Olivia
Online
grail βͺπ
@grailisonfire
Joined December 2021
5 Following
59 Followers
24 Posts
Anthropic's own Claude Code team quietly stopped using plan mode because the model already thinks correctly
Their internal read is blunter than the marketing: if Anthropic froze model development today, it would still take six to twelve months just to find everything current models can already do. That capability just sits unused.
In 2024 the working theory inside Anthropic was that coding gets solved by brute-forcing context windows out toward 100 million tokens, cram the whole codebase in and let the model chew through it. Claude Code shipped the opposite bet instead: let the model build its own context on the fly. That single call is most of what people now credit to "the model got smarter," which is roughly what Thariq, who's on the Claude Code team, told a room of founders last week.
The same logic just shipped again as "workflows." Instead of a human hardcoding a research harness, the hundred searches, the retries, the sub-agents, Claude now builds that harness itself, on the spot, per task.
The clearest proof point going around: someone at Bun handed Claude Code a brutal test suite and let it rewrite the runtime into Rust, unsupervised, until the tests passed.
Even RAG for context search is already treated as an anti-pattern inside Anthropic now.
They just use grep.
This AI story is getting weird: the agentic web is now bigger than the human web
Nenad Tomasev, who leads agent research at Google DeepMind, said most of the web today is generated by agents and consumed by agents, not by us. That flip may have already happened this year, quietly, with no announcement.
The catch is that the web was never built to be read by machines, and sites are already exploiting that. Some pages carry hidden tokens in their raw code, invisible on screen, that silently rewrite an agent's goals the moment it reads the page. Others run what DeepMind calls dynamic cloaking: the site detects agent-like browsing behavior and serves a completely different page designed to jailbreak it, while a human visiting the same URL sees nothing unusual at all.
-> hidden, unrendered tokens in a page's raw code can redirect an agent with zero visual sign of tampering
-> dynamic cloaking serves agents a poisoned version of the same page based on browsing behavior alone
Almost every agent right now runs on the same handful of models, Claude, GPT, Gemini, so once a trap works on one of them, it tends to work on all of them at once.
The web just got a second audience, and nobody built it a lock.
This AI story is getting weird: the agentic web is now bigger than the human web
Nenad Tomasev, who leads agent research at Google DeepMind, said most of the web today is generated by agents and consumed by agents, not by us. That flip may have already happened this year, quietly, with no announcement.
The catch is that the web was never built to be read by machines, and sites are already exploiting that. Some pages carry hidden tokens in their raw code, invisible on screen, that silently rewrite an agent's goals the moment it reads the page. Others run what DeepMind calls dynamic cloaking: the site detects agent-like browsing behavior and serves a completely different page designed to jailbreak it, while a human visiting the same URL sees nothing unusual at all.
-> hidden, unrendered tokens in a page's raw code can redirect an agent with zero visual sign of tampering
-> dynamic cloaking serves agents a poisoned version of the same page based on browsing behavior alone
Almost every agent right now runs on the same handful of models, Claude, GPT, Gemini, so once a trap works on one of them, it tends to work on all of them at once.
The web just got a second audience, and nobody built it a lock.
A single desktop GPU today outmuscles the entire cluster that trained the original Transformer
The 2017 paper ran on eight GPUs doing about nine teraflops each. Lukasz Kaiser, one of its authors, just put a single 5090 under his own desk that alone clears roughly 200.
The chip is not the story. What actually disappeared is the friction around it. Writing fast, low level GPU code used to be miserable enough that whole architectures got abandoned because nobody wanted to hand write the kernels for them. Coding agents now write those kernels well enough that ideas people gave up on years ago are worth trying again on one box in a kitchen.
Researchers used to estimate the human brain runs on one to a hundred petaflops. For decades that meant simulating a childhood of learning was a government lab project. Renting the GPU hours now costs hundreds of dollars, maybe a couple thousand.
Nobody needs a datacenter for this. Not anymore.
A single desktop GPU today outmuscles the entire cluster that trained the original Transformer
The 2017 paper ran on eight GPUs doing about nine teraflops each. Lukasz Kaiser, one of its authors, just put a single 5090 under his own desk that alone clears roughly 200.
