Joying.eth

TypeSafe AI just stripped agent stacks down to one frontier call A 26-year-old in Berlin clears $3,200 a month wiring Jev and Claude into one harness. Jev is a judgment model from TypeSafe AI that takes program state and returns typed decisions. No free-form text, no reasoning. Five gates wrap every agent step: 1. score files, keep 3 of 22 2. pick model tier: cheap, mid, frontier 3. write and run code 4. keep or drop the output 5. halt if the action is destructive Only gate 3 calls Claude. The rest are judgment calls at p50 44 ms, p95 61 ms. Jev scores the files, picks the tier and gates every output. Claude writes the code, runs it, passes tests. Confidence drops below 0.7 and Claude takes over. It rechecks whether the task is done. Splitting judgment from reasoning drops four of five calls off the frontier tier. That split used to cost a team. Now it costs one config.

holy sh*t. one memory field feeds ChatGPT, Claude and Codex. re-explaining is dead. 3,948 memories in a single field. ChatGPT, Claude and Codex all pull from the same store. 1. Nine signals feed one field 2. Clusters link every memory 3. Three models read same pool 4. One bad entry hits all three 5. No wall between any layer Telegram, GitHub, files, CRM, sales calls, email all pour straight in. The blast radius is not one wrong answer. It is every agent hallucinating off the same poisoned entry. Adding another pipe takes minutes. Adding a permission layer between them takes the rest of the quarter. The cluster map drifts and every thread tightens around its node ↓

Pure insanity. for 20 years you searched other people's thinking, karpathy drops one concept, 5 minutes of wiring gives you a search engine that only finds your own notes, and you never once built it andrej karpathy got 16M views on one post: spend more of your LLM budget building a second brain than writing code every article you read, every podcast you caught, every meeting you sat through, every dm thread you kept, every tweet you drafted. all searchable in one query, forever the ones already on it stopped opening chatgpt. they go to their own search bar now. that bar carries 11 years of their own thinking the stack: → brew install obsidian, mkdir ~/vault, cd ~/vault → open claude code in that folder, init CLAUDE.md as the rulebook, /raw for sources, /wiki for your pages → drop any file into /raw. transcript, pdf, screenshot, meeting notes. one word: ingest → claude writes a wiki page, links it to 5 existing ones, updates CLAUDE.md with every new pattern it catches in how you think → no blank chats anymore. every question runs through the vault. every answer pulls from your own sources with dates and links 5 minutes to set up. runs forever on the laptop sitting in front of you right now the alpha most skip: > query your vault in one line and the answer pulls 5 of your own past notes with dates. google returns strangers. your vault returns you > every ingest cross-links to 5 past pages on its own. after 90 days the vault answers questions you forgot you ever asked > CLAUDE.md learns how you reason, what you flip on, what you dodge. by month 3 it drafts posts in your voice better than you manage at 2am > flip two flags and the whole thing runs on local llama. zero API cost, zero vendor lock, your thinking never leaves the machine > compounds over 10 years. by year 3 your vault knows more about what you believe than you do. by year 5 recruiters ask to see parts of it the part most will scroll right past: a coworker wired this on a saturday in february by august his posts landed 4x the engagement on the exact same topics by october his first draft of any doc was already stronger than your third the gap is one 5 minute install and 90 days of feeding in what you already consume for free bookmark this