Grok 3 is out — breakdown
xAI quietly rolled out Grok 3. Here’s what’s happening:
𝟭. 𝗪𝗵𝗼 𝗰𝗮𝗻 𝘁𝗿𝘆 𝗶𝘁 𝗮𝗹𝗿𝗲𝗮𝗱𝘆?
Available to X Premium+ subscribers (price nearly doubled, but if you subscribed earlier — congrats, you won capitalism).
There are leaks about SuperGrok — a separate subscription without X features, purely for AI. Rumored to be $30/month or $300/year, but not launched yet.
On https://t.co/PuQWuFUygc, not everyone has access to the chat model Grok 3 or Reasoner yet. In short, it’s inconsistent.
𝟮. 𝗛𝗼𝘄’𝘀 𝘁𝗵𝗲 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲?
The Grok 3 chat model outperforms other chatbots, but lags behind the latest Reasoners.
Grok 3 Mini Reasoning is only behind OpenAI o3 (see the X graphs with reported results).
Few benchmarks available. How it performs on harder tests (Frontier Math, Humanity’s Last Exam) is still unclear.
Interestingly, Grok 3 Reasoning is weaker than Mini Reasoning. xAI explains this by saying that Mini was trained much longer. Expect performance gains as Grok 3 continues training.
𝟯. 𝗪𝗵𝗮𝘁’𝘀 𝗻𝗲𝘄?
Deep Search — a competitor to Deep Research from others.
Voice input — promised native support in a couple of weeks.
𝟰. 𝗧𝗵𝗲 𝗿𝗲𝗹𝗲𝗮𝘀𝗲 𝗶𝘀 𝗿𝗮𝘄 𝗮𝘀 𝘁𝗮𝗿𝘁𝗮𝗿𝗲
No blog post, no API.
Reasoner is being rolled out selectively, with no clear timeline.
Feels like they rushed to beat another release.
𝟱. 𝗕𝘂𝘁 — those who got access to Reasoner are generally impressed.
Here’s what Karpathy says: https://t.co/2ZJJKKkCeZ
Questions? Insights? Let’s discuss in the comments!
#ai #grok
I realized it’s easier to work with AI than to waste time on calls.
I used to tolerate it.
Endless meetings where people spend an hour discussing something that could be solved in five minutes.
Soft feedback that doesn’t hurt feelings - but also doesn’t change anything.
Tolerating slow work because “well, that’s just how the team operates.”
Then I started doing things through AI.
Faster. Clearer. More efficient.
I could get results in 1 day that my team would debate for a week.
I could test a hypothesis while people in Slack were still arguing over minor details (btw, Slack sucks and should be replaced).
I could just cut through the noise and see what was actually slowing things down.
And then it hit me: Isn’t it time to rethink all these communication norms we’ve been building since 2022?
For years, we prioritized psychological safety and inclusive communication. But the world has changed.
AI doesn’t ask, “Are you comfortable with this pace?” - it just gets things done.
AI doesn’t say, “Let’s discuss this again later.” - it delivers results.
The real question: Is the team ready to work at this speed?
Radical Candor 2.0 isn’t about being harsh - it’s about being fast.
- Not just “honest feedback” but immediate problem diagnosis and solutions.
- Not just “being kind” but eliminating wasted time.
- Not just “being direct” but instantly seeing who’s delivering and who isn’t.
I’m done wasting time on pointless discussions.
AI makes my work easier, faster, and more effective.
And if people can’t keep up with AI… what happens next?
🚀Exciting News!🚀
Visible Magic was featured on @QuickNode's prestigious #FeatureFridays! 🎉 This milestone highlights our mission to make blockchain & Web3 accessible for businesses of all sizes. 🌐
Big thanks to QuickNode for recognizing our efforts in building scalable, user-friendly Web3 solutions! 💡✨
8/ Whether you're dealing with resource constraints or large-scale projects, this tool simplifies the process. Plus, it's all under the Apache 2.0 license, so feel free to explore further: https://t.co/oNR0ilqmyK
1/ CogVideoX Factory is a game-changer for optimizing fine-tuning of CogVideoX video generation models (CogVideoX-2B, CogVideoX-5B). If you're working with limited GPU memory (24 GB), this tool offers an efficient way to train large models. #AI#MachineLearning#VideoGeneration
7/ CogVideoX Factory's documentation is a huge help, covering everything from dataset preparation to memory requirements. It ensures you can configure your training pipeline effectively.
6/ Curious for more details? Check out their blog post with cool demos, like visualizing the Mandelbrot set at crazy zoom levels: https://t.co/pMDqlsC9ME
1/ 🚀 NumPy QuadDType brings quadruple-precision floating-point to Python, no matter your platform. If you've struggled with np.longdouble before, this might be the solution you've been waiting for.
#Python#NumPy#AI#QuadruplePrecision
5/ They're also working on getting `numpy_quaddtype` onto PyPI and conda soon, plus they might integrate it officially into NumPy in the future. Stay tuned!