Vronzu retweeted
Christna Savoy caught her husband cheating while giving him head. The hair wasn't hers.
They'd been together since she was 19. She found a long black hair on him and didn't say a word that night.
Instead she quietly signed with an adult agency and told him she was flying to LA for paid photo shoots.
> her first scene went up on P*rnhub
> with the exact type of guy he never wanted her with
> she says she's making more money than ever
> when he finally found out, she laughed
Did he deserve it or did she take it too far?
Christna Savoy caught her husband cheating while giving him head. The hair wasn't hers.
They'd been together since she was 19. She found a long black hair on him and didn't say a word that night.
Instead she quietly signed with an adult agency and told him she was flying to LA for paid photo shoots.
> her first scene went up on P*rnhub
> with the exact type of guy he never wanted her with
> she says she's making more money than ever
> when he finally found out, she laughed
Did he deserve it or did she take it too far?
Vronzu retweeted
Sophie Rain was carrying plates fort tips. Last year OnlyF*ns paid her $83 million.
2 years at a mom and pop Italian place. Her friend had to beg her to quit.
First month online $100,000.
Last year: $83 million. Over $100 million total.
> 37% of it goes straight to federal tax
> the other $40 million goes into houses and a 20-acre farm
> 12 cows, and she wants a full cattle ranch
The richest waitress in America is buying cows.
Watch the host's face at the end when he hears the numbers.
Would you have quit the restaurant?
Sophie Rain was carrying plates fort tips. Last year OnlyF*ns paid her $83 million.
2 years at a mom and pop Italian place. Her friend had to beg her to quit.
First month online $100,000.
Last year: $83 million. Over $100 million total.
> 37% of it goes straight to federal tax
> the other $40 million goes into houses and a 20-acre farm
> 12 cows, and she wants a full cattle ranch
The richest waitress in America is buying cows.
Watch the host's face at the end when he hears the numbers.
Would you have quit the restaurant?
Vronzu retweeted
this Anthropic Fable 5.2 guide f*cking insane
they tested single sentences in the prompt and measured what each one changed.
they gave Fable an advisor. with no line saying when to ask, it was consulted 0 times out of 198.
a "never" rule with the exception in the same line got followed 9 times in 10. split into two lines: 1 in 10.
same model. one sentence apart.
so I put every line that moved the numbers into one copy-paste CLAUDE.md for Fable 5.2. every line, every number, all of it spelled out.
the crazy part is your old prompts make it worse. one stale "never retry" note cut retries to about 4 in 10.
99% of people will switch to Fable 5.2 and keep the CLAUDE.md they wrote for older models.
bookmark it before this thread gets buried.
then read the full article below ↓
Vronzu retweeted
MrBeast lost tens of millions making the $100 million Beast Games.
Episode one alone: over $15 million on 1,000 towers, each 10 feet tall.
Episode two: a $14 million city the contestants actually lived in.
Prize money on top: over $20 million.
Two episodes in, he was already past $50 million. The full budget? Over $100 million, he confirmed.
> the losses came out of his own pocket
> he called it "not a good financial decision"
> he says he would have more money if he never filmed it
He told the whole breakdown on The Diary of a CEO.
Would you lose tens of millions of your own money to make one season great?
MrBeast lost tens of millions making the $100 million Beast Games.
Episode one alone: over $15 million on 1,000 towers, each 10 feet tall.
Episode two: a $14 million city the contestants actually lived in.
Prize money on top: over $20 million.
Two episodes in, he was already past $50 million. The full budget? Over $100 million, he confirmed.
> the losses came out of his own pocket
> he called it "not a good financial decision"
> he says he would have more money if he never filmed it
He told the whole breakdown on The Diary of a CEO.
Would you lose tens of millions of your own money to make one season great?
Vronzu retweeted
holy f*ck. Snoop Dogg says a billion Spotify streams paid him under $45,000.
a billion. he asked his publisher to break it down, and that was the number. later he called it "robbery in broad daylight."
watch his face in the clip when he says it.