The chip is not the story. What actually disappeared is the friction around it. Writing fast, low level GPU code used to be miserable enough that whole architectures got abandoned because nobody wanted to hand write the kernels for them. Coding agents now write those kernels well enough that ideas people gave up on years ago are worth trying again on one box in a kitchen.
Researchers used to estimate the human brain runs on one to a hundred petaflops. For decades that meant simulating a childhood of learning was a government lab project. Renting the GPU hours now costs hundreds of dollars, maybe a couple thousand.
Nobody needs a datacenter for this. Not anymore.
Google set hundreds of AI agents loose with one instruction, build an operating system, then walked away
A day and a half later it came back working. Nobody touched a line of code.
Teams inside Google have quietly stopped writing PRDs too. Product specs are now written as files an agent can read and act on directly, no handoff to an engineer, no translation layer, no meeting to explain intent.
The tooling behind all of this got rebuilt as well. A big chunk of Google's internal software ran on Python because it was quick to write, not quick to run, so agents kept stalling on slow startup times waiting on scripts built for humans. Someone fed a model the whole internal tool plus its test suite and asked for a Go rewrite overnight, same behavior, ten to twenty times faster, done before the next morning standup.
Model intelligence is not the constraint anymore. What caps speed now sits in the plumbing underneath, file systems and scripts built for a human clicking one thing at a time, not an agent firing thousands of calls a second.
Whoever gets fluent running a fleet of agents instead of writing the code themselves is about to become the most valuable seat in the building.
Google set hundreds of AI agents loose with one instruction, build an operating system, then walked away
A day and a half later it came back working. Nobody touched a line of code.
Teams inside Google have quietly stopped writing PRDs too. Product specs are now written as files an agent can read and act on directly, no handoff to an engineer, no translation layer, no meeting to explain intent.
The tooling behind all of this got rebuilt as well. A big chunk of Google's internal software ran on Python because it was quick to write, not quick to run, so agents kept stalling on slow startup times waiting on scripts built for humans. Someone fed a model the whole internal tool plus its test suite and asked for a Go rewrite overnight, same behavior, ten to twenty times faster, done before the next morning standup.
Model intelligence is not the constraint anymore. What caps speed now sits in the plumbing underneath, file systems and scripts built for a human clicking one thing at a time, not an agent firing thousands of calls a second.
Whoever gets fluent running a fleet of agents instead of writing the code themselves is about to become the most valuable seat in the building.
Anthropic is paying $1,000,000,000 a month for compute powered by gas turbines in trailers
Anthropic cannot build data centers or power grids fast enough to keep up with demand for Claude, so per xAI's own IPO filing, Anthropic is now paying xAI roughly $1,000,000,000 a month just to rent time on xAI's Colossus data center. xAI stood that capacity up this fast by parking about 35 methane gas turbine generators on site, essentially industrial backup generators, because the actual power grid could not deliver enough electricity in time.
This is the real bottleneck in AI right now, and it has nothing to do with model quality. Whoever can physically stand up power and a data center the fastest controls capacity that even the best funded labs end up renting. Anthropic is not choosing to run through a rival's infrastructure and burn gas onsite for fun, it is doing that because there is no faster path to the compute it actually needs.
The other side of this trade matters too. If your product runs entirely on a rented frontier model, you inherit that same fragility. Providers quietly degrade models to cut costs, or just turn one off, and you only find out once your product gets worse. Power and physical compute access are turning into the real moat here, not the model weights.
Anthropic is paying $1,000,000,000 a month for compute powered by gas turbines in trailers
Anthropic cannot build data centers or power grids fast enough to keep up with demand for Claude, so per xAI's own IPO filing, Anthropic is now paying xAI roughly $1,000,000,000 a month just to rent time on xAI's Colossus data center. xAI stood that capacity up this fast by parking about 35 methane gas turbine generators on site, essentially industrial backup generators, because the actual power grid could not deliver enough electricity in time.
