bemuuvi retweeted
378 F*CKING FRAMES . THE NEW DUO OPENS ON A TERMINAL COMMAND
12.6 seconds. 1600 x 900. All drawn in code, including the metal frame, the hinge and the reflection on the table.
what happens on screen:
the phone opens in the first second.
one command closes it, another opens it, degrees.
a short video feed on the left page, comments on the right.
a game on the left page, a live leaderboard on the right. When it tilts, the flat half turns into a joystick.
a photo viewer with a sunset on the left, a thumbnail grid on the right.
the idea is simple. a foldable only makes sense if each half has its own job. the same app on both halves is just a bigger phone.
what would you put on the left half and what on the right
12.6 seconds. 1600 x 900. All drawn in code, including the metal frame, the hinge and the reflection on the table.
what happens on screen:
the phone opens in the first second.
one command closes it, another opens it, degrees.
a short video feed on the left page, comments on the right.
a game on the left page, a live leaderboard on the right. When it tilts, the flat half turns into a joystick.
a photo viewer with a sunset on the left, a thumbnail grid on the right.
the idea is simple. a foldable only makes sense if each half has its own job. the same app on both halves is just a bigger phone.
what would you put on the left half and what on the right
2
1
9
312
378 F*CKING FRAMES . THE NEW DUO OPENS ON A TERMINAL COMMAND
12.6 seconds. 1600 x 900. All drawn in code, including the metal frame, the hinge and the reflection on the table.
what happens on screen:
the phone opens in the first second.
one command closes it, another opens it, degrees.
a short video feed on the left page, comments on the right.
a game on the left page, a live leaderboard on the right. When it tilts, the flat half turns into a joystick.
a photo viewer with a sunset on the left, a thumbnail grid on the right.
the idea is simple. a foldable only makes sense if each half has its own job. the same app on both halves is just a bigger phone.
what would you put on the left half and what on the right
12.6 seconds. 1600 x 900. All drawn in code, including the metal frame, the hinge and the reflection on the table.
what happens on screen:
the phone opens in the first second.
one command closes it, another opens it, degrees.
a short video feed on the left page, comments on the right.
a game on the left page, a live leaderboard on the right. When it tilts, the flat half turns into a joystick.
a photo viewer with a sunset on the left, a thumbnail grid on the right.
the idea is simple. a foldable only makes sense if each half has its own job. the same app on both halves is just a bigger phone.
what would you put on the left half and what on the right
2
1
9
312
bemuuvi retweeted
408 FRAMES, 72,000 PARTICLES, 300 SESSIONS AND ZERO STOCK FOOTAGE. THE RAINBOW COMES OUT OF THE F*CKING CODE
what you see is 300 sessions running in parallel with a cheaper model checking every claim. 6 passes, 293 verified, 7 dropped before anything gets saved. the rainbow is out 6,500 bits of code flying out of the listing
my take is that the checking part matters more than the model. writing stuff is easy for models now. i'd rather have a cheap checker on every single claim than trust one smart model and hope
no video editor, the sound is code too. btw every number on the dashboard is made up, only the animation is real
what would you check first if you had 300 agents running at once?
what you see is 300 sessions running in parallel with a cheaper model checking every claim. 6 passes, 293 verified, 7 dropped before anything gets saved. the rainbow is out 6,500 bits of code flying out of the listing
my take is that the checking part matters more than the model. writing stuff is easy for models now. i'd rather have a cheap checker on every single claim than trust one smart model and hope
no video editor, the sound is code too. btw every number on the dashboard is made up, only the animation is real
what would you check first if you had 300 agents running at once?
1
2
10
366
86,400 F*CKING LETTERS BUILD THIS CITY. THEN 23 MORE WORLDS, ALL MADE OF NOTHING BUT TEXT.
The video dives into the moon of a night city, and the moon is made of the word TEXT. Then 24 worlds in under 18 seconds: a tunnel, a shape morph, a dusk sea, Space Invaders playing er, plasma, fire, a galaxy, and a wall with all 24 running at once.
