearded, calm, the kind of guy who's clearly walked a hundred students through this exact motion. he adjusts her strap so it won't catch on the stock, says something short, steps back out of frame. she shoulders the rifle, settles behind the sight, breathes, fires.

this is the sequence that's supposed to be the wall for generated video. everything before the trigger — the girl, the light, the two people talking, the dust in the air — is the easy part now, the stuff these models handle without much effort. the second a shot actually fires, four separate things all have to land at once or the whole clip falls apart: muzzle flash appearing in the right place on the barrel at the right size, recoil traveling correctly through shoulder and arm instead of just kicking the gun itself, brass ejecting on the correct arc, smoke drifting off instead of vanishing or freezing mid-air. miss any one of those and it reads instantly as fake, no matter how good everything around it looks.

this one holds all four. flash right on the barrel. recoil that actually propagates — shoulder first, then the arm catching up a beat later, the way a real shot kicks through an actual body rather than just rattling the prop. brass ejects on a real arc. smoke trails off to the side instead of just disappearing.

ngl the strap adjustment before the shot is the detail that got me, not the gunfire itself. that's a directing choice — a small two-person beat to warm the scene up and set the tone before the physics-heavy part kicks in, the same move a real director would make. whoever built this understood pacing, not just how to render a person and a gun.

where it actually slips is after the shot. her stance resets back into position a fraction too cleanly, too identical to where she started, like the whole burst looped back to frame one instead of genuinely recovering from the kick. a real shooter never lands in the exact same stance twice — there's always a small correction, a slight shift in weight. this one just snaps back.

no range. no rifle. no instructor. no shot ever actually fired — just $8,000 spent making sure it looked like one was.

[catazaurus @catazaurus]

THE MODEL GOT PAID $5,500. THE OPERATOR GOT $2,500. ONE VIDEO.

blue bikini, safety glasses pushed up into her hair, ear protection clamped over both sides of her head. an outdoor range, dust hanging low in the bright afternoon light, scrubland stretching out flat behind the firing line. an instructor steps in first — bearded, calm, the kind of guy who's clearly walked a hundred students through this exact motion. he adjusts her strap so it won't catch on the stock, says something short, steps back out of frame. she shoulders the rifle, settles behind the sight, breathes, fires.

this is the sequence that's supposed to be the wall for generated video. everything before the trigger — the girl, the light, the two people talking, the dust in the air — is the easy part now, the stuff these models handle without much effort. the second a shot actually fires, four separate things all have to land at once or the whole clip falls apart: muzzle flash appearing in the right place on the barrel at the right size, recoil traveling correctly through shoulder and arm instead of just kicking the gun itself, brass ejecting on the correct arc, smoke drifting off instead of vanishing or freezing mid-air. miss any one of those and it reads instantly as fake, no matter how good everything around it looks.

this one holds all four. flash right on the barrel. recoil that actually propagates — shoulder first, then the arm catching up a beat later, the way a real shot kicks through an actual body rather than just rattling the prop. brass ejects on a real arc. smoke trails off to the side instead of just disappearing.

ngl the strap adjustment before the shot is the detail that got me, not the gunfire itself. that's a directing choice — a small two-person beat to warm the scene up and set the tone before the physics-heavy part kicks in, the same move a real director would make. whoever built this understood pacing, not just how to render a person and a gun.

where it actually slips is after the shot. her stance resets back into position a fraction too cleanly, too identical to where she started, like the whole burst looped back to frame one instead of genuinely recovering from the kick. a real shooter never lands in the exact same stance twice — there's always a small correction, a slight shift in weight. this one just snaps back.

no range. no rifle. no instructor. no shot ever actually fired — just $8,000 spent making sure it looked like one was.

[Zasti @0xZasti]

https://t.co/PHzbWGORjL

[catazaurus @catazaurus]

A dead fly's brain is flying three drones in someone's office right now.

Not a simulation of a fly. The actual wiring — 166,700 real neurons, roughly 125 million real synapses, mapped one by one from electron-microscope images of an actual Drosophila brain. Published in 2026 by HHMI Janelia, Cambridge, and Google Research. Free, public, CC-BY.

Someone downloaded it and asked the obvious insane question: can this actually fly something.

No flight code written by a human. No "if obstacle, turn left." Just a camera feeding raw pixels into the fly's own visual neurons, exactly the pathway a living fly uses to see a predator or a wall, and the connectome deciding what to do with what it sees.

it works. that's the part that doesn't feel real even looking at the footage.

The drones sit on the floor of what looks like an ordinary office, monitors behind them glowing with a live render of the brain itself — tens of thousands of neurons lit up, firing in real time, every spark visible while the rotors respond underneath.