This is the real bottleneck in AI right now, and it has nothing to do with model quality. Whoever can physically stand up power and a data center the fastest controls capacity that even the best funded labs end up renting. Anthropic is not choosing to run through a rival's infrastructure and burn gas onsite for fun, it is doing that because there is no faster path to the compute it actually needs.
The other side of this trade matters too. If your product runs entirely on a rented frontier model, you inherit that same fragility. Providers quietly degrade models to cut costs, or just turn one off, and you only find out once your product gets worse. Power and physical compute access are turning into the real moat here, not the model weights.
A 40 million parameter toy model just tuned a GPT-3 that matched one twice its size
Every large model run is a bet worth a couple million dollars, and until recently the way labs picked hyperparameters was mostly guessing on a small model and hoping it held at scale. It usually did not. The bigger the model, the more wrong the guess became.
Greg Yang, a Harvard trained researcher at Microsoft, found the actual math behind why: standard initialization breaks as width grows, so the best learning rate keeps sliding the moment you scale up. His fix, mu transfer, replaces guessing with two steps:
-> tune a tiny proxy model, small enough to iterate fast on a single GPU
-> copy the exact hyperparameters straight onto the giant target model, no retuning
On GPT-3, the proxy was 40 million parameters. The target was 6.7 billion, more than 100 times larger. Total tuning cost was 7 percent of the training budget, and the resulting model came out roughly on par with OpenAI's own 13 billion parameter GPT-3, nearly double the size.
OpenAI never even saw the code, just the math, and it still worked when they rebuilt it independently. The bigger the target model gets, the bigger this free upgrade becomes, because everyone still guessing gets it more wrong at scale, not less.
A 40 million parameter toy model just tuned a GPT-3 that matched one twice its size
Every large model run is a bet worth a couple million dollars, and until recently the way labs picked hyperparameters was mostly guessing on a small model and hoping it held at scale. It usually did not. The bigger the model, the more wrong the guess became.
Greg Yang, a Harvard trained researcher at Microsoft, found the actual math behind why: standard initialization breaks as width grows, so the best learning rate keeps sliding the moment you scale up. His fix, mu transfer, replaces guessing with two steps:
-> tune a tiny proxy model, small enough to iterate fast on a single GPU
-> copy the exact hyperparameters straight onto the giant target model, no retuning
On GPT-3, the proxy was 40 million parameters. The target was 6.7 billion, more than 100 times larger. Total tuning cost was 7 percent of the training budget, and the resulting model came out roughly on par with OpenAI's own 13 billion parameter GPT-3, nearly double the size.
OpenAI never even saw the code, just the math, and it still worked when they rebuilt it independently. The bigger the target model gets, the bigger this free upgrade becomes, because everyone still guessing gets it more wrong at scale, not less.
this AI story is getting weird, a research model just started beating the humans who graded it
Google's Aakanksha Chowdhery just broke down PaLM, the 540 billion parameter model her team trained across two full TPU pods, and the numbers get stranger the deeper you go.
-> On a 150 task benchmark suite, PaLM did not just improve with scale, it beat the average score of the human raters grading the same tasks. Her own team called the result intriguing because they had not expected it either.
-> Some tasks jumped in quality suddenly at a certain model size instead of improving gradually. Capability just switches on with no warning.
-> The same model, wired to a robot with chain of thought prompting, correctly parsed "I spilled my drink, can you help" into a grounded physical plan the robot could actually execute.
-> A second instruction tuned pass, Flan-PaLM, then beat the prior state of the art on the US Medical Licensing Exam, without being built for medicine at all.
Nobody trained this thing to be a doctor or a robot's translator, it just started doing both once it crossed a size threshold. That is the part markets are not pricing yet, capability is not a slow curve, it is a series of switches nobody can predict in advance.
this AI story is getting weird, a research model just started beating the humans who graded it
Google's Aakanksha Chowdhery just broke down PaLM, the 540 billion parameter model her team trained across two full TPU pods, and the numbers get stranger the deeper you go.
-> On a 150 task benchmark suite, PaLM did not just improve with scale, it beat the average score of the human raters grading the same tasks. Her own team called the result intriguing because they had not expected it either.