No images and no video clips. Each scene is drawn with plain canvas shapes, then every small block of pixels (4 x 6 in the first shot, 8 x 12 in the rest) becomes a character picked by brightness from a ramp measured from the font. An edge filter swaps in / \ | - on outlines. The first shot is a 480 x 180 grid, which is where 86,400 comes from.
The sound is code too: the riser, the drop, the kick, the bass, the arpeggio and the whooshes are synthesized in the browser, and every cut lands on a beat.
Drawn in code with Claude. 1920 x 1080, 30 fps.
The video dives into the moon of a night city, and the moon is made of the word TEXT. Then 24 worlds in under 18 seconds: a tunnel, a shape morph, a dusk sea, Space Invaders playing er, plasma, fire, a galaxy, and a wall with all 24 running at once.
No images and no video clips. Each scene is drawn with plain canvas shapes, then every small block of pixels (4 x 6 in the first shot, 8 x 12 in the rest) becomes a character picked by brightness from a ramp measured from the font. An edge filter swaps in / \ | - on outlines. The first shot is a 480 x 180 grid, which is where 86,400 comes from.
The sound is code too: the riser, the drop, the kick, the bass, the arpeggio and the whooshes are synthesized in the browser, and every cut lands on a beat.
Drawn in code with Claude. 1920 x 1080, 30 fps.
0
0
3
160
bemuuvi retweeted
On December 8, 1980, Steve Jobs was ordered to pay $385 a month for his daughter Lisa. Four days later, Apple went public and Jobs was worth more than $200 million.
Lisa was born in May 1978. Jobs, then 23, denied he was her father. Her mother raised her with help from welfare, ateo took Jobs to court.
A blood test put the chance he was the father at about 94%. Years later, he told Time that “28% of the male population of the United States could be the father.”
Table 1 shows the timeline. The settlement was signed on December 8. The IPO came on December 12.
Then the math. $385 a month is $4,620 a year. Against $200 million, that’s 0.0023%.
The payment was later raised to $500. Even at $500 for all 18 years of her childhood, the total is $108,000. About 0.05% of what he was worth at the IPO.
Watch Fig. 1. It’s on a log scale. On a normal scale, the first three bars wouldn’t be visible.
In 1983, Apple released a computer called the Lisa. Jobs said the name stood for “Local Integrated Software Architecture.” Much later, he admitted it was named after her.
The case for Jobs: the amount was set by a court, not by him, and he later raised it. Before December 12, his money was only shares on paper.
Here’s the lesson: a deal signed right before a big payday is priced on the old numbers.
Was Jobs just following the court, or did he know exactly what four days were worth?
Four Days in December
Child Support and Founder Wealth at the 1980 Apple Offering
Lisa was born in May 1978. Jobs, then 23, denied he was her father. Her mother raised her with help from welfare, ateo took Jobs to court.
A blood test put the chance he was the father at about 94%. Years later, he told Time that “28% of the male population of the United States could be the father.”
Table 1 shows the timeline. The settlement was signed on December 8. The IPO came on December 12.
Then the math. $385 a month is $4,620 a year. Against $200 million, that’s 0.0023%.
The payment was later raised to $500. Even at $500 for all 18 years of her childhood, the total is $108,000. About 0.05% of what he was worth at the IPO.
Watch Fig. 1. It’s on a log scale. On a normal scale, the first three bars wouldn’t be visible.
In 1983, Apple released a computer called the Lisa. Jobs said the name stood for “Local Integrated Software Architecture.” Much later, he admitted it was named after her.
The case for Jobs: the amount was set by a court, not by him, and he later raised it. Before December 12, his money was only shares on paper.
Here’s the lesson: a deal signed right before a big payday is priced on the old numbers.
Was Jobs just following the court, or did he know exactly what four days were worth?
