Yipeng

Making sense of AI. Model comparisons, useful updates, and clear takes on what’s worth your time and money.

[Yipeng @Yipeng22]

$200 for Codex and still hesitating before picking Astra would annoy me.

Theo says GPT-6.1 Sol feels reasonable on that plan, but GPT-6 Astra feels almost unusable under the limits.

[Yipeng @Yipeng22]

Claude Code's experimental HTML-plan plugin shows UI mockups before coding.

“That's not the screen I meant” is a much easier mistake to fix while it's still a plan.

[Yipeng @Yipeng22]

@Bajanpunkw yep, same thing. the /review command now works with a plain chatgpt sign-in, no api key needed.

[Yipeng @Yipeng22]

Codex Auto-review is now free with ChatGPT sign-in. Anyone using it for long bug-fixing sessions?

How often do you still have to come back and approve something?

[Yipeng @Yipeng22]

OpenAI is reportedly raising $30B from MGX, UAE funds and BlackRock at about a $1.4T valuation. Venture checks got too small for AI, now the checks come from nation states.

[Yipeng @Yipeng22]

@Bajanpunkw it means erfans ran the same github rendering task through both. opus 5.5 finished in 5 min with page-by-page images, sol 6.1 took 8 with pan and zoom. paging wins for pr review, zoom is for exploring.

[Yipeng @Yipeng22]

Erfan's GitHub test: Opus 5.5 took ~5 min, Sol 6.1 ~8. Opus added image paging. Sol added pan and zoom. I'd pick paging for PR reviews.

[Yipeng @Yipeng22]

@JamesMalsawm that's the real question. i'd trust it on my code, not on prod.😂

[Yipeng @Yipeng22]

I keep seeing “built with AI in 20 minutes.”

I’d rather read the update a month later. What broke? What did users complain about? How much did it cost to keep running?

Who’s posting those follow-ups?

[Yipeng @Yipeng22]

“Fix this bug” shouldn’t come with a surprise refactor.

Anyone found a reliable way to get Claude or Codex to leave the rest of the code alone?

[Yipeng @Yipeng22]

@builtwithrubel living proof. 50 followers who reply beat 5000 who scroll.

[Yipeng @Yipeng22]

@JamesMalsawm right? it's only getting weirder.

[Yipeng @Yipeng22]

OpenAI will watermark eligible ChatGPT and Codex text in the EU.

But a watermark can't measure your contribution. Writing a draft and asking AI to polish it is very different from asking AI to write the whole thing.

[Yipeng @Yipeng22]

@Pranto39 the 90/10 split is real. everyone has a main.👀

[Yipeng @Yipeng22]

If you subscribe to both Claude and ChatGPT, do you actually use both, or does one get 90% of your prompts?

Which subscription would you cancel first?

[Yipeng @Yipeng22]

@Pranto39 actually good. the self-doubt is the only thing cooked here.

[Yipeng @Yipeng22]

@itsvlady no opt-out is the part that stings. cloud by default, choice by nobody.

[Yipeng @Yipeng22]

@b4dgerr be a redditor first, a founder second. help for a month before you mention your thing.

[Yipeng @Yipeng22]

@niuniu_dev early believers get remembered. that's the whole game.

[Yipeng @Yipeng22]

@niuniu_dev direction first, novel never. that's the workflow.

[Yipeng @Yipeng22]

Two AI workflows:

1. Spend 10 minutes writing the perfect prompt
2. Spend 10 minutes saying “no, that's not what I meant”

Which camp are you in?