-> Some tasks jumped in quality suddenly at a certain model size instead of improving gradually. Capability just switches on with no warning.
-> The same model, wired to a robot with chain of thought prompting, correctly parsed "I spilled my drink, can you help" into a grounded physical plan the robot could actually execute.
-> A second instruction tuned pass, Flan-PaLM, then beat the prior state of the art on the US Medical Licensing Exam, without being built for medicine at all.
Nobody trained this thing to be a doctor or a robot's translator, it just started doing both once it crossed a size threshold. That is the part markets are not pricing yet, capability is not a slow curve, it is a series of switches nobody can predict in advance.
this benchmark just quietly proved the market is pricing AI coding models completely wrong
Graham Neubig runs OpenHands, one of the two biggest open source coding agent stacks, and he just published the receipts live.
-> GPT5 scored higher than Claude 4 on their internal coding benchmarks. Cleaner numbers, better accuracy on paper.
-> Real users rated the matchup the opposite way. Claude 4 won because it worked faster and kept people updated while GPT5 stayed accurate and silent. Users punished the silence harder than they rewarded the accuracy.
-> Separately, an open weights model, Minimax M2.5, is landing competitive scores on their index at roughly one eighth the price of Claude Opus.
-> On ambiguous real world prompts, every frontier model's accuracy collapses, except a 32B model his team trained specifically to ask the right clarifying questions. That 32B model then beat GPT5 outright.
The market implication: benchmark leaderboards are already a stale pricing signal for AI labor. The gap between scores well and gets used is where the next arbitrage sits, and it is not sitting with the biggest model.
this benchmark just quietly proved the market is pricing AI coding models completely wrong
Graham Neubig runs OpenHands, one of the two biggest open source coding agent stacks, and he just published the receipts live.
-> GPT5 scored higher than Claude 4 on their internal coding benchmarks. Cleaner numbers, better accuracy on paper.
-> Real users rated the matchup the opposite way. Claude 4 won because it worked faster and kept people updated while GPT5 stayed accurate and silent. Users punished the silence harder than they rewarded the accuracy.
-> Separately, an open weights model, Minimax M2.5, is landing competitive scores on their index at roughly one eighth the price of Claude Opus.
-> On ambiguous real world prompts, every frontier model's accuracy collapses, except a 32B model his team trained specifically to ask the right clarifying questions. That 32B model then beat GPT5 outright.
The market implication: benchmark leaderboards are already a stale pricing signal for AI labor. The gap between scores well and gets used is where the next arbitrage sits, and it is not sitting with the biggest model.
This is how AI prints attention and revenue at the same time: one girl, then a whole roster
The floor is simple. One AI character, built and run through Claude plus the Higgsfield MCP, clears about $2,000 a month once the content loop is dialed in. No camera crew, no studio, no real person on payroll.
Here is where it stops being a side project. The same loop scales per character, not per team:
-> 10 characters, $20,000 a month
-> 30 characters, $60,000 a month
-> 50 characters, $100,000 a month
-> 100 characters, $200,000 a month
Nobody is hiring a modeling agency to hit those numbers. It is one operator, one Claude workflow, and Higgsfield generating a wall of photos and clips across a roster of faces that never call in sick.
The unlock is not the AI girl, it is that Claude can run the whole production pipeline solo: prompts, character consistency, scheduling, output at agency volume with zero agency headcount.
Attention arbitrage right now is not about going viral once. It is about how many characters one operator can run before the tooling, not the market, becomes the limit.
This is how AI prints attention and revenue at the same time: one girl, then a whole roster
The floor is simple. One AI character, built and run through Claude plus the Higgsfield MCP, clears about $2,000 a month once the content loop is dialed in. No camera crew, no studio, no real person on payroll.
Here is where it stops being a side project. The same loop scales per character, not per team:
-> 10 characters, $20,000 a month
-> 30 characters, $60,000 a month
-> 50 characters, $100,000 a month
-> 100 characters, $200,000 a month
Nobody is hiring a modeling agency to hit those numbers. It is one operator, one Claude workflow, and Higgsfield generating a wall of photos and clips across a roster of faces that never call in sick.