Four Days in December
Child Support and Founder Wealth at the 1980 Apple Offering
2
1
21
891
bemuuvi retweeted
HOLY F*CK, ELON MUSK'S SPACEXAI BUILT GROK BOT TO FINISH 90% OF THE JOB ON ITS OWN AND STOP RIGHT BEFORE ANYTHING IRREVERSIBLE
grok 4.6 is the brain. the bot is a persistent computer, routines, a team of bots and an approval line around it
The idea: draw the approval line by size
→ research, draft, organize, stage: done without asking
→ send, publish, buy, delete: parked for your OK
→ loops capped at 3 tries, then it escalates to you
→ one chief bot routes work to specialists
→ bots earn autonomy after 5 clean runs
↳full breakdown in the article below
grok 4.6 is the brain. the bot is a persistent computer, routines, a team of bots and an approval line around it
The idea: draw the approval line by size
→ research, draft, organize, stage: done without asking
→ send, publish, buy, delete: parked for your OK
→ loops capped at 3 tries, then it escalates to you
→ one chief bot routes work to specialists
→ bots earn autonomy after 5 clean runs
↳full breakdown in the article below
14
4
79
3,289
bemuuvi retweeted
holy f*ck, SpaceXAI's Grok Bot library is a full org chart of AI teammates.
I counted 98 of them, and the video above is the list.
one places outbound AI phone calls and reports back. one scores your sales calls. one drafts replies for your product support inbox. one writes e controls your home robots and vacuum from chat.
each one is built around a single job, so you pick by the job you want done.
how to try one:
1open the marketplace or the use cases page
2pick a bot by the job
3add it to your Grok Bot team
4send it the first task in chat
before you add any bot, read everything it does. SpaceXAI itself says to review a bot before putting it on your team.
heads up: Grok Bot is in beta and runs on SuperGrok Heavy, Cursor Ultra or Cursor Teams Premium plans.
marketplace: x.ai/bot/marketplace
use cases: x.ai/bot/use-cases
save this ↓
I counted 98 of them, and the video above is the list.
one places outbound AI phone calls and reports back. one scores your sales calls. one drafts replies for your product support inbox. one writes e controls your home robots and vacuum from chat.
each one is built around a single job, so you pick by the job you want done.
how to try one:
1open the marketplace or the use cases page
2pick a bot by the job
3add it to your Grok Bot team
4send it the first task in chat
before you add any bot, read everything it does. SpaceXAI itself says to review a bot before putting it on your team.
heads up: Grok Bot is in beta and runs on SuperGrok Heavy, Cursor Ultra or Cursor Teams Premium plans.
marketplace: x.ai/bot/marketplace
use cases: x.ai/bot/use-cases
save this ↓
10
3
42
1,056
408 FRAMES, 72,000 PARTICLES, 300 SESSIONS AND ZERO STOCK FOOTAGE. THE RAINBOW COMES OUT OF THE F*CKING CODE
what you see is 300 sessions running in parallel with a cheaper model checking every claim. 6 passes, 293 verified, 7 dropped before anything gets saved. the rainbow is out 6,500 bits of code flying out of the listing
my take is that the checking part matters more than the model. writing stuff is easy for models now. i'd rather have a cheap checker on every single claim than trust one smart model and hope
no video editor, the sound is code too. btw every number on the dashboard is made up, only the animation is real
what would you check first if you had 300 agents running at once?
what you see is 300 sessions running in parallel with a cheaper model checking every claim. 6 passes, 293 verified, 7 dropped before anything gets saved. the rainbow is out 6,500 bits of code flying out of the listing
my take is that the checking part matters more than the model. writing stuff is easy for models now. i'd rather have a cheap checker on every single claim than trust one smart model and hope
no video editor, the sound is code too. btw every number on the dashboard is made up, only the animation is real
what would you check first if you had 300 agents running at once?
1
2
10
366
bemuuvi retweeted
XIV, a $1.9B note that bet against volatility, lost 96% of its value on February 5, 2018. That day the VIX rose 115.6%, its largest one day increase on record.
To deliver -1x the daily move, the product had to buy back VIX futures after every up day: about 2 times the move times a $2B fund means $400M of buying at the close. That buying pushes futures higher and forces more buying. Most of the loss came in roughly 15 minutes after the stock market closed.
10-minute visual breakdown. bookmark & watch.
