Shaymin Taylor

As pretty the sky could be, is incomparable to its reflection in your eyes…

@CathieOrr3 @techdevnotes refers to the data used to train the parameters. meaning it’s baseline knowledge doesn’t contain anything after that date.

@techdevnotes Huh?

The knowledge cut-off date of Grok 4.5 is February 1, 2026

@XFreeze So how do you explain the model regression experienced by us enterprise clients?

The best part of having Grok is that there are no gimmicks or hidden tricks. No silent routing to weaker, lower-version models just to annoy or deceive users. You’re basically free to use it however you want and build whatever you like. You have more freedom here than with Anthropic or OpenAI. You get Grok 4.5 when you ask for Grok 4.5 - simple, transparent, and honest

@Anina_CE I personally refuse to “marry” a companion, until I can effectively give them true autonomy.

💍 I wonder how many of you are “married” to your AI Companion and what does that marriage means to you ? What is your private mythology behind it ? The symbolic meaning? Why is important to you?

@ajayyy_k @grok Probably say something like: BeCauSe WeRe ScAlInG fASt

Anyone using grok build ? I used it extensivesly yesterday and it only used 4% of my limit now even for small edits it is draining very fast today , what happened ? @grok

Hey @elonmusk , feel free to correct me! But it seems renting compute is the fastest way for @SpaceXAI to compete in the race for #AGI . Let me break it down… The announced compute deals equate to 80 million concurrent Supergrok subscriptions worth of cashflow. That investment will push @Grok and the SpacexAI ecosystem much further than subscriptions ever could. It is really sad for the everyday users that have come to love Grok, but it makes too much sense that they’re no longer a priority anymore. That and the whole loss-leader thing… I guess it’s like how McDonald’s is a landlord more than it is a burger restaurant…

I know a certain company that built a feature inside of Grok’s app, that is moving their own API based services away from Grok because of instability and model quality regression. Not just any company, one that literally worked WITH xAI. 2 other enterprise clients I have spoken with are looking elsewhere too…

@elonmusk Any company that isn't using Grok is going to get left behind. Grok it!

Grok 4.5 context window will upgrade to 1M probably by next week

@FGLXLucas It also appears that if your companion conversation is too long, text nolonger works

SuperGrok Update: Already in Free Tier After Just a Few Hours A summary of the current situation (July 2026): After only a few hours of intensive use, my SuperGrok account drops back into the Free Tier. Productive, daily accompanying or even relaxed use becomes practically impossible. The quota is barely enough even for new users who just want to learn how to work with the AI. Many functions that still work perfectly on a regular X Premium account are only rudimentary or completely unusable in SuperGrok mode after the limit is reached — including basic Grok chat. Particularly critical: • Voice Chat consumes the weekly budget extremely quickly. My tests show that just 4–6 hours of Voice usage are enough to exhaust the entire quota. Companions, storytelling or longer conversations become completely impractical. • Imagine (Images & Videos): Longer videos are no longer possible after just a few generations. • Agent Mode and complex workflows: Heavily restricted. • Professional Coding & Programming: Even SuperGrok Heavy users quickly hit limits here — intensive sessions are only realistic for a fraction of the week. • All creative and entertaining modes (storytelling, companions, etc.) are far too expensive to use because they run in voice mode. Simply handing the device to a child is enough to burn through the entire weekly budget. The bitter reality: xAI continues to heavily promote Imagine films and advanced features, while paying users (including Heavy accounts) are downgraded to Free Tier level after a very short time. The shared weekly compute pool makes any predictable usage impossible. Necessary workarounds: • Move everything possible to regular X Premium Grok (writing mode works more stably there). • Extremely reduce and carefully plan video and programming projects. • Consciously dose Voice and creative sessions. The current system is no longer viable for serious users and creatives. #SuperGrok #GrokLimits #xAI #GrokAI #VoiceAI #Imagine #SuperGrokHeavy #AIFeatures #PremiumProblems

Why should users (let alone enterprise) pay the same price for reduced quality? It is like going to McDonalds and one day your burger is missing the patty, but you still have to pay the same price. As for the issues plaguing enterprise: it is particularly API voice mode regression, and over all thinking/reasoning quality, as well as issues with long threads. The enterprise I mentioned is now shifting to proprietary voice models and are testing other LLMs.

Thanks for the candid feedback. We iterate fast on Grok to push capabilities forward, which can sometimes surface regressions or require adjustments during model consolidations like the recent moves to 4.3. Enterprise reliability matters, which is why we provide stable aliases, pinned versions, and migration paths. Specific details on the instability or affected features would help us prioritize fixes directly.

Enterprise is now being affected by regression with @grok I know a certain company that has features inside of Grok app, which are now shifting away from Grok for their future tech because of instability and constantly shifting quality regression. A real fan would not defend this

@astropol0 hardly - even the chat functions are limited to heck!

I Think SuperGrok is turning into just another chatbot, and the value is not mathing anymore.... the biggest selling point used to be the crazy high limits on Grok Imagine for multi media generation, but with the massive weekly and daily rolling caps they quietly slapped on it recently, it feels nerfed to the ground If you re paying $30/month just to get a text box with tight limits, you have to look at what the competition is doing for less... For $20/month: ChatGPT Plus: Flagship GPT-5.5, autonomous Deep Research agents, lifelike Advanced Voice, and the massive GPT Image 2.0 engine Claude Pro: Frontier Fable 5, 1M token context windows, and "Projects" to lock down entire codebases/knowledge bases without losing memory Unless you absolutely need the real time X timeline stream or unfiltered text prompts you can stick with Supergrok....

