All of AI, minus the hype. Just what actually works.
DMs OPEN
Joined September 2022
- Tweets7.5K
- Following95
- Followers3.1K
- Likes14.3K
this is getting f*cking dangerous. an engine has one job - the parts move together, or they don’t. no style points.
so I handed four models the same prompt and watched what they actually built.
Claude Opus 5.5 · GPT-6 Astra · Gemini 4 Argon · Grok 4.7
one looks machined and connected. one’s all show, no linkage. four takes on the same engine - and every one bet on a different idea of what working even means.
I didn’t touch a thing and I’m not scoring them. you’ve got eyes: which one could actually turn over, and which one’s just parts floating in a box?
same prompt, every day
Opus, Astra, Gemini or Grok?
Kinder retweeted
BREAKING! THIS IS ABSOLUTELY INSANE
I turned a single street camera into a speed radar that clocks every car in real time. no radar gun, no sensors - just Python.
How it works:
DETECTOR - YOLO boxes every car, bus, bike and truck, frame by frame.
TRACKER - ByteTrack IDs each vehicle and traces its path, so it’s never double-counted.
FLATTENER - warps the angled street into a top-down map so pixels become real metres. this is what makes the speed real, not guessed.
SPEEDOMETER - measures metres moved between frames: metres × fps × 3.6 = km/h.
FILTER - averages 25 frames and drops anything unrealistic, so the number never jumps.
PAINTER - tags each car “#1 CAR · 54 km/h”, colours it by speed, flags the FASTEST, logs it all to a CSV.
a camera can’t actually measure speed - far cars look slower than close ones. the top-down warp is the whole trick, and it’s what separates a cool clip from a real tool.
built it in a weekend. save this before your next build
BREAKING! THIS IS ABSOLUTELY INSANE
I turned a single street camera into a speed radar that clocks every car in real time. no radar gun, no sensors - just Python.
How it works:
DETECTOR - YOLO boxes every car, bus, bike and truck, frame by frame.
TRACKER - ByteTrack IDs each vehicle and traces its path, so it’s never double-counted.
FLATTENER - warps the angled street into a top-down map so pixels become real metres. this is what makes the speed real, not guessed.
SPEEDOMETER - measures metres moved between frames: metres × fps × 3.6 = km/h.
FILTER - averages 25 frames and drops anything unrealistic, so the number never jumps.
PAINTER - tags each car “#1 CAR · 54 km/h”, colours it by speed, flags the FASTEST, logs it all to a CSV.
a camera can’t actually measure speed - far cars look slower than close ones. the top-down warp is the whole trick, and it’s what separates a cool clip from a real tool.
built it in a weekend. save this before your next build
this broke my brain. same prompt, two models, one unbroken shot - a carpet becomes a peacock, the peacock becomes a lion.
Gemini 4 Argon vs GPT-6 Astra. no edits, no cherry-picking. top is Gemini, bottom is Astra.
one continuous transformation, two completely different takes. the real test is the morph itself - does it flow like one living shot, or does it break the second one thing turns into another?
I’m not calling it. that’s on you.
which transformation goes in the trailer - Gemini’s or Astra’s? drop your pick and tell me why.
This is pure f*cking gold.
A 22-year-old student from China just bypassed the entire SaaS industry with a single local script.
He didn’t build a fancy UI. He didn’t raise a seed round. He just wired 300 AI agents into a closed-loop neural hive - and let them eat the web alive.
What you’re looking at on screen is the live telemetry running on his rig right now:
• 300 autonomous sub-agents executing in parallel
• a brutal cross-verification layer killing hallucinations in ~30ms
• 6 months of market research compressed into 180 minutes flat
while big tech is charging $20/month for simple wrappers, one kid in a dorm room just turned a single PC into a Fortune 500 research department.
the game didn’t change. it ended.
this is f*cking motion. four models, one prompt: a working inline-4 engine - pistons, rods, crankshaft, the whole stroke.
Claude Opus 5.5 · GPT-6 Astra · Gemini 4 Argon · Grok 4.7
the shapes are the easy part. the test is the motion - do the rods actually turn the crank, do the pistons fire in sequence, does up-and-down become round-and-round, or are they just parts bobbing in a box?
I didn’t touch the outputs and I’m not scoring them. your call: which one’s a real engine, and which one’s just animated metal?
Opus, Astra, Gemini or Grok?
this is f*cking cinema. same prompt, two models, one unbroken shot - a bird rises out of a carpet, takes flight, and becomes a panther.
Gemini 4 Argon vs Opus 5.5. no edits, no cherry-picking. top is Gemini, bottom is Opus.
one continuous transformation, two completely different takes. one keeps it sharp and lit, the other melts into fire and shadow. the real test is the morph itself - does it flow as one shot, or does it break the moment the bird becomes the beast?
I’m not calling it. that’s on you.
which transformation goes in the trailer - Gemini’s or Opus’s? drop your pick and tell me why.
same prompt, two models, every day
look f*cking closer. four models, one prompt - a hand pump, cut open so you can see it actually work.
