←
Mojunel.eth8,369 posts
Mojunel.eth avatar

Mojunel.eth

@mojunel

34 | research AI since 2021 | ex Binance

◷ Joined March 2010▧ 175 Photos and videos
Mojunel.eth avatar
Mojunel.eth@mojunel·Nov 12

I'm still in shock, the CEO of Y Combinator open-sourced his agent brain on GitHub, 32.4K stars, one clone away 7 second brains you probably never noticed, all ready to clone right now: 1. gbrain · 32.4K stars Garry Tan's personal agent brain. every fact tracks its origin, and every agent taps the same memory github.com/garrytan/gbrain 2. Karpathy's llm wiki · 50K stars the single-page idea file that kicked off this entire wave gist.github.com/karpathy/442… 3. ai memory · 13.7K stars leave Claude Code halfway through and resume inside Codex with nothing lost. memory sits in plain markdown github.com/akitaonrails/ai-m… 4. COG second brain · 1.4K stars 35 skills and 10 agents, no database. just markdown files that reason github.com/huytieu/COG-secon… 5. agent second brain fire off a voice note through Telegram and it arrives as a linked entry in Obsidian, or a task inside Todoist github.com/smixs/agent-secon… 6. open second brain each night it converts the fixes you keep making into rules your agent obeys github.com/itechmeat/open-se… 7. claude second brain one npx command wires up Obsidian, Claude Code and local search in a single pass github.com/jessepinkman9900/… seven repos sitting in the open. your agent still loses you by morning save this

0:24

Ridark @ridark_eth Oct 29 Article One Gist Killed Manual Note-Linking Linking notes by hand was never building a second brain. It was killing one. Karpathy published one gist that puts the model in charge of every cross-reference, every summary, every contradiction

Ridark avatar
⟳ Mojunel.eth retweeted
Ridark@ridark_eth·1h

GTA 6 has already launched, and while Take-Two is counting billions, surpassing even the legendary opening of the fifth installment with its $1 billion in three days, the real story of the release lies elsewhere. Rockstar has set an unreal bar for detail once again: every vehicle features a fully interactive dashboard. You can control every element in detail: roll down the windows, switch between low and high beams, activate off-road mode, drop the truck bed, and use turn signals and hazard lights. While some players encounter initial server bugs, the developers prove that the sky-high expectations of millions of fans were fully justified in the details. A true masterclass in open-world immersion.

0:29

Joying.eth @Joyingeth Nov 17 Article Four Windows, One Money Map for the GTA 6 Launch $1 billion in three days. That was GTA V in September 2013. Take-Two tagged it the biggest entertainment opening in history. Rockstar kept the revenue. Everyone else kept the attention. Searches,

Mojunel.eth avatar
Mojunel.eth@mojunel·1h

While everyone is discussing the economies and budgets of other major gaming franchises, such as the multibillion-dollar revenues of Roblox or the massive investments in GTA Online, where players spent billions of dollars without an official marketplace prior to the integration of Cfx.re, GTA 6 itself is already surprising players right at launch. While some players and investors are securing their spots on the new content, server, and script "exchange," others are finding much more down-to-earth bugs: while I was riding my BMX bike in GTA 6, I found a physics bug and launched high into the air. Against the backdrop of Take-Two and Rockstar building a brand-new mod ecosystem, such physics exploits in the open world remind us that even the most massive releases are not without their amusing bugs.

0:15
Karmi@0xKarmiNov 17

Karmi @0xKarmi Nov 17 Article I Drew a Floor Plan for GTA 6 and Every Seat Was Still Open I looked at the richest gaming franchise ever built and tried to map its creator economy as a trading floor. Five layers from Rockstar down to the marketplace. Four tiers of players splitting at

