Codex CLI login blocked by phone verification rate limit — Pro subscriber

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #25737
• #25803
• #25185
• #25798
• #24990

Powered by Codex Action

Same issue

some tought - regarding training - and capcity

honestly its still amazing how there is space left for Ai training , considering how infrence is almost 30% power needed for training at this point , for classical machine vision its much less right? like 5% when you doing infrence -
when doing a queue system is acceptable for messages, right? you think can use the nvidia groq LPX, would assumed to enter production even with in 2026 ? https://www.nvidia.com/en-us/data-center/lpx/
but things will get better because the Grace Blackwell hardware are very powerful , hopefully in the future we could achieve whats is achievable now, but with much smaller models....with Vera Rubin, which would be shipped in the feature (more widely)
, right? so every new year 2x smaller model of what we already got , so every few months --- now models "brute force", every few months models would be able to "predict" in the true sense of the word thus 10x the input , of what we got today ... so right now models are very useful we dont need more than that, they can still predict.... us humans always think of the end result and progress along the way,, with this models , they act by logic and use logical templates (which some people do too, to make themselve efficent)
<img width="1600" height="810" alt="Image" src="https://github.com/user-attachments/assets/7831c0d9-7cd0-47b0-a8b3-2793d4bfdd6f" />
and right 5x what we got is not that attractive, even if we got 2x of what there is now does it free the same amount of compute for ,training, expirements, failed expiremmetns .... right because deploying 105 agents is a bit amuzing it proves the bruteforce point --- also why codex-openai is so great because the point is predicting , utalising past documnetation (right even pdf , which a lot of human knowledge is documented in that evil format --- also a lot of things should be scanned into datasets - like rare langiages - which you can argue doesnt mean a lot [dont get me wrong I used to be very into linguistics and languages ,, well 40-20 years im now 65 very soon 66] [right? artifacts being faked all the time , and a lot of artifacts either destoryed - or still exist but never will be documented --- like think how we could predict real means of acient languges , acient latin --- but again we all belong to the feature so better invest in interesting stuff like chemistry biology medince that being said its like a limited society, so not a lot of progress considering 4, 6, 8 billion people on the planet, I like to assumed we are all only 600million people rougly 10% of the real number ])

the gold standard is 2026 march - is the 60% 70% on some benchmarks - and keeping the very fast latency and speed these tasks are getting solved, having current hardware being able to run 10x the capacity is what needed? would be happy with those? and how do we become morally okay with 105 agents doing some task some people are never sastified ,, yet dont know what they aim to before the hit the prompt (but in many cases the ai doesnt able to sastify them) the AI / llm at this point feel perfect ,, its like the human factor being the issue (the ai made 10 different things I am too lazy to forget what it did / I forgot I fixed it asking the AI to do it again / not understanding the issue / dont look what the ai did - and have a memor issue [me the human] ) ai models only need to be stable - and not flip on themself too much (or spend to much resources when its not needed)

llm fun: "The Power Draw: The human brain continuously consumes about 12 to 20 Watts of power." 400 Watt-hours (i think much less, daily for actual parts in the human brain that think)

a rtx 3090 1,400 Wh, consume about ? yet its not 20x more efficnt than human for thinking - and at the very least it doesnt set goals like humans do 4x more power than human daily? yet you assumed even if they are not perfect to be atleast optiomized for prediction, not brute force.... but again as they work now its amazing they only need to be 4x times cheaper- to break profitablity in 4 years ? considering the real cost is the hardware you buy (for market share and longivity accross years)
think about this way about 40% is what LLM already are , they are 40% of a human - right many of us are slow to understnad how to adopt how to use this type of autocomplete -- and declare and ask it (like they are some bad boss - and boss who know whats next --- so its a skill )-- right now its a pivot - because there is the state of the art and good enough - but with good enough there is some things you dont want it to happen - so you dont pivor the maximum SoTA - because numbers get multiplied very quickly. -- so my prediction is that LLM can only get around 5x times better than they are now - in some ways -- even if they dont do too well on evals --- they only need to get cheaper - with the same feel (what we saw in gpt 4 and 5 ,, even back to gpt2 how hard that was to run -- and gpt3 , which now most of us understand is pretty primitve - and other models which had a very fast infrence --- yet you easily understand limits are higher? if you were to use chatgpt - free its pretty much unlimited in how much messages you can get - and yet there is Image generation (Sora 2 which was a big deal in eating compute) ) (if you were to use chatgpt-plus and codex - it gets you a lot of usage - and a lot more than other companies ) ---- everytime we see ratelimits higher - the ai labs are training (e,g temprature on the model, how much context - or if you are coding c++ which much external web knowledge you should use --- alltho maybe ai labs - now have some kind of way to create a secondary dataset - to not hammer all of the web too much --- when needing documentation etc === how truthful you going to be to the documenttion)

