这条说得太对了
我这边正好想弥补英文社区的这个空缺 - 不绕弯子,直接放数据:团队背景、资金费率数学、积分项目、真实交易数据,有邀请码也直接给
主要关注这几个:QFEX、BULK、RiseX、TxFlow、Variational - 数据优先,炒作靠后
很高兴认识中文社区的朋友们,以后也多交流 🤝
又来跟各位中国朋友打个招呼。我是 Muar,之后会不定期发中文内容。我挺喜欢中国,也喜欢跟这边的人交流。
先简单晒一下自己做过的一些 call,方便你们判断这个号值不值得看。
三月的时候我写过,Polymarket 上 Variational 500M FDV 的 Yes 盘基本就是白送。后来这笔大概 4 倍:
teapawt.net/muarmemuar/status/2084…
Robinhood Chain 上 meme 还没全面起势之前,我做了 UNI 和 ARB 的多单,当时也写过为什么这么做。这两笔加起来大概赚了 9 万人民币:
teapawt.net/muarmemuar/status/2101…
$LIT 还在 1.6B FDV 的时候,我给过 5B 的目标,后来差不多 3 倍:
teapawt.net/muarmemuar/status/2096…
七月初写过,AI 股票该空、crypto 该多。结果在这里:
teapawt.net/muarmemuar/status/2082…
StonkFun 生态还没被 Traction 起来的时候,我就说过值得看一眼。后面那些币大概走出了 3 到 7 倍:
teapawt.net/muarmemuar/status/2098…
TreadFi 还是早期的时候我也推过。我一个朋友在上面赚了大约 60 万人民币,后来因为这个 call 分给了我 3.2 万:
teapawt.net/muarmemuar/status/2100…
我还写过哪些 perp DEX 我会直接 skip。后面其中一部分关停了,有的被盗了。这个 call 其实帮不少人省了时间和钱:
teapawt.net/muarmemuar/status/2080…
Variational 还没被全网喊起来的时候,我就是比较早开始推这个 perp DEX 的 KOL 之一。现在我在上面大概有 6000 多积分。
我不觉得自己是什么 alpha caller,主要就是想认识一下各位。中国这边的 KOL 我一直看得挺有意思。
眼下我更盯那些还没有积分计划、用户注意力也还没过来的 perp DEX。我比较确定的一点是:Variational 空投落地之后,圈子会把目光转到这批盘上。这条建议最好先存一下。
原因很简单。一旦积分计划上线,积分本身就会被定价、被炒,用户会涌进来,supply 会被空投农民摊薄,后面再农就难了。没有积分计划的时候,没什么可炒的,也没多少人盯着,但这段早期交易往往会被追溯计入。
Hyperliquid 在 5800 万积分里,给了大约 1.1 万名早期用户 450 万积分,占总量 7.7%。积分计划上线前 9 个月的交易都被算进去了。
RiseX 在 1200 万积分里,给了不到 4000 名早期用户 160 万积分,占总量 13.3%。上线前 3 个月的交易被计入。
Variational 在 950 万积分里,给了不到 9000 名早期用户 300 万积分,占总量 31%。上线前 11 个月的交易被计入。
据我了解,下面这几个 perp DEX 大概会拿出总积分供给的 15%–20% 分给早期用户。这个比例很关键。积分计划大概率会在十月到十一月上线。
@QFEX
我挺确定,Variational 一 TGE,CT 就会涌到 QFEX,然后开始叫它 “Variational 2.0”。理由无非是 RWA 叙事,再加上团队背景。
QFEX 今年早些时候融了 950 万刀。有意思的是,创始人是在剑桥认识的,后来一起把 QFEX 做起来。Variational 也差不多:创始人在哥伦比亚大学认识,然后再把项目做出来。
你们可以记一下:后面会有人把 QFEX 叫成 “Variational 2.0”。但两边的成交模型完全不是一回事。Variational 是 RFQ,QFEX 是 CLOB。
邀请码(交易手续费 9 折):
7G6HRJ7K, XQTE418N, PGW3SKPV, 3M40WQQF, 6YFDESMD, 13FC90CV, AZWKB0PK, W35DAFQJ, J4MBY5B2, W5DVVZ15, 6Z0H1JM5, 2PVKPAJ3, NEVHBQ6B, 08NQGJSC, 585PS49A, AMHVE8K5, V8DYFG1V, SPEETBWS, MX44PVAN
@TxFlow_L1
BloFin 创始人现在也进 perp DEX 了,做的就是 TxFlow。BloFin 本身已经是一家比较成熟的 CEX。
这其实成了 CEX 的一条常规路径:Bybit 有 Apex,Binance 有 Aster,BloFin 则是一边做自己的链,一边做 TxFlow 这个 perp DEX。