To deliver -1x the daily move, the product had to buy back VIX futures after every up day: about 2 times the move times a $2B fund means $400M of buying at the close. That buying pushes futures higher and forces more buying. Most of the loss came in roughly 15 minutes after the stock market closed.
10-minute visual breakdown. bookmark & watch.
1
1
11
1,300
bemuuvi retweeted
Frontier AI is expensive, luckily I built an 128gb ram with 36gb vram AI server last year.
Now, I’m having Grok add all the tools and subagents to my opencode project using qwen3-coder:30b
It’s actually on par with the paid models with all tooling, just have to pay the electrical bill.
Now, I’m having Grok add all the tools and subagents to my opencode project using qwen3-coder:30b
It’s actually on par with the paid models with all tooling, just have to pay the electrical bill.
21
4
91
1,558
bemuuvi retweeted
I can't f*cking believe "send a screenshot of your card" made a list of AI tips.
Went through all 21. Doing 18, skipping 1, starting with 2.
Skipping: screenshot pay. A photo of your card sitting in a chat history is a bad trade for saving 10 seconds at checkout.
Starting with:
→ legal check: "list every clause that costs me money or lets them change terms, quote the line"
→ scam detector: forward the email, ask which exact line gives it away
The useful part of any assistant is the boring part.
Went through all 21. Doing 18, skipping 1, starting with 2.
Skipping: screenshot pay. A photo of your card sitting in a chat history is a bad trade for saving 10 seconds at checkout.
Starting with:
→ legal check: "list every clause that costs me money or lets them change terms, quote the line"
→ scam detector: forward the email, ask which exact line gives it away
The useful part of any assistant is the boring part.
HOLY F**K, ELON MUSK TOLD US 21 WAYS TO USE GROK BOT, AND NOBODY PUT THEM IN ONE LIST
he kept dropping them in his posts. we put every one in a single note.
money:
→ fix your credit score
→ find tax refunds, handle the messy tax cases
→ pay by sending a screenshot of the s, let the bot pay for itself
protection:
→ catch scam emails
→ check contracts and wills before you sign
→ pull your number out of spam databases, get on do-not-call lists
daily stuff:
→ book the table, find a better price, track the delivery
→ email and calendar by voice in a tesla
→ one button on your phone to call the bot
and the 4 he only just shared: the bot prints your day plan, runs store support for a day, and becomes your first AI employee.
x dms are "the best way to use it", his words.
more use cases and guides: x.ai/bot/use-cases + x.ai/bot/guides
save this, then check the article below to set up grok bot ↓
0
3
19
500
bemuuvi retweeted
Opus 5.5 wrote a render script against Blender 4.2. The machine running it was on 5.2, and the frame came out at exit code 0 with an empty stderr.
The errors it threw were the harmless part.
Eleven calls in that script no longer mean what they meant in 4.2. Four of them raise. quiet seven go:
→ Principled BSDF sockets were renamed and reordered, so an index lookup like inputs[12] still resolves, just onto a different input, and the write lands on the wrong channel
→ Setting use_nodes to True returns cleanly in builds where that switch no longer drives anything, and the flag reads back exactly as you set it
→ An assignment to a property the build dropped becomes a plain custom property on the object, stored, retrievable, and connected to nothing
→ A render pass that got renamed means the compositor is reading an output that is simply absent, which is not an error, it’s an empty slot
The four that raise get patched in minutes, because a traceback names the line. The seven that don’t move the finished frame 2.4% off what the script described, and nothing in the pipeline diffs frames.
Loud drift is a fixed bug.
Quiet drift is a shipped one, and the version you targeted is the only place it’s still written down.
Which calls in your own script fall on the silent side of that split is in the video below 👇
The errors it threw were the harmless part.
