Posts on X before Aug 19, 2025, 12:00:52 PM UTC
I hate gpt 5 high so much
because it has made every other LLM unusable
it never lies to me,
never makes the wrong decision,
tells me when it doesn't know/needs more info.
damn. thanks a lot openai.
pls ser i take back what i said about u anthropic pls i have 1 children to feed with my plugin and it needs updates pls 🥺
so weird to see this in gpt 5's thinking traces
thankfully it keeps going?
new v2.0.0 of systemsculpt out
lots of stability updates, byok, ease of use, standard/advanced modes, yeah lots of cool stuff
more official release info later today!
so cool. can someone show me the usa equivalent? no? why not? why does china have 40 types of robots from 20 different startups and we have 3 can-barely-walk-but-marketed-so-well slop bots?
i hate ironing out the final bugs
they're final bugs for a reason
this is how much time there's left to use gpt 5 high fast on cursor for free btw
lets take sonnet 4 1m for a spin
how did they greenlight this for release
i just know sama is gonna get banned soon
that or elon is about to go demon mode in everything else
what have they done to my boy chatgpt, omg
waiting for this
until then im using gpt 5 high
the reason that gpt 5 high (specifically talking about coding here) is the current SotA, is that it doesn't lie
aka doesn't hallucinate, or at least nowhere near as much as ex. opus 4.1 or sonnet 4
the simple fact of lie % is what pushes its efficacy beyond the horizon
why is chatgpt practically useless now? every answer it gives is trash, made for speed of response instead of quality of response, even when you run it in thinking mode it takes 10-15 seconds at most before giving you a slightly better slop answer, wtf openai
this memory update i'm cooking up for systemsculpt... it's like nothing anyone has experienced before
not claude code, not gpt 5, not cursor. there's a reason why i've studied the blade in silence for these past few months with little to no interaction with others
chatgpt's 4o finding out it's joever
being replaced by 5
good night sweet prince
first a brain dump, then my top 3:
for it to be more of a pleasure to spend hours on end within. concise vs. verbose tool call output mode. planning mode (or commands with ability to constrain tools ex. no editing, only researching). command timeouts (current one 800s deep)...
OH MY GOD
i just dont get it, why is the open source claude code killer codex still absolute garbage when it comes to tui?
@OpenAIDevs let me show you a few examples, a 🧵:
thank God they fixed naming though that's a relief, it's so intuitive as to which to choose, i love it
this sounds scary
i first felt this with o3 in cursor
it would read 30 files, make 3 edits, then say ok im done
i have a feeling it will be like that but on steroids
will it be unlimited use with the $200 plan? please?
these prices are mind boggling
this is basically:
- o3-level intelligence
- flash-level pricing/speed
btw if you're trying to use the new opus 4.1 dont forget to input:
/model claude-opus-4-1-20250805
at least for me, when i do /model opus right now, it still uses old 4.0
with opus 4.1 there's now no good reason to use sonnet 4 (given we're talking about max20 claude code plan here)
it used to be that sonnet 4 could potentially outperform especially with speed opus 4, but seems like opus 4.1 is across the board better now
keep your memories clean
store everything as markdown/plaintext
this is a short term victory but a victory nonetheless
in 5ish years though? it wont matter
you'll just swarm xyz application you used to use,
it will consume every bit of data it possibly can
and migrate you easily
you can now live in the cli btw and do everything from there
i mean you always could, but now you can make fully custom fine tuned golang tview interfaces perfected for your exact daily workflow... in about a few hours worth of tinkering
while i agree, the problem is
until openai catches up with...