用码 MUAR,手续费打 95 折。
@get_truenorth
这是我看得比较顺眼的项目之一,perp DEX 叙事叠 AI。
他们已经拿了 300 万刀 pre-seed。其中一个创始人以前做过 WOO X,那是很早期的一批永续合约交易所,所以团队其实懂交易员要什么。另一个创始人之前在中国两家头部 AI 芯片公司待过。
我觉得这套班底,是有机会用 AI 把 perp DEX 往前推一截的。
TrueNorth 给自己的定位是 AI-native 的 agentic brokerage,基本上就是第一家 “agentic broker”。它提供个性化的 AI 交易 agent,能帮忙做盘面分析、下单执行、风控,以及交易后的复盘。
等他们把还没发币的 perp DEX 接进去之后,理论上可以用 AI agent 在多个协议上自动跑 delta-neutral,同时被动农好几个盘的积分。
用码 MUAR,手续费 9 折,积分再加 15% boost。
@entropyIO
这是做 Hyperliquid HIP-3 上 RWA 的 perp DEX,刚刚拿了 Ribbit Capital 4500 万刀。
Ribbit 就是早年投 Coinbase、Robinhood 的那家。现在他们把注押到 perp DEX 上了:之前投过 Lighter,现在是 Entropy。
对我来说,听到这些已经够了,早期就该开始在这个盘上做交易。
手续费 12.5% off 的码:muarych
趁现在还早。再晚一点积分会更难农,潜在利润也会被摊薄。
I've spent enough hours across these perp DEXs to have real opinions now, not just "everything's a 10x" hopium. here's where I actually rank them and why
S-Tier
@QFEX - ex-Citadel/Tower Research quants, 24/7 leveraged trading on real US equities/commodities/FX, no broker. this is the closest thing to actual prime brokerage that's fully onchain
@TxFlow_L1 - first channel on a shared-liquidity L1, fully onchain CLOB with zero bridging. funding swings the widest of anything on this board, which is exactly why it's up here
A-Tier
@variational_io - $50M raise led by Dragonfly, team out of Google/Meta/Virtu/Jane Street, $200B+ processed. bringing commodities (gold, silver, oil) onchain through an RFQ model most people still don't understand
@bulktrade - mainnet live, survived an ADL event during the 89k BTC wick and made every wallet whole +10% same day. 5-20ms execution, portfolio-level risk engine, no single sequencer
@risextrade - fully onchain orderbook, matching runs inside the EVM not off it. stock perps trading 24/7 alongside crypto, $2B volume in 60 days with almost nobody talking about it yet
B-Tier
@hibachi_xyz - built by Hashflow's founder, $10B+ cumulative volume, ZK-CLOB doing ~6ms matching with actual on-chain privacy on positions and balances. $HEAT ticker confirmed, Playoffs season live at 1M points/week. solid infra, just hasn't separated itself from the pack yet the way the tiers above have
which one's misplaced?