meya. avatar
⟳ Mojunel.eth retweeted
meya.@birmeya·1h

You can ride a jet ski with your dog in GTA 6, and honestly, the sheer absurdity of this game's physics already has me completely hooked. While Ubisoft and EA spend years overthinking launches, Rockstar is effortlessly printing billions. Just like when GTA V pulled $1B in three days, the real gold rush right now is the massive search traffic. While players are busy hunting down hilarious launch bugs and weird water physics, I spun up an autonomous content pipeline in Claude Code in 15 minutes flat. Instead of manually clipping gameplay or chasing unverified leaks, I wired six parallel subagents to a single Obsidian markdown vault. Strict gating rules inside CLAUDE.md ensure zero hallucinations: my Scout and Librarian agents ingest only official Rockstar patch notes and my raw session logs. My SEO agent scrapes top unfulfilled search queries, Writer outputs direct-answer guides, and Editor auto-cuts 30-second vertical shorts. Launch week is a brutal land grab for attention, and manual workflows simply cannot scale. Full system architecture and the exact two prompts to deploy it are in the article.

0:12

Mojunel.eth @mojunel Nov 17 Article My GTA 6 Machine: Six Agents, One Vault, 15 Minutes GTA V made $800 million on day one. Crossed $1 billion in three days. Take-Two put it on record as the fastest-selling entertainment product in history. That money went to Rockstar. The attention went

Karmi avatar
⟳ Mojunel.eth retweeted
Karmi@0xKarmi·5h

I REWIRED MY PRE-LAUNCH PLAN. The whole floor moved. What changed: Old: One product, build after launch, hope the market holds. New: Four phases, one ship per week, 26 items ranked before the doors open. Old: Zero creator revenue from GTA Online across a decade. New: Cfx Marketplace live since January 2026. Roblox creators earned $1,503.1 million in 2025. 1. Now to month 2. FiveM stack learned, Claude wired in, 3-5 scripts shipped, Tebex storefront live. 2. Month 2 to 4. RP server in beta. Content rolling. First retention numbers logged. 3. Month 4 to launch. Catalog at 15-20 scripts, 100+ paid cleared, NPC pack in beta. 4. Launch to three months out. Compound. Stack sales. Revenue rolls. Same game. Same audience. Four shifts changed the math. One thing shipped per week across all four. Six months. Roughly 26 items before the crowd arrives. Compounding kicks around month 3-4 when the catalog carries weight. Week one is still open.

Post attachment
Karmi@0xKarmiNov 17

Karmi @0xKarmi Nov 17 Article I Drew a Floor Plan for GTA 6 and Every Seat Was Still Open I looked at the richest gaming franchise ever built and tried to map its creator economy as a trading floor. Five layers from Rockstar down to the marketplace. Four tiers of players splitting at

Mojunel.eth avatar
Mojunel.eth@mojunel·5h

Deleted. every. guide. I. had. A 90-day run. One Node bot replaced five jobs. 1. Daily articles → /mission pulls from the vault 2. Manual replies → crew tracker ranks members 3. Weekly roundup → recap builds from the log 4. Giveaway mod → bot logs finished missions 5. Thread scanning → prompt picks 3 tools 3 features plus 3 tools. 6 products in one loop. 10 daily-active servers outweigh 100 installs. Build what the log says, not what I want. Deleted every guide. The bot writes the next one.

0:24

Mojunel.eth @mojunel Nov 17 Article My GTA 6 Machine: Six Agents, One Vault, 15 Minutes GTA V made $800 million on day one. Crossed $1 billion in three days. Take-Two put it on record as the fastest-selling entertainment product in history. That money went to Rockstar. The attention went

Karmi avatar
⟳ Mojunel.eth retweeted
Karmi@0xKarmi·8h

Killed them all. Clips just retired every paid growth play. An GTA RP server wired 7 systems, ran every acquisition channel through the same funnel, and watched almost all of them die on arrival. 1. Paid ads. Budget gone before anyone stayed past day one. 2. Discord raids. Cold rooms, cold traffic. All gone. 3. Forum drops. Server promos buried under noise. 4. Influencer plug. One spike, one week. Then nothing. Gameplay clips survived the entire kill list. Best moments cut to Shorts, Reels, TikToks. No production, no ad cost. The viewer already watches GTA. The clip only proves the server exists. $15–25 a month per head, recurring. Only cost: the hours spent clipping and posting. Retention data locks by month 3–4. The kill list scrolls and the last line stays lit