right? there is something more than being able to create a good "vibe coded" website yet you need to sastify everything --- do we need to build everything from scratch (all of the ai models do it) or we need to take a template from paste memory -- on modify over it - saving compute --- you can be very aspiring for this kind of thing --- but due to expirements (and actual hardware limitation -- ) this is the only way to do things --- right --- if models are 6x as cheaper -- as of today 6-2026 --- then price per watt is less a crtical factor -- also thinking TPU and how much compute they consume -- and how much share they take in factory lines 300million dollar email to TSMC , ~130,000 wafers for TPU, 500,000ish for nvidia? TSMC at some point can make 2million wafers a year (which would make sense as chip die now are going to get bigger -- due to wishes of lower latency?? from 30x to 50x speedup in some cases)

yet we are here - and there is a lot of magic and wisdom to whats going on here

<img width="907" height="407" alt="Image" src="https://github.com/user-attachments/assets/086b642b-c6aa-4aef-85b9-1917f8dffc85" />
right openai sla is still very high - also think how much demand there is now to codex - yes its still fast -yet you can still use it - yet most of us get very high usage even on the chatgpt-plus plans , as things stand now what I predict that a 3x importvent in the models is needed - in the case things are going to be sustained... a lot of things happen -- when those ai labs , add datacenters, need to train (reboot/update models?idk? high latency when uploading 200gb file)
far as I concern openai is the only ai company the moment you fill a prompt the prompt doesnt crash in the middle

as per GROQ LPX , even if the models themselves get a 30% preformance degrading on evals --- nvidia having the GROQ team is great news... but right you cant really fit a 2 trillion paramter model onto it -- and most of the software
so its big news it can be a Cuda moment, a chatgpt moment for nvidia...
https://www.nvidia.com/en-us/data-center/lpx/

i recomend to see
https://en.wikipedia.org/wiki/AMD_FireStream#Limitations
which was a multi billion investment by AMD in 2006
the brook+ era https://github.com/hibengler/BrookPlus/blob/master/platform/brcc/src/BrookHighLevelView/HeaderFiles/brtkernel.h

<img width="1290" height="735" alt="Image" src="https://github.com/user-attachments/assets/0a6a66ae-3b63-4866-83fb-0f51bb8a40b0" />

same issue

<img width="762" height="518" alt="Image" src="https://github.com/user-attachments/assets/fdde006d-15a7-4042-9824-88080869abbf" />

same issue

<img width="952" height="399" alt="Image" src="https://github.com/user-attachments/assets/d157cc5f-f192-443a-a779-ff57c1aa9eea" />

@rizkidarmawan21 @inde3d100 Were you able to receive the SMS?

P.S. After some time, I received an SMS

@rizkidarmawan21 @inde3d100 Were you able to receive the SMS? P.S. After some time, I received an SMS

honestly now nvidia will be moving the SRAM and some other GROQ tech ,
which is assumed to increase some how the latency

honestly will all of this AI boom , Im suprise there is no new generation of SRAM , or some where to layer up more SRAM , reminder you can also use 60nm just for the SRAM part ,

nvidia is not stupid , that being said , they clearly put their their HBM ram next to the chips to cut on latency

whats funny is that NVIDIA new system racks takes a lot of ideas from phone design , no wires , compact , heat management , right ? we can afford those 20ms for data center called when it comes to do whatever

look at groq spamming a lot of patents now
https://patents.google.com/patent/US20260065098A1/en?assignee=GROQ%2c+INC&oq=GROQ%2c+INC&sort=new

that being said there is many buisness choises why not to fully ditch HBM (I like to call it poltical reason, like continue use SK HYNIX , and SAMSUNG because - because you dont want to fight years in the US COURTS)
Right , and nvidia is not perfect --- people praise CEREBRA for example , but they are limited by amount of wafers (e g , cant make 200k wafers ,, wafer ingots are very hard --- and even 60nm process is super complex)
and there is this korean company called rebellions_AI , which I suspect have some dealing with samsung and sx hynix, which they both have very great people

right? Nvidia might cry thaat it needed to only sell gpus, when they developed a lot of things for indurstry and cuda , and believed in it, before 2022 , training on AMD hardware is a nightmare