txflow is moving faster than most people are clocking
launched mainnet march 31. first app: CLOB perp dex invite-only still. second app dropped jul 14: Probly - a prediction market with 172 live markets, 7k events, rolling markets as short as 5 minutes
that's perps and predicts on the same chain, shared liquidity, no bridging
numbers from defillama: $1.28b in 30d perp volume, $14.1m OI, $12.8m TVL, $1.2m fees annualized - and the chart is still clearly going up, not sideways
one thing buried in their comms: they've said "zero insider token allocation" publicly. that phrase doesn't get written without a token being planned
still invite-only, 250k TPS, Matt Hu ex-BloFin is building this out of his own pocket
pulled up risex's dune dashboard, growth curve is a good tell
$3.34B cumulative volume, from basically zero in april. daily volume sat around $50-60m through may/june then broke $100m/day this month, right around when gold/silver went live
8m+ cumulative trades and the curve is still bending up, not flattening
what stands out more than the top-line number: liquidation fees are only $90.7k out of $895k total fees. that's not a venue getting torched by liquidations, that's actual two-sided flow
market split: BTC 56%, ETH 17%, HYPE 13%. non-majors pulling real share this early is a good sign
dune.com/rise_foundation/ris… if you want to dig in yourself
A Fortune 500 CTO flew to Chennai to inspect the "data center" hosting his company's AI workloads
He walked into a bedroom with a ceiling fan and beige curtains
Arjun is 29. He has been running the CTO's staging environment for 7 months. They pay him $8,400/month. They think he has a team
He does not
Along the back wall: three tower PCs forming his CPU cluster. On the front desk: a $210 Mac Mini running Kubernetes across the whole rack. A wicker basket in the corner
The CTO stood in the doorway for 40 seconds without saying a word
Voltage stabilizer with a red LED humming under load. Two 2TB enterprise SSDs. A 10 Gbps switch. Two managed switches for the internal network. Router, firewall, and storage stack outlined in green tape at the top of the rack
Build cost: $17,800
Cloud vendor quote for the same compute:
$192,000/year
Arjun did not design any of it. He asked Claude one question: "Design me a home rack that can host 10 client staging environments and 4 local AI training runs at the same time"
Claude spec'd the parts, priced the switches, wrote the Kubernetes manifests, wrote the network config, and wrote the failover plan for when the ceiling fan trips the breaker
He also runs 8 always-on Claude agents that scrape client dashboards and email weekly reports on Sunday nights. The CTO has been reading those reports for 4 months thinking a research team wrote them
The CTO finally spoke: "Where is the rest of the operation?"
Arjun pointed at the Mac Mini
The CTO did not cancel the contract. He tripled it. $28,000/month, on-prem exclusive, 18-month clause blocking Arjun from taking on more American clients.
He signed it at his kitchen table
His parents think he does freelance IT support. They stopped asking
Electricity bill last month: $113
The internet said serious AI work required a data center, a devops team, and a $2M cloud budget
He is running enterprise AI infrastructure from a bedroom, and the enterprise flew across the world to hand him a bigger check
The internet was renting when it should have been building
A 18-year-old built a 1-kilometer RC controller for $23 with Claude Opus 4.8
The commercial version costs $180
His name is Vinh. High school in Da Nang. He opened Claude on his phone and asked one question: "Can I build a 1km RC controller with an ESP32 and cheap radio modules?"
Claude said yes and wrote the whole build
Hand-soldered perfboard, 15 cm wide. Two joysticks. ESP32 in the middle with a red power LED. NRF24L01 radio standing vertical with a 2 cm antenna. Beside it, a receiver PCB with a servo spinning a white prop as Vinh moves the joystick
The full $23 Shopee list Claude generated: ESP32 $6.40, two NRF24L01+PA+LNA modules $8.80, two joysticks $2.20, perfboard $1, servo $2.60, wires $2
Claude wrote 340 lines of Arduino C++ in one session. Transmitter reads the joysticks and sends a 16-byte payload 50 times per second over 2.4 GHz. Receiver unpacks it, applies a lost-signal failsafe, drives the servo
Anti-jitter smoothing. Auto-reconnect. Channel hopping to avoid his neighbor's Wi-Fi
Vinh does not know what half of that means. He flashed both boards and the servo moved
He walked to the end of his street with his brother holding the receiver. Signal held for 1.1 km. The $180 commercial version quotes 1.0
The build hit a Vietnamese electronics forum. 40,000 views in 4 days. He started selling hand-soldered receivers for $19 each. 34 shipped so far
His physics teacher asked who wrote the firmware. Vinh opened the Claude conversation on his phone. The teacher scrolled it in silence
The internet said building custom RF hardware needed an EE degree
Vinh built it during lunch break with a model that explained every solder join
A 19-year-old turned his phone into the controller for a real drone in 4 hours with Claude
The drone was in his hand. The simulator was on his MacBook. The phone bridged them
His name is Kai. The drone is a $99 DJI Tello sitting on his desk. The MacBook is running a custom 3D drone simulator with a green wireframe grid. The phone is showing a virtual joystick interface he built that morning
The simulator on screen shows the drone hovering above the grid. Kai's thumb is on his phone. The drone on the MacBook tilts forward. The physical drone in his other hand spins its props at the same angle
One input. Two drones move. Sim and real, synced.