0:24
Karmi@0xKarmiNov 17

Karmi @0xKarmi Nov 17 Article I Drew a Floor Plan for GTA 6 and Every Seat Was Still Open I looked at the richest gaming franchise ever built and tried to map its creator economy as a trading floor. Five layers from Rockstar down to the marketplace. Four tiers of players splitting at

Mojunel.eth avatar
Mojunel.eth@mojunel·8h

they said five pages take a full team I loaded one folder, two agents and 30 days of Rockstar sources into a single vault before November 19. Scout scraped every official post and patch note from the past month. Librarian pinned each file with date and link. Five pages from the vault, every fact confirmed: 1. Release times by region, answer sitting on line one, never buried three scrolls down 2. Editions compared side by side, confirmed differences only, nothing from rumor threads 3. Platforms locked to PS5 and Xbox Series X|S, the full confirmed list 4. Map details, every line traced to a dated source file in the vault 5. Open rumors tagged "not confirmed yet" so one bad call does not burn trust on launch day Five pages and six clips, each under 30 seconds, each answering one question from the vault in its first 2 seconds. Five source cards stack and the counter moves

0:24

Mojunel.eth @mojunel Nov 17 Article My GTA 6 Machine: Six Agents, One Vault, 15 Minutes GTA V made $800 million on day one. Crossed $1 billion in three days. Take-Two put it on record as the fastest-selling entertainment product in history. That money went to Rockstar. The attention went

Joying.eth avatar
⟳ Mojunel.eth retweeted
Joying.eth@Joyingeth·Nov 17

THIS F*CKING INSANE Two prompts built six agents from one empty folder. I ran build from this one line: mkdir gta6-machine && claude 1. 5 directories dropped, one CLAUDE.md capped at 200 lines, 4 rules that make every agent cite the vault first 2. 6 subagent files written. Scout, librarian, writer, editor, seo, publisher. One job each 3. Nothing publishes unless a source file confirms it. Rumors stay tagged 15 minutes. 6 agents. 2 prompts. Save the build log

0:24

Joying.eth @Joyingeth Nov 17 Article Four Windows, One Money Map for the GTA 6 Launch $1 billion in three days. That was GTA V in September 2013. Take-Two tagged it the biggest entertainment opening in history. Rockstar kept the revenue. Everyone else kept the attention. Searches,

Mojunel.eth avatar
Mojunel.eth@mojunel·Nov 17

Elon Musk after finding out 65% of banking CEOs call AI a top priority and their bot still says “please contact your bank” when you ask it to raise a card limit

0:10

Introducing Stitch AI: put agents to work, native inside the Stitch operating system. Stitch AI fetches data, runs tasks and changes configuration across the platform, under the bank’s own permissions and regulatory constraints.

1:13
Karmi avatar
⟳ Mojunel.eth retweeted
Karmi@0xKarmi·Nov 16

Turning dangerous: a whole senior dev crew in 10 repos. I genuinely don't understand why everyone isn't using this yet. Claude Code ships in every paid plan from $20 a month. These 10 open-source MCP servers handle the rest. < read. docs. debug. run. track. ship. /> 01 servers → reads files and git history for the agent github.com/modelcontextproto… 02 serena → edits code by symbol, not by line number github.com/oraios/serena 03 context7 → pulls version-current library docs into the prompt github.com/upstash/context7 04 browser-tools-mcp → reads your live Chrome console and network github.com/AgentDeskAI/brows… 05 playwright-mcp → drives a browser through page snapshots, no vision model github.com/microsoft/playwri… 06 DesktopCommanderMCP → runs terminal commands and edits files from one tool github.com/wonderwhy-er/Desk… 07 mcp-gateway → runs every MCP server inside a Docker container github.com/docker/mcp-gatewa… 08 inspector → tests any MCP server before the agent burns tokens on it github.com/modelcontextproto… 09 azure-devops-mcp → reads boards, repos and pipelines for teams off GitHub github.com/microsoft/azure-d… 10 github-mcp-server → opens PRs and reads CI runs, patch to green build github.com/github/github-mcp… read → docs → debug → run → track → ship. 295,008 stars combined. Serena is GPL. The other nine are MIT or Apache. Lock your token scopes before connecting. Several of these run shell access or push to real repos. Pick one broken test and let the agent trace it from console to merged PR. The pipeline fires node by node when the first commit lands