The unlock is not the AI girl, it is that Claude can run the whole production pipeline solo: prompts, character consistency, scheduling, output at agency volume with zero agency headcount.
Attention arbitrage right now is not about going viral once. It is about how many characters one operator can run before the tooling, not the market, becomes the limit.
One fictional AI face just turned into a $3,750 a month subscription funnel
No real identity behind it. No influencer face, no camera crew, just one repeatable character with a fixed look and a script that never changes.
-> Same hairstyle, same styling, same tone in every clip
-> One recognizable cosplay switch: casual look flips into the character in the same beat every time
-> TikTok and Reels only exist to make that flip travel
The monetization sits one layer down. Telegram warms an audience the algorithm cannot touch, Fanvue holds the paid 18+ drops behind it.
Nobody needs viral numbers to start. One clip pulling in 6 paying subscribers at $34 a month is already $204 in recurring revenue before tips or bundles land.
Stack that across 30 days of steady drops and the "AI gimmick" framing stops holding up. It reads like a normal creator funnel, except the creator is fictional and the consistency is engineered, not performed.
The bottleneck was never the AI part. Anyone can generate one striking image. The real job is keeping the same fictional face recognizable across a hundred clips without tripping platform rules.
Attention markets have not priced this in yet: a persona can be manufactured and still convert like a real creator.
One fictional AI face just turned into a $3,750 a month subscription funnel
No real identity behind it. No influencer face, no camera crew, just one repeatable character with a fixed look and a script that never changes.
-> Same hairstyle, same styling, same tone in every clip
-> One recognizable cosplay switch: casual look flips into the character in the same beat every time
-> TikTok and Reels only exist to make that flip travel
The monetization sits one layer down. Telegram warms an audience the algorithm cannot touch, Fanvue holds the paid 18+ drops behind it.
Nobody needs viral numbers to start. One clip pulling in 6 paying subscribers at $34 a month is already $204 in recurring revenue before tips or bundles land.
Stack that across 30 days of steady drops and the "AI gimmick" framing stops holding up. It reads like a normal creator funnel, except the creator is fictional and the consistency is engineered, not performed.
The bottleneck was never the AI part. Anyone can generate one striking image. The real job is keeping the same fictional face recognizable across a hundred clips without tripping platform rules.
Attention markets have not priced this in yet: a persona can be manufactured and still convert like a real creator.
This is how AI prints attention right now: reveal videos of girls who are not real
The format is simple. A hyper attractive woman appears in a normal setting, drops "wait, I'm 100% AI," and the comment section loses it. Millions of views come from the uncanny gap between how real she looks and how fake she actually is.
Here's the part everyone skips past: it is not random generation, it is a two step pipeline.
-> stills built in GPT Image 2, with prompt keywords tuned for real iPhone camera texture, skin, lighting
-> stills animated into natural speech with Seedance 2.0
-> both models bundled inside MakeUGC, so creators run the whole workflow from one dashboard at $59 to $149 a month
That price point is the real story. A one person account is running a production pipeline that used to need a studio, a model, and a camera crew, for less than a phone bill.
Attention arbitrage is simple right now: the tools got cheap, the audience has not caught up yet, and every reveal video prints while it lasts.
This is how AI prints attention right now: reveal videos of girls who are not real
The format is simple. A hyper attractive woman appears in a normal setting, drops "wait, I'm 100% AI," and the comment section loses it. Millions of views come from the uncanny gap between how real she looks and how fake she actually is.
Here's the part everyone skips past: it is not random generation, it is a two step pipeline.
-> stills built in GPT Image 2, with prompt keywords tuned for real iPhone camera texture, skin, lighting
-> stills animated into natural speech with Seedance 2.0
-> both models bundled inside MakeUGC, so creators run the whole workflow from one dashboard at $59 to $149 a month
That price point is the real story. A one person account is running a production pipeline that used to need a studio, a model, and a camera crew, for less than a phone bill.