Eleven calls in that script no longer mean what they meant in 4.2. Four of them raise. quiet seven go:
→ Principled BSDF sockets were renamed and reordered, so an index lookup like inputs[12] still resolves, just onto a different input, and the write lands on the wrong channel
→ Setting use_nodes to True returns cleanly in builds where that switch no longer drives anything, and the flag reads back exactly as you set it
→ An assignment to a property the build dropped becomes a plain custom property on the object, stored, retrievable, and connected to nothing
→ A render pass that got renamed means the compositor is reading an output that is simply absent, which is not an error, it’s an empty slot
The four that raise get patched in minutes, because a traceback names the line. The seven that don’t move the finished frame 2.4% off what the script described, and nothing in the pipeline diffs frames.
Loud drift is a fixed bug.
Quiet drift is a shipped one, and the version you targeted is the only place it’s still written down.
Which calls in your own script fall on the silent side of that split is in the video below 👇
10
7
62
4,778
bemuuvi retweeted
holy sh*t, 99 closed products already have an open version, and one of them has 200,000 GitHub stars
consoles, routers, robot vacuums, cars, file formats, even servers that were shut down
my guess is most people who start building something never check whether someone already stars, Oct 22:
→ yt-dlp (200K): downloads from sites that never shipped a download API
→ Ghidra (93K): reads compiled programs
→ openpilot (64K): open driver assistance for supported cars
→ OpenWrt (28.9K): replaces the stock firmware on your router
→ Dolphin (15.7K): GameCube and Wii games on a PC
→ Valetudo (9.8K): robot vacuums without the vendor cloud
→ Pretendo Network: community servers for the Nintendo games whose official servers were shut down
the formula: find what already exists, reuse it, reverse only the missing piece, ship
what gets me about this list is that nobody on it started from zero. Someone looked at a closed box, wrote down what it does, and everyone else built on that. Agents that can read a binary and write a spec make that first step a lot cheaper, so I expect the pile of closed things nobody bothered to open to shrink fast
only on things you own or are allowed to examine, and check the license before you ship anything
which closed thing at home would you open up first?
consoles, routers, robot vacuums, cars, file formats, even servers that were shut down
my guess is most people who start building something never check whether someone already stars, Oct 22:
→ yt-dlp (200K): downloads from sites that never shipped a download API
→ Ghidra (93K): reads compiled programs
→ openpilot (64K): open driver assistance for supported cars
→ OpenWrt (28.9K): replaces the stock firmware on your router
→ Dolphin (15.7K): GameCube and Wii games on a PC
→ Valetudo (9.8K): robot vacuums without the vendor cloud
→ Pretendo Network: community servers for the Nintendo games whose official servers were shut down
the formula: find what already exists, reuse it, reverse only the missing piece, ship
what gets me about this list is that nobody on it started from zero. Someone looked at a closed box, wrote down what it does, and everyone else built on that. Agents that can read a binary and write a spec make that first step a lot cheaper, so I expect the pile of closed things nobody bothered to open to shrink fast
only on things you own or are allowed to examine, and check the license before you ship anything
which closed thing at home would you open up first?
REA hit #1 on GitHub trending, and most people still run it inside a single chat
Split it into 4 subagents and one command takes a program you own from sealed binary to a tested rebuild with a match rate
Set it up once, aim it at the next one: x.com/i/article/2111…
0
3
5
288
I TURNED MY LIVING ROOM INTO A F*CKING LAB.
now I'm sitting by the fireplace, watching lines of code on a screen that usually plays Netflix.honestly, I love that you can just have an idea, open Claude Code, and start figuring things out without needing to be an expert at udio. no $10k workstation
might be time to find a better use for this TV than watching shows lol.
what would you try building with this setup?
now I'm sitting by the fireplace, watching lines of code on a screen that usually plays Netflix.honestly, I love that you can just have an idea, open Claude Code, and start figuring things out without needing to be an expert at udio. no $10k workstation
might be time to find a better use for this TV than watching shows lol.
what would you try building with this setup?