He built all three pieces with Claude Opus 4.8 in one session. He typed three prompts. Total
"Build me a 3D drone simulator in the browser. Use Three.js. Quadcopter physics, gravity, drag, motor RPM per axis. Green wireframe environment."
"Now add a websocket server so a phone can control the drone in this sim. Phone shows a joystick UI. The drone in the browser reacts in real time."
"Now make the same websocket also send the same control values to a real drone over the Tello SDK. Make the sim and the real drone fly in parallel from the same input."
Claude wrote the Three.js sim. The websocket server. The mobile joystick UI. The UDP bridge to the physical drone. The drift correction loop. All of it
Kai did not write a line of code. Claude told him which files to paste into
4 hours from "I wonder if I could" to a phone flying a real drone through a wireframe world
The simulator alone would have been a $4,000 freelance project last year. The control bridge would have been another $6,000. Most senior devs would have quoted 3 weeks
Kai shipped it the same Saturday
He posted a 28-second clip to TikTok that night
1.9M views in 36 hours
4 drone hardware startups in his DMs by morning. Two offered to fly him out. One offered him equity
He is 19. He is still in college. He took the meeting from his dorm room with the same drone hovering behind him
His CS professor had told the class that semester that AI would "augment programmers, not replace the discipline."
Kai built more shippable software in one Saturday than he had built in 3 semesters of coursework
The internet said building real-world hardware control needed embedded engineers and months of firmware work
He did it on a school laptop with a model that wrote every line
A coffee shop in Seoul made $11,400 more in revenue last month after pointing an AI at its own ceiling
The owner did nothing. The AI did everything
The software is called NeuroSpot. One $290 CCTV dome on the ceiling. $340/month for the AI subscription
Nothing else
The AI is reading the room in real time. Green boxes around every customer. Floating timers above each one: 25 min, 5 min, 1 hour 15 min, 12 min. In the top-right corner, the AI has identified the baristas by face and labeled them live. Olga, 3 cups. Anna, 20 cups. A blue trail behind Anna shows every step she's taken in the last 4 hours
No one programmed those names. The AI learned them from the schedule and the camera in week one
The staff signed off on the face recognition when he installed it. They had one condition: they wanted to see the dashboard too. The AC vent fix was the first thing they cared about. They had been complaining about it for months
The dashboard flagged three insights the owner could never have seen on his own
One. The AI noticed customers were avoiding the back wall. Average sit-time there was 12 minutes vs 47 minutes by the window. The AI cross-referenced the timestamps with the kitchen log and found the AC vent was dripping onto table 7. The owner fixed the vent. Back-wall revenue went up 31%
Two. The AI showed Olga served 3 cups while Anna served 20 in the same shift. Same speed. Same skill. But the AI's movement map showed Olga spent 41 minutes walking to the storage shelf on the wrong side of the bar. Owner moved the shelf. Olga's output tripled
Three. The AI flagged 14 customers who sat over 90 minutes on one drink, then never came back. Regulars who would have ordered three drinks couldn't get a seat at peak. He posted a polite 90-minute limit on weekends. Weekend revenue went up 22%
The owner did not pull these insights. The AI did. He just opened the dashboard on Sunday and read what the model had already figured out
The internet said AI in physical stores would be a surveillance toy
This one is a $340/month consultant that never sleeps, never asks for equity, and just paid for a year of itself in 6 days
Every cafe running NeuroSpot is, without realizing it, recording the world's largest training dataset of humans making espresso
The next robot to pull a shot will be trained on Anna's wrist
She's already negotiating a percentage
A 24-year-old in Singapore sold the same people-counting tool to 14 shopping malls for $9,800 each
He built it in one Saturday with Claude
His name is Wei. He had never written a line of C# in his life. He opened Claude on a Saturday morning, typed one prompt, and watched the entire app appear on his screen by lunch
"Build me a Windows app that watches a camera feed, draws a line down the middle, and counts how many people walk past it in each direction."