Spotify says a billion streams pays rights holders millions. so the real question is who the rights holders are. usually it's not the guy who made the song.
now look at Toni Braxton. 15 million albums. two Grammys in one night. bankrupt by 1998.
the label fronted the studio, the videos, the wardrobe, the travel. then every dollar of it came back out of her royalties before she saw a cent. her "debt" was the bill for getting the music out.
two of the biggest names in music. both found out the hard way that the hit and the money go to different people.
that bill looks very different now:
> a home setup and AI mastering instead of months of booked studio time
> AI video tools instead of a label-funded music video
> selling straight to fans instead of waiting on a royalty statement
> seeing every sale yourself, which is exactly what Snoop asks for in the clip
none of this makes a hit for you. it just means the artist no longer has to borrow the machine from the people who own the masters.
if you could cut out the label tomorrow, what would you still pay someone else to do?
Toni Braxton's "Un-Break My Heart" spent 11 weeks at No. 1. Her records brought in a reported $170 million in sales. The check from her first record deal was $1,972. Then she filed for BANKRUPTCY, and Oprah grilled her about her Gucci flatware.
In the interview below, Braxton tells ABC's 20/20 where the money actually went.
By 1997 she was one of the biggest voices in the world. Her album Secrets sold more than 15 million copies worldwide, and she took home two Grammys in a single night.
But the label had fronted almost everything: studio time, videos, wardrobe, travel. Under her contract, all of it had to be paid back out of her royalties before she saw a cent.
"They'll give you an advancement on the next record and then the next record, so you kind of stay in debt."
Late in 1997 she sued LaFace to get out of the deal. In January 1998 she filed for Chapter 7.
Then came Oprah. Braxton later said Oprah told her: "I hear you have Gucci flatware. I'm Oprah Winfrey and I don't have Gucci flatware."
In the clip, Braxton doesn't deny it. Plates, housewares, sheets with a thousand thread count. "I'm a boring rock star," she says.
She settled with the label in 1999, signed a new deal worth millions and even performed at the Super Bowl.
In October 2010 she filed for bankruptcy again, listing debts of $10 million to $50 million.
The woman who sang "Un-Break My Heart" to the whole world couldn't unbreak her own bank account. Twice.
Was it the label's contract or her own spending that broke her?
holy f*ck. Snoop Dogg says a billion Spotify streams paid him under $45,000.
a billion. he asked his publisher to break it down, and that was the number. later he called it "robbery in broad daylight."
watch his face in the clip when he says it.
Spotify says a billion streams pays rights holders millions. so the real question is who the rights holders are. usually it's not the guy who made the song.
now look at Toni Braxton. 15 million albums. two Grammys in one night. bankrupt by 1998.
the label fronted the studio, the videos, the wardrobe, the travel. then every dollar of it came back out of her royalties before she saw a cent. her "debt" was the bill for getting the music out.
two of the biggest names in music. both found out the hard way that the hit and the money go to different people.
that bill looks very different now:
> a home setup and AI mastering instead of months of booked studio time
> AI video tools instead of a label-funded music video
> selling straight to fans instead of waiting on a royalty statement
> seeing every sale yourself, which is exactly what Snoop asks for in the clip
none of this makes a hit for you. it just means the artist no longer has to borrow the machine from the people who own the masters.
if you could cut out the label tomorrow, what would you still pay someone else to do?
Toni Braxton's "Un-Break My Heart" spent 11 weeks at No. 1. Her records brought in a reported $170 million in sales. The check from her first record deal was $1,972. Then she filed for BANKRUPTCY, and Oprah grilled her about her Gucci flatware.
In the interview below, Braxton tells ABC's 20/20 where the money actually went.
By 1997 she was one of the biggest voices in the world. Her album Secrets sold more than 15 million copies worldwide, and she took home two Grammys in a single night.
But the label had fronted almost everything: studio time, videos, wardrobe, travel. Under her contract, all of it had to be paid back out of her royalties before she saw a cent.
"They'll give you an advancement on the next record and then the next record, so you kind of stay in debt."
