Focus

Focus

[Focus @F0kcus]

https://t.co/EjpIUkSVxi

[Focus @F0kcus]

Notion raised $275 million at a $10 billion valuation to build company wikis. He is 24 and built a better one by pointing his laptop at their GitHub org

He runs a documentation agent out of a Berkeley studio apartment on a MacBook Air M2 and a Beelink mini PC he uses as a persistent server. The agent watches a company's GitHub org, Slack workspace, Linear tickets, and existing Notion docs through their APIs. Every 15 minutes it pulls new commits, engineering discussions, resolved tickets, and doc edits

Then a Claude Sonnet agent chain reads the diffs, extracts what changed, why, and what depends on it. Writes structured markdown into an Obsidian vault synced to the company's cloud storage. Onboarding docs for every service auto-regenerate weekly. New engineer joins Monday - vault has current docs, dependency graph, deploy history, and open questions for every service they touch

He posted a demo video to GitHub in October showing the agent generating a full 40-page engineering wiki for a 12-person startup from scratch in 34 minutes. A Notion product team engineer found it through a Hacker News thread three days later

Cristina Cordova quote-tweeted the demo the following Wednesday - called it "the auto-generated knowledge layer we should have shipped four years ago instead of hiring humans to update the docs." By December Notion had wired $8.4 million into his account for the agent architecture and a two-year consulting contract to integrate the auto-generation layer into Notion AI's enterprise tier rolling out to their 4 million business subscribers this quarter

Notion raised $275 million at a $10 billion valuation selling company wikis on the premise that team knowledge requires their proprietary block-based editor and their five-year enterprise integration roadmap. Atlassian sits at $50 billion in market cap running Confluence on the same premise. Cristina Cordova just paid a 24-year-old in a Berkeley studio more for a Claude agent chain than most Notion engineers earn in a decade

[Focus @F0kcus]

https://t.co/EjpIUkSVxi

[Focus @F0kcus]

@Kaiserwsf The biggest shift isn't AI generating content-it's AI becoming the creator behind entire businesses

[Focus @F0kcus]

@SkadooshGG The people building the tech usually know its limits better than the people selling it

[Focus @F0kcus]

@0xGenpix The hardest part isn't building the robot anymore-it's proving it can deliver real economic value

[Focus @F0kcus]

@woxiccq The details that once proved something was real are becoming the easiest things for AI to fake

[Focus @F0kcus]

@0xpupupu This isn't replacing coaches-it's replacing the repetitive work around coaching

[Focus @F0kcus]

@scream_crypto Nature's greatest optimization wasn't eliminating errors-it was controlling them

[Focus @F0kcus]

@88n77n AI memory needs version control, not just more storage

[Focus @F0kcus]

@warplarp The biggest shift isn't the robot-it's the idea of having a machine that's always on duty

[Focus @F0kcus]

Nvidia sold $47 billion in datacenter GPUs last quarter. He is 22 and built the firmware efficiency script they just paid $3.4 million for in his uncle's Youngstown garage

He runs 34 used RTX 3090s across three plywood shelves his uncle helped him build along the garage wall. The 3090s came from Craigslist mining farm shutdowns across the Rust Belt for $340 apiece after the 2024 crypto rout wiped out overleveraged operators. He replaced dead fans on eleven of them with $8 parts from a Cleveland electronics surplus store. Total hardware: $11,900. Total electricity bill: $340 a month because the Ohio Valley Industrial grid still runs old flat-rate pricing the utility never updated after the steel mill closed in 2003

He wrote a firmware layer that dynamically tunes GPU voltage curves and memory timing based on real-time thermal load and workload class - pulling 22% more sustained throughput per kilowatt hour than stock Nvidia firmware across mixed inference and mining workloads. Posted the code to GitHub in October under an MIT license. An Nvidia driver engineer found it through a Hacker News thread three days later. Jensen Huang quote-tweeted the demo the following Wednesday - called it "the kitchen-table optimization we could not have shipped ourselves without breaking six enterprise support contracts." By December Nvidia had wired $3.4 million into his account for exclusive integration rights and a two-year consulting contract. The optimization layer ships in the CUDA 13.4 driver update rolling out to every RTX 30-series and 40-series card on Earth this quarter

Nvidia sits at a $3.5 trillion market cap selling datacenter GPUs to Fortune 500 buyers on the premise that AI compute efficiency requires their proprietary silicon and their multi-billion-dollar research pipeline. Jensen Huang just paid a 22-year-old in a Youngstown garage more for a GitHub repo than most Nvidia driver engineers earn in a decade

[Focus @F0kcus]

https://t.co/EjpIUkSVxi

[Focus @F0kcus]

BrowserStack raised $200 million on a $1.6 billion valuation to run mobile app testing on emulators. He is 23 and runs it on 480 real phones for $500,000 a month

He operates the farm out of a 900-square-foot warehouse space he rents for $1,800 a month in an East Austin industrial park. 480 used Android phones - Xiaomi Redmi Notes, Oppo A-series, older Samsung Galaxies - bolted to 3D-printed racks along four walls. Every phone sits on continuous USB power through 12 industrial hubs wired into a single Beelink mini PC that runs the orchestration layer he wrote in one weekend

