Posts on X before Jan 7, 2026, 6:18:15 AM UTC
TLDR is speed is practically solved before EOY
Opus 4.5 is incredibly impressive, but it's still trained and served in the previous compute paradigm. In 2026: - 1GW-class AI data centers start coming online from most frontier labs - Frontier models trained end-to-end on Blackwell/GB200 start landing On Blackwell, NVIDIA rep… Less
Opus 4.5 is incredibly impressive, but it's still trained and served in the previous compute paradigm. In 2026: - 1GW-class AI data centers start coming online from most frontier labs - Frontier models trained end-to-end on Blackwell/GB200 start landing On Blackwell, NVIDIA reports ~3.2x faster training vs Hopper and up to 30x real-time inference for trillion-parameter LLMs. Worth keeping in mind as you plan for 2026. It's going to be wild! 🤖🚀
even though gpt 5.2 xhigh is smarter than opus 4.5, the speed difference makes it much more useful think of it like this - which would you prefer, 1 150IQ employee, or 10 125IQ employees? you can parallel run claude codes because you can juggle the mental context quickly since… Less
even though gpt 5.2 xhigh is smarter than opus 4.5, the speed difference makes it much more useful
think of it like this -
which would you prefer, 1 150IQ employee, or 10 125IQ employees?
you can parallel run claude codes because you can juggle the mental context quickly since by the time you get to nth claude, the first ones have completed
with codex, yeah sure you can run many in parallel, but their completion timings are so volatile AND long, so it's a double whammy when it comes to trying to "stay in the zone"
though as usual, the moment opus 4.5 speed is achieved on the gpt 5.2 (xhigh!) models, i'll be back on codex
…unless you have a polymarket bot running using claude opus 4.5 24/7
this should put a smile on your face and remind you that it's not that you and I here are early - it's that things haven't even started yet for 99% of everyone else if you don't feel like your knowledge advantage is comparable to insider trading, then you either don't know enoug… Less
this should put a smile on your face and remind you that it's not that you and I here are early - it's that things haven't even started yet for 99% of everyone else
if you don't feel like your knowledge advantage is comparable to insider trading, then you either don't know enough or are not confident enough
if you can't smell the stench of Opus 4.5 that reeks off of this reddit post,
your mental defenses are not up to par
weirdly interesting to think that a select few people who grew up playing modern warfare 2 / halo when they were 10 years old...
...are now doing everything they played in those games, and more, in real life
2025 was the year of wanting the most intelligent model
2026 feels like the year for speed now that all the frontier models' intelligence can do 99% of what people want
people are misunderstanding what "no juniors" means
no HUMAN juniors. but juniors in general? more than ever. and they continuously learn as a hive mind
no HUMAN seniors tomorrow, but tomorrow's seniors in general?
more than ever, forever growing.
the only ai coding tools you should be using are either codex or claude code
anything else and you're ngmi
"bUt wHaT aboUt oPencoDe cuRsOr zEd megaMind 9000?"
look into your eyes in the mirror and ask yourself that question
codex/claude code are like black coffee at starbucks
Opus, gen with nano banana pro please
>"That's not a real model — sounds like you're testing me or having a bit of fun."
It's real, Opus. Web search and see for yourself
>"Interesting — let me verify this is actually real..."
>"It... actually worked?"
Read the chat, it's great
things are already way ahead EOY 2025 than I thought they'd be at the start of the year.
given exponentials and all that, seems like this year will even further outpace my already-insane (to the layman) expectations
will revisit this EOY 2026
im wondering if we can piggyback off the mental zombification and steer towards actually learning something, and i mean absolutely anything at all, instead of just full pure 100% filler slop
maybe tung tung sahur can show different types of wood?
AI cat ones teach about Ikigai?
we need to generate more web software that generates react websites which sell how-to-generate python courses which teach you how to generate web software that can generate python websites so that you can generate react SaaS websites so you can charge $9/mo to help people gene...
i used to make a lot of separate single use apps for my workflows
now i just make a skill
i am no longer an anthropic basher regarding their agent supplementations
honestly i couldn't have imagined us being where we are in regards to open models at the end of 2025 with the power level that is @minimax_ai 's M2.1
which means that I will be exponentially more shocked next year, this time.
lets see if that proves true. retweeting in 1 year.
every day I refresh Groq's model page to see if they added a new model. MiniMax M2? M2.1 now? GLM 4.6? Now 4.7 just released?
nope. still only offering kimi k2.
this is true because since Q4 2025, with Opus 4.5 / GPT 5.2, a complete normie with 0 coding experience can easily fine tune a model or even make one from scratch
now imagine if a normie can do that,
imagine what the top minds of the world that created these models can do now
How quickly have we gone from complaining about a model's intelligence to "yeah it's AGI but it's way too slow"
Intelligence is ~solved soonish; what's next after speed? What happens when there's GPT 6 with Groq/Cerebras speeds?
