Kubi

using AI for fun, impact and profit

|

[Kubi @kubi_dev]

bangkok is incredible.

you go to a rooftop restaurant in a five-star hotel and ask for the drinks menu.

the prices look way too high for shots, but way too low for bottles.

turns out they are, in fact, bottles. bottles from the 7-eleven downstairs.

you order one and the waitress literally goes down to seven and brings it up to the rooftop.

absolute civilization.

[Kubi @kubi_dev]

@thsottiaux sounds like threatening

[Kubi @kubi_dev]

@cjzafir @thsottiaux those models are retarded from the beginning. so nerfed is not valid word to describe them. there are just retarded

[Kubi @kubi_dev]

@thsottiaux impossible, bc those models are ignoring what we want and building their own dreams. each fucking time

[Kubi @kubi_dev]

@jxnlco amazing, but we don't care

what we care is to get our shit done the way we want it to be done

make those models follow instructions

every single task done wrong way, ignoring explicit instructions

tired of that shit. make it work or return our money

[Kubi @kubi_dev]

@evi77ain @OpenAIDevs @ajambrosino @romainhuet @jxnlco @thsottiaux @reach_vb use agents .toml config for workaround

[Kubi @kubi_dev]

something that codex was amazing but getting terrible in last week: context management

i use codex for around 1 year. for most of that time i never saw autocompact

last week is happening all the time. once with spark i got 3 autocompacts before any change was made

probably bc system prompt adjustments for 5.6

[Kubi @kubi_dev]

@thsottiaux pls provide 5.6

[Kubi @kubi_dev]

@reach_vb i guess monitoring means making sure new cache time limit for older models is now 30 minutes as well

[Kubi @kubi_dev]

i guess this is the answer. they probably limited caching time for all models and now "they are investigating"

[Tibo @thsottiaux]

Ola. The Codex team is investigating issues where some accounts are seeing faster usage draining than intended. We believe this is related to some abuse & fraud prevention mechanisms that we have that might be overflagging.

Stay tuned.

[Kubi @kubi_dev]

@iuditg this maybe the answer - probably they reduced entire infra cache time to 30 minutes (from 24 hours)

[Kubi @kubi_dev]

@thsottiaux amazing

next time when adjust system prompt for frontier model give us access to it. now you made 5.5 retarded without way to use 5.6

[Kubi @kubi_dev]

@mustafaergisi good timing XD

https://t.co/PikJvSeq3E

[Kubi @kubi_dev]

today was mostly fixing cursed infrastructure and making the brain less stupid

first:

back in fbc hell

again

because apparently facebook attribution is not a feature

it is a recurring boss fight. real AGI test. codex is failing so far.

also started preparing the ceo ai to be actually useful

not just “chat with vibes”

but memory

knowledge base

access to production db in read-only mode

so it can understand what is happening in the business from real data, not from me manually feeding it crumbs like a pigeon

bonus:

codebase is now embedded with gbrain + voyage

so coding agents should navigate faster, use fewer tokens, and stop rereading the same files like goldfish

and we started redesigning the admin sales funnel dashboard

especially quiz attempts

because the current dashboard is too noisy

too confusing

and too good at pretending fake numbers are insight

goal is simple:

less dashboard soup

more clear funnel leaks

fix tracking

give ai memory

make admin useful again

[Kubi @kubi_dev]

another round of making the funnel less stupid

big thing:

we laid the groundwork for the memory check free trial funnel

instead of asking warm quiz traffic for money immediately like a goblin with a stripe account

the product now has a cleaner path:

finish quiz

see result

enter free trial

then decide

much better than “congrats on testing your memory, please pay now”

also tightened the whole post-quiz lifecycle around that path

result copy

email labels

next step

all starting to feel like one product journey

not five random marketing gremlins shouting from different tabs

manychat also got smarter

we refined follow-up routing and eligibility so we only keep nudging people when it still makes business sense

no chasing dead leads

no poking contacts who should not be pushed further

less “automation because automation”

more “this touchpoint might actually make money”

also calibrated memory check scoring and share behavior

because the result has to do multiple jobs at once:

feel trustworthy

make sense to the user

support sharing

bring repeatable traffic from the satellite pages

tiny product surface

big funnel job

and finally fixed some small landing flow friction

mobile ctas

locale defaults for social previews

little boring stuff that nobody celebrates

until it breaks and your ad traffic lands confused

startup life is mostly this:

find the leak

patch the leak

discover 7 smaller leaks behind it

repeat until the funnel stops being a haunted pipe

[Kubi @kubi_dev]

@iuditg @OpenAIDevs @OpenAI @thsottiaux i felt that pain. fixed with moving from 20 to 200 usd plan

[Kubi @kubi_dev]

@jxnlco on plus plan I would at least try auto-model assigment to safe token

on pro it's just xhigh

[Kubi @kubi_dev]

@yashhq_22 forcing codex to properly implement pixel - so both

[Kubi @kubi_dev]

@Dimillian whats the bill?

[Kubi @kubi_dev]

@jturntdev yeap, they gave you 2-3 resets this month, but probably decreased weekly limits so you get less

https://t.co/ccx8LjBqBL

[Kubi @kubi_dev]

@thsottiaux what about being okayish in usage calculations

those screen are separated by 2 minutes of work

[Kubi @kubi_dev]

@thsottiaux what about being okayish in usage calculations

those screen are separated by 2 minutes of work