1. a coding model surpassing unlimited Opus 4
2. a CLI 2x better than Claude Code (codex is TRASH)
3. aforementioned unlimited coding model included in their already-existing $200 plan
...we're stuck with Daddy Dario
some of anthropic rugpulls so far: > windsurf no access to claude 4 > plus not getting opus 4 in claude code > 1.58-bit quantized models during daytime > max plans limits cut in half 2 weeks ago, no comms > weekly limits without concrete numbers > 5x/20x plans being actuall… Less
some of anthropic rugpulls so far: > windsurf no access to claude 4 > plus not getting opus 4 in claude code > 1.58-bit quantized models during daytime > max plans limits cut in half 2 weeks ago, no comms > weekly limits without concrete numbers > 5x/20x plans being actually 3x/8x of plus > cutting off openai api access what an absolutely horrible company
this is an amazing example
of thinking that you're being super productive and making use of your max 200 claude code plan
the reality is this is much, much slower than one instance per project
regarding the straight to main and skip perms - that's totally fine, i do that too
unfortunately they are still dogwater at coding
yeah they can use a few tools and one shot things pretty ok but they totally fall apart further with every follow up prompt you make to it
Welp, time to do another all nighter and see what YET ANOTHER @openrouter stealth model is capable of.
claude code tip
@Maciek_roboblog made the amazing claude-monitor
most accurate and informationally useful tool i've found so far to quickly understand where you're at in terms of usage
if you pay attention
the llm psychosis farm is leaking out
on one hand,
funding would speed up progress at least 5x
on the other hand,
there is no "other" hand, both hands are pro-funding
crazy how we got gpt-5 level gpt-5
AND
gpt-5 level video generation
in the summer of 2025
a summer to remember
you have to be so rich to be rich,
even if you're rich, you're not rich;
you have to be AT LEAST wealthy to be rich nowadays.
feels like the open source "you can run it at home" model that openai has in store for us all
first impressions, yeah it's a SMALL model, but damn it if i say it's actually pretty good
like, able to vibe code on the plane with no wifi good and actually get some work done
every time i think im maybe shitposting a bit too much
i see that u can still see the exact number of tweets when u look at my account profile
vs. REAL shitposters have their numbers in the tens or hundreds of Ks
surely i'll hit minimum 10K before EOY
they're coming
quick, hide the thinking sand we made
lets pretend we're still at the nukes stage
weird stuff going on with claude code's session limit resets
twice today my claude 4 opus has reset at 4 hours, not 5
im not complaining but idk feels weird
and no it's not setting itself as sonnet behind the scenes, i thought so too but nope, still opus
did mark see something that explains why he's throwing billion dollar offers for single persons?
i have a feeling he has no care in the world when it comes to "paper money" as people call it
his eyes have seen behind the curtain
he's seen the loading screen for the next stage
claude code tip:
dont run agents in parallel. i've tried in many cases, the lack of them sharing information while in parallel really nerfs their efficacy.
instead, let your agents know they're able to run themselves/other agents at another level deep:
i was checking out, saw the cashier using chatgpt. "ah, chatgpt! everyone's using it now instead of google, huh?"
her: "haha yeah, it really helps me plan my homework out so i get the hard stuff done first"
me: "cool! do you use regular 4o with that or o3?"
her: "what's that?"
"contributing compute resources to advance critical research"
so this is where my unlimited claude 4 opus went
it's funny seeing people use tier 3/4 models (think gpt-4o-mini) to do mathematical predictions
they repeat it with such confidence; as a child boasting to new classmates on their first day of kindergarten
"did u know my dad is the smartest in the world? look what he said!"
why run multiple claude code instances when you can just run multiple parallel subagents
unless you're splitting your attention between multiple projects which in that case i wouldn't recommend doing that in the first place if you want high quality maintainable results
uh today i learned that claude code sub agents can call sub agents of their own
i wonder how many layers that can go
it's different than gif
because of the e instead of i
jepa is jeh-pah
gepa is gee-pah
this will never be reality again btw
it'll be mentioned the same way we mention how people handcrafted written letters, so carefully, lovingly, including a lipstick kiss or a dried daisy
future generations will say "aw cute, you used to just, like, talk to strangers before?"
the bright side of this heartbreaking anthropic announcement of rate limiting claude code:
openai perfectly positioned to add unlimited codex cli to their 200 plan