Late in 1997 she sued LaFace to get out of the deal. In January 1998 she filed for Chapter 7.
Then came Oprah. Braxton later said Oprah told her: "I hear you have Gucci flatware. I'm Oprah Winfrey and I don't have Gucci flatware."
In the clip, Braxton doesn't deny it. Plates, housewares, sheets with a thousand thread count. "I'm a boring rock star," she says.
She settled with the label in 1999, signed a new deal worth millions and even performed at the Super Bowl.
In October 2010 she filed for bankruptcy again, listing debts of $10 million to $50 million.
The woman who sang "Un-Break My Heart" to the whole world couldn't unbreak her own bank account. Twice.
Was it the label's contract or her own spending that broke her?
Vronzu retweeted
holy f*ck. Elon Musk just said his own Grok Bot will use Claude Opus 5.5 whenever it's the best model for the task.
Read that again. Elon is routing to his competitor's model.
so the model is no longer your edge. the system around it is: what the agent can touch, how long it keeps trying, and what checks its work.
that's why your agent passes every test and the issue is still open.
I turned the whole harness + loop + graph setup into one prompt for Opus 5.5. every layer, every stop rule, the full copy-paste template.
It's f*cking unreal...
Send it to Opus 5.5, and thank me later. Then read the article below ↓
Vronzu retweeted
Mark Zuckerberg admitted Meta could game the AI leaderboard. on camera.
he said a lot of AI benchmarks are "quite easily gameable." then he gave his own example.
his team tuned a version of Llama 4 Maverick that could sit way at the top of the Arena. the model they actually shipped had none of that tuning, so it landed further down.
same company. same model family. a completely different rank depending on which version gets tested.
now look at the TwIL-8B launch everyone is sharing:
> the 93% coding score comes from the "verbose" variant, not the balanced one
> every score is webAI's own run
> competitor numbers come from different test setups
that doesn't make it a bad model. it means the launch chart shows the best case.
before you build on any new model, check 5 things:
> which variant produced the headline number
> who ran the test, the lab or someone independent
> were rivals tested on the same setup
> what wasn't tested at all (agents, tool use, SWE-bench)
> does the license let you sell it
then run it on 10 real tasks from your product. that's the only score you can trust.
which model looked insane on launch day and fell apart on your real work?
DON’T BUILD YOUR STARTUP ON THE MODEL EVERYONE IS HYPING THIS WEEK.
TwIL-8B just dropped. 8.79B parameters, 4.98 GB, runs on a laptop, 93% of the best frontier model on LiveCodeBench. 577K people saw the launch post. the catch sits in the model card
Read this before you ship a single line on it:
→ the license is webAI Non-Commercial. test it all you want, but a paying customer can’t touch it
→ it burns tokens. on AIME it uses about 19.5k per problem. GPT-5.5 uses about 5k
→ the 93% coding score comes from the “verbose” variant, not the balanced one
→ every score is webAI’s own run, against competitor numbers from different test setups
→ hard science is still far off: GPQA Diamond 70.2 vs 93.6
→ tool use, agent tasks and SWE-bench weren’t tested at all
→ no extra safety tuning beyond the base model, the card says so itself
imagine building 3 months of product on it and finding the license on launch day
How to use it the smart way:
→ prototype locally for $0 in API calls
→ run it on 10 of your real tasks before you trust any benchmark
→ need a commercial 8B: it’s built on IBM’s Granite 4.2 8B, which is Apache 2.0
small models are closing the gap fast. the license decides what you can sell
model card in the reply ↓
Mark Zuckerberg admitted Meta could game the AI leaderboard. on camera.
he said a lot of AI benchmarks are "quite easily gameable." then he gave his own example.
his team tuned a version of Llama 4 Maverick that could sit way at the top of the Arena. the model they actually shipped had none of that tuning, so it landed further down.
same company. same model family. a completely different rank depending on which version gets tested.
now look at the TwIL-8B launch everyone is sharing:
> the 93% coding score comes from the "verbose" variant, not the balanced one
> every score is webAI's own run
> competitor numbers come from different test setups
that doesn't make it a bad model. it means the launch chart shows the best case.
before you build on any new model, check 5 things:
> which variant produced the headline number
> who ran the test, the lab or someone independent
> were rivals tested on the same setup
> what wasn't tested at all (agents, tool use, SWE-bench)
> does the license let you sell it
then run it on 10 real tasks from your product. that's the only score you can trust.
which model looked insane on launch day and fell apart on your real work?
