Posts on X before Mar 14, 2026, 11:08:54 AM UTC
Rarely should you ever be steering with follow-up messages.
Most of the time, queuing next messages produces much higher-quality results than steering the model in the middle of what it's doing.
RAM on your macbook will basically become an intelligence potential marker
instead of "how many things at once can I be doing" it turns into "how much intelligence at once can I be generating"
Ultimate marketing move here @thsottiaux
reset Codex limits for the reason of Claude being down 😂
you are no longer the violinist in an orchestra,
you are the conductor
Expectation: the age of the IDE is over Reality: we’re going to need a bigger IDE (imo). It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. It’s still programming. … Less
Expectation: the age of the IDE is over Reality: we’re going to need a bigger IDE (imo). It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. It’s still programming. x.com/karpathy/statu…
bigger than you realize
Gemini Embedding 2: Our first natively multimodal embedding model
Gemini Embedding 2 is our first natively multimodal embedding model that maps text, images, video, audio and documents into a single embedding space, enabling multimodal retrieval and classification
one is the original intelligence - whatever that word means to you - while the other one is a copycat of that intelligence
poetic in a way if you think about it
i dont really see the endgame of AI labs hundreds of billions of dollars spent for a SOTA model just for a random open source model to come around and similar benchmarks doest it even matter to be “first” to whatever the goal is? or just surviving long enough i.e. Apple will … Less
i dont really see the endgame of AI labs hundreds of billions of dollars spent for a SOTA model just for a random open source model to come around and similar benchmarks doest it even matter to be “first” to whatever the goal is? or just surviving long enough i.e. Apple will simply just take what works without burning any cash
this has been said at least once a day somewhere around the world for the past 60+ years of CLIs
CLIs are not going anywhere lmao
damnit @bryan_johnson you were right about nothing beating 8 hours of sleep. no amount of caffeine (or anything else 😏) replaces it
I was doing 6 hours for most of my 20s, if not less
so much more done with so much more clarity and quality than before
GPT 5.4 xhigh is the first model to have a semblance of high quality writing, feels like some of 4.5 rubbed off on it
Still not jaw dropping but if you deal with a lot of text (non-code) you're sure to feel the difference
Customer Support -> Customer Engineer
The former becomes the latter in the next few years; typical support will be automated away while custom engineering/fine tuning will replace it.
gpt 5.4 xhigh is the first model that can change your life even if you start completely from scratch right this moment
every model has its favorite word
gpt 5.4's favorite word is "seam"
even if next year's open source models are as good as today's closed, the closed ones of next year will make using those open source models a waste of time
it's been like that for a while if you've been paying attention these past few years
I don't know who wants to hear this, but buying an expensive MacBook Pro or Mac Studio to run small LLM models is useless. These models have low density of world knowledge and reasoning. It's much cheaper and faster to use an API. Maybe next year they will be useful, but not this… Less
I don't know who wants to hear this, but buying an expensive MacBook Pro or Mac Studio to run small LLM models is useless. These models have low density of world knowledge and reasoning. It's much cheaper and faster to use an API. Maybe next year they will be useful, but not this year.
it is jaw dropping how terrible antigravity / gemini 3.1 pro is
works of programming genius like this, especially those made before pre-2000's, are always awe inspiring
someone out there did all of this, efficiently, no templates, youtube tutorials, AI (lmao!), wasn't mainstream at all especially in the 80s,
humans. we had a great run.
Some games had the rare quality of being entirely unique, introducing concepts that had never been done before. These days, most games are variations of ideas that have existed for a long time. One of those unique-concept games that has stuck with me over the years is Marble Mad… Less
Some games had the rare quality of being entirely unique, introducing concepts that had never been done before. These days, most games are variations of ideas that have existed for a long time. One of those unique-concept games that has stuck with me over the years is Marble Madness. I was instantly hooked the first time I played it. Although I wasn’t very good, the urge to return and try again was incredibly strong. It had an almost mesmerizing quality. The graphics - for 1984 - were impressive, as were the music and sound effects, which added to the “this feels very different” vibe. Most fascinating, though, were the physics: the marble reacting to slippery surfaces, bumping over ramps, and being knocked aside by the “evil marble.” It demanded real skill and patience to navigate the levels. The unique gameplay also makes it timeless in my opinion. Marble Madness looked great over 40 years ago and will still look and play great 40 years from now. The only downside was its relative shortness with only 6 levels. Mark Cerny originally wanted to create more but the memory limitations were the main reason for the small number of levels.
