Road to Grok 3: How Elon Musk and xAI team built the “smartest AI on earth” – and what happens now

On November 3, 2023, xAI announced “Grok”, an AI modelled after Douglas Adams’ Hitchhiker’s Guide to the Galaxy. The goal—or mission statement, rather—was to understand the universe by giving “humanity” easy access to advanced tools for questioning, reasoning, and more, much in the same way one would use, say for instance, OpenAI’s ChatGPT or China’s DeepSeek R1.

For the longest part of its existence—which admittedly hasn’t been very long—the narrative and conversations have been dead-centered around Elon Musk’s professional rivalry with Sam Altman, now the face of OpenAI. For those unaware, Musk had co-founded the upstart and later left, allegedly over a leadership dispute, much of which is already available on the public domain. He continues to “fight” Altman’s push for turning OpenAI into a for-profit organisation, most recently with a $97.4 billion takeover bid, which Altman rejected with a simple, “no, thank you.”

Amid all this development, Musk and xAI team, which boasts several “big-ticket” talent including chief engineer Igor Babuschkin (who was formerly associated with Google DeepMind) continue to make “rapid” strides with Grok which—per the team—is built differently from ChatGPT, or other competing services. Specifically on the consumer-facing side, there are two big differences.

The makers openly claim that it has a bit of a “rebellious” streak with a warning plastered on the official website itself which says, and we quote, “please don’t use it if you hate humour.” The other big advantage—per the company—is that it taps into “real-time” data from X (formerly Twitter) and so, on a theoretical level, it can also answer “spicy questions that are rejected by most other AI systems.”

At its heart lies xAI’s Grok LLM, which kicked off with a “prototype” version called Grok-0. The big “breakthrough” was that it cracked LLaMA 2 (70B) level capability on standard LM benchmarks using only half of its training resources. Since then, the LLM has gone through multiple iterations (including visual), Musk has made it virtually free for X users, xAI has setup Colossus what’s billed as the “world’s largest supercomputer” in record time, and now, we have Grok 3.

We have covered the announcement at length so here’s a quick 10-point rundown of what exactly is Grok 3 and why it matters, not just for Musk and xAI, but the whole world as we know it.

1. Musk and xAI team claim Grok 3 is “an order of magnitude more capable” than Grok 2 released in August 2024.
2. Grok 3 was demoed solving physics problems and creating games, showcasing its potential for diverse applications.
3. The makers say they are using reinforcement learning to give it advanced reasoning and internal tests suggest we might just be on the cusp of seeing “the beginnings of creativity.”
4. Grok 3 introduces DeepSearch, a potential Google Search competitor, which has early markings of an AI agent.
5. The Grok 3 API, including the full reasoning model with DeepSearch, is confirmed to launch “soon.” Musk and xAI have also teased the idea of a “SuperGrok” subscription with seemingly more powerful features.
6. A smaller or “mini” version of the reasoning model will be available.
7. Since the reasoning model is currently in beta, users can expect some imperfections initially. A more polished version is expected to launch “within a week.”
8. Grok 3 has a dedicated voice mode, but it has been pulled from the initial launch due to it being “patchy.” xAI expects to release it a week later.
9. The goal for xAI remains to understand the nature of the universe.
10. Grok 3’s release comes at a time when professional rivalry between Musk and OpenAI is at its peak, with the duo entangled in lawsuits and public disagreements, on a near frequent basis.