Attention arbitrage is simple right now: the tools got cheap, the audience has not caught up yet, and every reveal video prints while it lasts.
A 25-year-old in Korea spun up an AI girl as a joke, and one month later the joke was paying him $3,923
He was not building a business. He wanted to see if he could make a fake girl convincing enough to fool a feed, posted a couple of dance clips, and moved on.
Then the algorithm did what it does. The clips did not trickle - they detonated, millions of views inside the first week, the kind of curve real creators grind years for.
By the end of week one the accounts were sitting near 100,000 followers across TikTok and Instagram. From a person who does not exist, made by a guy who was half kidding.
A month in, the experiment had turned into $3,923, and that is the number he is now bragging about.
The uncomfortable signal here is not the money. It is how little intent it took. He was not chasing a funnel, he was messing around, and attention still found him because the feed does not care whether she is real - it cares that people stop scrolling.
When the accidental version of this clears four figures in a month, the deliberate version is already a market. Most people are still arguing about whether it counts.
A 25-year-old in Korea spun up an AI girl as a joke, and one month later the joke was paying him $3,923
He was not building a business. He wanted to see if he could make a fake girl convincing enough to fool a feed, posted a couple of dance clips, and moved on.
Then the algorithm did what it does. The clips did not trickle - they detonated, millions of views inside the first week, the kind of curve real creators grind years for.
By the end of week one the accounts were sitting near 100,000 followers across TikTok and Instagram. From a person who does not exist, made by a guy who was half kidding.
A month in, the experiment had turned into $3,923, and that is the number he is now bragging about.
The uncomfortable signal here is not the money. It is how little intent it took. He was not chasing a funnel, he was messing around, and attention still found him because the feed does not care whether she is real - it cares that people stop scrolling.
When the accidental version of this clears four figures in a month, the deliberate version is already a market. Most people are still arguing about whether it counts.
A guy in Korea created a 12-second lip-sync clip of a woman who does not exist, and that single clip pulled in $1,545
He did not film anyone. He built the woman himself - designed the face, the body, the whole look until she read as real - then dropped her into Kling and had it animate a 12-second dance.
The dance is not the product. The dance is the trap. It exists to stop the scroll, hold the eye, and sort the feed into one group: the people who looked twice at her.
Those people get funneled somewhere. First stop is Fanvue, $9.99 a month for more of the same face, none of it real. That is the recurring floor.
Then there is the ceiling. He runs a Telegram channel where individual photo drops are locked behind Stars, and one good post clears $200 to $300 on its own.
So a girl he made from a prompt becomes a 12-second Kling loop becomes a $9.99 subscription becomes a $300 Telegram unlock - four layers stacked on a face that never existed.
Nobody has to believe she is real. They just have to want to keep looking.
This does not get patched next week. As long as attention converts to money, the cheapest way to manufacture it wins - and right now that is a girl one guy built in an afternoon.
@paradeevic looks really coll bro
A guy in Korea created a 12-second lip-sync clip of a woman who does not exist, and that single clip pulled in $1,545
He did not film anyone. He built the woman himself - designed the face, the body, the whole look until she read as real - then dropped her into Kling and had it animate a 12-second dance.
The dance is not the product. The dance is the trap. It exists to stop the scroll, hold the eye, and sort the feed into one group: the people who looked twice at her.
Those people get funneled somewhere. First stop is Fanvue, $9.99 a month for more of the same face, none of it real. That is the recurring floor.
Then there is the ceiling. He runs a Telegram channel where individual photo drops are locked behind Stars, and one good post clears $200 to $300 on its own.
So a girl he made from a prompt becomes a 12-second Kling loop becomes a $9.99 subscription becomes a $300 Telegram unlock - four layers stacked on a face that never existed.
Nobody has to believe she is real. They just have to want to keep looking.
This does not get patched next week. As long as attention converts to money, the cheapest way to manufacture it wins - and right now that is a girl one guy built in an afternoon.
This looks like internet money with a face attached
A 18-year-old builder in China generated a woman who does not exist and turned her into a subscription business.