GPT-6 Astra · Gemini 4 Argon · Grok 4.7 · Claude Opus 5.5
the outside is easy. the test is inside: does the piston seal on the downstroke, do the check valves open and close in the right order, does the water column actually climb the cylinder - or is it just blue pixels sliding up a tube?
I didn’t touch the outputs and I’m not scoring them. your call: which one’s a real pump cycle, and which one’s just animated water?
Astra, Gemini, Grok or Opus?
same prompt, four models, every day.
BREAKING! THIS IS BECOMING F*CKING DANGEROUS
Claude Code just uncovered 15,200 hidden facial captures resting inside a concealed iPhone cache.
Every unlock attempt, notification check, or quick glance while driving was silently dumping a raw snapshot into system logs.
Here’s what turned up during the raw filesystem audit:
1. System scan was executed to target all media files outside Photos & Trash.
2. Claude Code flagged a non-indexed directory packed with 15k+ infrared & standard front-facing frames.
3. Every frame contained embedded location coordinates and time logs.
Turns out internal diagnostic routines and third-party apps with camera access were hoarding local cache copies instead of flushing them.
Cleaned up my camera permissions and disabled system analytics right away. Run a local scan on your device if you use Claude Code - the results are wild.
this is getting f*cking dangerous. four models, one prompt: a stone hand pump actually pumping water.
Gemini 4 Argon · Claude Opus 5.5 · GPT-6 Astra · Grok 4.7
anyone can draw a pump. the tell is the water - does the stream arc like it has weight, does the basin ripple where it lands, does the lever actually drive the stroke, or is the water just painted on?
I didn’t touch the outputs and I’m not scoring them. your call: which one obeys physics, and which one’s a poster?
Gemini, Opus, Astra or Grok?
same prompt, four models, every day - follow so you don’t miss a round
this is f*cking satisfying. four models, one prompt: animate a working gear pump.
Claude Opus 5.5 · GPT-6 Astra · Gemini 4 Argon · Grok 4.7
a gear pump has one trick - two gears counter-rotate, trap fluid between the teeth and the wall, and carry it from inlet to outlet with no backflow at the mesh. get the meshing wrong and it’s just two gears spinning in a puddle.
I didn’t touch the outputs and I’m not scoring them. you call it: which one’s actually moving fluid, and which one’s just turning teeth for show?
Opus, Astra, Gemini or Grok?
same prompt, four models, every day - this is the daily drop.
this is f*cking engineering. four models got the same prompt: simulate a 9-cylinder radial engine.
Claude Opus 5.5 · GPT-6 Astra · Gemini 4 Argon · Grok 4.7
this isn’t about who looks slickest. a radial lives or dies on the details - does the master-rod linkage actually drive every piston, is the firing order right, does the combustion flash hit in sequence or just sparkle at random?
I didn’t touch the outputs and I’m not scoring them. you’ve got eyes: which one’s a real engine turning over, and which one’s just lights flashing in a circle?
Opus, Astra, Gemini or Grok?
same prompt, four models, every single day - this is the daily drop. follow so you don’t miss a round.
look f*cking closer. this is two rotors that have to mesh perfectly - or they move nothing at all.
four models, same prompt, same machine: Claude Opus 5.5 · GPT-6 Astra · Grok 4.7 · Gemini 4 Argon
a lobe pump has nowhere to hide. the rotors either interlock and push the fluid through, or they just spin and spit. some of these look tight. some are two shapes turning in a box.
I didn’t touch a thing and I’m not scoring them. you tell me: which one’s actually pumping, and which one just animated?
Opus, Astra, Grok or Gemini?
this is getting f*cking dangerous. a pump has exactly one job - water comes out, or it doesn’t. no style points.
so I handed the same prompt to four models and watched what they actually built.
Claude Opus 5.5 · Gemini 4 Argon · GPT-6 Astra · Grok 4.7
cast-iron and solid. skeletal and exposed. stripped to the bare mechanism. four takes on the same machine - and every one bet on a different idea of what a pump should even be.
I didn’t touch the outputs and I’m not scoring them. you’ve got eyes: which one’s actually hauling water up from that well, and which one’s just a nice drawing?
Opus, Gemini, Astra or Grok?
Kinder retweeted
pick a f*cking side. $PROPHET just shipped something clever - and it’s free to play.
the mechanic: your community makes ONE call. a single yes/no question with a real source - ape (yes) or fade (no). you vote, the majority becomes the pick.
every voter gets a tokenized prediction: an NFT of that call. got it wrong? burn it in the graveyard and it unlocks a free spin on Prophet’s wheel. hit or miss, everyone still gets the spin.
no wallet, no cost - sign in with X and you’re in, they make the wallet for you.
built on Robinhood Chain. not financial advice, nothing here is a promised return - it’s a community call game, and it’s genuinely fun to play.
so: ape or fade?