DON’T BUILD YOUR STARTUP ON THE MODEL EVERYONE IS HYPING THIS WEEK.
TwIL-8B just dropped. 8.79B parameters, 4.98 GB, runs on a laptop, 93% of the best frontier model on LiveCodeBench. 577K people saw the launch post. the catch sits in the model card
Read this before you ship a single line on it:
→ the license is webAI Non-Commercial. test it all you want, but a paying customer can’t touch it
→ it burns tokens. on AIME it uses about 19.5k per problem. GPT-5.5 uses about 5k
→ the 93% coding score comes from the “verbose” variant, not the balanced one
→ every score is webAI’s own run, against competitor numbers from different test setups
→ hard science is still far off: GPQA Diamond 70.2 vs 93.6
→ tool use, agent tasks and SWE-bench weren’t tested at all
→ no extra safety tuning beyond the base model, the card says so itself
imagine building 3 months of product on it and finding the license on launch day
How to use it the smart way:
→ prototype locally for $0 in API calls
→ run it on 10 of your real tasks before you trust any benchmark
→ need a commercial 8B: it’s built on IBM’s Granite 4.2 8B, which is Apache 2.0
small models are closing the gap fast. the license decides what you can sell
model card in the reply ↓
Vronzu retweeted
holy f*ck. Anthropic's Bun team ported 750,000 lines from Zig to Rust in 11 days with Claude Code.
99.8% of the old test suite still passed. hundreds of agents in parallel, two reviewers on every file.
the creator of Zig asked the obvious question: who checked a million lines no human read?
the answer is a loop inside every node and a graph between them. I turned it into one copy-paste prompt for Opus 5.5. every step, every formula, all of it spelled out.
99% of people still approve every step by hand and call it "using agents".
bookmark it before this thread gets buried.
then read the full article below ↓
Vronzu retweeted
holy f*ck. 17 claude code agents shipped a feature while i slept.
i woke up to a clean pull request, full test coverage and one message:
"Ready for your approval. Everything else is already done."
i used to write every line myself. then i spent three nights building a system instead of writing another prompt.
claude code built a team of 17 specialized agents:
> one only plans
> one only reviews diffs
> one only writes tests
> one only checks security
> one only makes architecture decisions
> twelve more do the actual work
and i gave them one hard rule. nobody wakes me unless the change is irreversible.
now the team runs without me. they research, write, test, fight with each other, fix bugs, and stop only when they need a final human signature.
i used to be the developer. now i'm the last person who has to say yes.
what's the one change you'd never let an agent ship without you?
holy f*ck. 17 claude code agents shipped a feature while i slept.
i woke up to a clean pull request, full test coverage and one message:
"Ready for your approval. Everything else is already done."
i used to write every line myself. then i spent three nights building a system instead of writing another prompt.
claude code built a team of 17 specialized agents:
> one only plans
> one only reviews diffs
> one only writes tests
> one only checks security
> one only makes architecture decisions
> twelve more do the actual work
and i gave them one hard rule. nobody wakes me unless the change is irreversible.
now the team runs without me. they research, write, test, fight with each other, fix bugs, and stop only when they need a final human signature.
i used to be the developer. now i'm the last person who has to say yes.
what's the one change you'd never let an agent ship without you?
Vronzu retweeted
holy f*ck. OpenAI co-founder Andrej Karpathy had Opus 5.5 chew through 43M comments and ship a full site in 2 hours.
The guy who coined "vibe coding" still has one rule for agents: keep AI on a tight leash. No 1,000-line diffs.