Claude wrote 340 lines. Picked the libraries. Wired the camera. Painted the interface. Wei did not touch a single function
That was the product
Pause at 0:11. The screen looks like something a bank used in 2004. Green title bar reading "PEOPLE COUNTING." Two video feeds with a blue line between them. OUT in red. IN in green. A woman with a red shopping cart crosses left. OUT clicks. Two more walk into the mall. IN clicks twice
That's it
Then Wei did the part nobody else does. He sold it
He found a shopping mall in his city losing customers to a competitor across the road. He walked into the manager's office with the WinForms window on his laptop and asked one question
"Do you know how many people walked past your storefront yesterday?"
The manager did not know
Wei asked the next question. "Do you know how many of them walked in?"
The manager did not know that either
Wei told him: "Without those two numbers, you cannot tell if your new window display is working. You cannot tell if your weekend foot traffic justifies a third security guard. You cannot tell if the ad you ran on the bus stop outside brought in one person or two hundred. You are spending $40,000 a month on rent and marketing, and you have no idea what's actually pulling people through the front door."
He charged $9,800 for the install. $400/month for support per location
The mall paid the same day
In the contract he called it an "AI-powered foot traffic intelligence platform." It is the Claude-built script from Saturday with the mall's logo pasted into the title bar in Microsoft Paint
Within two weeks the manager called him back. The mall across the road had heard about it. They wanted one too. Then another in the next district. Then a chain with 6 locations
Wei did not write a sales pitch. He let the first manager do the selling
14 malls in 7 months. $137,200 in setup fees
$5,600/month in recurring support
He changed three config lines per install. Camera IP. Store ID. Where the blue line sits. Then he zipped the folder, emailed it, and invoiced
One mall asked for a dashboard. He told them dashboards were a separate product. $4,400 extra. They paid. The dashboard was Claude wiring the same Postgres table to a Google Data Studio page in 40 minutes
His mom thinks he sells security software to corporations.
He does. The corporation is whoever signs the invoice. The software counts
The internet said the AI gold rush would be won by frontier labs and billion-dollar tooling startups
Wei is winning it by asking a shopping mall manager a question the manager cannot answer
A 22-year-old in Naples built a $14,500 restaurant website at 9:09AM in his pajamas
The pizzeria does not know he is not a studio
His name is Matteo. The restaurant is BRACE, a wood-fired Napoletana pizza place three streets from his apartment. They have one location, two ovens, and a 47-year-old owner who has never opened a website
The owner walked into Matteo's coffee shop two weeks ago and asked if "anyone knows someone who builds the modern ones." Matteo said he did. He quoted €13,400, around $14,500. The owner paid 50% on the spot in cash.
Matteo went home and opened Claude
Pause at 0:09. The hero is a top-down shot of a single Margherita on a black background. BRACE. in display serif. "// STONE-FIRED" underneath with a slash separator that no template would have chosen. Sub-tag: "NAPOLETANA · CRAFT · 03.01." Nav: Menu, Build, Story, Reserve. Bottom-left counter: 00:OPENING. Bottom-right: a tiny line graph that does nothing but look expensive
Section two: the pizza explodes. Ingredients suspended mid-air above a melted-cheese base. "BRACE // STONE-FIRED" in 280pt. The kind of art direction agencies in Milan charge €40,000 for
Section three: Signature Pizzas. Three cards on wooden boards. Margherita Verace €18. Prosciutto e Burrata €16. Tartufo Nero €24
Section four: Build Your Brace. An interactive pizza configurator. Pick a size. Pick toppings. Live preview updates as you click. Running total runs in real time
Section five: Reserve The Table. Calendar grid. Time slots
Note field. Functional booking flow that posts to a Google Sheet his cousin set up in four minutes
Five sections. One night. Zero designers
He used Claude Sonnet 4.6 with three prompts
"Brand site for a wood-fired Napoletana pizzeria. Dark mode. Editorial. Agency-grade. Reference: Pentagram for typography, Order for layout tension."
"Add an interactive pizza configurator. Toppings update price. Smooth rotation on the preview."
"Reservation system. Calendar plus time slots. POST to Google Sheets webhook."
The owner came back four days later. He brought his wife. They sat at the counter looking at the laptop in silence for six minutes
The wife said one sentence in Italian. "Sembra che siamo già a New York."
