MS.

Posts on X before Dec 21, 2025, 2:39:38 PM UTC

On X

@michaelstolarz ↗
@MichaelStolarz
I went back to using GPT 5.2 X-High instead of the new GPT 5.2 Codex X-High The Codex version feels much dumber, at least for the work that I am doing. GPT 5.2 -just gets it-, while the Codex variant needs to be handheld althroughout.
@MichaelStolarz
The only thing stopping me now from completely dropping Claude Opus 4.5 and going 100% GPT 5.2 Codex X-High is speed I've always been one of the "I don't care about speed as long as it's the smartest SOTA model"... but after fondling Opus 4.5 for a few days... I miss the speed.
@MichaelStolarz
Unlike trying to do creative writing with LLMs, coding is something that, for the most part, has a particular answer, if not an exact answer. Programming will be "solved" much sooner than creative writing getting to the point where you will preferably read AI-generated fiction.
@SamuelAlbanie
update to That Graph
Media attached to this post
@MichaelStolarz
Vibe code with opus 4.5 ultrathink Fix its tech debt with GPT 5.2 Codex xhigh Double check any missed tech debt with GPT 5.2 xhigh
@MichaelStolarz
I regret having tasted the speed and niche configurability of Claude Code + Opus 4.5. I know GPT 5.2 / GPT 5.2 Codex are smarter (xhigh versions), but damn it are they slow. By the time they are done, even though their output is perfect, I forget the direction of development.
@MichaelStolarz
when codex cli <-> atlas interoperability pls @ah20im @thsottiaux
@claudeai
Using the extension, Claude Code can test code directly in the browser to validate its work. Claude can also see client-side errors via console logs. Try it out by running /chrome in the latest version of Claude Code.
Media attached to this post
@MichaelStolarz
imagine mogging Dia Browser with: - a 14 second clip - of a vibecoded chrome plugin - it does your entire product's vision - executes it all flawlessly in a tab
@mattstallone
I shipped a feature to @posthog this week, except I don’t work at posthog and posthog didn’t approve it
Media attached to this post
@MichaelStolarz
one of my biggest issues with codex has been solved ever since gpt-5.2 xhigh. before: first prompts produced the highest quality results, but continuations felt rushed, guessed, non-important now: every prompt feels like the first prompt experience. i rarely /new anymore.
@MichaelStolarz
nvim and codex is all you need to be cool these days even cooler if you use codex headless feels like the same cool as using nvim in the 2010s and having a super riced out config and obscure plugins no one knows about but are secretly productivity gems cool != money though
@alxfazio
preferring nvim over vscode > cool preferring cc over cursor > not cool explain this. you can't
@MichaelStolarz
how has no one created a degradation benchmark yet? check how it performs at different times throughout 24 hours? on weekends? "lazy/holiday EU months"? I know there was a paper released roughly about this in 2023 but we're due for a way more tried and tested version now
@VictorTaelin
No there's no way this is the Opus 4.5 I used a few days ago. This thing is brain dead. It is completely moronic. I can't accept this is the same model.