Everyone's a founder now. Most will quit in 6 months — and it won't be the product's fault. AI removed the friction that used to filter people out early. So the real filter just moved downstream, to the part it can't fake for you: still showing up when the launch dopamine wears off. Building was always the easy 10%. Anyone disagree?

The math does not require cheating, but it also does not reward honesty. 15% per month compounds whatever thing you point it toward. A product that people love, and a product that people cannot put down, draw the same curve. So "clean" is not the default that the math hands to you. It is a choice you make early, about what you are optimizing, and then exponential growth makes this choice very big. PG is right that you can get there clean. The problem is that the engine will not keep you clean for you.

We work hard everyday on making Trellis even better github.com/mindfold-ai/Trell…

Ralph loop — a dead-simple way to let an AI coding agent run itself until the whole spec is done. It's the technique behind all the "loop engineering" noise right now. Ralph was coined by Geoffrey Huntley. Google's Addy Osmani gave the wave its name — loop engineering. Even Boris Cherny, who runs Claude Code at Anthropic, says he "doesn't prompt Claude anymore." One open-source take already has 20k stars. Here's the problem it solves: Coding with an AI today means babysitting it. It writes a chunk, stops, waits for you to read it and feed the next instruction. Two options, both bad — hover over it line by line, or walk away and come back to find it stalled halfway. Hands-off and trustworthy were a pick-one. Ralph kills that tradeoff. The core is almost insultingly simple — one line: while :; do cat PROMPT.md | claude-code ; done Write the spec to a file. Let the agent read it and work, over and over. Every pass starts on a fresh context; progress lives in files and git, not in the model's memory. Tests and lint are the brakes — get something wrong, the next loop fixes it. Huntley's rule: "sit on the loop, not in it." So a pile of things that used to need you now happen on their own: Drop in a PRD checklist, wake up to every box ticked The agent notices it's not done and keeps going — no nudging A bad pass doesn't need your rollback; failing tests just trigger a redo Your job shifts from writing prompts to designing the loop and watching it run Letting AI work unattended no longer means choosing between hands-off and reliable. Start here 👉 github.com/snwfdhmp/awesome-… Want to just run it 👉 github.com/snarktank/ralph #AIAgents #VibeCoding #ClaudeCode #OpenSource

A not-even-frontier model running on someone's own machine is doing work that used to bill by the hour. The trick isn't a stronger model. It's that someone mapped exactly where it's weak and built a system around those weak spots — a fallback for when it errs, an external memory for what it can't hold. The model is just the engine. The value is the car built around its blind spots. Staring at parameter counts misses the part that actually matters.

Everyone's first instinct for making an AI smarter is "retrain it." There's a cheaper move. Let it finish a task, reflect on what worked and what didn't, and write the lesson into a playbook. Next run, it reads the playbook first. No retraining. Just a notebook that keeps getting thicker — and an open small model starts closing the gap with the expensive closed ones. Turns out the thing agents lack usually isn't brains. It's a record of their own mistakes. Are you letting your agents keep one?

LMAOOO

Anthropic is now the new CLOSED AI

🛠️ Notion shipped an official hosted MCP server. It lets Claude, Cursor, or ChatGPT read and write your entire Notion workspace — no code, no copy-paste. For a while, getting AI into your Notion meant one of two things: → wire up the API yourself: create an integration, copy the secret token, paste it into a config file. Powerful, but a pain to set up and easy to break. → or just copy-paste pages into the chat. Simple, but the content is frozen — edit the page in Notion and your AI is already out of date. Connected vs. effortless. You had to pick one. Notion's official MCP kills that tradeoff. One OAuth click inside your AI tool and the whole workspace is wired in. The AI reads and writes live, exactly like you can — no token to copy, nothing to paste. So a bunch of things that used to be a chore are now just default: - Claude searches your entire workspace for you — you copy nothing - finish a research pass, tell it to write the results back as a clean new page - update project status and tasks right in the chat, never switching back to Notion - Cursor, ChatGPT, VS Code — same single auth, connect once and it works everywhere It's officially maintained, and Notion is now steering everyone here — they're even deprecating their older self-hosted MCP in favor of it. developers.notion.com/docs/m… #Notion #MCP #AItools

Trellis just crossed 10,000 stars on GitHub. We never optimized for stars. We optimized for one thing: your coding agent shouldn't start from scratch every session. That number is really thousands of devs who got tired of re-explaining their codebase to an AI — and decided specs, tasks and memory belong in the repo. In January this was an empty repo. Thank you all very much! What's next: docs.trytrellis.app/advanced… Check out Trellis here: github.com/mindfold-ai/Trell…

I used to measure progress by how fast the code shipped. I thought that was the win. Now I see it was the old game. The new game isn't writing code faster — it's how many AI agents you can keep pointed at the same target without them drifting apart. A year ago "AI-assisted" meant one engineer, one copilot, one file at a time. It felt unstoppable. Until the codebase got big enough that the agent forgot what it decided yesterday. The team of 2025: a handful of engineers, one copilot each, every session re-explaining the same context, work quietly redone. The team of 2026: one engineer, a swarm of agents, every one of them reading from the same spec. The bottleneck isn't typing code anymore. It's keeping agents that spawn agents aligned — so they stop undoing each other's work. Almost nobody has solved that. That's what we're building at Trellis. github.com/mindfold-ai/Trell…

"Would you pay for this?" is the most useless question in startups. People are polite about hypotheticals and honest about money they're already spending. So don't validate an idea — go find where people are already ranting, already overpaying to work around the problem. A complaint that already exists is worth more than a need you talked someone into. Your last "need" — did you ask it out of someone, or stumble onto the rant?

Fuck, we are doomed.

Learning UI design for my product Trellis, how do you think?

We're still in the primitive era for loop engineering lol

Speaking of the devil of loop engineering...

You're waiting for your feelings to give you permission. They never will. We treat emotions like a traffic light — anxious is red, motivated is green. So we spend our lives waiting for green. But your feelings' whole job is to keep you comfortable. They'll always find a reason for you not to do the hard thing. Stop asking "how do I feel?" Ask "what needs to be done?" — then do it, regardless of the feeling attached. How much of your life is waiting on a green light that's never going to come?

i don't think the value collapsed once generation got basically free, the scarce thing is trimming down, knowing which output is worth keeping. to me that's still intelligence, just the kind nobody bothered to benchmark. "human touch" is probably just what we call whoever still has working taste.

The things you listed are the execution layer. Automation eats execution first, but that also means whoever has taste about which experiment is worth 512 chips suddenly gets 100x more shots at it. Eval will get more valuable, not less.

One thing I might add is, the "spec" doesn't have to be a design doc. It can be any knowledge: domain rules, past decisions, review feedback, what broke last time. We've run this for ~5 months with Trellis, and claude code & codex are already capable of maintaining those doc themselves.