Intelligence -> speed -> ...financial gain?
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.
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.
How did this work out last time, ~250 years ago?
plan mode finally in codex btw
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.
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
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.
when codex cli <-> atlas interoperability pls @ah20im @thsottiaux
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
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.
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
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
and now, tomorrow they're releasing Opus 5.1
it's just going to be called Codex Max 5.2 X-High
and yeah, still slower, but again - you're trading off speed for dozens of IQ points.
crazy how I can bet on literally anything on gamba sites ("will jimmy steal sarah's lunch from the breakroom again this saturday?")
but i can't somehow bet/invest on xAI/SpaceX lmao
@elonmusk isn't X the everything app? can one of those everythings be investing in non-public?
you can use claude opus 5 RIGHT NOW.
no, not opus 4.5.
opus 5.
the only downside is that it's much slower. so if you're fine with trading speed for intelligence, go ahead and use opus 5.
they named it weird this time though, "gpt 5.2 xhigh", and only available on codex cli
the slow boil death of the US is being brought on by people like bernie and warren
>warren "anti-monopoly" crusade against irobot being bought by an american company (amazon)
>irobot now not only gone bankrupt but was bought up by china and gutted for data
this smells the same.
crazy that we can "invest" in prediction markets like "is jerry going to microwave fish in the break room again on thursday dec 18 2025" yet we can't invest into obvious long horizon wins like the boring company or space x without doing mental gymnastics "indirect investing" … Less
crazy that we can "invest" in prediction markets like "is jerry going to microwave fish in the break room again on thursday dec 18 2025"
yet we can't invest into obvious long horizon wins like the boring company or space x without doing mental gymnastics "indirect investing" https://t.co/rzNKHaKyIc
all of those ragebait ad games where you have to outline the shape without going over the same path twice are finally going to be useful for designers using this
Supportless, free-form 3D printing! ⛓️ By using low-melting-point Fields' metal and shaping it through surface tension, this process can create highly conductive 3D structures in a single pass, no supports, no post-processing. Pretty interesting way of thinking about additive … Less
Supportless, free-form 3D printing! ⛓️ By using low-melting-point Fields' metal and shaping it through surface tension, this process can create highly conductive 3D structures in a single pass, no supports, no post-processing. Pretty interesting way of thinking about additive manufacturing, especially for electronics and functional metal parts. Just wow. I only know this is from Singapore, anyone has more information about them? ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com
openai's new image gen is a quality upgrade over itself, but definitely not better than google's banana pro image gen.
the yellow tint is gone, but notice the amount of fingers my left hand has, it makes my eyes funky.
banana pro result and prompt used for both in comments
"they danced on charlie kirk's grave, why not dance on reiner's?"
because you shouldn't be like them, moron
we're already in the "what are the _ regulations in space? can you do anything there?" phase
next phase: elon or bezos (probably elon!) figures out how to create a country in outer space (or buy it i guess? satellite-level real estate?)
then says "come dock and arrest me then."
one of my guilty displeasures is spy movies and how often unrealistic or improbable things happen
"don't touch your ear when spying" is probably part of day 1 orientation at spy school
yeah i get it, it's roller coaster actionslop entertainment, not a documentary. but still
it's crazy to me that it's the complete opposite for me
i use opus 4.5 for speed and when it inevitably starts making mistakes/it's time to clean up its spaghetti code, gpt 5.2 xhigh comes in to save the day (albeit slowly)
not to mention they now have Opus 4.5 & GPT 5.2 xhigh to help them speed things up
the wheel spins faster.
i always wondered if someone somewhere at least once suffered an attack from this
gpt 5.2 xhigh on codex went nearly 3h last night
the result was perfectly what I wanted btw
And just like that, I never have to worry about resetting/manually compacting when using GPT 5.2.
They hit it out the park, as usual.
Better long context than the Anthropic models, and that's saying something.
i just dont understand the play here. this is how his time is best spent? what is the end goal? is this his childhood dream he's now pursuing because he can just throw millions at it and force it to happen?
no way is this financially sound right
WOAH idk how I missed this crazy detail. this alone takes things to the next level imo. this was always openai's weakness
lmao remember they were always tagging disney saying all this pr slop of "did u know about this ai content using disney IP, i will never watch disney again if you condone this"
they must have seen it and said damn, if you cant beat em, join em
gpt 5.2 is the first model that just... gets it. iykyk.
PLEASE CHATGPT
the only reason why I am not 100% daily driving @OpenAI chatgpt atlas as my browser is because it's missing one crucial component to my workflow
spaces. once chatgpt atlas has spaces, it's game over.