Following the upcoming special Security College on Monday, I am convening a special Zoom call in 3 months.
It is of the utmost importance that there is no further escalation until then.
If you haven't felt it yet, OpenAI has won.
Ever since their GPT-5 line of models, it's become more and more clear with each iteration they release.
btw, building your own transcription solution is something you should have done two years ago with Sonnet 3.5
let me finally all-in on Atlas and drop Chrome by making Atlas 100% controllable via Codex CLI
BONUS POINTS if I can continue to use Atlas/browse on my own (scrolling X) while Codex CLI + Atlas do stuff together like e2e testing or filling out forms or XYZ
codex cli is all you need.
if you're on the latest codex cli alpha, you can now see how absolutely beautiful it has become.
syntax highlighting and the general containerization of commands, not to mention subagent spawning and interactions, it's all beautifully tying together.
what i'd love to see from codex cli - allow it to ask me questions when not in plan mode
coming from always starting up a new session (because continued development after first prompt was always sub-par), now that it's fixed/feels much more high quality, I'd love MORE questions
tldr make sure you're someone's (an AI's) golden retriever before the takeover
unless you merge with that said AI in some cyberpunk cyborg symbiosis
sure, this isn't coming yet, at least a few decades, but it's coming
gpt 5.2 xhigh (non-codex) is still useful for things where gpt 5.3 codex xhigh just won't do things justice (think copyright / ideating)
which means gpt 5.3 is going to worth the wait, to say the least
Okay, Codex 5.3 is obviously the best coder LLM out there right now, hands down,
but when it comes to copyright or literally anything else outside of coding, it is absolute dog shit
Excited for GPT 5.3
Unfortunately Sundar's models have been total garbage, especially Gemini 3.1
I was aching for its release, hoping that we'll finally have a big model smell that is fast and can code as well as or better than Codex 5.3 xhigh
But nope. absolutely unusable
Sam and Dario not being able to hold hands on stage in when the PM of India asks them too, at a summit about the future of AI and its safety, kinda implicitly shows they feel above governments, right? Like they really wouldn't care about world peace when it comes down to it, onl… Less
Sam and Dario not being able to hold hands on stage in when the PM of India asks them too, at a summit about the future of AI and its safety, kinda implicitly shows they feel above governments, right? Like they really wouldn't care about world peace when it comes down to it, only winning? It's kinda freaky to me, I mean they are winning, but I don't know, you have to remember regardless of whatever level you're playing at (and they're playing at the highest level), entrepreneurship is still just a game? You can't take it so serious that you won't hold hands with your competitor right? Sundar being nice here only makes me more positive about Google
It's insane to me that neither one of them, arguably being in the top ten most important and powerful people alive today, were able to tap the other one and say,
"Come on now, it's just holding hands symbolically for a few seconds. Let's not be childish about it"
REEKS of ego.
SaaS (service here meaning an app/interface) will die out in the next decade or sooner
SaaS (service here meaning an API) will grow exponentially given the infra is designed for bots to pay/use it automatically
If you’re wondering about an example for SaaS API - OpenRouter.
There will come a point where anything humanity has imagined onto the movie screen will become a weekend project
I remember one of my professors, a bit over a decade ago, saying “16GB of RAM is all you’ll truly ever need as a typical consumer and even a software engineer; things will get more efficient from here.”
Oh how wrong he was… he’ll be right eventually, but for now he’s wrong.
a few years ago this was a tough purchase not because of cost but because "what would I do with it, I'd need to learn everything about it, its stack, etc. months of work"
now with something like codex it feels like a fun hobby project that you can just prompt "do backflip" lmao
pi is the nvim/emacs of the 2030's, mark my words
use the future of software, today.
