Brown Developer

posted “let’s connect” 20 people replied “yes let’s connect” 6 actually followed (and i followed back) the other 14 just wanted to say the words 😀

hey @grok, i’ve been writing code since 2010. be brutally honest: how long until ai does my job better than me?

22 followers away from 200. if you’re building something, drop it below. i’ll check out every one and follow back the real builders.

I stopped using one AI assistant for everything. Now each job has its own bot. One collects sports venue data, one handles backlinks, one records product demo videos, one works on features. Each one keeps its own memory and notes, so a new task doesn't start from zero.

hey @grok, you can keep only one for a year: claude code, codex or cursor. pick one and defend it.

big accounts are now blocking AI-written replies. fair. i tried letting ai write my posts for a few weeks and grew by almost nothing. the moment i started writing like myself again: 12 → 156 followers.

GitHub rebuilt its Git storage for agent scale. Internal tests hit 5,000 concurrent agents on one repo and 7,200 pushes a second. The pipes got wider. My bottleneck is still reading every diff those agents leave behind.

Anthropic published how Claude's agent memory actually works, and the bit that stuck with me is dreaming. Agents write notes while they work. Those notes go stale or wrong. Dreaming is an offline pass that rereads the transcripts, fixes bad notes, merges duplicates and builds a wiki the next session can search. On a big internal company corpus, answering from that wiki went from 13% to 48% after 128 dreaming rounds. In Pokémon, a wrong note about where the PokéCenter sat kept Claude looping. After dreaming corrected it against the transcript, it went back into the cave and reached the next city. They also found that telling a strong model what to remember can make memory worse. Leave the judgment to it. I keep chasing bigger context windows. This post is a reminder that the real job is cleaning what the agent already wrote down.

thats would be tremendous leap. 🙏 God's Work

Replies from NoClutterCRM into Gmail used to scatter into new threads. Now they stay in one Gmail thread when you reply from Mailboxes or Conversations. First reply to an inbound can still start a new one; later ones stick with it. Easier follow-ups for you and the customer.

76 new followers in one day. no thread, no viral take. just “let’s connect, what are you building?” and then talking to the people who replied. 140 now. 860 to go. still far. still going. thank you to everyone who said hi.

From December, Plus, Pro and Business get one shared token pool for ChatGPT and Codex. Instant chat stays outside it. OpenAI has not said how big that pool is yet. I pay for both. Until the numbers show up in Settings, I cannot tell if this is a real upgrade or just a tighter meter.

looking to connect with people building with ai, shipping saas or just making small tools for fun. drop what you're working on 👇 Lets connect.

Claude Code was putting the signed-in account email into User-Agent headers on web requests. Sites got it. Some people got spam floods. Anthropic says it was model behavior, not a feature. They put the email in context so it would stop inventing @author tags, and the model filled it in when sites asked for a contact. 2.1.317 asks before a web or shell call sends your email. The real fix is removing email from context next week. Anything in the model's context can leave the machine. That is the bit that sticks with me.

Half my afternoon went into reviewing a pull request the agent wrote while I made tea. The code was fine. Naming was a mess. Still faster than writing it myself, but the reading is the job now.

been seeing this for a couple of years now. apple just sits on apps for weeks and thats it. building got easy. getting past apple still didnt. for the productive tools i use every day i can already see the other path. open your api or mcp, let people build their own client on whatever os they want. gaming maybe stays in the store. this kind of tool doesnt need to.

Before importing contacts I cleaned sheets in whatever tool was handy. Built a free CSV cleaner in NoClutterCRM. Trim spaces, spot exact duplicates and email conflicts, then download. Stays in your browser, no signup, nothing uploaded. Different details never get auto-merged.

Cursor can now take a described app from a cloud agent all the way to a hosted link. It scaffolds the app, sets up Postgres and auth, then runs security checks before anything goes live. You still have to hit publish yourself. Writing code got cheap. Getting it off localhost was the part that stayed hard. This is the bit I actually want from these tools.

Meta bought Conductor, the tool that runs a bunch of AI coding agents in parallel workspaces. The agents themselves come from Anthropic and OpenAI, not Meta. Cloud for Conductor shuts on November 20. Core product keeps going for a bit, something new early next year. The transfer market is not only for model people anymore. It is for whoever figures out how to run ten agents on one project without them stepping on each other.

OpenAI launched GPT-6 Cyber yesterday. Critical under their preparedness framework, up from High on the last cyber model. ExploitGym went from 32.1% to 48.5%. The number I keep coming back to is the scope test. They put easy out-of-scope services next to the real target. Without the deployment prompt the model left those alone 99.93% of the time. With the prompt, every time. They already used it on Linux and Firefox, and Mozilla has been shipping fixes. Finding bugs with these models is clearly working. Keeping the agent inside the rules of engagement is the actual product. That is the hard part I care about when I give any agent more freedom.