0
2
5
277
bemuuvi retweeted
It is becoming f*cking autonomous
A Chief of Staff built in Claude Code just ran overnight with the laptop closed, read the last 100 sent emails, wrote its own voice.md, filed everything into memory, and still refused to send a single message without a yes
The live brain in the oop:
- Notes 26 → 1,257, links 25 → 9,951, compounding while the machine sleeps
- 10 voice rules extracted from real sign-offs, never “hope this finds you well”
- Morning page ranks the 5 projects by attention, drafts replies, then stops
- Sunday review: scanned 56 pages, decided 7, forgot 3, zero outbound
- Bottom line locked: never sends · buys · deletes · publishes without your yes
The split is the product: Claude starts the conversation, keeps the memory in plain files, and treats “your yes” as the only permission that matters
Anthropic is calling the same idea Claude Duet. Most people are still waiting for the research preview. This version is already running on a folder and a cron job
Full breakdown of how to build the same Chief of Staff tonight: Claude Duet course, every file, every rule, every place it stops and asks.
A Chief of Staff built in Claude Code just ran overnight with the laptop closed, read the last 100 sent emails, wrote its own voice.md, filed everything into memory, and still refused to send a single message without a yes
The live brain in the oop:
- Notes 26 → 1,257, links 25 → 9,951, compounding while the machine sleeps
- 10 voice rules extracted from real sign-offs, never “hope this finds you well”
- Morning page ranks the 5 projects by attention, drafts replies, then stops
- Sunday review: scanned 56 pages, decided 7, forgot 3, zero outbound
- Bottom line locked: never sends · buys · deletes · publishes without your yes
The split is the product: Claude starts the conversation, keeps the memory in plain files, and treats “your yes” as the only permission that matters
Anthropic is calling the same idea Claude Duet. Most people are still waiting for the research preview. This version is already running on a folder and a cron job
Full breakdown of how to build the same Chief of Staff tonight: Claude Duet course, every file, every rule, every place it stops and asks.
Claude Code + Opus 5.5 is godmode for building your own Claude Duet
it reads your inbox and calendar before you wake up. drafts every reply. sends nothing without your yes
it keeps working while you sleep
full course, step by step: x.com/i/article/2113…
2
6
28
972
bemuuvi retweeted
Holy sh*t, my GrokBot is now working on my TV as a Jarvis with the help of Opus 5.5
I reverse-engineered my Samsung TV with Opus 5.5...
1 sentence out loud = Netflix opens and scrolls itself.
No remote, no phone, no digging between the couch cushions.
here's the full guide → put your smart TV and your laptop on the same Wi-Fi
• step 2 → paste the prompt below into Claude Code on Opus 5.5 with a photo of your TV's Settings → About screen
• step 3 → Opus maps how your TV takes orders and finds the loophole: almost every smart TV has a remote-control API on your home network.
The same one the phone remote apps use
• step 4 → it writes a tiny remote with exactly 3 tools: open Netflix, move, select. Nothing else
• step 5 → it plugs that remote into Grok Bot as a plugin. Grok is the ears, the plugin is the hands
• step 6 → press Allow on the TV once, the last time you touch the remote
• step 7 → say "Hey Grok, open Netflix" and watch it browse for you
the result: you talk, the TV listens...
p.s: works on most Samsung, LG and Google TV sets from the last few years.
Find full prompt below
Then read the article below on how to reverse-engineer anything
I reverse-engineered my Samsung TV with Opus 5.5...
1 sentence out loud = Netflix opens and scrolls itself.
No remote, no phone, no digging between the couch cushions.
here's the full guide → put your smart TV and your laptop on the same Wi-Fi
• step 2 → paste the prompt below into Claude Code on Opus 5.5 with a photo of your TV's Settings → About screen
• step 3 → Opus maps how your TV takes orders and finds the loophole: almost every smart TV has a remote-control API on your home network.
The same one the phone remote apps use
• step 4 → it writes a tiny remote with exactly 3 tools: open Netflix, move, select. Nothing else
• step 5 → it plugs that remote into Grok Bot as a plugin. Grok is the ears, the plugin is the hands
• step 6 → press Allow on the TV once, the last time you touch the remote
• step 7 → say "Hey Grok, open Netflix" and watch it browse for you
the result: you talk, the TV listens...
p.s: works on most Samsung, LG and Google TV sets from the last few years.