It looks like we are already in New York
Matteo collected the remaining €6,700 the next morning.
His mother thinks he found a part-time job at a tech company
She refreshed the BRACE site eleven times last week. She asked her son who took the photos. He told her Claude generated them
She asked who Claude was
The internet said restaurants would always need agencies because they need taste
Matteo sold them taste from his bedroom three streets away
A 21-year-old in Toronto sold a $7,200 brand site to a ceramics studio that thought she ran a 4-person agency
She is one person. The agency is Claude
The studio's name is Oryzo. They make a single product: a terracotta mug warmer the size of a coaster. They had a logo, three product photos, and a Squarespace that loaded in 4.8 seconds
She quoted them $7,200 for a full rebrand and site. They paid 50% upfront. She opened Claude that night
Pause at 0:14. The hero on screen is what she shipped. ORYZO in display serif, 240pt. Tagline underneath: "Made for mugs. Built for tables." The product photo is the terracotta dish on a self-healing cutting mat with a sharpened pencil and a snap-knife next to it. Studio aesthetic. Award-show energy
She did not design a single pixel
She fed Claude the product photos, the brand brief, and one sentence: "Make this feel like it won a Cannes Lion before it shipped a unit."
Claude wrote the copy. Claude generated the layout in HTML and Tailwind. Claude picked the typography pairing (Editorial New display + Inter body). Claude wrote three full landing sections with different art direction in each. Same brand voice across all of them
She made one human edit. She changed a comma to a period
The studio's founder messaged her at 11:47pm the night she delivered. "this is better than what we paid $34,000 for last year. who else is on your team?"
She told him the truth. "Just me and the models I use."
He sent the second 50% an hour later
She has no team. She has Claude Sonnet 4.6, a Ryzen laptop, and a Notion doc with 14 brand briefs she already sent quotes on
3 more accepted this week. Total pipeline: $26,400
Her dad thinks she is freelancing on the side while applying for full-time jobs
She declined the last 4 interviews because the salary would not cover one Oryzo project
The internet said design agencies were safe because clients pay for taste
She is selling taste at agency rates without an agency
The internet was wrong about who owns the taste
A 26-year-old in his Seoul apartment ships features while he sleeps
Four AI agents do the work. None of them know each other
He calls it his dev team. The team is one MacBook, four Claude instances, and a shared folder named .pipeline.
The Planner. The Coder. The Tester. The Reviewer
The Planner takes a one-line feature request and writes a real engineering spec. File paths. Function signatures. Edge cases. Opus 4.8, because the quality of the plan sets the ceiling for everything downstream
The Coder reads the spec and builds it. Sonnet 4.6. Lower cost. The path is already laid out
The Tester reads what the Coder shipped and writes tests. Happy path. Edge cases. Failure modes
The Reviewer is read-only. It cannot touch a single file. It reads the spec, diffs the branch against main, and gives a verdict before anything merges
Pause at 0:52. The Reviewer's actual output is on screen. SPEC: RESET-PASSWORD.MD. Token expiry 15 min TTL single-use enforced. Bcrypt cost factor exactly 12. Invalidate all sessions on success. Return typed error codes TOKEN_EXPIRED, VALIDATION_ERROR. No plaintext password ever written to logs
The Reviewer responds: "Spec ingested. I have 5 requirements to verify. Running git diff main...feature/pw-reset now. I will not write or modify any file."