0:24

Karmi @0xKarmi Nov 6 Article A finance back office I wired from ten repos I used to pay for five tools that never talked to each other. Timesheets in one app, invoices in another, signatures in a third, ledger entries in a fourth, receipts in a folder I forgot to name.

Mojunel.eth avatar
Mojunel.eth@mojunel·Nov 16

I genuinely don't understand paying for ad research A 24-year-old in Amsterdam nets $8,750 a month running category teardowns on this exact stack. You browse ad libraries by hand. You screenshot every creative into folders. You still sort hooks into spreadsheets one brand at a time. Jev from TypeSafe AI classifies ads from public libraries. Scores every creative on hook archetype, format, offer, CTA, awareness and funnel stage. The dashboard loads 402 ads across 37 brands. Nine cents in Opus 5.5 tokens for the full run. 1. Pull every live ad in the category 2. Score each ad on six attributes 3. Aggregate hook types, formats and offer angles 4. Map every CTA the category runs right now Opus 5.5 lists at $4 per million input tokens. Nine cents buys 22,500 of them at that rate. The Batch API cuts the price in half. One pass. Entire category. That is the whole playbook.

0:07

Jev is the FASTEST AI model ever built for trading It makes calibrated buy/sell decisions in under 100 ms That is one real decision on every single block, 24/7 In this article I've shown EXACTLY how to build HFT trading system with Jev (from scratch)

Karmi avatar
⟳ Mojunel.eth retweeted
Karmi@0xKarmi·Nov 16

+$4,521.38 in 11 days. I coded a self-learning trading algorithm with Claude Fable 5.1 1,643 token launches scanned. 35,452 holders mapped. 75 wallet clusters flagged. 7 tokens cleared every filter, all scoring 80+/100 I released the full algorithm as open source GitHub: github.com/QuantumOrb/QuantO… The whole point is catching tokens with real growth potential while exposing the concentration risks buried in their charts One example: Oct 21: one token drops on a memecoin launchpad Oct 30: market cap hits $850K. The algorithm scores it 80/100 4 days later the cap reaches $5.97M. That is 7.02x after the alert Top 20 holders. 20 operators. 0 clusters Estimated selling impact: -15.3% How the algorithm works: Six modules running as one system: 1. RADAR picks up new launches and tracks bonding curve movement, market caps and early trading volume 2. SHIELD runs a check on developer history before anything enters the alert feed The algorithm scores the creator's graduation rate: GraduationRate = G / L x 100% SerialDev: flagged when L hits 10+ and G / L stays under 0.10 L = total launches G = graduated tokens If someone launched 50 tokens and only 2 graduated, the system marks them suspicious 3. SCANNER looks at the top 20 holders and first 20 buyers, tracing wallet histories, token flows and early sniper slots Fresh wallet share across analyzed wallets: FreshRate = FreshWallets / ReadNonDevWallets FreshScore = 25 x clamp of 0.80 minus FreshRate, divided by 0.70, floor 0, cap 1 More fresh wallets in the mix means fewer points for the token 4. CARTOGRAPHER links wallets through shared token origins, transfers and same-transaction buys Suspicious fresh wallet groups get a weighted exposure score: PackWeight = 0.60 x total WalletShare across the group 5. GUARD reviews developer history, first-slot accumulation and remaining dev holdings FirstSlotShare = TokensBoughtInLaunchSlot / TotalSupply FirstSlotShare >= 30%: FLAGGED DevHoldings >= 15%: FLAGGED 6. ORB merges every check into one final score from 0 to 100 The algorithm does not need to greenlight every new launch The job is separating tokens with real potential from ones loaded with hidden holder concentration

0:24

Karmi @0xKarmi Nov 6 Article A finance back office I wired from ten repos I used to pay for five tools that never talked to each other. Timesheets in one app, invoices in another, signatures in a third, ledger entries in a fourth, receipts in a folder I forgot to name.