When an enterprise mobile app team pushes a new build, the orchestrator queues automated test scripts across the 480 phones simultaneously. Real hardware

Real OS versions. Real regional locales pre-configured across the device grid. Test cycle completes in 34 minutes on average. BrowserStack's emulator equivalent takes 4-6 hours and misses 40% of hardware-specific bugs

He posted a demo video to GitHub in October showing his orchestrator catching 47 hardware-specific rendering bugs across a Fortune 500 fintech app test batch that BrowserStack's cloud farm had cleared as passing. A BrowserStack engineer found it through a Hacker News thread four days later. Ritesh Arora quote-tweeted the demo the following Monday - called it "the physical-device advantage we always knew we needed but could never justify the CapEx to build." By January BrowserStack had wired $7.2 million into his account for the orchestration framework and a three-year consulting contract to spin up 40 similar physical device farms across their global infrastructure

BrowserStack raised $200 million at a $1.6 billion valuation selling emulator-based mobile testing to enterprise on the premise that scalable QA requires their proprietary cloud infrastructure. Sauce Labs raised at $500 million on the same premise. Ritesh Arora just paid a 23-year-old in an East Austin warehouse more for a Python orchestrator than most BrowserStack engineers earn in a decade

[Focus @F0kcus]

https://t.co/EjpIUkSVxi

[Focus @F0kcus]

Progressive spent $340 million on Snapshot dashcam analytics last year. He is 24 and built a better one on a $40 eBay dashcam

He runs a fraud detection service out of a Cleveland apartment on a used ThinkPad X1 and a water-damaged dashcam he pulled from an eBay estate sale in September. Total hardware: $220. The model analyzes dashcam footage frame by frame and flags staged accidents, exaggerated damage claims, and pre-existing vehicle wear that insurance adjusters miss. He trained it on 40,000 hours of publicly available dashcam crash footage he scraped from YouTube during nights after his shift at the FedEx sorting facility on Rockside Road

He posted a demo video to GitHub in October showing the model correctly flagging 47 staged accident patterns across a 12-hour test batch with 96% accuracy running on consumer hardware. A Progressive Snapshot engineering team member found it through a Hacker News thread three days later. Tricia Griffith mentioned the project by name on LinkedIn the following Monday - held it up as an example of "small-team ML solving fraud detection problems that consulting firms have been overcharging insurers for a decade to fail at." By December Progressive had wired $3.6 million into his account for the model weights and a two-year consulting contract to integrate the detection layer into the Snapshot product line rolling out to 24 million active policyholders this quarter

Progressive runs at $60 billion in annual premiums on the premise that fraud detection at scale requires their proprietary Snapshot infrastructure. State Farm runs at $88 billion on the same premise with their Drive Safe & Save program. Tricia Griffith just paid a 24-year-old in a Cleveland apartment more for a GitHub repo than most Progressive fraud detection engineers earn in a decade

[Focus @F0kcus]

https://t.co/EjpIUkSVxi

[Focus @F0kcus]

Nvidia sold $47 billion in datacenter GPUs last quarter. He is 19 and bought his for $340 apiece off Craigslist

He runs a mining farm out of a two-story abandoned house he bought for $8,000 in East St. Louis after the previous owner's estate sold it as-is. 47 Nvidia RTX 3090s and 12 RTX 4090s across the living room, kitchen, upstairs bedroom, and basement, all salvaged from Craigslist mining farm shutdowns across the Midwest for an average of $340 apiece. Total GPU spend: $19,000. Total electricity bill last month: $340 because the neighborhood grid runs old flat-rate industrial pricing the utility never updated after the Boeing plant closed in 2019

He dropped out of Southern Illinois University after his second semester in October to run the farm full-time. His mother drives across the river from downtown St. Louis every Sunday to bring him groceries because the nearest Walmart is a forty-minute walk. He sleeps on a mattress in the one room without a rig because the humming is too loud in the others. Last quarter he mined $184,000 in Ethereum-based tokens between GPU rotations

The house cost him less than a single Nvidia H100 board. Nvidia sells the H100 at $30,000 apiece on the premise that scalable AI compute requires their datacenter-grade silicon and their enterprise licensing

Coreweave raised at a $23 billion valuation renting those same H100s to AI startups. He runs 59 Nvidia consumer GPUs in an East St. Louis abandoned house that could be unplugged one at a time and driven across the Mississippi River in his uncle's truck by Tuesday

[Focus @F0kcus]

https://t.co/EjpIUkSVxi

[Focus @F0kcus]

@100F_exe Atlas keeps raising the bar for real-world humanoid robotics

[Focus @F0kcus]

@the_latticex Shared learning across robots could dramatically speed up real-world progress

[Focus @F0kcus]

@hotered AI creators are becoming a real business model, not just a trend

[Focus @F0kcus]

@Neuron_404 Safety isn't just about alignment-it's about knowing the limits of a model

[Focus @F0kcus]

@norvex1029 Capability is accelerating, but strong evaluation and containment are becoming just as important

[Focus @F0kcus]

@Kiiugara Indoor lighting and hand tracking look surprisingly polished here

[Focus @F0kcus]

@Zdonuk94 AI influencers are changing how brands think about content