


Cognition Devin Ambassador | CS/Math grad | Dad x2, wife-powered. I teach AI to humans, not the other way around. Daily builds with agents, wins and fails.



Devin Bots Just Dropped. I'm Letting Them Run My Repos x.com/i/broadcasts/2zkZqZkQp…

Gemini 4 Carbon is now open to developers, enterprises, and Google AI Ultra subscribers. Google posted that on November 10. It is available through the API in Google AI Studio and Gemini for business, at $2 per million input tokens and $10 per million output tokens, the same introductory price as Gemini 4 Argon. For Ultra subscribers, it is rolling out in Google Antigravity, the Gemini app, and AI Mode in Search. I have not tested Gemini 4 Carbon. That's Google's announcement, not a review.


Most people will use this like a faster intern. Wrong move. You don't hand a Devin Bot a ticket. You hand it a job to own, and it spins up its own team of Devins to get it done. Going live tonight to build with these things. Real repos, real Bots, what works and what breaks.
Meet Devin Bots. From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.

Come join the space! Great discussions about what can be done with Bots of all shapes and sizes. Coolest host and cohost.

Almost all of my subscriptions have screwed me in some way, so one of my Grok Bots is Refund Reaper. Say it's $20 for a service you never got. They still charge it, and it just goes to their wallet. As everybody automates everything, I think more money gets taken and never finds a home. From my Monday morning live.

Excited to try 6.2 Sol Ultrafast. Lets burn some tokens.
6.2 Sol and Astra are out. These models can do amazing things and are very token-efficient; we hope people will get great value from them. In particular, 6.2 Sol ultrafast is something I think people will really enjoy.

X added a second scoring formula to its open source algorithm last night. It's switched off. But if they flip it, a like on a post people actually read counts for more than a like on one they skim. I went through the code to see what's real. Today your score is a sum. X guesses how likely each person is to like, reply, repost, stay, etc. Each guess gets a weight and it all gets added up. The default setting is still "value_model_mode: weighted". The new one multiplies three things: "skip_score * engage * time_factor" Engage is every other guess added up. Likes, replies, shares, follows. Blocks and reports count in there too, so a post people hate doesn't get saved by read time. Time is "1.0 + dwell_time" times a weight. Skip is supposed to grow the less likely people are to scroll past. Right now "not_dwelled_power_exponent" is 0.0, which makes skip a flat 1. Does nothing. The 3x claim is legit, but read the fine print. X's own test uses "cont_dwell_time: 0.1" and a 20 second read. 1 + 20 x 0.1 = 3. The live weight is 0.0075. With that, 20 seconds is 1.15x. You'd need about four and a half minutes of expected read time to triple. If they flip it on they'll probably retune it anyway. The direction is real though. Since Oct 13 likes went 0.5 to 0.708. Time per second went 0.004 to 0.0075. Clicks slid to 0.274. Copy link went from 20 to 0 on Nov 13. So what does it mean if you're small? Most of your reach is people who don't follow you, and those posts already get "OonWeightFactor" 0.75. You can't win on how many people see it yet. You can win on whether the ones who do stay. Multiply mode would make that the whole game. The small account boost is still there too. "ColdStartFollowerCap" 50000. Under 50k followers, a fresh original post in its first 2 hours with under 200 home views can get bumped to about slot 16. Early likes per view decide who gets it. Big accounts still have their own lane. "TOP_POSTING_AUTHORS_FRACTION" 0.00005, so the top 0.005% of posters by followers feed a popular posts pool. Nothing changes today. When it flips I'll post it that day. Until then write stuff people don't want to scroll past.

I pointed a Grok Bot at the exact phrases people post on X. On November 13, "ground stop" had about 50 times the posts of "flight delays." My post that night used neither phrase in the first line, and it still did about 6 times my usual reach. So the lesson is narrower than "the algorithm is search." Use the phrase people are actually posting, and don't guess the cousin.

Big accounts don't have big egos. They have fragile ones. A big ego would cheer on smaller creators. It has nothing to lose. A fragile ego guards every impression like it's the last one. No likes. No replies. No reposts for anyone who might catch up. Real confidence lifts people up. Insecurity gatekeeps. Watch who engages with small accounts. That tells you who's actually secure.

Small accounts on X: your posts aren't being read. They're being buried. If no big account touches your post in the first hour, it's dead. Good writing doesn't save it. And replies? They get almost zero home feed impressions. So why reply? Because replies are the only door left. They put you in front of someone else's audience, and that's where profile visits and follows come from. Those followers are what keep your next post alive. Being a reply guy isn't cringe. It's the only strategy that works when you're small. But here's the dark side. If early reactions from big accounts decide your reach, then a big account that follows you can sink a post just by seeing it and scrolling past. No like. No reply. Your post stalls. And if they actually cared, they'd turn on notifications for you. One, so they see your posts. Two, so they like them and your posts travel. It costs them nothing. They choose to be blind.



I built a fantasy league where the players are X accounts. Draft five public handles. Likes are runs. A 100k post is a home run. A log handicap lets a small account take a week off a whale. If you already treat X like a sport, this keeps score. xuideck.com