vote: graveyard.prophetthedog.xyz/…
🤝 Paid partnership
Papertrade has been proposed to be a $60M rug scheme
Its TVL experiences a loss of 70%, sliding down from $100M to $36M since launch date
Never was bullish on it, moreover after points system update (early users' efforts will be washed)
Here's how to profit from the TVL fall:
I made a prediction that Papertrade's TVL will not be above $35M on Tuesday on Prophet [graveyard.prophetthedog.xyz/…]
In return I get a tokenized prediction, an NFT that shows my picked side
> If my prediction wins - I get an airdrop in $PROPHET
> If not - I can burn my NFT and spin a wheel and still get some portion of tokens
That's basically rekt to riches mechanic
pick a f*cking side. $PROPHET just shipped something clever - and it’s free to play.
the mechanic: your community makes ONE call. a single yes/no question with a real source - ape (yes) or fade (no). you vote, the majority becomes the pick.
every voter gets a tokenized prediction: an NFT of that call. got it wrong? burn it in the graveyard and it unlocks a free spin on Prophet’s wheel. hit or miss, everyone still gets the spin.
no wallet, no cost - sign in with X and you’re in, they make the wallet for you.
built on Robinhood Chain. not financial advice, nothing here is a promised return - it’s a community call game, and it’s genuinely fun to play.
so: ape or fade?
vote: graveyard.prophetthedog.xyz/…
🤝 Paid partnership
Papertrade has been proposed to be a $60M rug scheme
Its TVL experiences a loss of 70%, sliding down from $100M to $36M since launch date
Never was bullish on it, moreover after points system update (early users' efforts will be washed)
Here's how to profit from the TVL fall:
I made a prediction that Papertrade's TVL will not be above $35M on Tuesday on Prophet [graveyard.prophetthedog.xyz/…]
In return I get a tokenized prediction, an NFT that shows my picked side
> If my prediction wins - I get an airdrop in $PROPHET
> If not - I can burn my NFT and spin a wheel and still get some portion of tokens
That's basically rekt to riches mechanic
this is f*cking wild. four models, one radial engine, four different brains.
Claude Opus 5.5 · GPT-6 Astra · Gemini 4 Argon · Grok 4.7
same words, four completely different machines. look at the cylinder count, the symmetry, how much each one packed in before calling it done. one over-engineers, one keeps it lean, one barely commits. it stopped being about accuracy - it’s how each model decides what “enough” is.
I’m not calling a winner. that’s on you.
which engine would you actually trust to turn over - Opus, Astra, Gemini or Grok? drop your pick and tell me why.
Kinder retweeted
this is f*cking gold.
Every time you explain your projects to Claude from scratch, you’re doing work your files could be doing for you.
Here’s a second-brain setup you can actually copy:
1. Create a folder called second-brain and open Claude Code inside it.
2. Add three folders:
• notes/ - facts, research, ideas
• people/ - contacts and meeting notes
• projects/ - one file per active project
3. Add three files:
• MEMORY.md - essential context about you
• decisions.md - decisions and the reasoning behind them
• LEARNINGS.md - mistakes, lessons, things to repeat
4. Create CLAUDE.md in the root folder. Describe where everything lives, then add these rules:
Read the relevant files before answering. Cite the file you used. If information is missing, say so. Keep one topic per file. Update existing notes instead of creating duplicates.
5. Start with ONE real project. Add its goal, current status, last decision, blocker, and next action.
6. Test it:
Where did I leave this project, why did I choose this approach, and what should I do next?
Check the answer against your files.
7. Before ending a work session, ask:
Save today’s decisions, update the project status, and record the next action.
That last step keeps the system useful tomorrow.
The video shows how these pieces connect. Save this checklist, then open the quoted guide for the full setup.
this is f*cking gold.
Every time you explain your projects to Claude from scratch, you’re doing work your files could be doing for you.
Here’s a second-brain setup you can actually copy:
1. Create a folder called second-brain and open Claude Code inside it.
2. Add three folders:
• notes/ - facts, research, ideas
• people/ - contacts and meeting notes
• projects/ - one file per active project
3. Add three files:
• MEMORY.md - essential context about you
• decisions.md - decisions and the reasoning behind them
• LEARNINGS.md - mistakes, lessons, things to repeat
4. Create CLAUDE.md in the root folder. Describe where everything lives, then add these rules:
Read the relevant files before answering. Cite the file you used. If information is missing, say so. Keep one topic per file. Update existing notes instead of creating duplicates.
5. Start with ONE real project. Add its goal, current status, last decision, blocker, and next action.
6. Test it:
Where did I leave this project, why did I choose this approach, and what should I do next?
Check the answer against your files.
7. Before ending a work session, ask:
Save today’s decisions, update the project status, and record the next action.
That last step keeps the system useful tomorrow.
The video shows how these pieces connect. Save this checklist, then open the quoted guide for the full setup.
this is f*cking surreal. same prompt, two models - and an entire galaxy ended up inside a falling raindrop.
GPT-6.1 Sol vs Opus 5.5. no edits, no cherry-picking. top is Sol, bottom is Opus.
three scenes, one prompt each - and the two read the same idea completely differently. neither is wrong, which is what makes it a hard call.
I’m not picking a side. that part’s on you.
which row goes in the trailer - Sol’s or Opus’s? drop your pick below, and tell me why.
next matchup drops here in a few days.