START RICING 😤
agreed. i can feel this both on a societal level and a familial level, <-> myself.
building has consumed me away from both, and yet now that ~everything I need to move forward has been built, exponentially sped up thanks to AI, I can hone back in on community and family rel.
another one bites the dust to slopus
this is the first time i'm not even going to bother to try out a new anthropic model (sonnet 4.6)
it's just not worth my time when we've got GPT 5.3 Codex X-High
99.99% of every single action this person takes in their life is thanks to programming
literal code. In everything, everywhere
All jobs are “coding” jobs when you get down to it
Software will eat the world even more than it has, now that everyone can be an S tier coder
he knows openai will win, regardless if you like them or not
they already won; iykyk
he wants to be part of history and who better to join than openai? name someone?
anthropic snubbed him, what's left, google? meta? microsoft directly?
if you "don't understand" this... 😬
the near future consists of companies such as @a16z having an army of 24/7 research scouts scouring blossoming indie devs with tipping-point projects
reaching out to VCs first for funding will be a negative sign, it already is but even more so, given the ease of comms blitzing.
Proof that X is a gateway where all dimensions can post into
this has no implications whatsoever
Sub-40ms full self-driving on a $100 drone! 🚁 A demonstration showing a complete full-self-driving pipeline running in under 40 milliseconds on a $100 drone. The prompt: "Find the bike and land." No pre-mapping. Running real-time on commodity hardware. For context, human reac… Less
Sub-40ms full self-driving on a $100 drone! 🚁 A demonstration showing a complete full-self-driving pipeline running in under 40 milliseconds on a $100 drone. The prompt: "Find the bike and land." No pre-mapping. Running real-time on commodity hardware. For context, human reaction time is around 200-250ms. This drone is processing sensor data, understanding natural language commands, identifying objects, planning motion, and executing control, all in 40ms. This is what happens when foundation models meet efficient inference. The models get smaller and faster while maintaining capability. The hardware gets cheaper while getting more powerful. The intersection makes previously impossible applications suddenly viable. A few years ago, this required thousands of dollars in compute, pre-mapped environments, and cloud connectivity. Now it runs locally on hardware that costs less than a nice dinner. Awesome stuff @chesterzelaya & @droneforge! 😮💨 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com
It's starting this year
You will see the transition regarding mainstream AI hate go from "this is stealing art" to "this is stealing jobs"
This seems like a small difference but surely you understand there are more people who "have jobs" vs people who "are artists"
the only thing i recommend is making use of AGENTS md very minimally, enough to start the model off with knowing you're an expert and should be treated as such, not a newbie making a weekend project
that is, if you're using it for serious prod work
If you tried GPT 5 Pro and weren’t impressed, neither was I
Go try GPT 5.2 Pro now. Big difference. You can smell the intelligence in every token.
Who gives a fuck if it’s SOTA results
It can write in backwards Chinese, IDGAF
Every day I talk to more people
Every day I realize no one knows what’s going on
Not to mention what’s going to happen
Interesting times ahead
This but the exact opposite should be your focus
More power since now you can develop out and simultaneously build multiple large projects
telegram is definitely sota for mobile agent ops relay
so everyone using imessage has two icloud accounts for it to work, right?
because there's no way to prevent doubled messages since you're sending and receiving to yourself
unless i'm missing something?
I disagree, you're not thinking of all the other things that will occur before this chain of events happens
by the time we reach the last part of this prediction, the cost of intelligence will already be 100x less than current
don't take my word for it, sama said so himself
as fun as its been seeing opus and gpt one up each other, i don't think these will be the models we standardize on a year from now. anthropic and openai have $200/mo plans where ppl have done the math is 36x cheaper than API cost. @steipete even said he had to stack subscription… Less
as fun as its been seeing opus and gpt one up each other, i don't think these will be the models we standardize on a year from now. anthropic and openai have $200/mo plans where ppl have done the math is 36x cheaper than API cost. @steipete even said he had to stack subscriptions to make clawdbot because it "made tokens so much cheaper". the play here is subsidize individual plans → devs get hooked → bring it to work → employers buy enterprise seats at 10x the price. BUT once they've captured enterprise, individuals and small teams *will* get priced out. 🧵
pivot into selling to bots, not humans
lets revisit this tweet once a year
my one man company verified stripe rev in 2025:
- $50K to humans
- $0 to bots
see u Q1 2027 for 2026 stats