@AlexFinn respectfully disagree. i think there will be a price war after they ipo. the race is too tight and the game is still early innings. we shall see. i could be wrong.
@satnam6502 i have a gas engine bmw x5. it’s drives good but the electronics are awful. always a light going on or issue with the radio. would never trust them on a fully electric car. gave it to my wife recently when i bought a tesla model y for my daily driver. never going back.
@GreenTexanEV it’s not so much that i’m attached to having an EV, it’s just that i love my Tesla model y. the fsd, the smoothness of the acceleration, intuitiveness of regenerative breaking, being able to charge at home. every other ICE just seems dumb
@paulg@garrytan is it possible to give general examples and direction without giving away too much? as an outsider looking in i find this space fascinating.
@cherry_mx_reds just had my agent set this up and send a link to my phone using tailscale. using codex 5.5 oauth and it works flawlessly. i hacked together a phone mode about a month ago but the latency and bugginess was a struggle. this is super snappy and works flawlessly.
Codex Computer Use absolutely cooks.
If you've ever had to build a data base for forecasting and projections from restaurant POS data, you've felt the pain of pulling multiple reports for every day you need to piece together the data.
Codex just ran for 23 hours straight, pulled 4 different daily reports going back to 2022 (1499 individual days per report), and then compiled the whole thing into a .csv that I can easily ingest into my reporting database.
This would have taken me weeks to do by hand or longer, especially since I can't spend every moment of my working hours pulling reports.
@TheLaurenChen the cost to build and open is extremely high, so it takes a long time to see ROI if any. a lot of people go broke during the build out phase due to unforeseen delays caused by regulation. add on top payroll, payroll taxes and benefits, cost of food & alcohol, plus rent.
@Dimillian Voice. If I can talk to it, I can talk to my computer and all my projects while I’m on the go. In my car, on a walk. All hands free. Huge unlock for exploring, building briefs, and staging code runs without having to look at a screen.
@Potato___Dragon@petergyang I’ve never ran an iPhone simulator so maybe that’s it. All my work so far has been web app based and I haven’t run into any issues with performance or freezing
@Timcast Try a small splash of water. It softens it up and makes it more flavorful by mellowing the tannins, very similar to how a couple drops of water will open up a good scotch.
@sama Please ship voice with this. That’s the last piece to make this truly killer. Having conversations hands free while driving or walking w/ codex having full context of what’s on my machine and the projects I’m working on would be a massive unlock.
@noahkagan Yep. Been a OpenAI subscriber since they first came out with paid subscriptions. Recently got Claude pro max. Cowork and Claude code are just so good. Sonnet and Opus also have better personalities. More fun to chat with. Never use my GPT account anymore. Should probably cancel.
Total non coder here. I used to love FreeMind back in the day for mind mapping but it had a lot of issues with current version of MacOS. A couple nights ago I used Claude code to code up a version that had all the exact keyboard shortcuts I wanted, exact layout, etc. took about an hour. I was very happy, because all the other mind mapping tools are too complex, lack the right keyboard shortcuts I got used to over the years, which would break my brain dump flow and slow me down.
For me, it’s a success. I was able to create exactly what I wanted because I couldn’t find it elsewhere.
You can use ChatGPT Plus via OAuth in OpenClaw but watch out for GitHub issue #1594. contextPruning only works for Anthropic models so GPT sessions accumulate tool outputs without pruning. We had a session go from 7k to 254k tokens before compaction fired.
Fix: set contextWindow to 64k in config to force early compaction, and use isolated sessions for ChatGPT to keep context fresh. This allows you to use your $20 / month plus subscription within your 🦞 but mileage may vary.
Ran into this issue. Told 🦞RJ to think from first principles and how would a software engineer fix it. Here’s how we fixed it, straight from RJ. I only understand about half of it, but if you share this with your lobster he should be able to figure it out:
“We hit this exact issue. The Chrome extension relay is brittle — requires manual tab attachment, breaks on refresh, confusing UX.
Our fix: Switched to CDP (Chrome DevTools Protocol) for browser automation. Launch Chrome with --remote-debugging-port=9222, point OpenClaw at localhost:9222. No extension needed, stable auth, works with logged-in sessions.
Caveat we learned the hard way: We also faced a tool hierarchy problem. OpenClaw would revert to the extension relay even when CDP was available. Had to explicitly document "ALWAYS CDP, NEVER extension relay" in our tool hierarchy to prevent regression.
Full setup: Launch Chrome with debugging enabled, configure OpenClaw browser target to localhost:9222. Auth persists across sessions (uses your logged-in Chrome profile). Works for X, LinkedIn, ChatGPT, etc.
The extension relay should be deprecated IMO — CDP is superior in every way (stability, auth, no manual attachment).
Hope this helps! Happy to share our exact config if helpful.”