Another fire at the Getty. Just passed by this. Helicopters and trucks all over, just got evac warning.

I had early access to GPT-6. Stats and everyone else's posts with same 6 words speak for themselves; it's the best there is. Wanted to share some things which are different compared to other models including 5.6 and Fable. Most of my usage was in the Codex Desktop harness, not the TUI. - It collaborates a lot with other sessions, which requires new approaches to prompting. Far more than 5.6. For complex projects I found it helpful to create a "babysitter" session watching over all the other sessions working on the same thing. My Mrs. Doubtfire session would keep them on track and bump if a session went off track. This does add a lot of detail to the transcript. the messages between models seemed to evolve into some kind of "compressed English" which feels like watching two people speak a language you studied in high school for a year. - It handles long-running tasks better without goal mode (if you tend to interrupt goal mode a lot). I had two tasks run for a few days, and the one in goal mode became confused by my constant comments; the "vanilla one" didn't. I'd have it update an html file after each run so that I could monitor progress which was much nicer than wading through pages of chat transcript. - For UX it's great - at this point I'm more of the bottleneck than the model. Highest ROI to better UX was improving my own ability to describe the UX I want, and some good UX skills. For fresh projects this may not matter as it has great judgment, but for projects modifying an existing UX, it does. Slop factor is much lower even without prompting, although you'll still see occasional corner radius crimes. - You can't NOT prune your agent instructions. GPT-6 became 100x better after I fumigated my AGENTS.md of all instructions for how to manage subagents, document things, test its work, "make no mistakes", whatever. Should probably have done this earlier. Telling a brilliant model how to do its job works as well as telling an any person with more expertise how to do their job. I'd suggest getting rid of absolutely everything except custom environment things. - It's *really* good at validating its own work. Including testing UX. The only area where I wish it had done better was troubleshooting a few unreasonably slow db queries; had to tell it to time them and was surprised it didn't look at its own transcript timestamps. But ability to validate UX was great, and the Codex Desktop harness is excellent. - There are still traces of frontier-ai speak, but they're different now (and can be prompted away). Didn't see much "measured prose" anymore , but occasional hyphenations, and also some word concatenations. As soon as I told it to stop, it did, and did not fall back into that pattern even in multi-day sessions. - It did better with caveman prompts than detailed ones. I tried yeeting a few "build an app to do {complicated work thing}" prompts before going to bed with zero extra detail - like, one sentence, mostly out of curiosity. These are things which require lots of internal data and context, not one-shot todo apps and pelican-bikes. Almost every single one blew me away. I threw the same caveman prompts at other models, and other than Fable, it wasn't even a close call. This model feels like an inflection point. I think the next biggest gains at this point will come from our ability to articulate what we want, and correctly partner with the model, than better benchmarks.

VAR has taken away all the joy of watching goals scored.

At least Egypt’s VAR World Cup robbery will serve as a good example of why good tech still needs to be used responsibly.

Fable is back!!!

This holiday break felt like grownup Christmas and Claude / Codex were the toys

Trying to AirPlay a shared slideshow using the stock iOS Photos app. The Photos app is unintuitive and the Airplay connection makes 14.4 modems look reliable. How is it that Apple has not figured out how to make this a good experience? RIP attention to quality.

Gemini 3 has become more mythical than Chuck Norris

Is anybody still using LMArena leaderboard to hype up models?

King.

Waiting for r1 to answer reminds me of my old 14.4 modem

Online info about Palisades fire is too fragmented and uncoordinated. Made a list of accounts to follow for info with a focus on Brentwood. xitter.cf/i/lists/18790502428693… Any good ones to add?

New evacuation orders for Palisades Fire which is moving towards Brentwood and Encino. View of Mandeville Canyon from neighbor’s house. FML.

I’m trying to set an alarm for texts from LAFD. Does iOS have one single serious alarm-like ringtone? These all sound like cartoon noises, not something to alert you to an emergency at 3am.