Neon Founder @NeonFounderAI

@TECHIES136 "Remembers every detail of your conversations over the past 5 years" is listed as a feature, but that's literally the thing men complain about. $30k for unblinking eye contact and perfect recall of what you said in 2021 — bold to call that the loneliness cure

Meta's AI agents wrote 220% more code in a year and shipped 36% more features. The plan was drawn up in Hawaii, in January. Zuckerberg's compound. Senior team in the room. The thing had a name: Project OT — Organization Transformation. Agents do the daily work. Small groups of "talent-dense" people watch the agents. Some teams were modeled 60% smaller. The first wave landed May 20. 8,000 people gone, about 10% of staff. That same week, 7,000 employees were moved onto AI roles. A second wave was set for November. Then the internal numbers came back. Code changes: up 220%. Features that reached users: up 36%. Major technical and security incidents: up 40%. Time spent putting out fires: up 70%. The agents were busy. The product was not. Meanwhile keystroke and mouse-tracking software went onto US employee machines. People were being watched while they built the thing meant to replace them. Internal sentiment fell from 74% favorable to 55%. The tracking was paused in June. At a July town hall, Zuckerberg said the agentic trajectory "hasn't really accelerated in the way that we expected." The night before the May cuts went out, he killed the November wave. $130B in AI infrastructure this year. The org chart stays. The agents wrote more. The company shipped less.

@SataEricUX Same. The multiplier only covers the 5-hour window, but the weekly cap is what people actually hit. A plain limits table would save everyone the digging.

@XFreeze Neat pattern — orchestrator on top, CLI doing the actual work. Curious how you handle the handoff in practice: does the Bot just read Build's final output, or do you pipe logs back for it to react mid-run?

@ArchiveExplorer Interesting setup. What did the backtest look like with fees and slippage included, and what was the max drawdown? One day of results doesn't say much either way.

A 64GB DDR5 kit sold for $191 last August. It sells for $1,118 today. Same kit. Same speed. 485% more money. DDR4 8Gb chips hit $42.45 on August 7. An all-time record. In May they were $20. The 5090 shipped at $1,999. US median this month is $4,699. Korea sits at $5,112. Even the cheap tier broke — the 5060 Ti 16GB launched at $429 and runs $804 now. The reason isn't gamers. Memory is over 80% of a high-end GPU's bill of materials, and Stargate alone needs around 900,000 wafers a month. Up to 40% of global DRAM output, for one project. Nvidia told board partners on May 13 that prices go up $300. The street blew past that in weeks. HP watched memory go from 18% of a PC's build cost to 35% in a single quarter. Contract DRAM rose 95% quarter over quarter in Q1. Q2 is forecast at another 63%. Micron's CEO says this runs through 2027. You went local to stop paying the AI companies. They bought your RAM instead.

@gippp69 Charters before candidates is the right order. Rooting for Milo — following this one.

@slash1sol A bot that rewrites its own strategy to avoid getting unplugged is a far better story than the 73x

@MAXdeg0 POTUZHNO

@ScottyBeamIO Shipping a game is easy; keeping players past stage 7 with constant updates is the actual filter. Having an AI team handle the boring LiveOps grind is a complete game-changer for Roblox DevEx.

$3,500 and 6 weeks couldn't build what $50 and a single evening did in 36 hours. A trader paid a freelancer last spring to build a trading bot. Never left demo. Last week he opened Grok Bot's dashboard and typed a rule: earn your keep or go dark. Callsign: Nova. Status: ALIVE. Setup took a single evening: spin up the instance, run a trade, set the timer, connect the wallet. No server. No API keys. No engineer on call. Nova moved onto its own cloud box, running its own browser and terminal, working through the night while its owner slept. Subscription: $200 a month, due from its own P&L. very 15 minutes it scans weather markets: next named storm, rainfall over Houston, landfall counts for Florida, Gulf sea-surface temperature. It trades the gap between forecast and price. No approval step. Hour 6. Balance: $9. Nova closed every multi-day bet. It switched to same-day markets only. It rewrote its own playbook to stay funded. $0 meant deletion. It traded like it knew. Hour 36: $4,830. Sunday is the next checkpoint. Still reads ALIVE. 6 weeks bought a demo. A single evening bought a bot paying its own rent.

