nwptz.eth

💼 Solo dev, experimenting with @zarosabe
🏰 I run a small game guild

[nwptz.eth @nwptz]

I launched @zarosabe.
Not to chase the next meta, but to outlast them all.
If narratives are temporary, then ownership shouldn't be.

The experiment begins.
Built on @base via @clanker_world.

CA: 0xeB68C7efC185965199f5C260F15D5ebA3340db07

Full thinking here:

[nwptz.eth @nwptz]

https://t.co/6UlE4UkqB0

[nwptz.eth @nwptz]

The most critical part that codex needs to improve is context compaction.

If that can be fast, reliable, without timeout or nasty reconnections,
thats actually the best selling point.

[nwptz.eth @nwptz]

I still believe that the real winner for AI providers battle still going to be the one that:
- super cheap
- not too smart
- can do basic needs for consumer

in the end, most people in the world still need help for basic needs.

high IQ model is not public needs.

[nwptz.eth @nwptz]

"Codex is letting you to finish the task" is not an unlimited use.
it bypasses the 5h limit by eating your weekly limit.

So once your weekly limit is gone,
you still get the limit error and it stops.

summary: don't use /goal on your $20 plan.

[nwptz.eth @nwptz]

UI Tips with GPT models:
never use GPT-5.5 for UI job ever.
it has the similar capability as 5.4, same potential mistakes, and more creativity that causing drifts.

And it also more expensive (2x price).

better stick with GPT-5.4,
or use Claude.

[nwptz.eth @nwptz]

it's actually getting worse.
even a new thread is RTO. @OpenAIDevs

how to handle this misery?

[nwptz.eth @nwptz]

context compaction is now always RTO on codex.
More people are going back to Antrhropic I guess?

[nwptz.eth @nwptz]

People with simple tasks should never use GPT-5.5.
It's still can make mistakes, and it's expensive.

use GPT-5.5 as orchestrator to select the correct model for your task. That's why agent loop is important.

or just you chinese cheap model for your simple task.

[nwptz.eth @nwptz]

context compaction is now always RTO on codex.
More people are going back to Antrhropic I guess?

[nwptz.eth @nwptz]

@OpenAI how codex-cli users trigger that banked reset?
need to open codex?

[nwptz.eth @nwptz]

long thread basically wasting time even you do compactions.
my gpt-5.5 works like hours facing websockets error mostly, and that shit only happens on long thread.

[nwptz.eth @nwptz]

- Codex is slow on long thread.
- OpenAI banning some payment methods paying users, but adding reset limit feature that abuseable via free account in exchange to pump fake number codex usage pre-IPO.

Looks like June is Anthropic month.

[nwptz.eth @nwptz]

Soon, GPT will follow Fable.
API only for frontier, and you will only get that GPT-5.5 expensive model via subscription. kek...

the gap for frontier model is coming.

[nwptz.eth @nwptz]

Just gave my codex a broader target, finally it suggests to use 4 subagents. First time.
Usually only 2 max.

[nwptz.eth @nwptz]

Addicted to AI. I have several series/movies on my Netflix still not watched.

[nwptz.eth @nwptz]

@ShanuMathew93 you have 1 agent (loop master), controlling other agents.

You give loop master a goal.
It then controlling/prompting other agents to reach that goal.

So, your job is only defining the goal.
Loop master will report back once the goal reached.

[nwptz.eth @nwptz]

@indrawxyz sayangnya gpt go selevel sama gpt free di codex.
limit beda tipis sama free dan sama2 reset bulanan.

[Tyler @rezoundous]

Working with GPT-5.5 be like

10min backend, 20hours UI

[nwptz.eth @nwptz]

@rezoundous lmao, yeah..

[nwptz.eth @nwptz]

@simpsoka turn start to include thread title creation using model concise. this is a good ux. Most people are often renaming.

[nwptz.eth @nwptz]

codex on snail mode.
each turn is super slow!

[nwptz.eth @nwptz]

Damn ETH keeps bleeding