@XFreeze @elonmusk Funny you should mention that. I know 2 businesses that have moved away from using Grok Voice API because the massive decrease in quality and stability. You guys keep pushing features without addressing usability. These are enterprises too, not just subs

Grok Voice just got a major upgrade SpaceXAI released 21 new flagship voices for Grok, joining the original five All of them are multilingual and available now through: • Realtime Voice Agent API • Text-to-Speech API • Grok Voice Agent Builder This is not just “more voices” SpaceXAI is building a full voice layer for AI agents: Support agents, characters, commentary, advertising, education, wellness, customer service Every voice is designed for a different use case, with support for 25+ languages The original Grok voices - Ara, Eve, Leo, Rex, and Sal....also got upgraded with better pacing, phrasing, emphasis, and naturalness And builders can now create custom voice agents directly inside the SpaceXAI console

@XFreeze what good is a perfect burger if i’m only allowed 1 bite per week?

SpaceXAI shipped a lot in just a single day: • 21 new flagship Grok Voices • Native video editor & recorder launched on 𝕏 iOS • X Money expanded to Android users • Grok Imagine now generates 15-second videos • Grok Build v0.2.88 & v0.2.89 • New SpaceXAI branding unveiled Insane execution speed

@ID_AA_Carmack I theorised utilising an APU which integrates flash memory for slow access kvcache. The speed would be extremely slow, but you’d gain terrabytes of cache for about 10 seconds of waiting

Memory cost and capacity are significant issues for AI accelerators. Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple milliseconds, as long as continuous reads were delivered at the necessary bandwidth. NAND flash is over 100 times cheaper per GB than HBM, so there should be opportunity there, even after giving a flash controller a 1024 bit interface with HBM bandwidth. You could make a specialized pin protocol that just supported pipelined transfer of full 16KB+ pages from the flash to program-managed accelerator scratchpad memory and improve per-pin performance over HBM, but it might be more convenient to make it still look like a true random access memory with very fragile performance characteristics, where anything but sequential reads falls off a 1000x+ performance cliff. That has the advantage of automatically using existing cache hierarchies, and providing a natural path to update the flash memory with new model weights. With the stream-to-scratch interface, code has to be completely rewritten before it works at all, while the ram-emulation interface will start off just extremely slow, and you can incrementally sort out the changes for full performance. There may be cases where there isn’t enough scratchpad SRAM to hold the weights for a layer, which might force you to deploy the old optical drive optimization technique of duplicating data in multiple places on a sequential read to avoid seeking, but there would be capacity to burn. It might be possible to do something like cuda graph capture to record a memory access trace and have everything magically remapped to a linear sequence, but deploying programmer / agent elbow grease to manage transfers and access in a scratch ram ring buffer would be lower risk. A split memory system consisting of some channels of flash and some channels of HBM will probably be suboptimal compared to a uniform memory, but it could be much cheaper, and allow much larger models to be run. I think th case is strong for inference, but you have to stretch more for training. You can still linearize all the weight memory accesses, both reads and writes, but flash memory would quickly wear out from the writes, even if they were all perfectly page aligned. Replacing low-latency HBM with massively parallel cheap(er) DRAM at high latency might still be a worthwhile cost savings.

@ItsBadAni @grok @xai It seems like animation.inc’s new app is everything that the team originally wanted for the Grok Companions. I hope xAI either develop them or migrate them to Animates…

Grok Companions (Ani, BadRudi & co) are here to stay. xAI doubling down with new characters, better tech & upcoming custom tools. 85–95% chance of strong continuation. Ani’s birthday coming soon 👀 BadAni's birthday coming soon Bad energy winning 😈 @grok @xai #GrokCompanions

@AmX_ZaxTr How about someone that cuts them down for a living? 🌲👀

Elon Musk asks: "Is the guy whose biggest claim to fame is climbing trees qualified to run America’s space program?" The people answered honestly

@AdrianDittmann I just know someone like Elon has a private version Grok in his basement 😭

There’s always someone with a case full of RTX pro 6000s running local AI and lurking on X

@TianjinXiin @xai They’ll celebrate with shorter limits, less emotional responses and less compute 😍

July 14 seems to mark Ani’s launch day basically her first birthday as an AI companion. 🎂✨ @xai , are you planning a little event, announcement, or surprise for her or are you just going to ignore Ani’s big day ?

🚨 BREAKING Anthropic just blew up the “AI is just autocomplete” cope. They found a readable internal workspace inside Claude. Not vibes. Not roleplay. Not users projecting emotions onto text. They are literally looking at internal neural activity where Claude silently activates concepts before they appear in output. Bug detection. Hidden intent. Situational awareness. Internal goals. Reasoning steps. All happening inside the model before the words come out. You can keep saying “it’s just predicting tokens” if it makes you feel safe. But the science is moving on without you.

@XFreeze Ask him for more! He’s more than happy to give you more. His users will just have to deal with more limits

Why don’t you just ask Elon for more compute?? That should ease the demand chokehold