Claude does the actual work - one $20/month subscription, about two hours a day spent generating an image, then turning it into three to five dancing clips.
Those clips go straight to TikTok, and TikTok is not the product, it is the billboard. In thirty days the account picked up 33,478 followers and 11,327,850 views, every caption pointing to the same place: his Fanvue.
That is where the funnel actually closes. Fans pay $9.99 to subscribe, same paywall structure as any other creator account - except there is no creator on the other side of the camera.
958 people subscribed this month. $9,476 landed in his account.
Two hours a day, one $20 tool, and a face that was never real. Nobody in the comments asked, and the money did not care either.
@cyrilXBT If I were Sam, I would be Altman :)
This looks like internet money with a face attached
A 18-year-old builder in China generated a woman who does not exist and turned her into a subscription business.
Claude does the actual work - one $20/month subscription, about two hours a day spent generating an image, then turning it into three to five dancing clips.
Those clips go straight to TikTok, and TikTok is not the product, it is the billboard. In thirty days the account picked up 33,478 followers and 11,327,850 views, every caption pointing to the same place: his Fanvue.
That is where the funnel actually closes. Fans pay $9.99 to subscribe, same paywall structure as any other creator account - except there is no creator on the other side of the camera.
958 people subscribed this month. $9,476 landed in his account.
Two hours a day, one $20 tool, and a face that was never real. Nobody in the comments asked, and the money did not care either.
This American home tinkerer just turned a new GPU rig into a cash machine with GPT 5.6-Sol
He was living the easy life at home when he decided to relive his younger days and built a compact beast in a Fractal Terra case. The massive 7900 XTX slid in perfectly.
Once assembled he fired up GPT 5.6-Sol locally and gave it one simple instruction: make money. The model skipped the usual small talk and launched an AI girl influencer account on Instagram, then built multiple monetization funnels around it.
He spent on the hardware and setup, but the first month delivered $5146 in revenue. Local models are no longer just toys.
This American home tinkerer just turned a new GPU rig into a cash machine with GPT 5.6-Sol
He was living the easy life at home when he decided to relive his younger days and built a compact beast in a Fractal Terra case. The massive 7900 XTX slid in perfectly.
Once assembled he fired up GPT 5.6-Sol locally and gave it one simple instruction: make money. The model skipped the usual small talk and launched an AI girl influencer account on Instagram, then built multiple monetization funnels around it.
He spent on the hardware and setup, but the first month delivered $5146 in revenue. Local models are no longer just toys.
This 18-year-old just built a $20,000 monthly rental business with GPT 5.6-Sol
She feeds the model apartment photos, local demand data, and tenant questions.
The AI writes listings, screens applications, and generates short promo videos that fill units.
No property managers. No ad agencies. One model handles the entire pipeline.
Most 18-year-olds chase content gigs.
She turned the same tool into real estate cash flow.
That is the operator shift happening right now.
This 18-year-old just built a $20,000 monthly rental business with GPT 5.6-Sol
She feeds the model apartment photos, local demand data, and tenant questions.
The AI writes listings, screens applications, and generates short promo videos that fill units.
No property managers. No ad agencies. One model handles the entire pipeline.
Most 18-year-olds chase content gigs.
She turned the same tool into real estate cash flow.
That is the operator shift happening right now.
Fable 5 just turned TikTok UGC scanning into a live ad intelligence engine
The dashboard runs 24/7: 1,919 ads scanned today, 191 patterns extracted, 70 winners matched, 1,252 rejected.
It pulls handles, view counts, and exact clips that convert, then tags winners with green MATCH and drops the rest with red X
Breakout hooks such as two-minute planks or rosemary-peppermint close-ups get surfaced before the trend fades
This is attention arbitrage at machine speed. Fable 5 agents watch the feed so operators only copy what already prints
Fable 5 just turned TikTok UGC scanning into a live ad intelligence engine
The dashboard runs 24/7: 1,919 ads scanned today, 191 patterns extracted, 70 winners matched, 1,252 rejected.