It read its own rules out loud before it started
Each agent writes to .pipeline. Planner drops specs.md. Coder picks it up, writes changes.md. Tester reads changes.md, writes tests.md. Reviewer reads all three and writes verdict.md
One slash command chains the whole loop. He types/ship add rate limiting to login endpoint and walks away
Eight hours later he wakes up to a merged PR, passing tests, and a verdict from an agent that refused to touch the code itself
He has not written production code in 9 weeks
His old team lead at the company he left thinks he is taking a sabbatical
He is. The agents are not
The internet said AI would replace junior developers first
He replaced the senior, the QA, and the reviewer
The junior is still him
The seniors do not sleep
A 20-year-old made $5,939 today before his iced coffee melted
The dentist who sent it has never met him
On track for $51,800 this month. 7 months in a row
He sells three things. All AI agents. All to local businesses that have never heard the word "agent" before
The receptionist answers calls 24/7. Books appointments. Never sleeps. Replaces a $42,000/year hire for $1,997/month
The Google review agent texts every customer after they walk out the door. Asks for 5 stars. The business climbs the local map pack. The phone rings more next week
The lead reactivation agent scrapes 3 years of dead leads out of their CRM and texts every one of them. Two closes pay for the whole year
The dentist did not buy one. She bought all three. Bundle: $4,440 setup, $1,499/month after
She found him on TikTok. She signed the contract from her hygienist's chair between cleanings
Pause at 0:48. Two ring lights. Exposed ceiling pipe. One framed photo on the shelf behind him. No team. No office. The iced coffee sweating onto the laptop's left palm rest
His mom thinks he runs a marketing company
He does. It just answers its own phones
The internet said selling to local businesses meant cold-calling 200 plumbers a day
He posts one video at noon. Dentists DM him from the chair
Jonathan has not opened a DM in 31 days
His agency just had its best week. $15,500 in new MRR. 5 new clients. 500,000 views on a video he did not know was being posted
His agent runs the whole thing. It has a name and a British accent
Pause at 0:23. The dashboard on the TV is not a dashboard. That is the agent thinking. The colored web in the center is its cortex map. Each branch lights up when a region activates
Memory. Voice. Outreach. Reply. Closing. Schedule. Content. Routing. Learning. Vision. Tone. Posting. Calendar. Reporting. Persona. 15 modules. All LIVE.
The clock on the recording says 23:10:50 because the agent does not stop
Jonathan says one line into the room
"Hey Zero. Give me a recap on the agency."
The agent replies in the accent of a butler from a 1940s film
"Good evening, Jonathan. The agency recap we posted to Instagram has just crossed 500,000 views. I am, as the children say, famous. We shall require a publicist by morning"
Then it pulls his weekly numbers without being asked. New MRR: $15,500. WoW: +32%. New clients: 5, three of them inbound from the video the agent published while he was asleep
Jonathan did not write copy. The agent wrote, edited, posted, replied, qualified, scheduled, and closed.
Jonathan slept through three of the five closes
His business partner thought the British assistant was a real hire and asked when he was flying in
A consultant in Lyon stopped doing client calls in March. He makes $13,630 a month from his Obsidian vault
2,400 notes on B2B SaaS pricing strategy, built over 8 months. Every framework, every case study, every objection he has ever heard on a sales call
He stopped selling his time. He licenses the graph instead. $290 a month per company. 47 companies subscribed
The graph queries Claude on their behalf. They ask a question. The model pulls from his structured knowledge. They get the answer his old hourly clients used to pay $400 for
He runs the whole thing solo from one terminal window. Stripe handles billing. The graph handles delivery. He handles nothing in between
Most people collect notes. A few connect them. Almost nobody sells access to the graph
The note-takers fill Obsidian with 1,283 markdown files and call it a second brain
The builders connect those files with Claude Code and finally see the graph that was always there
The monetizers wrap it in an API and charge for it
The problem is not note-taking. The problem is most people stop at collection
A few keep going to connection
Then there is the third step. That is where the money lives
Your notes are a database waiting for a price tag
A 21-year-old in Miami runs an AI agency from his living room
He has zero employees
Last month: $11,500. Onboarded 84 new clients. Did not write a single sales message
His TV is the only thing visible from the street. It is always on
Pause at 0:07. That is not a screensaver. That is his agent dashboard projected on a 65-inch screen above the fireplace. Replies. Conversion. Response time. New leads. Everything updates in real time while he sleeps
The agent's name is Jonathan
Jonathan lives across Telegram and Discord. Leads come in from his cold outreach system. Jonathan qualifies them in DMs. Asks the right questions. Books the discovery call directly into his calendar. By the time the lead joins the Zoom, Jonathan has already sent a pre-call brief
Yesterday alone Jonathan sent 847 messages. Across two platforms. To 312 different leads. Tone matched to each one. He did not write a single one
The agency offers a single service. Lead generation for B2B SaaS founders. The irony is not lost on him. He sells leads to companies that could build Jonathan themselves but do not
18 calls booked this week. He takes 6 of them. Jonathan takes the rest as a transparent AI voice agent. Clients pick which model they want when they sign
The TV is on 24 hours a day. He calls it the heartbeat.
