Grok4 : The end of Human Intelligence is near
Grok4 released by Elon Musk
Elon Musk’s company, xAI, has released Grok-4.
They’re calling it the smartest AI in the world.
That kind of claim usually means nothing. This time, it might actually mean something.
https://medium.com/media/2b7e52bde1fd9f3822803f6ed423b547/href
This model isn’t just faster or bigger. It’s better at solving problems, deep, complicated ones, across subjects most people don’t even try to touch. Tests, reasoning tasks, live challenges, even subjective judgment.
It does them all. And in most cases, it wins.
Grok4 is dangerously smart
Grok-4 doesn’t need training wheels. If you handed it the SAT, it would get a perfect score. Every time. Not because it’s memorized the questions, but because it knows how to reason through ones it hasn’t seen before.
- It also scores near-perfect on GRE-level tests across every subject.
- Not just math and logic but languages, humanities, literature, engineering, you name it.
Think of the best student from each department in your university. Now imagine an AI doing better than all of them, at the same time.
It’s not just smart in one way. It reasons. It plans. It adjusts. The team behind it says they expect it to start discovering new technologies by next year. And possibly new laws of physics within two years. Right now, it’s still missing some common sense. Doesn’t always “get” the world the way people do. But they’re not worried. That part, they say, is just a matter of time.
Heavy compute used
Each Grok model got trained with 10x more compute than the one before it. Grok-3 already used a massive amount of training power. Grok-4 used 10x that. Which means Grok-4 had 100 times the training compute of Grok-2.
And this time, the focus wasn’t just on giving it more data. They shifted towards reasoning and reinforcement learning.
That means Grok-4 was trained not just to predict the next word, but to work through problems, make mistakes, and fix them. That kind of training costs a lot more.
The whole thing ran on Colossus, xAI’s custom-built supercomputer running 200,000 H100 GPUs.
It’s the biggest known training setup anyone’s attempted outside of governments.
Training Philosophy: Build, Test, Fail, Fix
Their process had three parts:
- Build the best base model, using clean data and optimized infrastructure.
- Give it rewards for solving real problems, not just right answers, but problems where the outcome can be verified. This teaches it to reason from scratch.
- Invent new challenges, as the model gets smarter, it stops learning from basic tasks. So they built systems to find harder ones and create stronger feedback loops.
Reinforcement learning at this scale is hard. The more intelligent the model becomes, the harder it is to “teach” it new things. But xAI pushed through that wall by designing new ways to test and reward the model’s thinking, not just its output.
Benchmarks
All this training wasn’t just theoretical. Grok-4 went head-to-head with other models on some of the toughest AI tests out there. And won.
- Humanities Last Exam (HLE): This one is brutal, 2,500 PhD-level questions written by experts. Humans score around 5%. Grok-4 scored 25%, with no tools.
- ARC-AGI (v2 private subset): This benchmark is the holy grail for measuring general intelligence in machines. For three months, no model crossed 10%. Grok-4 just did 15.8%, in less than 12 hours. Claude 3 Opus, which was the best before, is now far behind.
- VendingBench: Simulates running a vending machine company, think pricing, inventory, long-term planning. Grok-4 made double the money of the next-best model by building a working strategy and sticking to it. People joked that xAI could now pay for GPUs by launching a million vending machines run by Grok.
- Math & Logic: On AIME (a high-level math exam), Grok-4 Heavy scored 100%. It also beat every other model on coding tasks, live problem-solving, and physics tests.
- Efficiency: They even showed a chart of “intelligence per dollar.” Grok-4 was in its own category, smarter and more efficient than anything else.
Grok-4 Heavy: The elder brother
The standard Grok-4 is already strong. But there’s another version — Grok-4 Heavy.
- This one works like a team. It spawns multiple AI agents to tackle the same task, each working independently. They compare notes, share tricks, argue a little, and decide on the best answer.
- It’s not a voting system. Often, one of them finds a breakthrough, a clever trick, and helps the others see it. It’s like a study group where the smart one helps the rest get there too.
- This approach adds more compute at inference time, but also makes Grok-4 Heavy much more capable on hard, multi-step problems.
Real-World Demos: Not Just Theory
They didn’t just talk numbers. They showed what Grok-4 can actually do:
- Black Hole Simulation: It visualized two black holes merging. Not just visuals — it used real physics, and even cited which textbook it was drawing from.
- Betting Prediction: Grok-4 Heavy scanned live odds for the MLB, compared them to its own calculations, and predicted that the Dodgers had a 21.6% chance of winning. Took about 4.5 minutes.
- Weird Photo Hunt: They gave it a vague task: “Find the employee with the weirdest profile picture.” It searched the web, looked through the xAI team, and picked one (Greg Yang). Funny, but also showed how it could take subjective instructions and act on them.
Voice Mode Got a Major Upgrade
Grok-4 doesn’t just write. It speaks. And now it sounds more like a person.
Lower latency: Everything runs faster now. Responses feel natural.
New voices: One called Eve can whisper, sing, and even talk in rhythm. Another one, Sal, sounds like a deep movie-trailer voice.
Real use cases: They demoed Eve singing an opera about Diet Coke, playing a quick game, and carrying natural conversations.
People seem to like it, usage of the voice mode went up 10x in two months.
API’s Ready
The API is live. It supports 256K tokens of context. That’s a lot, more than enough to process large documents, complex workflows, or codebases.
Some early adopters:
- ARC Institute: Using it for CRISPR research. Letting Grok read millions of logs and flag useful insights.
- Ender Labs: Building and testing business simulations with it.
- Finance companies: Using it for real-time tools and decision-making.
- Also, Grok-4 will soon be available on cloud platforms like AWS, GCP, etc. The xAI enterprise team is up and running now.
What’s Next
They’re not slowing down. In fact, they’re speeding up.
- New coding model: A fast, smart model for developers is coming in a few weeks.
- Multimodal Grok: Version 7 will let Grok see and hear — addressing its current blind spot in vision.
- Video generation: Full training starts soon on 100,000 GB200s. The goal? A model that can take in raw pixels and output new video. Frame by frame. AI movies, games, shows — made from scratch.
Expected:
- First AI-generated TV show in 2025.
- First full AI-generated game and movie in 2026.
- Eventually: A feed on X where you scroll, watch AI content, then jump in and change the outcome yourself. Like Choose Your Own Adventure, but powered by real-time AI.
Final Word
Grok-4 isn’t just a step forward. It’s a signal that the old rules are fading.
We’re used to models being “better at writing,” “faster at summarizing,” “good for customer support.” This thing isn’t that. It’s thinking across disciplines, inventing, testing, reasoning, and doing it better than most people.
Whether it’s good or bad for us long-term, hard to say.
But one thing’s clear: AI isn’t crawling forward anymore. It’s sprinting. Grok-4 is proof of that.
Grok4 : The end of Human Intelligence is near was originally published in Data Science in Your Pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.