Samsung has a lot of engineers working on HBM , for example in 2021 and 2023 -- where they gave a lot of attention to a product, which you can admit and they make --- so it comes to a relization -- worked on SDRAM --- we worked on SSDs --- so all as the same roots

In many ways , AMD win on some topics --- no one does it better than AMD , when it comes to video proccesing cards
its almost like they got a policy of do not have too many ideas in nvidia that comes to hardware --- only 3 products lines of something (and heavily invest in software)

https://www.amd.com/en/products/accelerators/alveo/ma35d.html
see the "alveo" line, a so called Streaming Media Accelerator

nvidia cannot touched this Streaming Media Proccesing field ,,,
right? pretty sure Twitch uses this Alveo - media proccessing line
right, you see where I am going , nvidia cant enter this market (or perhaps you say , someone wont know there is an AISC non gpu way,,, and buy more gpu orders --- which is buissness we know how sales work sometimes )

in theory , nvidia could have done something to increase infrence --- like making a new chip line . way back in 2023 for example --- and they likely did something (which is the Grace Blackwell line)

right? in few years we can going to have hybrid systmes of gpu and lpu -- and nvidia going to crash everyone -- honestly I dont see AMD going to build computer clusters --- those new Vera Rubin serious , is built like an iphone stack

I kindda like it how phones are built in 10 minutes , and its not remaking the classical server computing --- I wondered once , why they need all of this cabels - and create a standard where things are more clustered

its a big deal or whatever, im still impressed that those "clusters" able to run for so long
another thing ,,, think how many of those AI outages --- are datacenters forcing to shut down because -- they took too much of the grid -- causing preformance outages

I love the idea that AI , is going to make people jump from more topics to another topics -- and not go on small technicality ideas --- and this is what llms solve for now

another big deal , is the whole idea of model confidentioly --- which is a big deal for nvidia executives --- and also and think if there is some engineers - and need to review all of the stack - and this kind of design is slow by nature - because the idea is complex

the idea of model confidentially , is actually pretty cool - where the compute cluster, because a baremetal idea (im sure they want to explore the idea of many different type of models - but whatever)

the current step is , make gpus faster , make cpus - more alligned , and removed things that arent made for what you want

right? think where we are
6.4 Gb/s , on one pin..... right???? right?
Right , when it comes to HBM pins , they tick? whats happening here?
but lets be fair -- its not an invention but its is a relization imagine yourself working on SSD - trying to back and rack more and more bytes --- right , so what drove it was SSD development --- not the wish for highest and highest RAM usage --- because let be honest DDR3 and DDR2 , are just as fast in first glance -- I dont think you need more? hey? its like as they arent fast already.... LOLLOLOL the ddr3 and ddr2

Right we per "Research" sub 1nm transsistors achieved in lap settings , let me fair,,, right but our knowledge is only refined to 14nm in production (you see where I am going? we are always going to be limited as time progress --- we aim higher than we should-- - but the knowledge to 7nm is unknown -- right, you see what I mean? we know very well 7nm is possible)

right? should the memory be optimized a bit -- or we should do some type of encryptoin -- so AWS or AZURE wont modify the boards . for "model confidentially" right? even small data encrpytion are problematic (and the last thing is add more copper -- like some clever say - add another chip that does the number shuffeling or encrpyiton)

but lets be honest - we should be very thankful cpus and gpus can work with HBM , (the modern successor of SDRAM ,,, aimming to the idea HBM 1, was released in 2013 --- also notice - this is right about the time , nvidia started making great chips -- so we should thank SK HYNIX , and SAMSUNG - not nvidia , nvidia had made a great bet - but making their hardware so portable - right and little hardware release --- and a lot of software --- so a nvidia is first a software company )

in the end , we know nvidia new gpus are slightly bigger chips . over hopper -- think how little chemsitry is known is 1870 - and barely to 1920 - and people understand how to tinker with atoms and thoerize - how to harness energy over it --- and start view matter as radio singals ? Right? light can pass tru the lizzard skin but its solid - the water is solid because I get wet so it exist - but I can see tru it ... I feel the wind but cant see it ? what is wind - is it a something - does it have a weight ?