This wiring diagram spent something like 350 million years getting refined by evolution to do exactly one job: keep a body airborne and out of danger with almost no processing power at all. No training run. No GPU cluster. A brain smaller than a grain of rice, built for a bug that lives about two weeks, now sitting behind a camera lens making real-time flight decisions.

The fly that owned this brain is dead. The wiring just got its first job since.

[catazaurus @catazaurus]

A professor stands in a parking lot with a jetpack strapped to his back, throttle building, thrusters warming up while a crowd gathers to watch.

The rig itself looks nothing like a product. Boxy, metallic casing, a couple of small digital displays bolted onto the side flashing readouts nobody in the crowd can actually interpret. Two hand-held thrusters, one in each fist, connected back to the main unit by thick cables and hose lines running along his arms like something between scuba gear and a fire extinguisher rig. It looks assembled, not manufactured — the kind of thing built piece by piece in a workshop, tested, adjusted, tested again.

People pull out their phones. Nobody's fully sure this is actually going to work.

Then the flames kick out the bottom, loud, orange light flashing across the pavement, and he lifts clean off the ground. Not a wobble. A real, sustained hover, rising above the heads of the crowd he built this in front of.

He holds there for a few seconds, drifting slightly, adjusting with small movements of the hand thrusters, the whole rig roaring under him while everyone tilts their heads back to follow him up.

Then he eases back down onto the asphalt, boots touching pavement like he's done this exact landing a hundred times before.

Nobody stops him. Nobody seems surprised. Like this is just what he does now.

A professor built a jetpack that actually works, then used it to answer the one question every open house crowd actually came to ask without saying a word — not what does he teach, but can he actually do it.

[catazaurus @catazaurus]

https://t.co/XEncOeQQ9V

[catazaurus @catazaurus]

A Stanford professor took two average stocks, rebalanced them daily, and turned $100,000 into $7 million without predicting a single price.

Thomas Cover built it on the board with two names most people have never heard of. One goes up 4x over twenty years by itself. The other goes up 8x by itself. Both fine. Neither remarkable. Nobody would call either a genius pick.

Rebalance between them the entire time — sell some of whichever one just grew, buy back into whichever one just shrank — and the combined position goes up 75x. Not the average of 4 and 8. Not close to either number. Fifteen times better than the better stock on its own.

he calls it making money out of nothing. because that's exactly what it is.

The mechanism has nothing to do with picking winners. It's the opposite. Take a stock that does nothing at all — doubles half the time, gets cut in half the other half, dead even, going nowhere over the long run — pair it with cash that just sits there, and rebalance between the two every period. The combined portfolio grows at 6% per period with probability one. Not on average. Essentially guaranteed.

You're living off the fluctuations themselves. The volatility everyone else is trying to avoid is the entire source of the return.

He runs the same framework against real market data — thirteen actual months leading into a crash. Buy and hold the index during that stretch, and you lose money. His method, applied blind with zero prediction, still comes out ahead.

Nobody predicted a single price. The prediction was never the point.

[Quiq @xQu1q]

A DEAD FISH JUST SWAM UPSTREAM. NO PULSE, NO MUSCLE, NO NERVOUS SYSTEM.

George Lauder's lab at Harvard tied a dead rainbow trout to a slack string, dropped it into a water tunnel behind a rock, and watched the vortices peeling off the back of the rock grab its body and push it forward. Against the current. Zero biological input. MIT professor Russ Tedrake shows this footage in his robotics lectures — not his lab's experiment, but the one that reframes his entire field.

Live trout swim upstream every mating season, and they hang out behind rocks doing it. Lauder's team wanted to know if that was effort or physics. So they killed the fish and removed the effort entirely. What was left still moved upstream.

That's the same failure mode humanoid robots have been stuck in for over 20 years. Honda's ASIMO — the most advanced walking robot on the planet in 2009 — moves like, in Tedrake's words, "a guy in a space suit." Knees permanently bent, one foot always flat on the ground, fighting its own dynamics with motors at every single step instead of working with them. It burns 20 times the energy of a human walking the same distance and its battery dies in 26 minutes.

A Cornell team built the opposite: a passive dynamic walker with no motors, no computer, no control system whatsoever. Just sticks and hinges with ball bearings at the joints. Put it on a small ramp, give it one push, and gravity alone carries it down the slope in a gait that looks more human than ASIMO's does.

The same idea runs through the whole lecture. A fighter jet holds a rigid, conservative flight envelope to stay controllable. A falcon does the opposite on purpose — stalls its wings at 240 mph to generate drag and stick a landing on a branch, a maneuver Tedrake calls harder than landing on an aircraft carrier. An albatross burns the same energy sitting on a beach as it does flying across an ocean for days without a single wingbeat.

Tedrake's term for all of it is "underactuated" — a system that doesn't have enough motors to just force its way through the physics. You can't override the dynamics with more torque. You have to work with what the machine already has, the same way a dead fish works with a current it can't even feel.