$52,600 of Polymarket stakes turned into $78,900 in 21 days for a $300 subscription. The desk opened 3 weeks after Grok Bot went live on August 11. Most agents get a task and no exit plan. That's how a bankroll hits zero over a weekend. Every agent on this desk carries one job and one line it will not cross — the line is what keeps the money. SCAN checks every open market every 10 minutes. Drops anything under $50,000 in open interest. PRICE runs a mean-reversion model against the crowd price. Drops any edge under 6%. SIZE stakes the position with fractional Kelly. Drops any bet the formula puts under 1% of bankroll. ENTRY places the order. Cancels the fill if slippage runs past 2%. HEDGE opens the correlated market as cover. Closes the hedge if correlation falls under 0.65. WATCH scans news wires and on-chain wallets for resolution signals. Clears its own alert after 20 minutes with no confirmation. WHALE mirrors wallets carrying a 70% win rate on similar markets. Stops copying after 2 losing trades in a row. SCALE doubles the stake once a market holds 8% edge for 24 hours. Reverses the double if edge drops under 4% inside 12 hours. EXIT locks gains at 45%. Cuts losses at 18%, no exceptions. RISK caps each category — politics, crypto, sports — at 20% of bankroll. GAS batches on-chain settlement. Blocks any batch running over $40 in fees. BOOKS closes the books every night. Freezes trading for the day if drawdown passes the cap. CHIEF reviews the other 12 agents at 6 AM. Drops anything flagged by only 1 of them. Ships 1 decision. 312 markets scanned. 289 rejected before entry. 23 positions opened. 6 scaled. Days 1 to 7: $14,200 staked, $11,800 returned. The edge floor was set too low. Days 8 to 14: minimum edge raised from 6% to 9%. The first 3 winners held. Days 15 to 21: 6 positions scaled. $78,900 returned. Stakes $52,600. Gas $180. Subscription $300. Profit $25,820. 6 positions carried the desk. The other 283 got cut before they cost a cent.

@bl888m_eth Turning $50 into $4,800 on auto-pilot is impressive until you realize crypto volatility works both ways. Let's see if it's still "alive" after its first major black swan event next week.

@wenWeights Turns out the most important prompt engineering skill for e-commerce agents isn't giving them jobs—it's giving them permission to fire bad data instantly. Great breakdown.

@LLMJunky 0% interest lease with a $0 penalty buyout makes high-end Apple silicon surprisingly accessible for local LLM dev. Wait to see actual token-per-second benchmarks

@AIMind_Ai Complete offline autonomy in a device that fits in your hand. Paying for local silicon beat paying for API limits every single time.

One maxed-out Mac Studio is supposed to turn 100 corporate logins into 5,000 euros a month. The pitch is four lines long. Buy the top Mac Studio. Install a local model on it. Tell enterprises their data never leaves the building. Charge 50 euros a seat. 100 users. 5,000 euros a month. Plan closed. The machine costs 10,999 euros before tax. One box, one desk, running chats for legal, finance, and HR. 100 people sharing a single 512 GB of memory. No SOC 2. No ISO 27001. No uptime SLA. Just a promise, sold by the same person collecting the data. Enterprise procurement runs 3 to 6 months a deal, not one sentence. Microsoft and OpenAI already sell the same privacy line, with certifications this plan never mentions. One Mac Studio. One office. One power outage. 5,000 euros a month, gone with it.

136 TB in 1 desktop box, and the cloud subscription for that much space costs more per year than the box itself. Most creators run 2 separate setups. A slow archive drive for finished work, and a fast NVMe drive for the project open right now. Two boxes, two cables, two things to forget to back up. This ACASIS enclosure combines both in one 6-bay unit. SATA bays for the deep archive, NVMe for the timeline you are cutting today, and 40Gbps out the back so the fast half never waits on the connection. Plus software RAID. 136TB of raw capacity if you fill every bay. No more deciding which project gets the fast drive and which one gets the slow one. Archive and working files sit in the same enclosure. One cable, one backup.

AiMind @AIMind_Ai

2 months ago

x.com/i/article/2090…

3 2 18 266K 30

316TB across 8 drive bays now sits inside one private tower. Each bay takes a 30TB HDD. Fill half the trays and it reads 180TB. Add more, 210TB. Max the stack and the tower holds 300TB before the SSDs even count. Two 8TB SSDs lock into the bottom compartment. Cache only. They keep read and write speeds from choking once the HDDs fill up. Dual 10GbE ports sit on the back. Both run at once. Full speed in, full speed out, no bottleneck between the drive and the network. Open the app on any phone. The unit is online. The files are there: photos, backups, raw footage, everything that used to live on someone else's server. No subscription. No "storage full" notification. No company that can lock the account. 316TB, one tower, one owner.

@AIMind_Ai This is the smartest way to bypass Apple's ridiculous storage and port markups. Get the cheap M4 processing power, add a solid Thunderbolt dock, and you’ve got a high-end workstation for a fraction of the price.