future prediction are is that electricity wont matter --- but its needsly impotent that it has high value and expensive now (because we need to get efficent and fast) -- in the future I have no reason to argue -- it wont be powered by neacular power

a patetic 835-megawatt output for neacular power plant is funny -- when we think about 5gw per datacenter
(another cool thing -- they want to make servers less noisy -- that cool , I like to believe its because of regulation -- so they invest in novel ideas like better water cooling --- heating issues of the literal copper wires [lol so you removed the wires] )
but its not as that unfunny --- if you think we will see 90x improvement on infrence -- because the trejctury is nvidia cleary going to make aother type of chip, for transformers -- also deal with the massive amounts of data needed to be dealt with
would be cool , if future investment would be --- packed in with a power source (could be coal , could be alternatives like neacular power )--- that combines we ai agents going to be even smaller lol ... who knows ??? lets be fair there is a lot of great ideas trying to be added to LLMs but a lot of good ideas get ejected -- because of literal hardware limitation --- cant push too much data etc --- weird latency issues -- ai labs research dont know how they come

and smaller chips are likely the answer as much as I would like to see 100nm higher quality wafers (another things, cerebras chips EAT A LOT OF ELECTRICTY YOU ARE LIMITED BY THE GRID )

<img width="976" height="832" alt="Image" src="https://github.com/user-attachments/assets/964c9c95-37b1-4b4f-aaad-0b4d02d2ba44" />

also
HBM is a legal warzone

also pretty funny if nvidia , pay SK HYNIX

also think how much AMD spent R&D over making HBM --- and nvidia didnt had a role in it ---
yet HBM is one of the biggest kickstarters for the GPU power ramp up thinks how GPUs were in 2014 , and 2016 how great they because -- and 2016 is when AI was easier to train

(side note: NVIDIA introduced HBM2 in their Tesla P100 GPU in 2016. Equipped with 16GB of HBM2 memory) --- but the developer of HBM and differnet types of RAMS is what kick start lal of this AI race
(also it exlpains AMD push for 3D stacking technology - v-cache ---- you see where I am going --- alll of this developement for ram ---
so AMD is a powerful game in the RAM race -- and AMD r&d is very strong --- plus their cpus and gpus - dont brick after a month nowadays so they improved a lot (but you can also argue they took risk --- by trying different IC packaging or whatever - or expriemental semiconductor and nanoelectronics ideas )
on the other hand NVIDIA only release hardware when there is some demand for it - again , nvidia is a software company - because less hardware
)
https://www.sec.gov/Archives/edgar/data/917273/000091727315000015/R21.htm#

---
i know its a bit off topic - just few ideas on the economics of things ---
but also think how GPUs are not the only answer for infrences for llms -- they can b run on GPUs and they can be run ARM cpus ,, they can be run on modified RISC V - with the idea of some how making them stronger for llm tasks

in one hand ARM has a lot of adoption (the phone market)
in other hand - they dont have adoption in servers --- so a lot expriment with ARM ---

pretty funny how they didnt let NVIDA buy ARM --- but ARM literally allow different manufactours access the design and optimize for their needs --- right? so they ended up , paying much less than they would gain from ARM --- yet gain all of the options to use the chips design so they pay less for that , that being said NVIDIA could have buy ARMand they the rights to do so -- also its funny because ARM is not really a suplier --- if anything they closed a different deal with ARM , so it doesnt really matter (only on the software scope) you could htink ARM could been sold later for a bit more ---
could have done with ARM influence -- and guilding where ARM is going --- or they can partner with MediaTek -- which now made those Vera Rubin look like a smartphone ((ANd MEdiaTek previous bet in 2023 about the automotive ideas ---that automotives comes the whole field of Heat management right if you think about it? and how you stack cars fast?))

think how we are not at Huly 2026 .... but we cant predict how edge llm can be much better in the future --- but as of today we are still very capibale --- but we dont know where to inovate - and how much time to give

a lot of things are still 2or10 µm, just so everyone knows