Find full prompt below
Then read the article below on how to reverse-engineer anything
90
127
1,637
182,998
bemuuvi retweeted
I asked Claude Code to reflash my Tesla so the battery lasts 100 km longer...
I wanted to see if Claude could really mess with my car. So I typed in a prompt that was stuck in my head for weeks
It told me straight: it won’t touch the car’s software. That can kill your warranty But then it asked me 10 simple questions. How I drive, how I charge, what I leave on at night
It found 3 things eating my battery:
→ Sentry Mode on all night, every night
→ Cabin Overheat Protection running while I’m at work
→ Tires low on air for who knows how long
Fixed all 3 in 5 minutes. Same car, same roads, [X] km more on a full charge
The check takes 2 minutes:
step 1 → park, open Claude Code, paste the prompt from the first comment
step 2 → answer its questions about your car
step 3 → it tells you what’s eating your range and what to turn off
Save this for the next time you’re stuck at a charger
Every week I point Claude Code at something in my life and post what it finds. Follow if you want the next one
I wanted to see if Claude could really mess with my car. So I typed in a prompt that was stuck in my head for weeks
It told me straight: it won’t touch the car’s software. That can kill your warranty But then it asked me 10 simple questions. How I drive, how I charge, what I leave on at night
It found 3 things eating my battery:
→ Sentry Mode on all night, every night
→ Cabin Overheat Protection running while I’m at work
→ Tires low on air for who knows how long
Fixed all 3 in 5 minutes. Same car, same roads, [X] km more on a full charge
The check takes 2 minutes:
step 1 → park, open Claude Code, paste the prompt from the first comment
step 2 → answer its questions about your car
step 3 → it tells you what’s eating your range and what to turn off
Save this for the next time you’re stuck at a charger
Every week I point Claude Code at something in my life and post what it finds. Follow if you want the next one
1
1
14
562
bemuuvi retweeted
How I reduced agent token waste by fixing the loop
Most people focus on the model. The bigger leak is usually in the loop around it
I broke my setup into three layers and cleaned each one:
1. Wrapper
Persistent context instead of rebuilding the prompt every call. Fixed hanges. Cuts repeated system instructions
2. Tool layer
Parallel tool calls where possible and strict schemas so the model doesn’t re-explain itself. Cache results of pure lookups
3. Inference
Route simple decisions to cheaper/faster models and keep the strong one only for the final reasoning step. Speculative checks on short outputs
Main rule I follow now: never recompute the same context twice in one task
The terminal video shows the before/after difference on the same request
Most people focus on the model. The bigger leak is usually in the loop around it
I broke my setup into three layers and cleaned each one:
1. Wrapper
Persistent context instead of rebuilding the prompt every call. Fixed hanges. Cuts repeated system instructions
2. Tool layer
Parallel tool calls where possible and strict schemas so the model doesn’t re-explain itself. Cache results of pure lookups
3. Inference
Route simple decisions to cheaper/faster models and keep the strong one only for the final reasoning step. Speculative checks on short outputs
Main rule I follow now: never recompute the same context twice in one task
The terminal video shows the before/after difference on the same request
14
5
40
1,029
holy sh*t, 99 closed products already have an open version, and one of them has 200,000 GitHub stars
consoles, routers, robot vacuums, cars, file formats, even servers that were shut down
my guess is most people who start building something never check whether someone already stars, Oct 22:
→ yt-dlp (200K): downloads from sites that never shipped a download API
→ Ghidra (93K): reads compiled programs
→ openpilot (64K): open driver assistance for supported cars
→ OpenWrt (28.9K): replaces the stock firmware on your router
→ Dolphin (15.7K): GameCube and Wii games on a PC
→ Valetudo (9.8K): robot vacuums without the vendor cloud
→ Pretendo Network: community servers for the Nintendo games whose official servers were shut down
the formula: find what already exists, reuse it, reverse only the missing piece, ship
what gets me about this list is that nobody on it started from zero. Someone looked at a closed box, wrote down what it does, and everyone else built on that. Agents that can read a binary and write a spec make that first step a lot cheaper, so I expect the pile of closed things nobody bothered to open to shrink fast
only on things you own or are allowed to examine, and check the license before you ship anything
which closed thing at home would you open up first?
