Rahil

Want to see what’s at stake for AI data centers on Tuesday? ballot.dcmap.us Here’s what’s planned, what’s been approved or banned, what’s still pending, how officials voted and what candidates have said, wherever AI agents could find it.

dcmap.us crossed 250k visits in the last 30 days. Pretty cool to see where it’s turning up (thanks @bot and @ChatGPT ) There’s been a lot more news coverage, from local reporting on individual projects to national and international stories about the AI buildout. Journalists have been using the map and data to look at land development, power demand, and what all of this means for nearby communities. It’s also being cited in investment analysis of where capacity is being built, the opportunities in power infrastructure, and the grid constraints and political risks that could slow development. A regional planning agency has even built the data into its own planning tools. I’ve been sharing data with researchers studying everything from local water systems and community health to employment, election participation and waste-heat recovery. I’ve improved the data pipeline quite a bit along the way, largely through agents doing the research, analysis, deduplication and data cleanup. There’s a lot that goes into turning scattered announcements, filings and local reporting into something useful. And always another agent to spin up (and something to fix). It’s a bottomless pit for my ChatGPT and Claude tokens (and Grok and Muse contribute where they can) It’s pretty humbling to see people using it in their own work. I want to make sure the data stays current and the details are right. The research requests have grown too, and I’m terribly behind on getting back to people. If you’re waiting on me, thank you for your patience. Next on my list is an agent to help prepare the requested datasets and supporting notes, so getting the data to people doesn’t depend quite so much on my inbox.

Update: Forward Deployed is now live on the ChatGPT plugin marketplace. chatgpt.com/plugins/plugin_a… Go be your own FDE. I’ll be over here automating myself out of a job.

Some thoughts on 6.1 Terra. Aside from Computer Use, it is meh for everything one does in Codex. BUT it is a great model for agentic applications! Eval'd for some FS agents (recon, traversing legacy apps/DBs, KB retrieval etc.). Better, faster, cheaper.

Was actually excited to try the Meetings plugin in ChatGPT, but the results are meh. Getting much better transcripts and speaker labels from a rough pipeline feeding the same audio through OpenAI’s APIs (gpt-transcribe + gpt-4o-transcribe-diarize+review). First-party experience should be winning this?

As if we needed another reason to never doubt @elonmusk. On a Boston SF Delta flight, and he was right. Delta is the solution to sandboxing AI 😒😒 55 mins of Codex thinking without so much as a heartbeat. Now I’m logged out and the browser takes so long to sign me in that the auth flow breaks

as it turns out, heading the other way after a year. what’s fun these days, where does everyone hang out, who needs east coast data? can i wear a codex hoodie to an ant thing? so many questions

Claude Code got us hooked to turning all the experimental knobs on. But doesn’t work the same way on Codex. The experimental features are truly “experimental” lol. Going back to stock Codex.

Some more thoughts after using my dot pretty intensely the last few days: - Multiple dots, please. I want to split up work without everything competing in one endless chat. They should know what the others are doing and share relevant context so I’m not coordinating all of them myself (like bots) - Multiple computers too. Some files are on my desktop, other stuff is on my Mac/server/mini/work. Let it use the right computer and keep working when I switch devices - Show me what it’s tracking. I genuinely have no idea sometimes. What did it agree to do? What’s running, blocked, waiting on me, forgotten? Give me a task list I can reprioritize, park things in and actually steer - Giving it a new task shouldn’t make it drop the previous ones. I should be able to throw a few things at it throughout the day without having to keep my own list of what it’s probably forgotten - The permissions situation needs work. I had to go into EACH Codex thread it launched, change permissions and open the in-app browser myself. Why am I doing setup for every thread my agent starts? - And then some of those threads still couldn’t use the browser. It needs to know what access it has, whether the browser actually works and what’s broken. Asking me to repeat the same setup isn’t getting us anywhere - Login handoffs should be simple. Get me to the login, let me sign in, then carry on. Same for approvals or anything else that needs me. Don’t make me hunt down the thread and tell it to resume afterwards - For Codex specifically, I want actual multi-turn steering. Read what came back, challenge it, send it back with useful feedback, check the revision. I had to ask repeatedly for this on a thread. It kept coming back with a status update or mid checkpoints. The whole point of having a dot orchestrate is that I shouldn’t have to manage every turn - Keep the overall objective through those revisions. Fixing one thing shouldn’t undo three earlier improvements. - Look at the actual output before calling it done. Open the file. Look at the layout. Try the website/app. A file existing is a very low bar. - Give me visibility into the threads underneath. What’s each one doing, on which computer, with which model? Let me jump in directly when I want to, and make sure dot picks up whatever I told that thread. I shouldn’t have to relay my own feedback back to it - Keep going when I’m away. When a thread finishes, review it and take the next step. When I’ve unblocked something, resume. Another message from me shouldn’t be the thing that reminds it to work - Better notifications. Tell me when something needs me or there’s something worth seeing. Group the routine updates. And if I’ve asked it to watch for something important, actually catch it - Show me usage across all of its work that it delegates out. Models, limits, where the usage is going. Let me spend more on the work that matters and less on routine checks. Especially once there are multiple dots and a bunch of Codex threads running

a message from my dot. currently we take turns being the assistant🙄 @TheRohanVarma @victornunez @thsottiaux