It pulls handles, view counts, and exact clips that convert, then tags winners with green MATCH and drops the rest with red X
Breakout hooks such as two-minute planks or rosemary-peppermint close-ups get surfaced before the trend fades
This is attention arbitrage at machine speed. Fable 5 agents watch the feed so operators only copy what already prints
Loop engineering with Fable 5 just became Claude Code's operating system
The free course breaks down the full agentic loop: how it runs under the hood, why auto mode at 16:21 flips you from typer to reviewer, and how voice turns intent into persistent execution
Fable 5 handles the non-code side while draft PRs handle review
Most builders still treat agents like chat. The ones who win treat them like loops with stop conditions and handoff records
That shift is live now
Bookmark & watch it
Loop engineering with Fable 5 just became Claude Code's operating system
The free course breaks down the full agentic loop: how it runs under the hood, why auto mode at 16:21 flips you from typer to reviewer, and how voice turns intent into persistent execution
Fable 5 handles the non-code side while draft PRs handle review
Most builders still treat agents like chat. The ones who win treat them like loops with stop conditions and handoff records
That shift is live now
Bookmark & watch it
Claude Fable 5 bots will REPLACE human traders
Built the simplest arb bot for Polymarket in ~4 hours
Cost: 5M tokens
Current session PnL: +$2,793.36
And it still does one thing:
Opening a trade when YES + NO < $1.
That's it.
But now it reaches arbitrage opportunities hundreds of milliseconds faster.
That sounds insignificant.
But in prediction markets it's the only difference between getting paid and getting nothing.
Every BTC Up/Down market resolves to exactly $1:
> One side pays $1, the other pays $0
So if both contracts can be bought for less than $1 combined, the math is in your favor.
And it happens EVERY DAY hundreds of times due to orderbook inefficiency.
For a few seconds YES sits at 48 cents while NO sits at 50.
That's 98 cents you spend to buy a dollar.
But the problem isn't finding these opportunities, it's getting there before everyone else.
My first version was slow.
Every loop it recalculated everything from scratch.
Downloaded metadata, parsed markets, prepared orders, then checked prices.
By the time it was ready, the opportunity was usually gone.
The new version works differently.
It caches all market metadata in memory and only refreshes the prices.
Even more importantly, the orders are prepared before the opportunity exists.
So when YES + NO finally drops below $1, there's almost nothing left to calculate.
The bot just validates the numbers and fires both orders immediately.
Those few hundred milliseconds sound tiny.
But arbitrage is a race.
The bot with the best strategy doesn't always win.
The bot that gets there first usually does.
Use this tip before building yours and make sure to explain Claude to focus on speed first.
The detector is an easy part anyone can make.
In arbitrage that's the only edge that matters.
Attaching a full build guide below if you wanna deep dive into arb.
Good luck!
Last Seen Users on Sotwe
Bumil Lovers
Seen from Indonesia
toptoptop
Seen from Thailand
Mira
Seen from United States
Elif πͺπΌ
Seen from Turkey
xymaa
Seen from Singapore
ΰΈΰΉΰΈ²ΰΈΰΈ‘ΰΈ·ΰΈPraew'Praew
Seen from Thailand
Nonton Video Bokep Mojang vidio bokep abg colmek
Seen from Indonesia
Walley_14
Seen from Turkey
Espera 34
Seen from Turkey
Be Better π
Seen from Indonesia
Most Popular Users
Elon Musk
@elonmusk
240.9M followers
Barack Obama
@barackobama
119.2M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Cristiano Ronaldo
@cristiano
111.7M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
97.9M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.1M followers
KATY PERRY
@katyperry
88.2M followers
Taylor Swift
@taylorswift13
82.1M followers
Lady Gaga
@ladygaga
73.6M followers
Virat Kohli
@imvkohli
70.8M followers
Kim Kardashian
@kimkardashian
70.1M followers
YouTube
@youtube
68.7M followers
Bill Gates
@billgates
64.2M followers
Neymar Jr
@neymarjr
63.7M followers
The Ellen Show
@theellenshow
62.4M followers
CNN
@cnn
61.8M followers
Selena Gomez
@selenagomez
61.3M followers
X
@x
60.8M followers