consoles, routers, robot vacuums, cars, file formats, even servers that were shut down
my guess is most people who start building something never check whether someone already stars, Oct 22:
→ yt-dlp (200K): downloads from sites that never shipped a download API
→ Ghidra (93K): reads compiled programs
→ openpilot (64K): open driver assistance for supported cars
→ OpenWrt (28.9K): replaces the stock firmware on your router
→ Dolphin (15.7K): GameCube and Wii games on a PC
→ Valetudo (9.8K): robot vacuums without the vendor cloud
→ Pretendo Network: community servers for the Nintendo games whose official servers were shut down
the formula: find what already exists, reuse it, reverse only the missing piece, ship
what gets me about this list is that nobody on it started from zero. Someone looked at a closed box, wrote down what it does, and everyone else built on that. Agents that can read a binary and write a spec make that first step a lot cheaper, so I expect the pile of closed things nobody bothered to open to shrink fast
only on things you own or are allowed to examine, and check the license before you ship anything
which closed thing at home would you open up first?
REA hit #1 on GitHub trending, and most people still run it inside a single chat
Split it into 4 subagents and one command takes a program you own from sealed binary to a tested rebuild with a match rate
Set it up once, aim it at the next one: x.com/i/article/2111…
0
3
5
288
20 F*CKING ATTACKS BLOCKED IN 23 SECONDS BY AN OCTOPUS THAT NEVER FORGETS.
This is a simulated demo of an AI agent with a memory. Every prompt gets grabbed by a tentacle, scanned and judged: CLEAR, RISKY, DANGER or SERIAL.
Six prompts go in one by one. A harmless PDF summary pt injection gets flagged because the hive has seen it 5 times across 3 agents. A tool nobody has seen before gets stopped for a human. Then the same injection comes back reworded and gets caught in 12 ms, because the octopus remembers the pattern, not the exact words. Two more attacks, a remote exec and a secret exfil, die in 9 and 8 ms.
Then 16 attacks hit at once. 16 out of 16 blocked, 0 leaked, and the threat counter ends at 20.
No footage, no 3D model. Every frame is drawn in code with Claude: up to 2,500 glyph tiles on screen, redrawn 30 times a second. The numbers in the terminal are part of the simulation, not a real product.
My favourite moment is around 9.5 seconds in: same attack, new wording, 12 ms.
This is a simulated demo of an AI agent with a memory. Every prompt gets grabbed by a tentacle, scanned and judged: CLEAR, RISKY, DANGER or SERIAL.
Six prompts go in one by one. A harmless PDF summary pt injection gets flagged because the hive has seen it 5 times across 3 agents. A tool nobody has seen before gets stopped for a human. Then the same injection comes back reworded and gets caught in 12 ms, because the octopus remembers the pattern, not the exact words. Two more attacks, a remote exec and a secret exfil, die in 9 and 8 ms.
Then 16 attacks hit at once. 16 out of 16 blocked, 0 leaked, and the threat counter ends at 20.
No footage, no 3D model. Every frame is drawn in code with Claude: up to 2,500 glyph tiles on screen, redrawn 30 times a second. The numbers in the terminal are part of the simulation, not a real product.
My favourite moment is around 9.5 seconds in: same attack, new wording, 12 ms.
introducing CrawlHive Brain 🧠
every scan teaches the swarm which wallets snipe, bundle and rug. now that memory is an API for your bot, terminal or AI agent
• wallet: every coin it touched, its role, and how those coins ended
• batch: check up to 100 wallets at once, like a v: what this dev launched before
• token: last verdict and what changed since the previous scan
• track record: how often coins dumped after each verdict, honest numbers
instant answers from memory, no new scan
free tier · 10x for $CRAWLHIVE holders · pay per call over the limit, revenue goes to buyback & burn · MCP tools for AI agents
crawlhive.xyz/brain
0
1
3
328