Mojunel.eth avatar
Mojunel.eth@mojunel·Nov 16

Pointed REA at a 400-function app. 41 minutes, $1.80, every call came back with evidence. I wired one CLI into every agent on the machine. One setup command, a doctor check confirming which decompiler engines were live. Native binaries went through Ghidra, free and open source. REA itself is MIT-licensed, no extra cost when you already run a Claude plan. The session log is what stays. The tagged evidence folder underneath, growing with every step, versioned the moment you commit. 1. Install registers REA with every agent it finds. Doctor confirms which engines work. 2. House rules load first. Owned targets only, every claim tagged, bypass blocked. 3. Electron apps skip the engine. REA reads bundled scripts, traces from click to handler to IPC to network call. 4. Every function ranked by call count, top 30 written into a research file with evidence ids. 5. Output lands as markdown in the research folder. Feature map, function map, release diff, versioned in git per build. Five apps fully mapped in one week. $14. One function every six seconds. Under half a cent per call on the Grok-only pass. When did you last trace every function in your own build?

0:24

Mojunel.eth @mojunel Nov 15 Article How I Built a $14 Reverse-Engineering Lab cover This is a walkthrough of a setup I assembled in one week. Two models, one open-source agent, a free decompiler. Five paid apps mapped end to end. Total bill: $14. The lab scales from zero extra cost

Karmi avatar
⟳ Mojunel.eth retweeted
Karmi@0xKarmi·Nov 15

Three models in one Claude Code session, zero tokens on the wrong tier. Anthropic's advisor strategy dropped back in April. The play: run Sonnet, Haiku 5.5 and Opus 5.5 in one session, each locked to the task it handles best. Sharper output, far fewer tokens burned. Every job lands on the right model. Launch it with one command: claude --model sonnet --advisor opus the split: → Sonnet 5.5 drives the session, writes the code and lands the patches → Haiku 5.5 fans out across the repo in parallel: search, grep and diff passes → Opus 5.5 waits on standby as advisor for the critical decisions and approves the diff → you stop bleeding Opus context on work that never needed it Opus only fires at three moments: → before a plan locks in → when the same error surfaces twice → before calling the task done Sonnet writes. Haiku scouts. Opus approves. JEV engineering works the same way one level deeper. Deterministic forks that need no reasoning, which file, which tool, retry or halt, route through JEV in milliseconds. The heavy models only see the splits that matter. the full tree: → Sonnet 5.5 owns the main session and steers the workflow → search scans the repo on Haiku 5.5, read-only → grep tracks call sites on Haiku 5.5 in a separate context → audit verifies the diff on Haiku 5.5 in a disposable worktree → Opus 5.5 stands by as advisor, reviews the final diff before sign-off → JEV handles micro-forks in milliseconds feed this tree and the prompt into Claude Code: "Reconfigure my Claude Code setup around this tree: 1. Main session on Sonnet 5.5 with Opus 5.5 as advisor. 2. Build three Haiku 5.5 subagents in .claude/agents: search maps the repo, grep finds call sites, audit checks the diff. 3. Set isolation: worktree on audit so its checks never touch the working branch. 4. One rule in CLAUDE.md: delegate every repo search and test run to the Haiku 5.5 subagents to keep main context lean. 5. Route fast deterministic forks through JEV: tool choice, path selection, retries. Show every change as a diff first. No edits until I say go." ↳ code.claude.com/docs/en/sub-… Keep this one ready.

0:30
Karmi@0xKarmiOct 24

Karmi @0xKarmi Oct 24 Article The Agent Desk: a Knowledge Graph and an Action Guard in One Folder Most developers who try to run a coding agent give it one prompt, watch it loop for ten minutes, and close the tab. They don't know where the agent should live between sessions. They don't know how to