Wharton tested what happens without one. People accepted wrong AI answers 80% of the time.
I turned his loop into one prompt for Claude Code. Small diffs, an autonomy slider, evidence you can look at, and nothing ships you can't explain.
It's f*cking unreal...
Send it to Opus 5.5, and thank me later. Then read the article below ↓
holy f*ck. claude built a 150-servo robot bee with zero 3d models.
no blender. no asset store. not a single downloaded model.
one html file and three.js. every plate, every gear, every cable in this clip is code.
and it almost didn't work.
the gears came together. the optics came together. the full teardown came together.
the wings didn't.
4 tries. 3 of them looked like a desk fan.
on the 4th, the blur finally looked like real bee wings.
watch the 19 seconds and look for:
> white armor plates with the mechanics showing through the gaps
> a chest motor with an orange ring and spinning gears
> optics pulling apart into labeled layers. iris, inner element, outer glass, lens barrel
> a wing lifting off its shoulder mount and rotating into view
> a tail with 7 glass vanes fanning open
> the full teardown, then every part snaps back and the wings beat again
the hud is fake. it still moves like a real flight sim. wingbeat readout, battery cells, servo loads, a flight log scrubber.
and the whole thing renders straight from the browser.
what should it take apart next?
$0 spent on 3d models. this robotic hummingbird is one html file and every part of it is code
no blender no asset store just three.js and claude
what's inside:
1. a slim body built from white armor plates with the mechanics showing through the gaps
2. 150 small servos, blocks and cables packed under the shell
3. a chest motor with an orange ring and spinning gears
4. optics that pull forward into labeled layers. iris, inner element, outer glass, lens barrel
5. a wing that lifts off its shoulder mount and rotates into view
6. a shuttlecock tail with 7 glass vanes that fans open
7. a full teardown where every part floats around the frame, then snaps back and the wings beat again
the hud is fake but it moves like a real flight sim. wingbeat readout, battery cells, servo loads, a flight log scrubber
the whole 19 second clip renders straight from the browser
honestly the hardest part wasn't the bird. it was getting the wing blur to look like a real hummingbird and not a fan
what would you make it take apart next?
7
11
535
Vronzu retweeted
this Anthropic research f*cking insane
they studied 132 of their own engineers and 200,000 Claude Code transcripts and named the trap everyone is walking into: the paradox of supervision.
to supervise Claude you need the exact skills that heavy Claude use wears down.
so I turned the fix into one prompt for Opus 5.5. every step, every formula, the full copy-paste template, all of it spelled out.
the crazy part is the fix doesn't mean using AI less. the agent proves its own work and you stay the one who can explain it.
99% of people approve a 600-line PR because the tests are green and can't explain a single line of it a week later.
bookmark it before this thread gets buried.
then read the full article below ↓
Vronzu retweeted
$300 a month and Claude Code still forgot my shit every morning.
I spent 4 hours a month re-teaching it the same things. None of it was new work. It was the same rules I'd already explained the day before.
Every morning it was a brand new intern who had never met me.
Same folder structure.
Same naming rules.
Same "no, not like that, like THIS."
$300 in tokens just to start from zero again.
Then I found out Claude Code has 4 memory layers. Almost everyone uses one. I did too.
You keep stuffing that one file with every rule, every preference, every correction. It gets longer and longer, Claude starts ignoring half of it, and you decide AI memory is broken.
Turns out I'd just never set up the other three.
The setup took me 10 minutes. Here's what each layer does:
> one short map file. Claude reads it every session, so it only holds what matters every time
> rules that only load when Claude touches the files they're about
> skills that sit quiet until you call them, so they don't eat context the rest of the time
> auto-memory that saves your corrections, so "no, like THIS" happens once
Since then I haven't explained the same thing twice. The folder structure, the naming rules, the "like THIS" moments. Claude already knows them when I open it.
The full breakdown and the exact copy-paste setup prompt are in the post below. 10 minutes once, instead of 4 hours every month.