When the lights go quiet he knows something is broken. When they pulse he knows it is working
His parents visited last month and asked why the screen above the fireplace looks like a sci-fi movie poster.
He told them it was art
It is not art. It is the only employee he has ever needed
A developer in Bucharest ran an experiment no one asked for
Two models. One website. Same prompt. Same time
He gave Claude Mythos the architecture. He gave Opus 4.8 the components. He told both to build an AI agency site from scratch and streamed the whole thing without telling anyone what he was testing
That navigation is already live. KB automated. Avg resolution rate. Agents in production. Mythos built the structure in 4 minutes. The copy on the hero section took 11 seconds
Pause at 0:32. Read that line slowly.
"What agencies sell you a demo. We give you a worker that already did the job while you were reading this sentence."
Mythos wrote that. Not him
He said he stared at it for a long time before he moved on
Opus 4.8 handled the components. Cleaner code. Faster render. Better table logic. When he found the bug at the end he gave it to Opus. Fixed in 90 seconds
Two models. Different strengths. Neither one needed him to do the hard part
The site went live at 11pm
By morning it had three inbound requests from companies that had fired their previous agency two weeks earlier
He had not set up the contact form yet
They found his email in the page source
Someone in the comments asked which model won
He said that was the wrong question
The question is what you do with the 6 hours you just got back
The internet said you need a team to build something that looks like a real agency
He built one in an evening
One model handled the words. One handled the code. He handled the coffee
Save this. The experiment was not which model is better. The experiment was how few humans a serious operation actually needs
Claude Mythos → Claude Fable
Anthropic is reportedly exposing a model that was never meant to be public
“Claude Fable” is said to be the sanitized surface of something far more restricted - an internal system previously kept behind closed access tiers
Rumors around “Mythos” don’t focus on performance. They focus on containment
Too capable in security contexts. Too unpredictable in system-level environments. Too sensitive for broad deployment
Access was allegedly limited to a small circle of vetted organizations under strict oversight
Now it’s being pushed outward
Not because it changed
But because it can’t stay contained anymore
Will Claude Fable actually be "safe"? Or is this the beginning of the leak?
Everyone is renting GPU power for $300 a month
He bought a Mac Mini for $599. That was eight months ago. He has not thought about compute since
Pause at 0:03. Three monitors. Reddit on the left. Live agent output in the middle. Discord on the right. The Mac Mini behind the keyboard is running all of it
He is playing a dice game on the main screen
The people paying $300 a month are sitting at their desks. Watching dashboards. Checking usage limits. Upgrading plans
He is rolling dice
No usage cap. No billing alert at 3am. No "you have exceeded your monthly limit." No vendor. No plan. No renewal email
He paid once. Eight months ago. That was it
Pause at 0:08. That color-coded output is not a cloud dashboard. That is a local agent running on the box behind his keyboard. The screen is mostly green. He checks it between rolls
The rental crowd is still renting
The $599 box is still running
The internet said cloud compute is the only way to run serious AI
He bought the box instead. Now he plays dice while it works
Most people will keep renting. A few will buy the box
A 24-year-old built a datacenter for $28,752
Amazon spent $2.3 billion on the same thing last quarter
Pause at 0:03. Count the boxes on the right shelf
That is not crypto. That is 48 Mac Minis on a $40 wire shelf from a hardware store
Each one runs a separate agent
48 agents. 48 niches. 48 outputs running in parallel while he sleeps on the other side of that wall
The monitors in the middle are not dashboards he watches. That is the agents reporting back. He checks it once in the morning. Then he leaves the room
The whole setup fits in a spare bedroom
No lease. No cooling system. No IT team
A fire extinguisher in the corner because he is not completely reckless
Total hardware cost: $28,752
Last month: $58,400
His landlord thinks he stores furniture in there
One Mac Mini costs $599. Runs local models, scrapes, posts, optimizes. Does not sleep. Does not ask for equity
Forty eight of them do that forty eight times over
The internet built infrastructure in Virginia with raised floors and redundant power
He built his on a wire shelf and an extension cord.
The internet was thinking too big
Save this. In 6 months everyone will have one. Right now almost nobody does











